Skip to content

L
lwir-nn
Commit b1111bca authored by Andrey Filippov's avatar Andrey Filippov
Browse files
Options

Trying versions and working on new evaluation criteria for FG/BG 

features
parent 2c44046d
Show whitespace changes
Inline Side-by-side
Showing with 9208 additions and 44 deletions
explore_data6.py
View file @ b1111bca
#!/usr/bin/env python3
#from numpy import float64
from tensorflow.contrib.image.ops.gen_distort_image_ops import adjust_hsv_in_yiq
#from tensorflow.contrib.image.ops.gen_distort_image_ops import adjust_hsv_in_yiq
__copyright__ = "Copyright 2018, Elphel, Inc."
__license__ = "GPL-3.0+"
......@@ -161,7 +161,7 @@ def writeTFRecordsFromImageSet(
s_gt, # wght, # will be modified in place
w_diag = 0.7,
w_reduce = 0.7,
num_pass = 10,
num_pass = 50,
eps = 1E-6)
if debug > 1:
......@@ -281,7 +281,8 @@ def fillGapsByLaplacian(
w_diag = 0.7,
w_reduce = 0.7,
num_pass = 10,
eps = 1E-6):
eps = 1E-6,
debug_level = 0):
dirs = ((-1,0), (-1,1), (0,1), (1,1), (1,0), (1,-1), (0,-1), (-1,-1))
wneib = ( 1.0, w_diag, 1.0, w_diag, 1.0, w_diag, 1.0, w_diag)
gap_tiles = []
......@@ -304,7 +305,7 @@ def fillGapsByLaplacian(
valn = np.copy(val)
wghtn = np.copy(wght)
achange = eps * np.max(wght)
for _ in range (num_pass):
for npass in range (num_pass):
num_new = 1
max_diff = 0.0;
for tile, neibs in zip (gap_tiles, gap_neibs):
......@@ -326,6 +327,8 @@ def fillGapsByLaplacian(
max_diff = max(max_diff, wdiff)
np.copyto(val, valn)
np.copyto(wght, wghtn)
if (debug_level > 3):
print("Pass %d, max_diff = %f"%(npass, max_diff))
if (num_new == 0) and (max_diff < achange):
break
......@@ -536,13 +539,23 @@ class ExploreData:
ds_pair[:,:,:,2] = gtaux[:,:,:, ijt.IJFGBG.AUX_DISP] # 8
for nf in range (ds_pair.shape[0]):
if (self.debug_level > 3):
print ("---- nf=%d"%(nf,))
fillGapsByLaplacian(
ds_pair[nf,:,:,0], # val, # will be modified in place
ds_pair[nf,:,:,1], # wght, # will be modified in place
w_diag = 0.7,
w_reduce = 0.7,
num_pass = 20,
eps = 1E-6)
eps = 1E-6,
debug_level = self.debug_level)
if (self.debug_level > 0):
print ("---- nf=%d min = %f mean = %f max = %f"%(
nf,
ds_pair[nf,:,:,0].min(),
ds_pair[nf,:,:,0].mean(),
ds_pair[nf,:,:,0].max()))
print("zero strength",np.nonzero(ds_pair[nf,:,:,1]==0.0))
return ds_pair
......@@ -847,6 +860,10 @@ class ExploreData:
min_neibs = None, # Minimal number of valid tiles to include
use_split = False, # Select y single/multi-plane tiles (center only)
keep_split = False):# When sel_split, keep only multi-plane tiles (false - only single-plane)
#for file names:
self.min_neibs = min_neibs
self.use_split = use_split
self.keep_split = keep_split
if data_ds is None:
data_ds = self.train_ds
......@@ -903,6 +920,7 @@ class ExploreData:
'''
def augmentBatchFileIndices(self,
seed_index,
seed_list = None,
min_choices=None,
max_files = None,
set_ds = None
......@@ -915,14 +933,21 @@ class ExploreData:
set_ds = self.train_ds
full_num_choices = self.num_batch_tiles[seed_index].copy()
flist = [seed_index]
all_choices = list(range(self.num_batch_tiles.shape[0]))
all_choices.remove(seed_index)
if seed_list is None:
seed_list = list(range(self.num_batch_tiles.shape[0]))
all_choices = list(seed_list) # a copy of seed list
all_choices.remove(seed_index) # seed_list made unique by the caller
### list(filter(lambda a: a != seed_index, all_choices)) # remove all instances of seed_index
for _ in range (max_files-1):
if full_num_choices.min() >= min_choices:
break
if len(all_choices) == 0:
print ("Nothing left in all_choices!")
break
findx = np.random.choice(all_choices)
flist.append(findx)
all_choices.remove(findx)
all_choices.remove(findx) # seed_list made unique by the caller
### list(filter(lambda a: a != findx, all_choices)) # remove all instances of findx
full_num_choices += self.num_batch_tiles[findx]
file_tiles_sparse = [[] for _ in set_ds] #list of empty lists for each train scene (will be sparse)
......@@ -1012,6 +1037,7 @@ ValueError: need at least one array to concatenate
def prepareBatchData(self,
ml_list,
seed_index,
seed_list,
min_choices=None,
max_files = None,
ml_num = None,
......@@ -1036,7 +1062,12 @@ ValueError: need at least one array to concatenate
width_m1 = width-1
height_m1 = height-1
corr_layers = ['hor-pairs', 'vert-pairs','diagm-pair', 'diago-pair']
flist,tiles = self.augmentBatchFileIndices(seed_index, min_choices, max_files, set_ds)
flist,tiles = self.augmentBatchFileIndices(
seed_index,
seed_list,
min_choices,
max_files,
set_ds)
ml_all_files = self.getBatchData(
flist,
ml_list,
......@@ -1137,6 +1168,7 @@ ValueError: need at least one array to concatenate
writer = tf.io.TFRecordWriter(tfr_filename)
if num_scenes is None:
num_scenes = len(files_list)
'''
if len(files_list) <= num_scenes:
#create and shuffle repetitive list of files of num_scenes.length
seed_list = np.arange(num_scenes) % len(files_list)
......@@ -1147,11 +1179,25 @@ ValueError: need at least one array to concatenate
np.random.shuffle(seed_list)
seed_list = seed_list[:num_scenes]
np.random.shuffle(seed_list)
'''
augment_list = []
for seed_indx in np.arange(len(files_list)):
if self.num_batch_tiles[seed_indx].sum() >0:
augment_list.append(seed_indx)
seed_list = list(augment_list) # seed list will be modified while augment_list will have unique/full list of suitable files
while len(seed_list) < num_scenes:
seed_list.append(np.random.choice(seed_list))
np.random.shuffle(seed_list)
if len(seed_list) >= num_scenes:
seed_list = seed_list[:num_scenes]
cluster_size = (2 * radius + 1) * (2 * radius + 1)
for nscene, seed_index in enumerate(seed_list):
corr2d_batch, target_disparity_batch, gt_ds_batch = ex_data.prepareBatchData( #'hor-pairs' is not in list
ml_list,
seed_index,
augment_list,
min_choices=None,
max_files = None,
ml_num = None,
......@@ -1195,6 +1241,7 @@ ValueError: need at least one array to concatenate
sweep_files,
sweep_disparities,
seed_index,
seed_list,
min_choices=None,
max_files = None,
set_ds = None,
......@@ -1210,8 +1257,6 @@ ValueError: need at least one array to concatenate
min_choices = self.min_batch_choices #10
if max_files is None:
max_files = self.max_batch_files #10
### if ml_num is None:
### ml_num = self.files_per_scene #5 ????
if set_ds is None:
set_ds = self.train_ds
tiles_in_sample = (2 * radius + 1) * (2 * radius + 1)
......@@ -1223,6 +1268,7 @@ ValueError: need at least one array to concatenate
flist0, tiles0 = self.augmentBatchFileIndices(
seed_index,
seed_list,
min_choices,
max_files,
set_ds)
......@@ -1306,6 +1352,13 @@ ValueError: need at least one array to concatenate
rnd_plate = 0.0):## disparity random for each plate (now 25 tiles)
# open the TFRecords file
fb = ""
if self.use_split:
fb = ["-FB1","-FB2"][self.keep_split] # single plane - FB1, split FG/BG planes - FB2
tfr_filename+="-RT%1.2f-RP%1.2f-M%d-NB%d%s"%(rnd_tile,rnd_plate,self.fgbg_mode,self.min_neibs, fb)
if not '.tfrecords' in tfr_filename:
tfr_filename += '.tfrecords'
......@@ -1324,10 +1377,11 @@ ValueError: need at least one array to concatenate
#skip writing if file exists - it will be possible to continue or run several instances
if os.path.exists(tfr_filename):
print(tfr_filename+" already exists, skipping generation. Please remove and re-run this program if you want to regenerate the file")
# return # Temporary disable
return # Temporary disable
writer = tf.io.TFRecordWriter(tfr_filename)
if num_scenes is None:
num_scenes = len(files_list)
'''
if len(files_list) <= num_scenes:
#create and shuffle repetitive list of files of num_scenes.length
seed_list = np.arange(num_scenes) % len(files_list)
......@@ -1337,7 +1391,19 @@ ValueError: need at least one array to concatenate
seed_list = np.arange(len(files_list))
np.random.shuffle(seed_list)
seed_list = seed_list[:num_scenes]
'''
augment_list = []
for seed_indx in np.arange(len(files_list)):
if self.num_batch_tiles[seed_indx].sum() >0:
augment_list.append(seed_indx)
seed_list = list(augment_list) # seed list will be modified while augment_list will have unique/full list of suitable files
while len(seed_list) < num_scenes:
seed_list.append(np.random.choice(seed_list))
np.random.shuffle(seed_list)
if len(seed_list) >= num_scenes:
seed_list = seed_list[:num_scenes]
cluster_size = (2 * radius + 1) * (2 * radius + 1)
for nscene, seed_index in enumerate(seed_list):
corr2d_batch, target_disparity_batch, gt_ds_batch = ex_data.prepareBatchDataLwir( #'hor-pairs' is not in list
......@@ -1345,9 +1411,9 @@ ValueError: need at least one array to concatenate
sweep_files = sweep_files,
sweep_disparities = sweep_disparities,
seed_index = seed_index,
seed_list = augment_list,
min_choices = None,
max_files = None,
### ml_num = None,
set_ds = set_ds, #DS data from all GT_AX files scanned
radius = radius,
rnd_tile = rnd_tile, ## disparity random for each tile
......@@ -1463,7 +1529,7 @@ if __name__ == "__main__":
## topdir_train = "/data_ssd/data_sets/train_mlr32_18d"
## topdir_train = '/data_ssd/data_sets/test_only'# ''
### topdir_train = '/data_ssd/data_sets/train_set2'# ''
topdir_train = '/data_ssd/lwir_sets/lwir_train1'# ''
topdir_train = '/data_ssd/lwir_sets/lwir_train2'# ''
# tf_data_5x5_main_10_heur
try:
......@@ -1472,8 +1538,8 @@ if __name__ == "__main__":
# topdir_test = "/mnt/dde6f983-d149-435e-b4a2-88749245cc6c/home/eyesis/x3d_data/data_sets/test"#test" #all/"
# topdir_test = "/data_ssd/data_sets/test_mlr32_18d"
## topdir_test = '/data_ssd/data_sets/test_only'
### topdir_test = '/data_ssd/data_sets/test_set2'
topdir_test = '/data_ssd/lwir_sets/lwir_test1'
### topdir_test = '/data_ssd/data_sets/test_set21'
topdir_test = '/data_ssd/lwir_sets/lwir_test2'
try:
......@@ -1482,7 +1548,7 @@ if __name__ == "__main__":
# pathTFR = "/mnt/dde6f983-d149-435e-b4a2-88749245cc6c/home/eyesis/x3d_data/data_sets/tf_data_3x3b" #no trailing "/"
# pathTFR = "/home/eyesis/x3d_data/data_sets/tf_data_5x5" #no trailing "/"
### pathTFR = "/data_ssd/data_sets/tf_data_5x5_main_13_heur"
pathTFR = '/data_ssd/lwir_sets/tf_data_5x5_01'
pathTFR = '/data_ssd/lwir_sets/tf_data_5x5_2'
## pathTFR = "/data_ssd/data_sets/tf_data_5x5_main_11_rnd"
## pathTFR = "/data_ssd/data_sets/tf_data_5x5_main_12_rigrnd"
......@@ -1589,7 +1655,7 @@ if __name__ == "__main__":
FRAC_NEIBS_VALID = 0.55# 8 #LWIR new
MIN_NEIBS = (2 * RADIUS + 1) * (2 * RADIUS + 1) # All tiles valid == 9
MIN_NEIBS = round (MIN_NEIBS * FRAC_NEIBS_VALID)
VARIANCE_THRESHOLD = 0.4 # 1.5
VARIANCE_THRESHOLD = 0.8 # 0.4 # 1.5
VARIANCE_SCALE_DISPARITY = 5.0 #Scale variance if average is above this
NUM_TRAIN_SETS = 32 # 8
......@@ -1600,6 +1666,9 @@ if __name__ == "__main__":
RND_AMPLIUDE_TRAIN_PLATE = 0.5 # train with corr2d rendered +/- this far from the GT - common for each (5x5) plate component
MAX_MAIN_OFFSET = 2.5 # do not use tile for training if MAIN camera (AUX for LWIR) differs more from GT
MODEL_ML_DIR = "ml32" # subdirectory with the ML disparity sweep files
USE_SPLIT = False # True, # Select y single/multi-plane tiles (center only)
KEEP_SPLIT = False # When sel_split, keep only multi-plane tiles (false - only single-plane)
if not topdir_train:
NUM_TRAIN_SETS = 0
......@@ -1618,14 +1687,14 @@ if __name__ == "__main__":
''' Prepare full image for testing '''
"""
for model_ml_path in test_sets:
writeTFRecordsFromImageSet(
model_ml_path, # model/version/ml_dir
FGBGMODE_TEST, # 0, # expot_mode, # 0 - GT average, 1 - GT FG, 2 - GT BG, 3 - AUX disparity
RND_AMPLIUDE_TEST, # random_offset, # for modes 0..2 - add random offset of -random_offset to +random_offset, in mode 3 add random to GT average if no AUX data
pathTFR) # TFR directory
"""
# disp_bins = 20,
......@@ -1641,13 +1710,13 @@ if __name__ == "__main__":
ml_pattern = ml_pattern,
max_main_offset = MAX_MAIN_OFFSET,
latest_version_only = LATEST_VERSION_ONLY,
debug_level = 3, #1, #3, ##0, #3,
debug_level = 1, #3, #1, #3, ##0, #3,
disparity_bins = 50, #100 #200, #1000,
strength_bins = 50, #100
disparity_min_drop = -0.1,
disparity_min_clip = -0.1,
disparity_max_drop = 5.0, #100.0,
disparity_max_clip = 5.0, #100.0,
disparity_max_drop = 8.0, #100.0,
disparity_max_clip = 8.0, #100.0,
strength_min_drop = 0.02, # 0.1,
strength_min_clip = 0.02, # 0.1,
strength_max_drop = 0.3, # 1.0,
......@@ -1743,13 +1812,21 @@ if __name__ == "__main__":
max_var = VARIANCE_THRESHOLD, # Maximal tile variance to include
scale_disp = VARIANCE_SCALE_DISPARITY,
min_neibs = MIN_NEIBS, # Minimal number of valid tiles to include
use_split = True, # Select y single/multi-plane tiles (center only)
keep_split = False) # When sel_split, keep only multi-plane tiles (false - only single-plane)
use_split = USE_SPLIT, # Select y single/multi-plane tiles (center only)
keep_split = KEEP_SPLIT) # When sel_split, keep only multi-plane tiles (false - only single-plane)
pass
for train_var in range (NUM_TRAIN_SETS):
fpath = train_filenameTFR+("%03d"%(train_var,))
ex_data.writeTFRewcordsEpoch(fpath, ml_list = ml_list_train, files_list = ex_data.files_train, set_ds= ex_data.train_ds)
ex_data.writeTFRewcordsEpochLwir(
fpath,
sweep_files = ex_data.train_sweep_files,
sweep_disparities = ex_data.train_sweep_disparities,
files_list = ex_data.files_train,
set_ds = ex_data.train_ds,
radius = ex_data.radius,
rnd_tile = ex_data.rnd_tile,
rnd_plate = ex_data.rnd_plate)
list_of_file_lists_test, num_batch_tiles_test = ex_data.makeBatchLists( # results are also saved to self.*
data_ds = ex_data.test_ds,
......@@ -1759,12 +1836,21 @@ if __name__ == "__main__":
min_var = 0.0, # Minimal tile variance to include
max_var = VARIANCE_THRESHOLD, # Maximal tile variance to include
min_neibs = MIN_NEIBS, # Minimal number of valid tiles to include
use_split = True, # Select y single/multi-plane tiles (center only)
keep_split = False) # When sel_split, keep only multi-plane tiles (false - only single-plane)
use_split = USE_SPLIT, # Select y single/multi-plane tiles (center only)
keep_split = KEEP_SPLIT) # When sel_split, keep only multi-plane tiles (false - only single-plane)
fpath = test_filenameTFR # +("-%03d"%(train_var,))
ex_data.writeTFRewcordsEpoch(fpath, ml_list = ml_list_test, files_list = ex_data.files_test, set_ds= ex_data.test_ds, num_scenes = num_test_scenes)
pass
ex_data.writeTFRewcordsEpochLwir(
fpath,
sweep_files = ex_data.test_sweep_files,
sweep_disparities = ex_data.test_sweep_disparities,
files_list = ex_data.files_test,
set_ds = ex_data.test_ds,
radius = ex_data.radius,
num_scenes = num_test_scenes,
rnd_tile = ex_data.rnd_tile,
rnd_plate = ex_data.rnd_plate)
else: # RADIUS > 0
# test
list_of_file_lists_test, num_batch_tiles_test = ex_data.makeBatchLists( # results are also saved to self.*
......@@ -1775,8 +1861,8 @@ if __name__ == "__main__":
min_var = 0.0, # Minimal tile variance to include
max_var = VARIANCE_THRESHOLD, # Maximal tile variance to include
min_neibs = MIN_NEIBS, # Minimal number of valid tiles to include
use_split = True, # Select y single/multi-plane tiles (center only)
keep_split = False) # When sel_split, keep only multi-plane tiles (false - only single-plane)
use_split = USE_SPLIT, # Select y single/multi-plane tiles (center only)
keep_split = KEEP_SPLIT) # When sel_split, keep only multi-plane tiles (false - only single-plane)
num_le_test = num_batch_tiles_test.sum()
print("Number of <= %f disparity variance tiles: %d (est)"%(VARIANCE_THRESHOLD, num_le_test))
......@@ -1802,8 +1888,8 @@ if __name__ == "__main__":
min_var = VARIANCE_THRESHOLD, # Minimal tile variance to include
max_var = 1000.0, # Maximal tile variance to include
min_neibs = MIN_NEIBS, # Minimal number of valid tiles to include
use_split = True, # Select y single/multi-plane tiles (center only)
keep_split = False) # When sel_split, keep only multi-plane tiles (false - only single-plane)
use_split = USE_SPLIT, # Select y single/multi-plane tiles (center only)
keep_split = KEEP_SPLIT) # When sel_split, keep only multi-plane tiles (false - only single-plane)
num_gt_test = num_batch_tiles_test.sum()
high_fract_test = 1.0 * num_gt_test / (num_le_test + num_gt_test)
......@@ -1831,8 +1917,8 @@ if __name__ == "__main__":
min_var = 0.0, # Minimal tile variance to include
max_var = VARIANCE_THRESHOLD, # Maximal tile variance to include
min_neibs = MIN_NEIBS, # Minimal number of valid tiles to include
use_split = True, # Select y single/multi-plane tiles (center only)
keep_split = False) # When sel_split, keep only multi-plane tiles (false - only single-plane)
use_split = USE_SPLIT, # Select y single/multi-plane tiles (center only)
keep_split = KEEP_SPLIT) # When sel_split, keep only multi-plane tiles (false - only single-plane)
num_le_fake = num_batch_tiles_fake.sum()
print("Number of <= %f disparity variance tiles: %d (test)"%(VARIANCE_THRESHOLD, num_le_fake))
......@@ -1858,8 +1944,8 @@ if __name__ == "__main__":
min_var = VARIANCE_THRESHOLD, # Minimal tile variance to include
max_var = 1000.0, # Maximal tile variance to include
min_neibs = MIN_NEIBS, # Minimal number of valid tiles to include
use_split = True, # Select y single/multi-plane tiles (center only)
keep_split = False) # When sel_split, keep only multi-plane tiles (false - only single-plane)
use_split = USE_SPLIT, # Select y single/multi-plane tiles (center only)
keep_split = KEEP_SPLIT) # When sel_split, keep only multi-plane tiles (false - only single-plane)
num_gt_fake = num_batch_tiles_fake.sum()
high_fract_fake = 1.0 * num_gt_fake / (num_le_fake + num_gt_fake)
......@@ -1888,8 +1974,8 @@ if __name__ == "__main__":
min_var = 0.0, # Minimal tile variance to include
max_var = VARIANCE_THRESHOLD, # Maximal tile variance to include
min_neibs = MIN_NEIBS, # Minimal number of valid tiles to include
use_split = True, # Select y single/multi-plane tiles (center only)
keep_split = False) # When sel_split, keep only multi-plane tiles (false - only single-plane)
use_split = USE_SPLIT, # Select y single/multi-plane tiles (center only)
keep_split = KEEP_SPLIT) # When sel_split, keep only multi-plane tiles (false - only single-plane)
num_le_train = num_batch_tiles_train.sum()
print("Number of <= %f disparity variance tiles: %d (train)"%(VARIANCE_THRESHOLD, num_le_train))
......@@ -1914,8 +2000,8 @@ if __name__ == "__main__":
min_var = VARIANCE_THRESHOLD, # Minimal tile variance to include
max_var = 1000.0, # Maximal tile variance to include
min_neibs = MIN_NEIBS, # Minimal number of valid tiles to include
use_split = True, # Select y single/multi-plane tiles (center only)
keep_split = False) # When sel_split, keep only multi-plane tiles (false - only single-plane)
use_split = USE_SPLIT, # Select y single/multi-plane tiles (center only)
keep_split = KEEP_SPLIT) # When sel_split, keep only multi-plane tiles (false - only single-plane)
num_gt_train = num_batch_tiles_train.sum()
high_fract_train = 1.0 * num_gt_train / (num_le_train + num_gt_train)
......
explore_data7.py 0 → 100644
View file @ b1111bca
This source diff could not be displayed because it is too large. You can view the blob instead.
explore_data8.py 0 → 100644
View file @ b1111bca
This source diff could not be displayed because it is too large. You can view the blob instead.
explore_data9.py 0 → 100644
View file @ b1111bca
This source diff could not be displayed because it is too large. You can view the blob instead.
imagej_tiffwriter.py 0 → 100644
View file @ b1111bca
#!/usr/bin/env python3
'''
/**
* @file imagej_tiffwriter.py
* @brief save tiffs for imagej (1.52d+) - with stacks and hyperstacks
* @par <b>License</b>:
* This program is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
'''
__copyright__ = "Copyright 2018, Elphel, Inc."
__license__ = "GPL-3.0+"
__email__ = "oleg@elphel.com"
'''
Usage example:
import imagej_tiffwriter
import numpy as np
Have a few images in the form of numpy arrays np.float32:
- (h,w)
- (n,h,w)
Labels can be provided as a list: ['label1','label2', etc.]
No list length check against number of images
imagej_tiffwriter.save(path,images,labels)
'''
import numpy as np
import struct
import tifffile
import math
# from here: https://stackoverflow.com/questions/50258287/how-to-specify-colormap-when-saving-tiff-stack
def imagej_metadata_tags(metadata, byteorder):
"""Return IJMetadata and IJMetadataByteCounts tags from metadata dict.
The tags can be passed to the TiffWriter.save function as extratags.
"""
header = [{'>': b'IJIJ', '<': b'JIJI'}[byteorder]]
bytecounts = [0]
body = []
def writestring(data, byteorder):
return data.encode('utf-16' + {'>': 'be', '<': 'le'}[byteorder])
def writedoubles(data, byteorder):
return struct.pack(byteorder+('d' * len(data)), *data)
def writebytes(data, byteorder):
return data.tobytes()
metadata_types = (
('Info', b'info', 1, writestring),
('Labels', b'labl', None, writestring),
('Ranges', b'rang', 1, writedoubles),
('LUTs', b'luts', None, writebytes),
('Plot', b'plot', 1, writebytes),
('ROI', b'roi ', 1, writebytes),
('Overlays', b'over', None, writebytes))
for key, mtype, count, func in metadata_types:
if key not in metadata:
continue
if byteorder == '<':
mtype = mtype[::-1]
values = metadata[key]
if count is None:
count = len(values)
else:
values = [values]
header.append(mtype + struct.pack(byteorder+'I', count))
for value in values:
data = func(value, byteorder)
body.append(data)
bytecounts.append(len(data))
body = b''.join(body)
header = b''.join(header)
data = header + body
bytecounts[0] = len(header)
bytecounts = struct.pack(byteorder+('I' * len(bytecounts)), *bytecounts)
return ((50839, 'B', len(data), data, True),
(50838, 'I', len(bytecounts)//4, bytecounts, True))
#def save(path,images,force_stack=False,force_hyperstack=False):
def save(path,images,labels=None,label_prefix="Label "):
'''
labels a list or None
'''
'''
Expecting:
(h,w),
(n,h,w) - just create a simple stack
'''
# Got images, analyze shape:
# - possible formats (c == depth):
# -- (t,z,h,w,c)
# -- (t,h,w,c), t or z does not matter
# -- (h,w,c)
# -- (h,w)
# 0 or 1 images.shapes are not handled
#
# (h,w)
if len(images.shape)==2:
images = images[np.newaxis,...]
# now the shape length is 3
if len(images.shape)==3:
# tifffile treats shape[0] as channel, need to expand to get labels displayed
#images = images[images.shape[0],np.newaxis,images.shape[1],images.shape[2]]
images = np.reshape(images,(images.shape[0],1,images.shape[1],images.shape[2]))
labels_list = []
if labels is None:
for i in range(images.shape[0]):
labels_list.append(label_prefix+str(i+1))
else:
labels_list = labels
ijtags = imagej_metadata_tags({'Labels':labels_list}, '<')
with tifffile.TiffWriter(path, bigtiff=False,imagej=True) as tif:
for i in range(images.shape[0]):
tif.save(images[i], metadata={'version':'1.11a','loop': False}, extratags=ijtags)
# Testing
if __name__ == "__main__":
def hamming_window(x,N):
y = 0.54 - 0.46*math.cos(2*math.pi*x/(N-1))
return y
hw = hamming_window
NT = 5
NX = 512
NY = 512
images = np.empty((NT,NY,NX),np.float32)
import time
print(str(time.time())+": Generating test images")
for t in range(NT):
images[t,:,:] = np.array([[(255-t*25)*hw(i,512)*hw(j,512) for i in range(NX)] for j in range(NY)],np.float32)
print(str(time.time())+": Test images generated")
print("Images shape: "+str(images.shape))
v = save("tiffwriter_test.tiff",images)
nn_ds_neibs30.py 0 → 100644
View file @ b1111bca
#!/usr/bin/env python3
__copyright__ = "Copyright 2018-2019, Elphel, Inc."
__license__ = "GPL-3.0+"
__email__ = "andrey@elphel.com"
#python3 nn_ds_neibs30.py /data_ssd/lwir_sets/conf/qcstereo_lwir01.xml /data_ssd/lwir_sets/
#tensorboard --logdir="nn_ds_neibs30_graph13-9RNSWLAM0.5SLAM0.1SCLP0.2_nG_nI_HF_CP0.3_S0.03" --port=7001
import os
import sys
import numpy as np
import time
import shutil
from threading import Thread
import qcstereo_network
import qcstereo_losses
import qcstereo_functions as qsf
import tensorflow as tf
qsf.TIME_START = time.time()
qsf.TIME_LAST = qsf.TIME_START
IMG_WIDTH = 20 # 324 # tiles per image row
DEBUG_LEVEL= 1
try:
conf_file = sys.argv[1]
except IndexError:
print("Configuration path is required as a first argument. Optional second argument specifies root directory for data files")
exit(1)
try:
root_dir = sys.argv[2]
except IndexError:
root_dir = os.path.dirname(conf_file)
print ("Configuration file: " + conf_file)
parameters, dirs, files, _ = qsf.parseXmlConfig(conf_file, root_dir)
"""
Temporarily for backward compatibility
"""
if not "SLOSS_CLIP" in parameters:
parameters['SLOSS_CLIP'] = 0.5
print ("Old config, setting SLOSS_CLIP=", parameters['SLOSS_CLIP'])
"""
Defined in config file
"""
TILE_SIDE, TILE_LAYERS, TWO_TRAINS, NET_ARCH1, NET_ARCH2 = [None]*5
ABSOLUTE_DISPARITY,SYM8_SUB, WLOSS_LAMBDA, SLOSS_LAMBDA, SLOSS_CLIP = [None]*5
SPREAD_CONVERGENCE, INTER_CONVERGENCE, HOR_FLIP, DISP_DIFF_CAP, DISP_DIFF_SLOPE = [None]*5
CLUSTER_RADIUS = None
PARTIALS_WEIGHTS, MAX_IMGS_IN_MEM, MAX_FILES_PER_GROUP, BATCH_WEIGHTS, ONLY_TILE = [None] * 5
USE_CONFIDENCE, WBORDERS_ZERO, EPOCHS_TO_RUN, FILE_UPDATE_EPOCHS = [None] * 4
LR600,LR400,LR200,LR100,LR = [None]*5
SHUFFLE_FILES, EPOCHS_FULL_TEST, SAVE_TIFFS = [None] * 3
CHECKPOINT_PERIOD = None
TRAIN_BUFFER_GPU, TRAIN_BUFFER_CPU = [None]*2
TEST_TITLES = None
LOGFILE="results.txt"
"""
Next gets globals from the config file
"""
globals().update(parameters)
TRAIN_BUFFER_SIZE = TRAIN_BUFFER_GPU * TRAIN_BUFFER_CPU # in merged (quad) batches
#exit(0)
WIDTH = 20 # 324
HEIGHT = 15 # 242
TILE_SIZE = TILE_SIDE* TILE_SIDE # == 81
FEATURES_PER_TILE = TILE_LAYERS * TILE_SIZE# == 324
BATCH_SIZE = ([1,2][TWO_TRAINS])*2*1000//25 # == 80 Each batch of tiles has balanced D/S tiles, shuffled batches but not inside batches
SUFFIX=(str(NET_ARCH1)+'-'+str(NET_ARCH2)+
(["R","A"][ABSOLUTE_DISPARITY]) +
(["NS","S8"][SYM8_SUB])+
"WLAM"+str(WLOSS_LAMBDA)+
"SLAM"+str(SLOSS_LAMBDA)+
"SCLP"+str(SLOSS_CLIP)+
(['_nG','_G'][SPREAD_CONVERGENCE])+
(['_nI','_I'][INTER_CONVERGENCE]) +
(['_nHF',"_HF"][HOR_FLIP]) +
('_CP'+str(DISP_DIFF_CAP)) +
('_S'+str(DISP_DIFF_SLOPE))
)
NN_LAYOUT1 = qcstereo_network.NN_LAYOUTS[NET_ARCH1]
NN_LAYOUT2 = qcstereo_network.NN_LAYOUTS[NET_ARCH2]
USE_PARTIALS = not PARTIALS_WEIGHTS is None # False - just a single Siamese net, True - partial outputs that use concentric squares of the first level subnets
# Tiff export slice labels
SLICE_LABELS = ["nn_out_ext","hier_out_ext","gt_disparity","gt_strength",
"cutcorn_cost_nw","cutcorn_cost",
"gt-avg_dist","avg8_disp","gt_disp","out-avg"]
##############################################################################
cluster_size = (2 * CLUSTER_RADIUS + 1) * (2 * CLUSTER_RADIUS + 1)
center_tile_index = 2 * CLUSTER_RADIUS * (CLUSTER_RADIUS + 1)
qsf.prepareFiles(dirs, files, suffix = SUFFIX)
#copy config to results directory
print ("Copying config files to results directory:\n ('%s' -> '%s')"%(conf_file,dirs['result']))
try:
os.makedirs(dirs['result'])
except:
pass
shutil.copy2(conf_file,dirs['result'])
LOGPATH = os.path.join(dirs['result'],LOGFILE)
if TEST_TITLES is None:
TEST_TITLES = qsf.defaultTestTitles(files)
partials = None
partials = qsf.concentricSquares(CLUSTER_RADIUS)
PARTIALS_WEIGHTS = [1.0*pw/sum(PARTIALS_WEIGHTS) for pw in PARTIALS_WEIGHTS]
if not USE_PARTIALS:
partials = partials[0:1]
PARTIALS_WEIGHTS = [1.0]
qsf.evaluateAllResults(result_files = files['result'],
absolute_disparity = ABSOLUTE_DISPARITY,
cluster_radius = CLUSTER_RADIUS,
labels = SLICE_LABELS,
logpath= LOGPATH)
image_data = qsf.initImageData(
files = files,
max_imgs = MAX_IMGS_IN_MEM,
cluster_radius = CLUSTER_RADIUS,
tile_layers = TILE_LAYERS,
tile_side = TILE_SIDE,
width = IMG_WIDTH,
replace_nans = True)
corr2d_len, target_disparity_len, gtds_len, _ = qsf.get_lengths(CLUSTER_RADIUS, TILE_LAYERS, TILE_SIDE)
train_next, dataset_train, datasets_test= qsf.initTrainTestData(
files = files,
cluster_radius = CLUSTER_RADIUS,
buffer_size = TRAIN_BUFFER_SIZE * BATCH_SIZE, # number of clusters per train
test_titles = TEST_TITLES)
corr2d_train_placeholder = tf.compat.v1.placeholder(dataset_train.dtype, (None,FEATURES_PER_TILE * cluster_size)) # corr2d_train.shape)
target_disparity_train_placeholder = tf.compat.v1.placeholder(dataset_train.dtype, (None,1 * cluster_size)) #target_disparity_train.shape)
gt_ds_train_placeholder = tf.compat.v1.placeholder(dataset_train.dtype, (None,2 * cluster_size)) #gt_ds_train.shape)
dataset_tt = tf.data.Dataset.from_tensor_slices({
"corr2d": corr2d_train_placeholder,
"target_disparity": target_disparity_train_placeholder,
"gt_ds": gt_ds_train_placeholder})
tf_batch_weights = tf.compat.v1.placeholder(shape=(None,), dtype=tf.float32, name = "batch_weights") # way to increase importance of the high variance clusters
feed_batch_weights = np.array(BATCH_WEIGHTS*(BATCH_SIZE//len(BATCH_WEIGHTS)), dtype=np.float32)
feed_batch_weight_1 = np.array([1.0], dtype=np.float32)
dataset_test_size = len(datasets_test[0])
dataset_test_size //= BATCH_SIZE
dataset_img_size = len(image_data[0]['corr2d'])
dataset_img_size //= BATCH_SIZE
dataset_tt = dataset_tt.batch(BATCH_SIZE)
dataset_tt = dataset_tt.prefetch(BATCH_SIZE)
iterator_tt = dataset_tt.make_initializable_iterator()
next_element_tt = iterator_tt.get_next()
result_dir = './attic/result_neibs_'+ SUFFIX+'/'
checkpoint_dir = './attic/result_neibs_'+ SUFFIX+'/'
save_freq = 500
def debug_gt_variance(
indx, # This tile index (0..8)
center_indx, # center tile index
gt_ds_batch # [?:9:2]
):
with tf.name_scope("Debug_GT_Variance"):
d_gt_this = tf.reshape(gt_ds_batch[:,2 * indx],[-1], name = "d_this")
d_gt_center = tf.reshape(gt_ds_batch[:,2 * center_indx],[-1], name = "d_center")
d_gt_diff = tf.subtract(d_gt_this, d_gt_center, name = "d_diff")
d_gt_diff2 = tf.multiply(d_gt_diff, d_gt_diff, name = "d_diff2")
d_gt_var = tf.reduce_mean(d_gt_diff2, name = "d_gt_var")
return d_gt_var
target_disparity_cluster = tf.reshape(next_element_tt['target_disparity'], [-1,cluster_size, 1], name="targdisp_cluster")
corr2d_Nx325 = tf.concat([tf.reshape(next_element_tt['corr2d'],[-1,cluster_size,FEATURES_PER_TILE], name="coor2d_cluster"),
target_disparity_cluster], axis=2, name = "corr2d_Nx325")
if SPREAD_CONVERGENCE:
outs, inp_weights = qcstereo_network.networks_siam(
input = corr2d_Nx325,
input_global = target_disparity_cluster,
layout1 = NN_LAYOUT1,
layout2 = NN_LAYOUT2,
inter_convergence = INTER_CONVERGENCE,
sym8 = SYM8_SUB,
only_tile = ONLY_TILE, #Remove/put None for normal operation
partials = partials,
use_confidence= USE_CONFIDENCE,
cluster_radius = CLUSTER_RADIUS)
else:
outs, inp_weights = qcstereo_network.networks_siam(
input_tensor= corr2d_Nx325,
input_global = None,
layout1 = NN_LAYOUT1,
layout2 = NN_LAYOUT2,
inter_convergence = False,
sym8 = SYM8_SUB,
only_tile = ONLY_TILE, #Remove/put None for normal operation
partials = partials,
use_confidence= USE_CONFIDENCE,
cluster_radius = CLUSTER_RADIUS)
tf_partial_weights = tf.constant(PARTIALS_WEIGHTS,dtype=tf.float32,name="partial_weights")
G_losses = [0.0]*len(partials)
target_disparity_batch= next_element_tt['target_disparity'][:,center_tile_index:center_tile_index+1]
gt_ds_batch_clust = next_element_tt['gt_ds']
gt_ds_batch = gt_ds_batch_clust[:,2 * center_tile_index: 2 * (center_tile_index +1)]
G_losses[0], _disp_slice, _d_gt_slice, _out_diff, _out_diff2, _w_norm, _out_wdiff2, _cost1 = qcstereo_losses.batchLoss(
out_batch = outs[0], # [batch_size,(1..2)] tf_result
target_disparity_batch= target_disparity_batch, # next_element_tt['target_disparity'][:,center_tile_index:center_tile_index+1], # target_disparity_batch_center, # next_element_tt['target_disparity'], # target_disparity, ### target_d, # [batch_size] tf placeholder
gt_ds_batch = gt_ds_batch, # next_element_tt['gt_ds'][:,2 * center_tile_index: 2 * (center_tile_index +1)], # gt_ds_batch_center, ## next_element_tt['gt_ds'], # gt_ds, ### gt, # [batch_size,2] tf placeholder
batch_weights = tf_batch_weights,
disp_diff_cap = DISP_DIFF_CAP,
disp_diff_slope= DISP_DIFF_SLOPE,
absolute_disparity = ABSOLUTE_DISPARITY,
use_confidence = USE_CONFIDENCE, # True,
lambda_conf_avg = 0.01,
## lambda_conf_pwr = 0.1,
conf_pwr = 2.0,
gt_conf_offset = 0.08,
gt_conf_pwr = 2.0,
error2_offset = 0, # 0.0025, # (0.05^2)
disp_wmin = 1.0, # minimal disparity to apply weight boosting for small disparities
disp_wmax = 8.0, # maximal disparity to apply weight boosting for small disparities
use_out = False) # use calculated disparity for disparity weight boosting (False - use target disparity)
G_loss = G_losses[0]
for n in range (1,len(partials)):
G_losses[n], _, _, _, _, _, _, _ = qcstereo_losses.batchLoss(
out_batch = outs[n], # [batch_size,(1..2)] tf_result
target_disparity_batch= target_disparity_batch, #next_element_tt['target_disparity'][:,center_tile_index:center_tile_index+1], # target_disparity_batch_center, # next_element_tt['target_disparity'], # target_disparity, ### target_d, # [batch_size] tf placeholder
gt_ds_batch = gt_ds_batch, # next_element_tt['gt_ds'][:,2 * center_tile_index: 2 * (center_tile_index +1)], # gt_ds_batch_center, ## next_element_tt['gt_ds'], # gt_ds, ### gt, # [batch_size,2] tf placeholder
batch_weights = tf_batch_weights,
disp_diff_cap = DISP_DIFF_CAP,
disp_diff_slope= DISP_DIFF_SLOPE,
absolute_disparity = ABSOLUTE_DISPARITY,
use_confidence = USE_CONFIDENCE, # True,
lambda_conf_avg = 0.01,
# lambda_conf_pwr = 0.1,
conf_pwr = 2.0,
gt_conf_offset = 0.08,
gt_conf_pwr = 2.0,
error2_offset = 0, # 0.0025, # (0.05^2)
disp_wmin = 1.0, # minimal disparity to apply weight boosting for small disparities
disp_wmax = 8.0, # maximal disparity to apply weight boosting for small disparities
use_out = False) # use calculated disparity for disparity weight boosting (False - use target disparity)
tf_wlosses = tf.multiply(G_losses, tf_partial_weights, name = "tf_wlosses")
G_losses_sum = tf.reduce_sum(tf_wlosses, name = "G_losses_sum")
if SLOSS_LAMBDA > 0:
S_loss, rslt_cost_nw, rslt_cost_w, rslt_d , rslt_avg_disparity, rslt_gt_disparity, rslt_offs = qcstereo_losses.smoothLoss(
out_batch = outs[0], # [batch_size,(1..2)] tf_result
target_disparity_batch = target_disparity_batch, # [batch_size] tf placeholder
gt_ds_batch_clust = gt_ds_batch_clust, # [batch_size,25,2] tf placeholder
clip = SLOSS_CLIP,
absolute_disparity = ABSOLUTE_DISPARITY, #when false there should be no activation on disparity output !
cluster_radius = CLUSTER_RADIUS)
GS_loss = tf.add(G_losses_sum, SLOSS_LAMBDA * S_loss, name = "GS_loss")
else:
S_loss = tf.constant(0.0, dtype=tf.float32,name = "S_loss")
GS_loss = G_losses_sum # G_loss
if WLOSS_LAMBDA > 0.0:
W_loss = qcstereo_losses.weightsLoss(
inp_weights = inp_weights[0], # inp_weights - list of tensors, currently - just [0]
tile_layers= TILE_LAYERS, # 4
tile_side = TILE_SIDE, # 9
wborders_zero = WBORDERS_ZERO)
GW_loss = tf.add(GS_loss, WLOSS_LAMBDA * W_loss, name = "GW_loss")
else:
GW_loss = GS_loss # G_loss
W_loss = tf.constant(0.0, dtype=tf.float32,name = "W_loss")
GT_variance = debug_gt_variance(indx = 0, # This tile index (0..8)
center_indx = 4, # center tile index
gt_ds_batch = next_element_tt['gt_ds'])# [?:18]
tf_ph_G_loss = tf.compat.v1.placeholder(tf.float32,shape=None,name='G_loss_avg')
tf_ph_G_losses = tf.compat.v1.placeholder(tf.float32,shape=[len(partials)],name='G_losses_avg')
tf_ph_S_loss = tf.compat.v1.placeholder(tf.float32,shape=None,name='S_loss_avg')
tf_ph_W_loss = tf.compat.v1.placeholder(tf.float32,shape=None,name='W_loss_avg')
tf_ph_GW_loss = tf.compat.v1.placeholder(tf.float32,shape=None,name='GW_loss_avg')
tf_ph_sq_diff = tf.compat.v1.placeholder(tf.float32,shape=None,name='sq_diff_avg')
tf_gtvar_diff = tf.compat.v1.placeholder(tf.float32,shape=None,name='gtvar_diff')
tf_img_test0 = tf.compat.v1.placeholder(tf.float32,shape=None,name='img_test0')
tf_img_test9 = tf.compat.v1.placeholder(tf.float32,shape=None,name='img_test9')
"""
with tf.name_scope('sample'):
tf.compat.v1.summary.scalar("GW_loss", GW_loss)
tf.compat.v1.summary.scalar("G_loss", G_loss)
tf.compat.v1.summary.scalar("S_loss", S_loss)
tf.compat.v1.summary.scalar("W_loss", W_loss)
tf.compat.v1.summary.scalar("sq_diff", _cost1)
tf.compat.v1.summary.scalar("gtvar_diff", GT_variance)
"""
with tf.name_scope('epoch_average'):
for i in range(tf_ph_G_losses.shape[0]):
tf.compat.v1.summary.scalar("G_loss_epoch_"+str(i), tf_ph_G_losses[i])
tf.compat.v1.summary.scalar("GW_loss_epoch", tf_ph_GW_loss)
tf.compat.v1.summary.scalar("G_loss_epoch", tf_ph_G_loss)
tf.compat.v1.summary.scalar("S_loss_epoch", tf_ph_S_loss)
tf.compat.v1.summary.scalar("W_loss_epoch", tf_ph_W_loss)
tf.compat.v1.summary.scalar("sq_diff_epoch", tf_ph_sq_diff)
tf.compat.v1.summary.scalar("gtvar_diff", tf_gtvar_diff)
tf.compat.v1.summary.scalar("img_test0", tf_img_test0)
tf.compat.v1.summary.scalar("img_test9", tf_img_test9)
trainable_vars= tf.trainable_variables()
lr= tf.compat.v1.placeholder(tf.float32)
G_opt= tf.compat.v1.train.AdamOptimizer(learning_rate=lr).minimize(GW_loss)
ROOT_PATH = './attic/nn_ds_neibs30_graph'+SUFFIX+"/" # for tensorboard
TT_SUBDIRS = ['train'] #,'test0','test1','test2','test3']
for i,_ in enumerate(datasets_test):
# TT_SUBDIRS.append('test%d'%(i))
TT_SUBDIRS.append(TEST_TITLES[i].replace(' ','_'))
TT_PATHS = [ROOT_PATH + p for p in TT_SUBDIRS]
# CLEAN OLD STAFF
shutil.rmtree(ROOT_PATH, ignore_errors=True)
#for p in TT_PATHS:
# shutil.rmtree(p, ignore_errors=True)
#seems that runs use directory creation time to order graphs
#for p in TT_PATHS:
# os.makedirs(p)
# time.sleep(1.5) # reduce later
num_train_subs = len(train_next) # number of (different type) merged training sets
dataset_train_size = TRAIN_BUFFER_GPU * num_train_subs # TRAIN_BUFFER_SIZE
tt_summaries = [0.0 for e in TT_SUBDIRS]
tt2_avg = [0.0 for e in TT_SUBDIRS]
tt_gw_avg = [0.0 for e in TT_SUBDIRS]
tt_g_avgs = [[0.0]*len(partials) for e in TT_SUBDIRS]
tt_w_avg = [0.0 for e in TT_SUBDIRS]
tt_s_avg = [0.0 for e in TT_SUBDIRS]
tt_gtvar_avg = [0.0 for e in TT_SUBDIRS]
saver=tf.compat.v1.train.Saver(trainable_vars)
saver_def = saver.as_saver_def()
# The name of the tensor you must feed with a filename when saving/restoring.
print ('saver_def.filename_tensor_name=',saver_def.filename_tensor_name)
# The name of the target operation you must run when restoring.
print ('saver_def.restore_op_name=',saver_def.restore_op_name)
# The name of the target operation you must run when saving.
print ('saver_def.save_tensor_name=',saver_def.save_tensor_name)
try:
os.makedirs(os.path.dirname(files['checkpoints']))
print ("Created directory ",os.path.dirname(files['checkpoints']))
# os.makedirs(files['checkpoints'])
except:
pass
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
merged = tf.summary.merge_all()
tt_writers = []
for p in TT_PATHS:
tt_writers.append(tf.summary.FileWriter(p, sess.graph))
print ("Adding delay to make directory creation time different: "+p)
time.sleep(2.0) # reduce later
loss_gw_train_hist= np.empty(dataset_train_size, dtype=np.float32)
loss_g_train_hists= [np.empty(dataset_train_size, dtype=np.float32) for p in partials]
loss_s_train_hist= np.empty(dataset_train_size, dtype=np.float32)
loss_w_train_hist= np.empty(dataset_train_size, dtype=np.float32)
loss_gw_test_hist= np.empty(dataset_test_size, dtype=np.float32)
loss_g_test_hists= [np.empty(dataset_test_size, dtype=np.float32) for p in partials]
loss_s_test_hist= np.empty(dataset_test_size, dtype=np.float32)
loss_w_test_hist= np.empty(dataset_test_size, dtype=np.float32)
loss2_train_hist= np.empty(dataset_train_size, dtype=np.float32)
loss2_test_hist= np.empty(dataset_test_size, dtype=np.float32)
gtvar_train_hist= np.empty(dataset_train_size, dtype=np.float32)
gtvar_test_hist= np.empty(dataset_test_size, dtype=np.float32)
gtvar_train = 0.0
gtvar_test = 0.0
img_gain_test0 = 1.0
img_gain_test9 = 1.0
thr=None
thr_result = None
trains_to_update = [train_next[n_train]['more_files'] for n_train in range(len(train_next))]
for epoch in range (EPOCHS_TO_RUN):
"""
update files after each epoch, all 4.
Convert to threads after testing
"""
if (FILE_UPDATE_EPOCHS > 0) and (epoch % FILE_UPDATE_EPOCHS == 0):
if not thr is None:
if thr.is_alive():
qsf.print_time("***WAITING*** until tfrecord gets loaded", end=" ")
else:
qsf.print_time("tfrecord is ***ALREADY LOADED*** ", end=" ")
thr.join()
qsf.print_time("Done")
qsf.print_time("Inserting new data", end=" ")
for n_train in range(len(trains_to_update)):
if trains_to_update[n_train]:
qsf.add_file_to_dataset(dataset = dataset_train,
new_dataset = thr_result[n_train],
train_next = train_next[n_train])
qsf.print_time("Done")
thr_result = []
fpaths = []
for n_train in range(len(trains_to_update)):
if trains_to_update[n_train]:
fpaths.append(files['train'][n_train][train_next[n_train]['file']])
qsf.print_time("Will read in background: "+fpaths[-1])
thr = Thread(target=qsf.getMoreFiles, args=(fpaths,thr_result, CLUSTER_RADIUS, HOR_FLIP, TILE_LAYERS, TILE_SIDE))
thr.start()
train_buf_index = epoch % TRAIN_BUFFER_CPU # GPU memory from CPU memory (now 4)
if epoch >=600:
learning_rate = LR600
elif epoch >=400:
learning_rate = LR400
elif epoch >=200:
learning_rate = LR200
elif epoch >=100:
learning_rate = LR100
else:
learning_rate = LR
if (train_buf_index == 0) and SHUFFLE_FILES:
qsf.print_time("Shuffling how datasets datasets_train_lvar and datasets_train_hvar are zipped together", end="")
qsf.shuffle_in_place(
dataset_data = dataset_train, #alternating clusters from 4 sources.each cluster has all needed data (concatenated)
period = num_train_subs)
qsf.print_time(" Done")
sti = train_buf_index * dataset_train_size * BATCH_SIZE # TRAIN_BUFFER_GPU * num_train_subs
eti = sti+ dataset_train_size * BATCH_SIZE# (train_buf_index +1) * TRAIN_BUFFER_GPU * num_train_subs
sess.run(iterator_tt.initializer, feed_dict={corr2d_train_placeholder: dataset_train[sti:eti,:corr2d_len],
target_disparity_train_placeholder: dataset_train[sti:eti,corr2d_len:corr2d_len+target_disparity_len],
gt_ds_train_placeholder: dataset_train[sti:eti,corr2d_len+target_disparity_len:corr2d_len+target_disparity_len+gtds_len] })
for i in range(dataset_train_size):
# try:
_, GW_loss_trained, G_losses_trained, S_loss_trained, W_loss_trained, output, disp_slice, d_gt_slice, out_diff, out_diff2, w_norm, out_wdiff2, out_cost1, gt_variance = sess.run(
[ G_opt,
GW_loss,
G_losses,
S_loss,
W_loss,
outs[0],
_disp_slice,
_d_gt_slice,
_out_diff,
_out_diff2,
_w_norm,
_out_wdiff2,
_cost1,
GT_variance
],
feed_dict={tf_batch_weights: feed_batch_weights,
lr: learning_rate
}) # previous value of *_avg #Fetch argument 0.0 has invalid type <class 'float'>, must be a string or Tensor. (Can not convert a float into a Tensor or Operation.)
loss_gw_train_hist[i] = GW_loss_trained
for nn, gl in enumerate(G_losses_trained):
loss_g_train_hists[nn][i] = gl
loss_s_train_hist[i] = S_loss_trained
loss_w_train_hist[i] = W_loss_trained
loss2_train_hist[i] = out_cost1
gtvar_train_hist[i] = gt_variance
# except tf.errors.OutOfRangeError:
# print("****** NO MORE DATA! train done at step %d"%(i))
# break
tt_gw_avg[0] = np.average(loss_gw_train_hist).astype(np.float32)
for nn, lgth in enumerate(loss_g_train_hists):
tt_g_avgs[0][nn] = np.average(lgth).astype(np.float32)
tt_s_avg[0] = np.average(loss_s_train_hist).astype(np.float32)
tt_w_avg[0] = np.average(loss_w_train_hist).astype(np.float32)
tt2_avg[0] = np.average(loss2_train_hist).astype(np.float32)
tt_gtvar_avg[0] = np.average(gtvar_train_hist).astype(np.float32)
for ntest,dataset_test in enumerate(datasets_test):
sess.run(iterator_tt.initializer, feed_dict={corr2d_train_placeholder: dataset_test[:, :corr2d_len], #['corr2d'],
target_disparity_train_placeholder: dataset_test[:, corr2d_len:corr2d_len+target_disparity_len], # ['target_disparity'],
gt_ds_train_placeholder: dataset_test[:, corr2d_len+target_disparity_len:] }) # ['gt_ds']})
"""
TODO: Make it possible to have different length dataset_test arrays to mix different length test files
"""
for i in range(dataset_test_size):
# for i in range(dataset_test.shape[0]):
# try:
GW_loss_tested, G_losses_tested, S_loss_tested, W_loss_tested, output, disp_slice, d_gt_slice, out_diff, out_diff2, w_norm, out_wdiff2, out_cost1, gt_variance = sess.run(
[GW_loss,
G_losses,
S_loss,
W_loss,
outs[0],
_disp_slice,
_d_gt_slice,
_out_diff,
_out_diff2,
_w_norm,
_out_wdiff2,
_cost1,
GT_variance
],
feed_dict={tf_batch_weights: feed_batch_weight_1 , # feed_batch_weights,
lr: learning_rate
}) # previous value of *_avg
loss_gw_test_hist[i] = GW_loss_tested
for nn, gl in enumerate(G_losses_tested):
loss_g_test_hists[nn][i] = gl
loss_s_test_hist[i] = S_loss_tested
loss_w_test_hist[i] = W_loss_tested
loss2_test_hist[i] = out_cost1
gtvar_test_hist[i] = gt_variance
# except tf.errors.OutOfRangeError:
# print("test done at step %d"%(i))
# break
tt_gw_avg[ntest+1] = np.average(loss_gw_test_hist).astype(np.float32)
for nn, lgth in enumerate(loss_g_test_hists):
tt_g_avgs[ntest+1][nn] = np.average(lgth).astype(np.float32)
tt_s_avg[ntest+1] = np.average(loss_s_test_hist).astype(np.float32)
tt_w_avg[ntest+1] = np.average(loss_w_test_hist).astype(np.float32)
tt2_avg[ntest+1] = np.average(loss2_test_hist).astype(np.float32)
tt_gtvar_avg[ntest+1] = np.average(gtvar_test_hist).astype(np.float32)
if (((epoch + 1) == EPOCHS_TO_RUN) or (((epoch + 1) % EPOCHS_FULL_TEST) == 0)) and (len(image_data) > 0) :
lf = None
if (epoch + 1) == EPOCHS_TO_RUN: # last
print("Last epoch, removing train/test datasets to reduce memory footprint")
del(dataset_train)
del(dataset_test)
if LOGPATH:
lf=open(LOGPATH,"w") #overwrite previous (or make it "a"?
last_epoch = (epoch + 1) == EPOCHS_TO_RUN
ind_img = [0]
if last_epoch:
ind_img = [i for i in range(len(image_data))]
###################################################
# Read the full image
###################################################
## test_summaries_img = [0.0]*len(ind_img) # datasets_img)
disp_out= np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_cost_nw= np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_cost_w= np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_d= np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_avg_disparity = np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_gt_disparity = np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_offs = np.empty((WIDTH*HEIGHT), dtype=np.float32)
for ntest in ind_img: # datasets_img):
dataset_img = qsf.readImageData(
image_data = image_data,
files = files,
indx = ntest,
cluster_radius = CLUSTER_RADIUS,
tile_layers = TILE_LAYERS,
tile_side = TILE_SIDE,
width = IMG_WIDTH,
replace_nans = True)
sess.run(iterator_tt.initializer, feed_dict={corr2d_train_placeholder: dataset_img['corr2d'],
target_disparity_train_placeholder: dataset_img['target_disparity'],
gt_ds_train_placeholder: dataset_img['gt_ds']})
for start_offs in range(0,disp_out.shape[0],BATCH_SIZE):
end_offs = min(start_offs+BATCH_SIZE,disp_out.shape[0])
# try:
output, cost_nw, cost_w, dd, avg_disparity, gt_disparity, offs = sess.run(
[outs[0], # {?,1]
rslt_cost_nw, #[?,]
rslt_cost_w, #[?,]
rslt_d, #[?,]
rslt_avg_disparity,
rslt_gt_disparity,
rslt_offs
],
feed_dict={
tf_batch_weights: feed_batch_weight_1 # feed_batch_weights,
}) # previous value of *_avg
# except tf.errors.OutOfRangeError:
# print("test done at step %d"%(i))
# break
# try:
disp_out[start_offs:end_offs] = output.flatten()
dbg_cost_nw[start_offs:end_offs] = cost_nw.flatten()
dbg_cost_w [start_offs:end_offs] = cost_w.flatten()
dbg_d[start_offs:end_offs] = dd.flatten()
dbg_avg_disparity[start_offs:end_offs] = avg_disparity.flatten()
dbg_gt_disparity[start_offs:end_offs] = gt_disparity.flatten()
dbg_offs[start_offs:end_offs] = offs.flatten()
# except ValueError:
# print("dataset_img_size= %d, i=%d, output.shape[0]=%d "%(dataset_img_size, i, output.shape[0]))
# break;
pass
result_file = files['result'][ntest] # result_files[ntest]
try:
os.makedirs(os.path.dirname(result_file))
except:
pass
rslt = np.concatenate(
[disp_out.reshape(-1,1),
dataset_img['t_disps'], #t_disps[ntest],
dataset_img['gtruths'], # gtruths[ntest],
dbg_cost_nw.reshape(-1,1),
dbg_cost_w.reshape(-1,1),
dbg_d.reshape(-1,1),
dbg_avg_disparity.reshape(-1,1),
dbg_gt_disparity.reshape(-1,1),
dbg_offs.reshape(-1,1)],1)
np.save(result_file, rslt.reshape(HEIGHT,WIDTH,-1))
rslt = qsf.eval_results(result_file, ABSOLUTE_DISPARITY, radius=CLUSTER_RADIUS, logfile=lf)
img_gain_test0 = rslt[0][0]/rslt[0][1]
img_gain_test9 = rslt[9][0]/rslt[9][1]
if SAVE_TIFFS:
qsf.result_npy_to_tiff(result_file, ABSOLUTE_DISPARITY, fix_nan = True,labels=SLICE_LABELS, logfile=lf)
"""
Remove dataset_img (if it is not [0] to reduce memory footprint
"""
if ntest > 0:
image_data[ntest] = None
if lf:
lf.close()
# tensorboard scalars
tt_summaries[0] = sess.run([merged],
feed_dict={ tf_ph_GW_loss: tt_gw_avg[0],
tf_ph_G_loss: tt_g_avgs[0][0], #train_g_avg,
tf_ph_G_losses: tt_g_avgs[0],
tf_ph_S_loss: tt_s_avg[0],
tf_ph_W_loss: tt_w_avg[0],
tf_ph_sq_diff: tt2_avg[0], # train2_avg,
tf_gtvar_diff: tt_gtvar_avg[0],
tf_img_test0: img_gain_test0,
tf_img_test9: img_gain_test9}) # previous value of *_avg #Fetch argument 0.0 has invalid type <class 'float'>, must be a string or Tensor. (Can not convert a float into a Tensor or Operation.)
for ntest, _ in enumerate(datasets_test):
tt_summaries[ntest+1] = sess.run([merged],
feed_dict={
tf_ph_GW_loss: tt_gw_avg[ntest+1],
tf_ph_G_loss: tt_g_avgs[ntest+1][0],
tf_ph_G_losses: tt_g_avgs[ntest+1], # train_g_avgs, # temporary, there is o data fro test
tf_ph_S_loss: tt_s_avg[ntest+1],
tf_ph_W_loss: tt_w_avg[ntest+1],
tf_ph_sq_diff: tt2_avg[ntest+1], #test2_avg,
tf_gtvar_diff: tt_gtvar_avg[ntest+1],
tf_img_test0: img_gain_test0,
tf_img_test9: img_gain_test9})
for n,tt_writer in enumerate(tt_writers):
## tt_writer.add_summary(tt_summaries[n],epoch)
tt_writer.add_summary(tt_summaries[n][0],epoch)
# if epoch ==0 :
# print ("adding delay to make directory creation time different")
# time.sleep(2.0) # reduce later
qsf.print_time("==== %04d:%03d -> %.4f %.4f %.4f %.4f %.4f (%.4f %.4f %.4f %.4f %.4f) ===="%(
epoch,i,
tt_gw_avg[0], tt_gw_avg[1], tt_gw_avg[2], tt_gw_avg[3], tt_gw_avg[4],
tt2_avg[0], tt2_avg[1], tt2_avg[2], tt2_avg[3], tt2_avg[4]))
if (not CHECKPOINT_PERIOD is None) and (((epoch + 1) % CHECKPOINT_PERIOD) == 0):
print("Saving periodic checkpoint (trained variables only) to %s, global_step = %d"%(os.path.dirname(files['checkpoints']), epoch),end=" => ")
print(saver.save(sess, files['checkpoints'], global_step=epoch, write_meta_graph=False))
# Close writers
for tt_writer in tt_writers:
try:
tt_writer.close()
except:
print ("Could not close tt_writer: ",tt_writer)
print("Saving final checkpoint (trained variables only) to %s"%(files['checkpoints']),end=" => ")
print(saver.save(sess, files["checkpoints"]))
print("All done")
exit (0)
"""
Traceback (most recent call last):
File "nn_ds_neibs30.py", line 721, in <module>
tt2_avg[0], tt2_avg[1], tt2_avg[2], tt2_avg[3], tt2_avg[4]))
ValueError: unsupported format character ' ' (0x20) at index 20
"""
\ No newline at end of file
nn_ds_neibs31.py 0 → 100644
View file @ b1111bca
#!/usr/bin/env python3
__copyright__ = "Copyright 2018-2019, Elphel, Inc."
__license__ = "GPL-3.0+"
__email__ = "andrey@elphel.com"
#python3 nn_ds_neibs31.py /data_ssd/lwir_sets/conf/qcstereo_lwir05.xml /data_ssd/lwir_sets/
#tensorboard --logdir="nn_ds_neibs30_graph13-9RNSWLAM0.5SLAM0.1SCLP0.2_nG_nI_HF_CP0.3_S0.03" --port=7001
import os
import sys
import numpy as np
import time
import shutil
from threading import Thread
import qcstereo_network
import qcstereo_losses
import qcstereo_functions as qsf
import tensorflow as tf
qsf.TIME_START = time.time()
qsf.TIME_LAST = qsf.TIME_START
IMG_WIDTH = 20 # 324 # tiles per image row
DEBUG_LEVEL= 1
try:
conf_file = sys.argv[1]
except IndexError:
print("Configuration path is required as a first argument. Optional second argument specifies root directory for data files")
exit(1)
try:
root_dir = sys.argv[2]
except IndexError:
root_dir = os.path.dirname(conf_file)
print ("Configuration file: " + conf_file)
parameters, dirs, files, _ = qsf.parseXmlConfig(conf_file, root_dir)
"""
Temporarily for backward compatibility
"""
if not "SLOSS_CLIP" in parameters:
parameters['SLOSS_CLIP'] = 0.5
print ("Old config, setting SLOSS_CLIP=", parameters['SLOSS_CLIP'])
"""
Defined in config file
"""
TILE_SIDE, TILE_LAYERS, TWO_TRAINS, NET_ARCH1, NET_ARCH2 = [None]*5
ABSOLUTE_DISPARITY,SYM8_SUB, WLOSS_LAMBDA, SLOSS_LAMBDA, SLOSS_CLIP = [None]*5
SPREAD_CONVERGENCE, INTER_CONVERGENCE, HOR_FLIP, DISP_DIFF_CAP, DISP_DIFF_SLOPE = [None]*5
CLUSTER_RADIUS = None
PARTIALS_WEIGHTS, MAX_IMGS_IN_MEM, MAX_FILES_PER_GROUP, BATCH_WEIGHTS, ONLY_TILE = [None] * 5
USE_CONFIDENCE, WBORDERS_ZERO, EPOCHS_TO_RUN, FILE_UPDATE_EPOCHS = [None] * 4
LR600,LR400,LR200,LR100,LR = [None]*5
SHUFFLE_FILES, EPOCHS_FULL_TEST, SAVE_TIFFS = [None] * 3
CHECKPOINT_PERIOD = None
TRAIN_BUFFER_GPU, TRAIN_BUFFER_CPU = [None]*2
TEST_TITLES = None
LOGFILE="results.txt"
"""
Next gets globals from the config file
"""
globals().update(parameters)
TRAIN_BUFFER_SIZE = TRAIN_BUFFER_GPU * TRAIN_BUFFER_CPU # in merged (quad) batches
#exit(0)
WIDTH = 20 # 324
HEIGHT = 15 # 242
TILE_SIZE = TILE_SIDE* TILE_SIDE # == 81
FEATURES_PER_TILE = TILE_LAYERS * TILE_SIZE# == 324
BATCH_SIZE = ([1,2][TWO_TRAINS])*2*1000//25 # == 80 Each batch of tiles has balanced D/S tiles, shuffled batches but not inside batches
SUFFIX=(str(NET_ARCH1)+'-'+str(NET_ARCH2)+
(["R","A"][ABSOLUTE_DISPARITY]) +
(["NS","S8"][SYM8_SUB])+
"WLAM"+str(WLOSS_LAMBDA)+
"SLAM"+str(SLOSS_LAMBDA)+
"SCLP"+str(SLOSS_CLIP)+
(['_nG','_G'][SPREAD_CONVERGENCE])+
(['_nI','_I'][INTER_CONVERGENCE]) +
(['_nHF',"_HF"][HOR_FLIP]) +
('_CP'+str(DISP_DIFF_CAP)) +
('_S'+str(DISP_DIFF_SLOPE))
)
NN_LAYOUT1 = qcstereo_network.NN_LAYOUTS[NET_ARCH1]
NN_LAYOUT2 = qcstereo_network.NN_LAYOUTS[NET_ARCH2]
USE_PARTIALS = not PARTIALS_WEIGHTS is None # False - just a single Siamese net, True - partial outputs that use concentric squares of the first level subnets
# Tiff export slice labels
SLICE_LABELS = ["nn_out_ext","hier_out_ext","gt_disparity","gt_strength",
"cutcorn_cost_nw","cutcorn_cost",
"gt-avg_dist","avg8_disp","gt_disp","out-avg"]
##############################################################################
cluster_size = (2 * CLUSTER_RADIUS + 1) * (2 * CLUSTER_RADIUS + 1)
center_tile_index = 2 * CLUSTER_RADIUS * (CLUSTER_RADIUS + 1)
qsf.prepareFiles(dirs, files, suffix = SUFFIX)
#copy config to results directory
print ("Copying config files to results directory:\n ('%s' -> '%s')"%(conf_file,dirs['result']))
try:
os.makedirs(dirs['result'])
except:
pass
shutil.copy2(conf_file,dirs['result'])
LOGPATH = os.path.join(dirs['result'],LOGFILE)
if TEST_TITLES is None:
TEST_TITLES = qsf.defaultTestTitles(files)
partials = None
partials = qsf.concentricSquares(CLUSTER_RADIUS)
PARTIALS_WEIGHTS = [1.0*pw/sum(PARTIALS_WEIGHTS) for pw in PARTIALS_WEIGHTS]
if not USE_PARTIALS:
partials = partials[0:1]
PARTIALS_WEIGHTS = [1.0]
qsf.evaluateAllResults(result_files = files['result'],
absolute_disparity = ABSOLUTE_DISPARITY,
cluster_radius = CLUSTER_RADIUS,
labels = SLICE_LABELS,
logpath= LOGPATH)
image_data = qsf.initImageData(
files = files,
max_imgs = MAX_IMGS_IN_MEM,
cluster_radius = CLUSTER_RADIUS,
tile_layers = TILE_LAYERS,
tile_side = TILE_SIDE,
width = IMG_WIDTH,
replace_nans = True)
corr2d_len, target_disparity_len, gtds_len, _ = qsf.get_lengths(CLUSTER_RADIUS, TILE_LAYERS, TILE_SIDE)
train_next, dataset_train, datasets_test= qsf.initTrainTestData(
files = files,
cluster_radius = CLUSTER_RADIUS,
buffer_size = TRAIN_BUFFER_SIZE * BATCH_SIZE, # number of clusters per train
test_titles = TEST_TITLES)
corr2d_train_placeholder = tf.compat.v1.placeholder(dataset_train.dtype, (None,FEATURES_PER_TILE * cluster_size)) # corr2d_train.shape)
target_disparity_train_placeholder = tf.compat.v1.placeholder(dataset_train.dtype, (None,1 * cluster_size)) #target_disparity_train.shape)
gt_ds_train_placeholder = tf.compat.v1.placeholder(dataset_train.dtype, (None,2 * cluster_size)) #gt_ds_train.shape)
dataset_tt = tf.data.Dataset.from_tensor_slices({
"corr2d": corr2d_train_placeholder,
"target_disparity": target_disparity_train_placeholder,
"gt_ds": gt_ds_train_placeholder})
tf_batch_weights = tf.compat.v1.placeholder(shape=(None,), dtype=tf.float32, name = "batch_weights") # way to increase importance of the high variance clusters
feed_batch_weights = np.array(BATCH_WEIGHTS*(BATCH_SIZE//len(BATCH_WEIGHTS)), dtype=np.float32)
feed_batch_weight_1 = np.array([1.0], dtype=np.float32)
dataset_test_size = len(datasets_test[0])
dataset_test_size //= BATCH_SIZE
dataset_img_size = len(image_data[0]['corr2d'])
dataset_img_size //= BATCH_SIZE
dataset_tt = dataset_tt.batch(BATCH_SIZE)
dataset_tt = dataset_tt.prefetch(BATCH_SIZE)
iterator_tt = dataset_tt.make_initializable_iterator()
next_element_tt = iterator_tt.get_next()
result_dir = './attic/result_neibs_'+ SUFFIX+'/'
checkpoint_dir = './attic/result_neibs_'+ SUFFIX+'/'
save_freq = 500
def debug_gt_variance(
indx, # This tile index (0..8)
center_indx, # center tile index
gt_ds_batch # [?:9:2]
):
with tf.name_scope("Debug_GT_Variance"):
d_gt_this = tf.reshape(gt_ds_batch[:,2 * indx],[-1], name = "d_this")
d_gt_center = tf.reshape(gt_ds_batch[:,2 * center_indx],[-1], name = "d_center")
d_gt_diff = tf.subtract(d_gt_this, d_gt_center, name = "d_diff")
d_gt_diff2 = tf.multiply(d_gt_diff, d_gt_diff, name = "d_diff2")
d_gt_var = tf.reduce_mean(d_gt_diff2, name = "d_gt_var")
return d_gt_var
target_disparity_cluster = tf.reshape(next_element_tt['target_disparity'], [-1,cluster_size, 1], name="targdisp_cluster")
corr2d_Nx325 = tf.concat([tf.reshape(next_element_tt['corr2d'],[-1,cluster_size,FEATURES_PER_TILE], name="coor2d_cluster"),
target_disparity_cluster], axis=2, name = "corr2d_Nx325")
if SPREAD_CONVERGENCE:
outs, inp_weights = qcstereo_network.networks_siam(
input = corr2d_Nx325,
input_global = target_disparity_cluster,
layout1 = NN_LAYOUT1,
layout2 = NN_LAYOUT2,
inter_convergence = INTER_CONVERGENCE,
sym8 = SYM8_SUB,
only_tile = ONLY_TILE, #Remove/put None for normal operation
partials = partials,
use_confidence= USE_CONFIDENCE,
cluster_radius = CLUSTER_RADIUS)
else:
outs, inp_weights = qcstereo_network.networks_siam(
input_tensor= corr2d_Nx325,
input_global = None,
layout1 = NN_LAYOUT1,
layout2 = NN_LAYOUT2,
inter_convergence = False,
sym8 = SYM8_SUB,
only_tile = ONLY_TILE, #Remove/put None for normal operation
partials = partials,
use_confidence= USE_CONFIDENCE,
cluster_radius = CLUSTER_RADIUS)
tf_partial_weights = tf.constant(PARTIALS_WEIGHTS,dtype=tf.float32,name="partial_weights")
G_losses = [0.0]*len(partials)
target_disparity_batch= next_element_tt['target_disparity'][:,center_tile_index:center_tile_index+1]
gt_ds_batch_clust = next_element_tt['gt_ds']
gt_ds_batch = gt_ds_batch_clust[:,2 * center_tile_index: 2 * (center_tile_index +1)]
G_losses[0], _disp_slice, _d_gt_slice, _out_diff, _out_diff2, _w_norm, _out_wdiff2, _cost1 = qcstereo_losses.batchLoss(
out_batch = outs[0], # [batch_size,(1..2)] tf_result
target_disparity_batch= target_disparity_batch, # next_element_tt['target_disparity'][:,center_tile_index:center_tile_index+1], # target_disparity_batch_center, # next_element_tt['target_disparity'], # target_disparity, ### target_d, # [batch_size] tf placeholder
gt_ds_batch = gt_ds_batch, # next_element_tt['gt_ds'][:,2 * center_tile_index: 2 * (center_tile_index +1)], # gt_ds_batch_center, ## next_element_tt['gt_ds'], # gt_ds, ### gt, # [batch_size,2] tf placeholder
batch_weights = tf_batch_weights,
disp_diff_cap = DISP_DIFF_CAP,
disp_diff_slope= DISP_DIFF_SLOPE,
absolute_disparity = ABSOLUTE_DISPARITY,
use_confidence = USE_CONFIDENCE, # True,
lambda_conf_avg = 0.01,
## lambda_conf_pwr = 0.1,
conf_pwr = 2.0,
gt_conf_offset = 0.08,
gt_conf_pwr = 2.0,
error2_offset = 0, # 0.0025, # (0.05^2)
disp_wmin = 1.0, # minimal disparity to apply weight boosting for small disparities
disp_wmax = 8.0, # maximal disparity to apply weight boosting for small disparities
use_out = False) # use calculated disparity for disparity weight boosting (False - use target disparity)
G_loss = G_losses[0]
for n in range (1,len(partials)):
G_losses[n], _, _, _, _, _, _, _ = qcstereo_losses.batchLoss(
out_batch = outs[n], # [batch_size,(1..2)] tf_result
target_disparity_batch= target_disparity_batch, #next_element_tt['target_disparity'][:,center_tile_index:center_tile_index+1], # target_disparity_batch_center, # next_element_tt['target_disparity'], # target_disparity, ### target_d, # [batch_size] tf placeholder
gt_ds_batch = gt_ds_batch, # next_element_tt['gt_ds'][:,2 * center_tile_index: 2 * (center_tile_index +1)], # gt_ds_batch_center, ## next_element_tt['gt_ds'], # gt_ds, ### gt, # [batch_size,2] tf placeholder
batch_weights = tf_batch_weights,
disp_diff_cap = DISP_DIFF_CAP,
disp_diff_slope= DISP_DIFF_SLOPE,
absolute_disparity = ABSOLUTE_DISPARITY,
use_confidence = USE_CONFIDENCE, # True,
lambda_conf_avg = 0.01,
# lambda_conf_pwr = 0.1,
conf_pwr = 2.0,
gt_conf_offset = 0.08,
gt_conf_pwr = 2.0,
error2_offset = 0, # 0.0025, # (0.05^2)
disp_wmin = 1.0, # minimal disparity to apply weight boosting for small disparities
disp_wmax = 8.0, # maximal disparity to apply weight boosting for small disparities
use_out = False) # use calculated disparity for disparity weight boosting (False - use target disparity)
tf_wlosses = tf.multiply(G_losses, tf_partial_weights, name = "tf_wlosses")
G_losses_sum = tf.reduce_sum(tf_wlosses, name = "G_losses_sum")
if SLOSS_LAMBDA > 0:
S_loss, rslt_cost_nw, rslt_cost_w, rslt_d , rslt_avg_disparity, rslt_gt_disparity, rslt_offs = qcstereo_losses.smoothLoss(
out_batch = outs[0], # [batch_size,(1..2)] tf_result
target_disparity_batch = target_disparity_batch, # [batch_size] tf placeholder
gt_ds_batch_clust = gt_ds_batch_clust, # [batch_size,25,2] tf placeholder
clip = SLOSS_CLIP,
absolute_disparity = ABSOLUTE_DISPARITY, #when false there should be no activation on disparity output !
cluster_radius = CLUSTER_RADIUS)
GS_loss = tf.add(G_losses_sum, SLOSS_LAMBDA * S_loss, name = "GS_loss")
else:
S_loss = tf.constant(0.0, dtype=tf.float32,name = "S_loss")
GS_loss = G_losses_sum # G_loss
if WLOSS_LAMBDA > 0.0:
W_loss = qcstereo_losses.weightsLoss(
inp_weights = inp_weights[0], # inp_weights - list of tensors, currently - just [0]
tile_layers= TILE_LAYERS, # 4
tile_side = TILE_SIDE, # 9
wborders_zero = WBORDERS_ZERO)
GW_loss = tf.add(GS_loss, WLOSS_LAMBDA * W_loss, name = "GW_loss")
else:
GW_loss = GS_loss # G_loss
W_loss = tf.constant(0.0, dtype=tf.float32,name = "W_loss")
GT_variance = debug_gt_variance(indx = 0, # This tile index (0..8)
center_indx = 4, # center tile index
gt_ds_batch = next_element_tt['gt_ds'])# [?:18]
tf_ph_G_loss = tf.compat.v1.placeholder(tf.float32,shape=None,name='G_loss_avg')
tf_ph_G_losses = tf.compat.v1.placeholder(tf.float32,shape=[len(partials)],name='G_losses_avg')
tf_ph_S_loss = tf.compat.v1.placeholder(tf.float32,shape=None,name='S_loss_avg')
tf_ph_W_loss = tf.compat.v1.placeholder(tf.float32,shape=None,name='W_loss_avg')
tf_ph_GW_loss = tf.compat.v1.placeholder(tf.float32,shape=None,name='GW_loss_avg')
tf_ph_sq_diff = tf.compat.v1.placeholder(tf.float32,shape=None,name='sq_diff_avg')
tf_gtvar_diff = tf.compat.v1.placeholder(tf.float32,shape=None,name='gtvar_diff')
tf_img_test0 = tf.compat.v1.placeholder(tf.float32,shape=None,name='img_test0')
tf_img_test9 = tf.compat.v1.placeholder(tf.float32,shape=None,name='img_test9')
"""
with tf.name_scope('sample'):
tf.compat.v1.summary.scalar("GW_loss", GW_loss)
tf.compat.v1.summary.scalar("G_loss", G_loss)
tf.compat.v1.summary.scalar("S_loss", S_loss)
tf.compat.v1.summary.scalar("W_loss", W_loss)
tf.compat.v1.summary.scalar("sq_diff", _cost1)
tf.compat.v1.summary.scalar("gtvar_diff", GT_variance)
"""
with tf.name_scope('epoch_average'):
for i in range(tf_ph_G_losses.shape[0]):
tf.compat.v1.summary.scalar("G_loss_epoch_"+str(i), tf_ph_G_losses[i])
tf.compat.v1.summary.scalar("GW_loss_epoch", tf_ph_GW_loss)
tf.compat.v1.summary.scalar("G_loss_epoch", tf_ph_G_loss)
tf.compat.v1.summary.scalar("S_loss_epoch", tf_ph_S_loss)
tf.compat.v1.summary.scalar("W_loss_epoch", tf_ph_W_loss)
tf.compat.v1.summary.scalar("sq_diff_epoch", tf_ph_sq_diff)
tf.compat.v1.summary.scalar("gtvar_diff", tf_gtvar_diff)
tf.compat.v1.summary.scalar("img_test0", tf_img_test0)
tf.compat.v1.summary.scalar("img_test9", tf_img_test9)
trainable_vars= tf.trainable_variables()
lr= tf.compat.v1.placeholder(tf.float32)
G_opt= tf.compat.v1.train.AdamOptimizer(learning_rate=lr).minimize(GW_loss)
ROOT_PATH = './attic/nn_ds_neibs30_graph'+SUFFIX+"/" # for tensorboard
TT_SUBDIRS = ['train'] #,'test0','test1','test2','test3']
for i,_ in enumerate(datasets_test):
# TT_SUBDIRS.append('test%d'%(i))
TT_SUBDIRS.append(TEST_TITLES[i].replace(' ','_'))
TT_PATHS = [ROOT_PATH + p for p in TT_SUBDIRS]
# CLEAN OLD STAFF
shutil.rmtree(ROOT_PATH, ignore_errors=True)
#for p in TT_PATHS:
# shutil.rmtree(p, ignore_errors=True)
#seems that runs use directory creation time to order graphs
#for p in TT_PATHS:
# os.makedirs(p)
# time.sleep(1.5) # reduce later
num_train_subs = len(train_next) # number of (different type) merged training sets
dataset_train_size = TRAIN_BUFFER_GPU * num_train_subs # TRAIN_BUFFER_SIZE
tt_summaries = [0.0 for e in TT_SUBDIRS]
tt2_avg = [0.0 for e in TT_SUBDIRS]
tt_gw_avg = [0.0 for e in TT_SUBDIRS]
tt_g_avgs = [[0.0]*len(partials) for e in TT_SUBDIRS]
tt_w_avg = [0.0 for e in TT_SUBDIRS]
tt_s_avg = [0.0 for e in TT_SUBDIRS]
tt_gtvar_avg = [0.0 for e in TT_SUBDIRS]
saver=tf.compat.v1.train.Saver(trainable_vars)
saver_def = saver.as_saver_def()
# The name of the tensor you must feed with a filename when saving/restoring.
print ('saver_def.filename_tensor_name=',saver_def.filename_tensor_name)
# The name of the target operation you must run when restoring.
print ('saver_def.restore_op_name=',saver_def.restore_op_name)
# The name of the target operation you must run when saving.
print ('saver_def.save_tensor_name=',saver_def.save_tensor_name)
try:
os.makedirs(os.path.dirname(files['checkpoints']))
print ("Created directory ",os.path.dirname(files['checkpoints']))
# os.makedirs(files['checkpoints'])
except:
pass
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
merged = tf.summary.merge_all()
tt_writers = []
for p in TT_PATHS:
tt_writers.append(tf.summary.FileWriter(p, sess.graph))
print ("Adding delay to make directory creation time different: "+p)
time.sleep(2.0) # reduce later
loss_gw_train_hist= np.empty(dataset_train_size, dtype=np.float32)
loss_g_train_hists= [np.empty(dataset_train_size, dtype=np.float32) for p in partials]
loss_s_train_hist= np.empty(dataset_train_size, dtype=np.float32)
loss_w_train_hist= np.empty(dataset_train_size, dtype=np.float32)
loss_gw_test_hist= np.empty(dataset_test_size, dtype=np.float32)
loss_g_test_hists= [np.empty(dataset_test_size, dtype=np.float32) for p in partials]
loss_s_test_hist= np.empty(dataset_test_size, dtype=np.float32)
loss_w_test_hist= np.empty(dataset_test_size, dtype=np.float32)
loss2_train_hist= np.empty(dataset_train_size, dtype=np.float32)
loss2_test_hist= np.empty(dataset_test_size, dtype=np.float32)
gtvar_train_hist= np.empty(dataset_train_size, dtype=np.float32)
gtvar_test_hist= np.empty(dataset_test_size, dtype=np.float32)
gtvar_train = 0.0
gtvar_test = 0.0
img_gain_test0 = 1.0
img_gain_test9 = 1.0
thr=None
thr_result = None
trains_to_update = [train_next[n_train]['more_files'] for n_train in range(len(train_next))]
for epoch in range (EPOCHS_TO_RUN):
"""
update files after each epoch, all 4.
Convert to threads after testing
"""
if (FILE_UPDATE_EPOCHS > 0) and (epoch % FILE_UPDATE_EPOCHS == 0):
if not thr is None:
if thr.is_alive():
qsf.print_time("***WAITING*** until tfrecord gets loaded", end=" ")
else:
qsf.print_time("tfrecord is ***ALREADY LOADED*** ", end=" ")
thr.join()
qsf.print_time("Done")
qsf.print_time("Inserting new data", end=" ")
for n_train in range(len(trains_to_update)):
if trains_to_update[n_train]:
qsf.add_file_to_dataset(dataset = dataset_train,
new_dataset = thr_result[n_train],
train_next = train_next[n_train])
qsf.print_time("Done")
thr_result = []
fpaths = []
for n_train in range(len(trains_to_update)):
if trains_to_update[n_train]:
fpaths.append(files['train'][n_train][train_next[n_train]['file']])
qsf.print_time("Will read in background: "+fpaths[-1])
thr = Thread(target=qsf.getMoreFiles, args=(fpaths,thr_result, CLUSTER_RADIUS, HOR_FLIP, TILE_LAYERS, TILE_SIDE))
thr.start()
train_buf_index = epoch % TRAIN_BUFFER_CPU # GPU memory from CPU memory (now 4)
if epoch >=600:
learning_rate = LR600
elif epoch >=400:
learning_rate = LR400
elif epoch >=200:
learning_rate = LR200
elif epoch >=100:
learning_rate = LR100
else:
learning_rate = LR
if (train_buf_index == 0) and SHUFFLE_FILES:
qsf.print_time("Shuffling how datasets datasets_train_lvar and datasets_train_hvar are zipped together", end="")
qsf.shuffle_in_place(
dataset_data = dataset_train, #alternating clusters from 4 sources.each cluster has all needed data (concatenated)
period = num_train_subs)
qsf.print_time(" Done")
sti = train_buf_index * dataset_train_size * BATCH_SIZE # TRAIN_BUFFER_GPU * num_train_subs
eti = sti+ dataset_train_size * BATCH_SIZE# (train_buf_index +1) * TRAIN_BUFFER_GPU * num_train_subs
sess.run(iterator_tt.initializer, feed_dict={corr2d_train_placeholder: dataset_train[sti:eti,:corr2d_len],
target_disparity_train_placeholder: dataset_train[sti:eti,corr2d_len:corr2d_len+target_disparity_len],
gt_ds_train_placeholder: dataset_train[sti:eti,corr2d_len+target_disparity_len:corr2d_len+target_disparity_len+gtds_len] })
for i in range(dataset_train_size):
# try:
_, GW_loss_trained, G_losses_trained, S_loss_trained, W_loss_trained, output, disp_slice, d_gt_slice, out_diff, out_diff2, w_norm, out_wdiff2, out_cost1, gt_variance = sess.run(
[ G_opt,
GW_loss,
G_losses,
S_loss,
W_loss,
outs[0],
_disp_slice,
_d_gt_slice,
_out_diff,
_out_diff2,
_w_norm,
_out_wdiff2,
_cost1,
GT_variance
],
feed_dict={tf_batch_weights: feed_batch_weights,
lr: learning_rate
}) # previous value of *_avg #Fetch argument 0.0 has invalid type <class 'float'>, must be a string or Tensor. (Can not convert a float into a Tensor or Operation.)
loss_gw_train_hist[i] = GW_loss_trained
for nn, gl in enumerate(G_losses_trained):
loss_g_train_hists[nn][i] = gl
loss_s_train_hist[i] = S_loss_trained
loss_w_train_hist[i] = W_loss_trained
loss2_train_hist[i] = out_cost1
gtvar_train_hist[i] = gt_variance
# except tf.errors.OutOfRangeError:
# print("****** NO MORE DATA! train done at step %d"%(i))
# break
tt_gw_avg[0] = np.average(loss_gw_train_hist).astype(np.float32)
for nn, lgth in enumerate(loss_g_train_hists):
tt_g_avgs[0][nn] = np.average(lgth).astype(np.float32)
tt_s_avg[0] = np.average(loss_s_train_hist).astype(np.float32)
tt_w_avg[0] = np.average(loss_w_train_hist).astype(np.float32)
tt2_avg[0] = np.average(loss2_train_hist).astype(np.float32)
tt_gtvar_avg[0] = np.average(gtvar_train_hist).astype(np.float32)
for ntest,dataset_test in enumerate(datasets_test):
sess.run(iterator_tt.initializer, feed_dict={corr2d_train_placeholder: dataset_test[:, :corr2d_len], #['corr2d'],
target_disparity_train_placeholder: dataset_test[:, corr2d_len:corr2d_len+target_disparity_len], # ['target_disparity'],
gt_ds_train_placeholder: dataset_test[:, corr2d_len+target_disparity_len:] }) # ['gt_ds']})
"""
TODO: Make it possible to have different length dataset_test arrays to mix different length test files
"""
for i in range(dataset_test_size):
# for i in range(dataset_test.shape[0]):
# try:
GW_loss_tested, G_losses_tested, S_loss_tested, W_loss_tested, output, disp_slice, d_gt_slice, out_diff, out_diff2, w_norm, out_wdiff2, out_cost1, gt_variance = sess.run(
[GW_loss,
G_losses,
S_loss,
W_loss,
outs[0],
_disp_slice,
_d_gt_slice,
_out_diff,
_out_diff2,
_w_norm,
_out_wdiff2,
_cost1,
GT_variance
],
feed_dict={tf_batch_weights: feed_batch_weight_1 , # feed_batch_weights,
lr: learning_rate
}) # previous value of *_avg
loss_gw_test_hist[i] = GW_loss_tested
for nn, gl in enumerate(G_losses_tested):
loss_g_test_hists[nn][i] = gl
loss_s_test_hist[i] = S_loss_tested
loss_w_test_hist[i] = W_loss_tested
loss2_test_hist[i] = out_cost1
gtvar_test_hist[i] = gt_variance
# except tf.errors.OutOfRangeError:
# print("test done at step %d"%(i))
# break
tt_gw_avg[ntest+1] = np.average(loss_gw_test_hist).astype(np.float32)
for nn, lgth in enumerate(loss_g_test_hists):
tt_g_avgs[ntest+1][nn] = np.average(lgth).astype(np.float32)
tt_s_avg[ntest+1] = np.average(loss_s_test_hist).astype(np.float32)
tt_w_avg[ntest+1] = np.average(loss_w_test_hist).astype(np.float32)
tt2_avg[ntest+1] = np.average(loss2_test_hist).astype(np.float32)
tt_gtvar_avg[ntest+1] = np.average(gtvar_test_hist).astype(np.float32)
if (((epoch + 1) == EPOCHS_TO_RUN) or (((epoch + 1) % EPOCHS_FULL_TEST) == 0)) and (len(image_data) > 0) :
lf = None
if (epoch + 1) == EPOCHS_TO_RUN: # last
print("Last epoch, removing train/test datasets to reduce memory footprint")
del(dataset_train)
del(dataset_test)
if LOGPATH:
lf=open(LOGPATH,"w") #overwrite previous (or make it "a"?
last_epoch = (epoch + 1) == EPOCHS_TO_RUN
ind_img = [0]
if last_epoch:
ind_img = [i for i in range(len(image_data))]
###################################################
# Read the full image
###################################################
## test_summaries_img = [0.0]*len(ind_img) # datasets_img)
disp_out= np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_cost_nw= np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_cost_w= np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_d= np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_avg_disparity = np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_gt_disparity = np.empty((WIDTH*HEIGHT), dtype=np.float32)
dbg_offs = np.empty((WIDTH*HEIGHT), dtype=np.float32)
for ntest in ind_img: # datasets_img):
dataset_img = qsf.readImageData(
image_data = image_data,
files = files,
indx = ntest,
cluster_radius = CLUSTER_RADIUS,
tile_layers = TILE_LAYERS,
tile_side = TILE_SIDE,
width = IMG_WIDTH,
replace_nans = True)
sess.run(iterator_tt.initializer, feed_dict={corr2d_train_placeholder: dataset_img['corr2d'],
target_disparity_train_placeholder: dataset_img['target_disparity'],
gt_ds_train_placeholder: dataset_img['gt_ds']})
for start_offs in range(0,disp_out.shape[0],BATCH_SIZE):
end_offs = min(start_offs+BATCH_SIZE,disp_out.shape[0])
# try:
output, cost_nw, cost_w, dd, avg_disparity, gt_disparity, offs = sess.run(
[outs[0], # {?,1]
rslt_cost_nw, #[?,]
rslt_cost_w, #[?,]
rslt_d, #[?,]
rslt_avg_disparity,
rslt_gt_disparity,
rslt_offs
],
feed_dict={
tf_batch_weights: feed_batch_weight_1 # feed_batch_weights,
}) # previous value of *_avg
# except tf.errors.OutOfRangeError:
# print("test done at step %d"%(i))
# break
# try:
disp_out[start_offs:end_offs] = output.flatten()
dbg_cost_nw[start_offs:end_offs] = cost_nw.flatten()
dbg_cost_w [start_offs:end_offs] = cost_w.flatten()
dbg_d[start_offs:end_offs] = dd.flatten()
dbg_avg_disparity[start_offs:end_offs] = avg_disparity.flatten()
dbg_gt_disparity[start_offs:end_offs] = gt_disparity.flatten()
dbg_offs[start_offs:end_offs] = offs.flatten()
# except ValueError:
# print("dataset_img_size= %d, i=%d, output.shape[0]=%d "%(dataset_img_size, i, output.shape[0]))
# break;
pass
result_file = files['result'][ntest] # result_files[ntest]
try:
os.makedirs(os.path.dirname(result_file))
except:
pass
rslt = np.concatenate(
[disp_out.reshape(-1,1),
dataset_img['t_disps'], #t_disps[ntest],
dataset_img['gtruths'], # gtruths[ntest],
dbg_cost_nw.reshape(-1,1),
dbg_cost_w.reshape(-1,1),
dbg_d.reshape(-1,1),
dbg_avg_disparity.reshape(-1,1),
dbg_gt_disparity.reshape(-1,1),
dbg_offs.reshape(-1,1)],1)
np.save(result_file, rslt.reshape(HEIGHT,WIDTH,-1))
rslt = qsf.eval_results(result_file, ABSOLUTE_DISPARITY, radius=CLUSTER_RADIUS, logfile=lf)
img_gain_test0 = rslt[0][0]/rslt[0][1]
img_gain_test9 = rslt[9][0]/rslt[9][1]
if SAVE_TIFFS:
qsf.result_npy_to_tiff(result_file, ABSOLUTE_DISPARITY, fix_nan = True,labels=SLICE_LABELS, logfile=lf)
"""
Remove dataset_img (if it is not [0] to reduce memory footprint
"""
if ntest > 0:
image_data[ntest] = None
if lf:
lf.close()
# tensorboard scalars
tt_summaries[0] = sess.run([merged],
feed_dict={ tf_ph_GW_loss: tt_gw_avg[0],
tf_ph_G_loss: tt_g_avgs[0][0], #train_g_avg,
tf_ph_G_losses: tt_g_avgs[0],
tf_ph_S_loss: tt_s_avg[0],
tf_ph_W_loss: tt_w_avg[0],
tf_ph_sq_diff: tt2_avg[0], # train2_avg,
tf_gtvar_diff: tt_gtvar_avg[0],
tf_img_test0: img_gain_test0,
tf_img_test9: img_gain_test9}) # previous value of *_avg #Fetch argument 0.0 has invalid type <class 'float'>, must be a string or Tensor. (Can not convert a float into a Tensor or Operation.)
for ntest, _ in enumerate(datasets_test):
tt_summaries[ntest+1] = sess.run([merged],
feed_dict={
tf_ph_GW_loss: tt_gw_avg[ntest+1],
tf_ph_G_loss: tt_g_avgs[ntest+1][0],
tf_ph_G_losses: tt_g_avgs[ntest+1], # train_g_avgs, # temporary, there is o data fro test
tf_ph_S_loss: tt_s_avg[ntest+1],
tf_ph_W_loss: tt_w_avg[ntest+1],
tf_ph_sq_diff: tt2_avg[ntest+1], #test2_avg,
tf_gtvar_diff: tt_gtvar_avg[ntest+1],
tf_img_test0: img_gain_test0,
tf_img_test9: img_gain_test9})
for n,tt_writer in enumerate(tt_writers):
## tt_writer.add_summary(tt_summaries[n],epoch)
tt_writer.add_summary(tt_summaries[n][0],epoch)
# if epoch ==0 :
# print ("adding delay to make directory creation time different")
# time.sleep(2.0) # reduce later
qsf.print_time("==== %04d:%03d -> %.4f %.4f %.4f %.4f %.4f (%.4f %.4f %.4f %.4f %.4f) ===="%(
epoch,i,
tt_gw_avg[0], tt_gw_avg[1], tt_gw_avg[2], tt_gw_avg[3], tt_gw_avg[4],
tt2_avg[0], tt2_avg[1], tt2_avg[2], tt2_avg[3], tt2_avg[4]))
if (not CHECKPOINT_PERIOD is None) and (((epoch + 1) % CHECKPOINT_PERIOD) == 0):
print("Saving periodic checkpoint (trained variables only) to %s, global_step = %d"%(os.path.dirname(files['checkpoints']), epoch),end=" => ")
print(saver.save(sess, files['checkpoints'], global_step=epoch, write_meta_graph=False))
# Close writers
for tt_writer in tt_writers:
try:
tt_writer.close()
except:
print ("Could not close tt_writer: ",tt_writer)
print("Saving final checkpoint (trained variables only) to %s"%(files['checkpoints']),end=" => ")
print(saver.save(sess, files["checkpoints"]))
print("All done")
exit (0)
"""
Traceback (most recent call last):
File "nn_ds_neibs30.py", line 721, in <module>
tt2_avg[0], tt2_avg[1], tt2_avg[2], tt2_avg[3], tt2_avg[4]))
ValueError: unsupported format character ' ' (0x20) at index 20
"""
\ No newline at end of file
qcstereo_functions.py 0 → 100644
View file @ b1111bca
#!/usr/bin/env python3
__copyright__ = "Copyright 2018, Elphel, Inc."
__license__ = "GPL-3.0+"
__email__ = "andrey@elphel.com"
import os
import numpy as np
import tensorflow as tf
import xml.etree.ElementTree as ET
import time
import imagej_tiffwriter
TIME_LAST = 0
TIME_START = 0
class bcolors:
HEADER = '\033[95m'
OKBLUE = '\033[94m'
OKGREEN = '\033[92m'
WARNING = '\033[38;5;214m'
FAIL = '\033[91m'
ENDC = '\033[0m'
BOLD = '\033[1m'
BOLDWHITE = '\033[1;37m'
UNDERLINE = '\033[4m'
def print_time(txt="",end="\n"):
global TIME_LAST
t = time.time()
if txt:
txt +=" "
print(("%s"+bcolors.BOLDWHITE+"at %.4fs (+%.4fs)"+bcolors.ENDC)%(txt,t-TIME_START,t-TIME_LAST), end = end, flush=True)
TIME_LAST = t
DEFAULT_TITLES = [
['test_lvar', 'Test_flat_heuristic'],
['test_hvar', 'Test_edge_heuristic'],
['test_lvar1', 'Test_flat_random'],
['test_hvar1', 'Test_edge_random'],
['fake_lvar', 'Fake_flat_heuristic'],
['fake_hvar', 'Fake_edge_heuristic'],
['fake_lvar1', 'Fake_flat_random'],
['fake_hvar1', 'Fake_edge_random']]
def parseXmlConfig(conf_file, root_dir):
tree = ET.parse(conf_file)
root = tree.getroot()
parameters = {}
for p in root.find('parameters'):
## print ("p.tag=%s, p.text.stri[p()=%s"%(p.tag,p.text.strip()))
parameters[p.tag]=eval(p.text.strip())
# globals
dirs={}
for p in root.find('directories'):
dirs[p.tag]=eval(p.text.strip())
if not os.path.isabs(dirs[p.tag]):
dirs[p.tag] = os.path.join(root_dir, dirs[p.tag])
files={}
for p in root.find('files'):
files[p.tag]=eval(p.text.strip())
dbg_parameters = {}
for p in root.find('dbg_parameters'):
dbg_parameters[p.tag]=eval(p.text.strip())
return parameters, dirs, files, dbg_parameters
def defaultTestTitles(files):
test_titles = []
for f, n in DEFAULT_TITLES:
if f in files:
test_titles.append(n)
else:
test_titles.append(None)
return test_titles
def prepareFiles(dirs, files, suffix):
#MAX_FILES_PER_GROUP
for i, path in enumerate(files['train_lvar']):
files['train_lvar'][i]=os.path.join(dirs['train_lvar'], path)
for i, path in enumerate(files['train_hvar']):
files['train_hvar'][i]=os.path.join(dirs['train_hvar'], path)
for i, path in enumerate(files['train_lvar1']):
files['train_lvar1'][i]=os.path.join(dirs['train_lvar1'], path)
for i, path in enumerate(files['train_hvar1']):
files['train_hvar1'][i]=os.path.join(dirs['train_hvar1'], path)
for i, path in enumerate(files['test_lvar']):
files['test_lvar'][i]=os.path.join(dirs['test_lvar'], path)
for i, path in enumerate(files['test_hvar']):
files['test_hvar'][i]=os.path.join(dirs['test_hvar'], path)
if ('test_lvar1' in files) and ('test_lvar1' in dirs):
for i, path in enumerate(files['test_lvar1']):
files['test_lvar1'][i]=os.path.join(dirs['test_lvar1'], path)
if ('test_hvar1' in files) and ('test_hvar1' in dirs):
for i, path in enumerate(files['test_hvar1']):
files['test_hvar1'][i]=os.path.join(dirs['test_hvar1'], path)
if ('fake_lvar' in files) and ('fake_lvar' in dirs):
for i, path in enumerate(files['fake_lvar']):
files['fake_lvar'][i]=os.path.join(dirs['fake_lvar'], path)
if ('fake_hvar' in files) and ('fake_hvar' in dirs):
for i, path in enumerate(files['fake_hvar']):
files['fake_hvar'][i]=os.path.join(dirs['fake_hvar'], path)
if ('fake_lvar1' in files) and ('fake_lvar1' in dirs):
for i, path in enumerate(files['fake_lvar1']):
files['fake_lvar1'][i]=os.path.join(dirs['fake_lvar1'], path)
if ('fake_hvar' in files) and ('fake_hvar' in dirs):
for i, path in enumerate(files['fake_hvar1']):
files['fake_hvar1'][i]=os.path.join(dirs['fake_hvar1'], path)
result_files=[]
for i, path in enumerate(files['images']):
result_files.append(os.path.join(dirs['result'], path+"_"+suffix+'.npy'))
files['result'] = result_files
if not 'checkpoints' in files:
files['checkpoints'] = 'checkpoints'
if not 'checkpoints' in dirs:
dirs['checkpoints'] = dirs['result']
files['checkpoints'] = os.path.join(dirs['checkpoints'], files['checkpoints'])
if not 'inference' in files:
files['inference'] = 'inference'
if not 'inference' in dirs:
dirs['inference'] = dirs['result']
files['inference'] = os.path.join(dirs['inference'], files['inference'])
if not 'exportdir' in files:
files['exportdir'] = 'exportdir'
if not 'exportdir' in dirs:
dirs['exportdir'] = dirs['result']
files['exportdir'] = os.path.join(dirs['exportdir'], files['exportdir'])
if not 'figures' in dirs:
dirs['figures'] = os.path.join(dirs['result'],"figs")
files['train'] = [files['train_lvar'],files['train_hvar'], files['train_lvar1'], files['train_hvar1']]
files['test'] = [files['test_lvar'], files['test_hvar']]
if 'test_lvar1' in files:
files['test'].append(files['test_lvar1'])
if 'test_hvar1' in files:
files['test'].append(files['test_hvar1'])
# should be after result files
for i, path in enumerate(files['images']):
files['images'][i] = os.path.join(dirs['images'], path+'.tfrecords')
def readTFRewcordsEpoch(train_filename, cluster_radius):
if not '.tfrecords' in train_filename:
train_filename += '.tfrecords'
npy_dir_name = "npy"
dirname = os.path.dirname(train_filename)
npy_dir = os.path.join(dirname, npy_dir_name)
filebasename, _ = os.path.splitext(train_filename)
filebasename = os.path.basename(filebasename)
file_all = os.path.join(npy_dir,filebasename + '.npy')
if os.path.exists(file_all):
data = np.load (file_all)
else:
record_iterator = tf.python_io.tf_record_iterator(path=train_filename)
corr2d_list=[]
target_disparity_list=[]
gt_ds_list = []
extra_list = []
for string_record in record_iterator:
example = tf.train.Example()
example.ParseFromString(string_record)
corr2d_list.append (np.array(example.features.feature['corr2d'].float_list.value, dtype=np.float32))
target_disparity_list.append (np.array(example.features.feature['target_disparity'].float_list.value, dtype=np.float32))
gt_ds_list.append (np.array(example.features.feature['gt_ds'].float_list.value, dtype= np.float32))
try:
extra_list.append (np.array(example.features.feature['extra'].float_list.value, dtype= np.float32))
except:
pass
corr2d= np.array(corr2d_list)
target_disparity = np.array(target_disparity_list)
gt_ds = np.array(gt_ds_list)
if len(extra_list):
extra = np.array(extra_list)
else:
extra = None
try:
os.makedirs(os.path.dirname(file_all))
except:
pass
if cluster_radius > 0:
reformat_to_clusters(
corr2d,
target_disparity,
gt_ds,
extra,
cluster_radius)
if not extra is None:
data = np.concatenate(
[corr2d, target_disparity, gt_ds, extra],
axis = 1)
else:
data = np.concatenate(
[corr2d, target_disparity, gt_ds],
axis = 1)
np.save(file_all, data)
return data
def getMoreFiles(fpaths,rslt, cluster_radius, hor_flip, tile_layers, tile_side):
for fpath in fpaths:
dataset = readTFRewcordsEpoch(fpath, cluster_radius)
if hor_flip:
if np.random.randint(2):
print_time("Performing horizontal flip", end=" ")
flip_horizontal(dataset, cluster_radius, tile_layers, tile_side)
print_time("Done")
rslt.append(dataset)
#from http://warmspringwinds.github.io/tensorflow/tf-slim/2016/12/21/tfrecords-guide/
def read_and_decode(filename_queue, featrures_per_tile):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
# Defaults are not specified since both keys are required.
features={
'corr2d': tf.FixedLenFeature([featrures_per_tile],tf.float32), #string),
'target_disparity': tf.FixedLenFeature([1], tf.float32), #.string),
'gt_ds': tf.FixedLenFeature([2], tf.float32) #.string)
})
corr2d = features['corr2d'] # tf.decode_raw(features['corr2d'], tf.float32)
target_disparity = features['target_disparity'] # tf.decode_raw(features['target_disparity'], tf.float32)
gt_ds = tf.cast(features['gt_ds'], tf.float32) # tf.decode_raw(features['gt_ds'], tf.float32)
in_features = tf.concat([corr2d,target_disparity],0)
corr2d_out, target_disparity_out, gt_ds_out = tf.train.shuffle_batch( [in_features, target_disparity, gt_ds],
batch_size=1000, # 2,
capacity=30,
num_threads=2,
min_after_dequeue=10)
return corr2d_out, target_disparity_out, gt_ds_out
def add_margins(npa,radius, val = np.nan):
npa_ext = np.empty((npa.shape[0]+2*radius, npa.shape[1]+2*radius, npa.shape[2]), dtype = npa.dtype)
npa_ext[radius:radius + npa.shape[0],radius:radius + npa.shape[1]] = npa
npa_ext[0:radius,:,:] = val
npa_ext[radius + npa.shape[0]:,:,:] = val
npa_ext[:,0:radius,:] = val
npa_ext[:, radius + npa.shape[1]:,:] = val
return npa_ext
def add_neibs(npa_ext,radius):
height = npa_ext.shape[0]-2*radius
width = npa_ext.shape[1]-2*radius
side = 2 * radius + 1
# size = side * side
npa_neib = np.empty((height, width, side, side, npa_ext.shape[2]), dtype = npa_ext.dtype)
for dy in range (side):
for dx in range (side):
npa_neib[:,:,dy, dx,:]= npa_ext[dy:dy+height, dx:dx+width]
return npa_neib.reshape(height, width, -1)
def extend_img_to_clusters(datasets_img,radius, width): # = 324):
# side = 2 * radius + 1
# size = side * side
if len(datasets_img) ==0:
return
num_tiles = datasets_img[0]['corr2d'].shape[0]
height = num_tiles // width
for rec in datasets_img:
if not rec is None:
rec['corr2d'] = add_neibs(add_margins(rec['corr2d'].reshape((height,width,-1)), radius, np.nan), radius).reshape((num_tiles,-1))
rec['target_disparity'] = add_neibs(add_margins(rec['target_disparity'].reshape((height,width,-1)), radius, np.nan), radius).reshape((num_tiles,-1))
rec['gt_ds'] = add_neibs(add_margins(rec['gt_ds'].reshape((height,width,-1)), radius, np.nan), radius).reshape((num_tiles,-1))
try:
rec['extra'] = add_neibs(add_margins(rec['extra'].reshape((height,width,-1)), radius, np.nan), radius).reshape((num_tiles,-1))
except:
pass
pass
def reformat_to_clusters_rec(datasets_data, cluster_radius):
cluster_size = (2 * cluster_radius + 1) * (2 * cluster_radius + 1)
# Reformat input data
for rec in datasets_data:
rec['corr2d'] = rec['corr2d'].reshape( (rec['corr2d'].shape[0]//cluster_size, rec['corr2d'].shape[1] * cluster_size))
rec['target_disparity'] = rec['target_disparity'].reshape((rec['target_disparity'].shape[0]//cluster_size, rec['target_disparity'].shape[1] * cluster_size))
rec['gt_ds'] = rec['gt_ds'].reshape( (rec['gt_ds'].shape[0]//cluster_size, rec['gt_ds'].shape[1] * cluster_size))
try:
rec['extra'] = rec['extra'].reshape( (rec['extra'].shape[0]//cluster_size, rec['extra'].shape[1] * cluster_size))
except:
pass
def reformat_to_clusters(
corr2d,
target_disparity,
gt_ds,
extra, # may be None
cluster_radius):
cluster_size = (2 * cluster_radius + 1) * (2 * cluster_radius + 1)
# Reformat input data
corr2d.shape = ((corr2d.shape[0]//cluster_size, corr2d.shape[1] * cluster_size))
target_disparity.shape = ((target_disparity.shape[0]//cluster_size, target_disparity.shape[1] * cluster_size))
gt_ds.shape = ((gt_ds.shape[0]//cluster_size, gt_ds.shape[1] * cluster_size))
if not extra is None:
extra.shape = ((extra.shape[0]//cluster_size, extra.shape[1] * cluster_size))
def get_lengths(
cluster_radius,
tile_layers,
tile_side):
cluster_side = 2 * cluster_radius + 1
cl = cluster_side * cluster_side * tile_layers * tile_side * tile_side
tl = cluster_side * cluster_side
gl = cluster_side * cluster_side * 2 # disparity+strength, resto goes to extra
return cl, tl, gl, cluster_side
def flip_horizontal(dataset, cluster_radius, tile_layers, tile_side):
cl, tl, gl, cluster_side = get_lengths(cluster_radius, tile_layers, tile_side)
corr2d = dataset[:,:cl] .reshape([dataset.shape[0], cluster_side, cluster_side, tile_layers, tile_side, tile_side])
target_disparity = dataset[:,cl:cl+tl].reshape([dataset.shape[0], cluster_side, cluster_side, -1])
gt_ds = dataset[:,cl+tl:cl+tl+gl] .reshape([dataset.shape[0], cluster_side, cluster_side, -1])
# no extra here !
"""
Horizontal flip of tiles
"""
corr2d = corr2d[:,:,::-1,...]
target_disparity = target_disparity[:,:,::-1,...]
gt_ds = gt_ds[:,:,::-1,...]
corr2d[:,:,:,0,:,:] = corr2d[:,:,:,0,::-1,:] # flip vertical layer0 (hor)
corr2d[:,:,:,1,:,:] = corr2d[:,:,:,1,:,::-1] # flip horizontal layer1 (vert)
corr2d_2 = corr2d[:,:,:,3,::-1,:].copy() # flip vertical layer3 (diago)
corr2d[:,:,:,3,:,:] = corr2d[:,:,:,2,::-1,:] # flip vertical layer2 (diago)
corr2d[:,:,:,2,:,:] = corr2d_2
"""
pack back into a single (input)array
"""
dataset[:,:cl] = corr2d.reshape((corr2d.shape[0],-1))
dataset[:,cl:cl+tl] = target_disparity.reshape((target_disparity.shape[0],-1))
dataset[:,cl+tl:] = gt_ds.reshape((gt_ds.shape[0],-1))
def replace_nan(datasets_data): # , cluster_radius):
# Reformat input data
for rec in datasets_data:
if not rec is None:
np.nan_to_num(rec['corr2d'], copy = False)
np.nan_to_num(rec['target_disparity'], copy = False)
if 'gt_ds' in rec:
np.nan_to_num(rec['gt_ds'], copy = False)
if 'extra' in rec:
np.nan_to_num(rec['extra'], copy = False)
def permute_to_swaps(perm):
pairs = []
for i in range(len(perm)):
w = np.where(perm == i)[0][0]
if w != i:
pairs.append([i,w])
perm[w] = perm[i]
perm[i] = i
return pairs
def shuffle_in_place(dataset_data, #alternating clusters from 4 sources.each cluster has all needed data (concatenated)
period):
for i in range (period):
np.random.shuffle(dataset_data[i::period])
def add_file_to_dataset(dataset, new_dataset, train_next):
train_next['file'] = (train_next['file']+1)%train_next['files']
l = new_dataset.shape[0] * train_next['step']
if (train_next['entry'] + l) < (train_next['entries']+train_next['step']):
dataset[train_next['entry']:train_next['entry']+l:train_next['step']] = new_dataset
train_next['entry'] += l
if (train_next['entry'] >= train_next['entries']):
train_next['entry'] -= train_next['entries']
return True
else:
return False
else: # split it two parts
l = (train_next['entries'] - train_next['entry'] + (train_next['step']-1)) // train_next['step']
dataset[train_next['entry']::train_next['step']] = new_dataset[:l]
train_next['entry'] = (train_next['entry'] + l * train_next['step']) % train_next['entries'] #0,1,2,3
l1 = new_dataset.shape[0] - l # remainder
ln = train_next['entry'] + l1 * train_next['step']
dataset[train_next['entry']:ln:train_next['step']] = new_dataset[l:]
train_next['entry'] = ln
return True
"""
train_next[n_train]
Read as many files as needed, possibly repeating, until each buffer is f
"""
def initTrainTestData(
files,
cluster_radius,
buffer_size, # number of clusters per train
test_titles = None
):
"""
Generates a single np array for training with concatenated cluster of corr2d,
cluster of target_disparity, and cluster of gt_ds for convenient shuffling
"""
num_trains = len(files['train'])
num_entries = num_trains * buffer_size
dataset_train_merged = None
train_next = [None]*num_trains
for n_train, f_train in enumerate(files['train']):
train_next[n_train] = {'file':0, 'entry':n_train, 'files':len(f_train), 'entries': num_entries, 'step':num_trains, 'more_files':False}
buffer_full = False
while not buffer_full:
for fpath in f_train:
print_time("Importing train data "+(["low variance","high variance", "low variance1","high variance1"][n_train]) +" from "+fpath, end="")
new_dataset = readTFRewcordsEpoch(fpath, cluster_radius)
if dataset_train_merged is None:
dataset_train_merged = np.empty([num_entries,new_dataset.shape[1]], dtype =new_dataset.dtype)
## print("\nbefore add_file_to_dataset: train_next["+str(n_train)+"]=",train_next[n_train])
rollover = add_file_to_dataset(
dataset = dataset_train_merged,
new_dataset = new_dataset,
train_next = train_next[n_train])
## print("after add_file_to_dataset: train_next["+str(n_train)+"]=",train_next[n_train])
print_time(" Done")
if rollover:
buffer_full = True
train_next[n_train][ 'more_files'] = train_next[n_train][ 'file'] < train_next[n_train][ 'files'] # Not all files used, need to load during training
break
if test_titles is None:
test_titles = defaultTestTitles(files)
datasets_test = []
for t,v in zip(test_titles,DEFAULT_TITLES):
if not t is None:
grp = v[0]
for fpath in files[grp]:
print_time("Importing test data ("+grp+") from "+fpath, end="")
new_dataset = readTFRewcordsEpoch(fpath, cluster_radius)
datasets_test.append(new_dataset)
print_time(" Done")
"""
for grp in ['test_lvar','test_hvar','test_lvar1','test_hvar1']:
if grp in files:
for fpath in files[grp]:
print_time("Importing test data ("+grp+") from "+fpath, end="")
new_dataset = readTFRewcordsEpoch(fpath, cluster_radius)
datasets_test.append(new_dataset)
print_time(" Done")
"""
return train_next, dataset_train_merged, datasets_test
def get_full_tile_indices2d(height,width):
a = np.empty([height,width,2], dtype=np.int32)
a[...,0] = np.arange(height).reshape([height,1])
a[...,1] = np.arange(width)
return a.reshape(-1,2)
def get_full_tile_indices(height,width):
return np.arange(height*width).reshape(-1,1)
def readImageData(image_data,
files,
indx,
cluster_radius,
tile_layers,
tile_side,
width,
replace_nans,
infer = False,
keep_gt = False):
cl, tl, gl, _ = get_lengths(0, tile_layers, tile_side)
if image_data[indx] is None:
dataset = readTFRewcordsEpoch(
files['images'][indx],
cluster_radius = 0)
corr2d = dataset[:,:cl]
target_disparity = dataset[:,cl:cl+tl]
if infer:
image_data[indx] = {
'corr2d': corr2d,
'target_disparity': target_disparity,
'xy': get_full_tile_indices2d(corr2d.shape[0]//width, width),
'ntile': get_full_tile_indices(corr2d.shape[0]//width, width)}
if keep_gt:
gt_ds = dataset[:,cl+tl:cl+tl+gl]
image_data[indx]["gt_ds"] = gt_ds
image_data[indx]["gtruths"]= gt_ds.copy()
image_data[indx]["t_disps"]= target_disparity.reshape([-1,1]).copy()
else:
gt_ds = dataset[:,cl+tl:cl+tl+gl]
extra = dataset[:,cl+tl+gl:]
image_data[indx] = {
'corr2d': corr2d,
'target_disparity': target_disparity,
"gt_ds": gt_ds,
"gtruths": gt_ds.copy(), # never used?
"t_disps": target_disparity.reshape([-1,1]).copy(),
"extra": extra, # will be increased by 25
"t_extra": extra.copy() } #will still be (ntiles,3)
if cluster_radius > 0:
extend_img_to_clusters(
[image_data[indx]],
cluster_radius,
width)
if replace_nans:
replace_nan([image_data[indx]])
return image_data[indx]
def initImageData(files,
max_imgs,
cluster_radius,
tile_layers,
tile_side,
width,
replace_nans,
infer = False,
keep_gt = False):
# no_train = False):
num_imgs = len(files['images'])
img_data = [None] * num_imgs
for nfile in range(min(num_imgs, max_imgs)):
print_time("Importing test image data from "+ files['images'][nfile], end="")
readImageData(img_data,
files,
nfile,
cluster_radius,
tile_layers,
tile_side,
width,
replace_nans,
infer = infer,
keep_gt = keep_gt)
print_time(" Done")
return img_data
def evaluateAllResults(result_files, absolute_disparity, cluster_radius, labels=None, logpath=None):
if logpath:
lf=open(logpath,"w")
else:
lf = None
for result_file in result_files:
try:
print_time("Reading resuts from "+result_file, end=" ")
eval_results(result_file, absolute_disparity, radius=cluster_radius, logfile=lf)
except:
print_time(" - does not exist")
continue
print_time("Done")
print_time("Saving resuts to tiff", end=" ")
result_npy_to_tiff(result_file, absolute_disparity, fix_nan = True, labels=labels)
print_time("Done")
if lf:
lf.close()
def result_npy_prepare(npy_path, absolute, fix_nan, insert_deltas=True,labels=None):
"""
@param npy_path full path to the npy file with 4-layer data (242,324,4) - nn_disparity(offset), target_disparity, gt disparity, gt strength
data will be written as 4-layer tiff, extension '.npy' replaced with '.tiff'
@param absolute - True - the first layer contains absolute disparity, False - difference from target_disparity
@param fix_nan - replace nan in target_disparity with 0 to apply offset, target_disparity will still contain nan
@parame insert_deltas: +1 - add delta layers, +2 - add variance (max - min of this and 8 neighbors)
"""
data = np.load(npy_path) #(324,242,4) [nn_disp, target_disp,gt_disp, gt_conf]
if labels is None:
labels = ["chn%d"%(i) for i in range(data.shape[2])]
# labels = ["nn_out","hier_out","gt_disparity","gt_strength"]
nn_out = 0
# target_disparity = 1
gt_disparity = 2
gt_strength = 3
heur_err = 7
min_heur_err = 0.001
height = data.shape[0]
width = data.shape[1]
nocenter9 = np.array([[[1,1,1,1,np.nan,1,1,1,1]]], dtype = data.dtype)
if not absolute:
if fix_nan:
data[...,nn_out] += np.nan_to_num(data[...,1], copy=True)
else:
data[...,nn_out] += data[...,1]
if (insert_deltas & 1):
np.nan_to_num(data[...,gt_strength], copy=False)
data = np.concatenate([data[...,0:4],data[...,0:2],data[...,0:2],data[...,4:]], axis = 2) # data[...,4:] may be empty
labels = labels[:4]+["nn_out","hier_out","nn_err","hier_err"]+labels[4:]
data[...,6] -= data[...,gt_disparity]
data[...,7] -= data[...,gt_disparity]
for l in [2, 4, 5, 6, 7]:
if l < data.shape[2]:
data[...,l] = np.select([data[...,gt_strength]==0.0, data[...,gt_strength]>0.0], [np.nan,data[...,l]])
# All other layers - mast too
for l in range(8,data.shape[2]):
data[...,l] = np.select([data[...,gt_strength]==0.0, data[...,gt_strength]>0.0], [np.nan,data[...,l]])
"""
Calculate bad tiles where ggt was used as a master, to remove them from the results (later versions add random error)
"""
bad1 = abs(data[...,heur_err]) < min_heur_err
bad1_ext = np.concatenate([bad1 [0:1,:], bad1 [0:1,:], bad1[:,:], bad1 [-1:height,:], bad1 [-1:height,:]],axis = 0)
bad1_ext = np.concatenate([bad1_ext[:,0:1], bad1_ext[:,0:1], bad1_ext[:,:], bad1_ext[:,-1:width], bad1_ext[:,-1:width]], axis = 1)
bad25 = np.empty(shape=[height, width, 25], dtype=bad1.dtype)
bm25=np.array([[[1,1,1,1,1, 1,1,1,1,1, 1,1,1,1,1, 1,1,1,1,1, 1,1,1,1,1]]])
bm09=np.array([[[0,0,0,0,0, 0,1,1,1,0, 0,1,1,1,0, 0,1,1,1,0, 0,0,0,0,0]]])
bm01=np.array([[[0,0,0,0,0, 0,0,0,0,0, 0,0,1,0,0, 0,0,0,0,0, 0,0,0,0,0]]])
for row in range(5):
for col in range(5):
pass
bad25 [...,row*5+col]= bad1_ext[row:height+row, col:width+col]
bad_num1=(np.sum(bad25*bm25,axis=2) > 0).astype(data.dtype)
bad_num2=(np.sum(bad25*bm09,axis=2) > 0).astype(data.dtype)
bad_num3=(np.sum(bad25*bm01,axis=2) > 0).astype(data.dtype)
bad_num = bad_num1 + bad_num2 + bad_num3
if (insert_deltas & 2):
wo = 0.7 # ortho
wc = 0.5 #corner
w8=np.array([wc,wo,wc,wo,0.0,wo,wc,wo,wc], dtype=data.dtype)
w8/=np.sum(w8) #normalize
gt_ext = np.concatenate([data[0:1,:,gt_disparity],data[:,:,gt_disparity],data[-1:height,:,gt_disparity]],axis = 0)
gt_ext = np.concatenate([gt_ext[:,0:1], gt_ext[:,:], gt_ext[:,-1:width]],axis = 1)
gs_ext = np.concatenate([data[0:1,:,gt_strength], data[:,:,gt_strength], data[-1:height,:,gt_strength]],axis = 0)
gs_ext = np.concatenate([gs_ext[:,0:1], gs_ext[:,:], gs_ext[:,-1:width]],axis = 1)
data9 = np.empty(shape=[height, width, 9], dtype=data.dtype)
weight9 = np.empty(shape=[height, width, 9], dtype=data.dtype)
for row in range(3):
for col in range(3):
pass
data9 [...,row*3+col]= gt_ext[row:height+row, col:width+col]
weight9[...,row*3+col]= gs_ext[row:height+row, col:width+col]
data9 *= weight9/weight9 # make data=nan where wigth is 0
# data = np.concatenate([data[...],np.empty_like(data[...,-1])], axis = 2) # data[...,4:] may be empty
data = np.concatenate([data[...],np.empty(shape=[height,width,4],dtype=data.dtype)], axis = 2) # data[...,4:] may be empty
data[...,-4] = np.nanmax(data9*nocenter9, axis=2)-np.nanmin(data9*nocenter9,axis=2)# will ignore nan
np.nan_to_num(data9,copy=False) # replace all nan in data9 with 0.
weight9 *= w8
w_center = np.sum(weight9, axis=2)
dw_center = np.sum(data9*weight9, axis=2)
dw_center /= w_center # now dw_center - weighted average in the center
data[...,-3] = np.abs(data[...,gt_disparity]- dw_center)
# data[...,-2] = data[...,gt_disparity]- dw_center
#data[...,-3] *= (data[...,-4] < 1.0) # just temporary
#data[...,-3] *= (data[...,gt_disparity] < 5) #just temporary
data[...,-2] =bad_num.astype(data.dtype)
data [...,-1]= np.sum(np.nan_to_num(weight9/weight9),axis=2).astype(data.dtype)
# data[...,-1] = dw_center
labels +=["max-min","abs-center","badness","neibs"]
#neib = np.concatenate([gt_ext[:height,:width,:],],axis = )
pass
return data, labels
def result_npy_to_tiff(npy_path,
absolute,
fix_nan,
insert_deltas=True,
labels = None,
logfile = None):
"""
@param npy_path full path to the npy file with 4-layer data (242,324,4) - nn_disparity(offset), target_disparity, gt disparity, gt strength
data will be written as 4-layer tiff, extension '.npy' replaced with '.tiff'
@param absolute - True - the first layer contains absolute disparity, False - difference from target_disparity
@param fix_nan - replace nan in target_disparity with 0 to apply offset, target_disparity will still contain nan
"""
data,labels = result_npy_prepare(npy_path, absolute, fix_nan, insert_deltas, labels=labels)
tiff_path = npy_path.replace('.npy','.tiff')
data = data.transpose(2,0,1)
print("Saving results to TIFF: "+tiff_path)
if (logfile):
print("Saving results to TIFF: "+tiff_path,file=logfile)
imagej_tiffwriter.save(tiff_path,data,labels=labels)
def eval_results(rslt_path, absolute,
min_disp = -0.1, #minimal GT disparity
max_disp = 20.0, # maximal GT disparity
max_ofst_target = 1.0,
max_ofst_result = 1.0,
str_pow = 2.0,
radius = 0,
logfile = None):
variants = [[ -0.1, 5.0, 0.5, 0.5, 1.0],
[ -0.1, 5.0, 0.5, 0.5, 2.0],
[ -0.1, 5.0, 0.2, 0.2, 1.0],
[ -0.1, 5.0, 0.2, 0.2, 2.0],
[ -0.1, 20.0, 0.5, 0.5, 1.0],
[ -0.1, 20.0, 0.5, 0.5, 2.0],
[ -0.1, 20.0, 0.2, 0.2, 1.0],
[ -0.1, 20.0, 0.2, 0.2, 2.0],
[ -0.1, 20.0, 1.0, 1.0, 1.0],
[min_disp, max_disp, max_ofst_target, max_ofst_result, str_pow]]
rslt = np.load(rslt_path)
not_nan = ~np.isnan(rslt[...,0])
not_nan &= ~np.isnan(rslt[...,1])
not_nan &= ~np.isnan(rslt[...,2])
not_nan &= ~np.isnan(rslt[...,3])
not_nan_ext = np.zeros((rslt.shape[0] + 2*radius,rslt.shape[1] + 2 * radius),dtype=np.bool)
not_nan_ext[radius:-radius,radius:-radius] = not_nan
for dy in range(2*radius+1):
for dx in range(2*radius+1):
not_nan_ext[dy:dy+not_nan.shape[0], dx:dx+not_nan.shape[1]] &= not_nan
not_nan = not_nan_ext[radius:-radius,radius:-radius]
if not absolute:
rslt[...,0] += rslt[...,1]
nn_disparity = np.nan_to_num(rslt[...,0], copy = False)
target_disparity = np.nan_to_num(rslt[...,1], copy = False)
gt_disparity = np.nan_to_num(rslt[...,2], copy = False)
gt_strength = np.nan_to_num(rslt[...,3], copy = False)
rslt = []
print ("--------------- %s ---------------"%(rslt_path))
if logfile:
print ("--------------- %s ---------------"%(rslt_path), file=logfile)
for min_disparity, max_disparity, max_offset_target, max_offset_result, strength_pow in variants:
good_tiles = not_nan.copy();
good_tiles &= (gt_disparity >= min_disparity)
good_tiles &= (gt_disparity <= max_disparity)
good_tiles &= (target_disparity != gt_disparity)
good_tiles &= (np.abs(target_disparity - gt_disparity) <= max_offset_target)
good_tiles &= (np.abs(target_disparity - nn_disparity) <= max_offset_result)
gt_w = gt_strength * good_tiles
gt_w = np.power(gt_w,strength_pow)
sw = gt_w.sum()
diff0 = target_disparity - gt_disparity
diff1 = nn_disparity - gt_disparity
diff0_2w = gt_w*diff0*diff0
diff1_2w = gt_w*diff1*diff1
rms0 = np.sqrt(diff0_2w.sum()/sw)
rms1 = np.sqrt(diff1_2w.sum()/sw)
print ("%7.3f<disp<%7.3f, offs_tgt<%5.2f, offs_rslt<%5.2f pwr=%05.3f, rms0=%7.4f, rms1=%7.4f (gain=%7.4f) num good tiles = %5d"%(
min_disparity, max_disparity, max_offset_target, max_offset_result, strength_pow, rms0, rms1, rms0/rms1, good_tiles.sum() ))
if logfile:
print ("%7.3f<disp<%7.3f, offs_tgt<%5.2f, offs_rslt<%5.2f pwr=%05.3f, rms0=%7.4f, rms1=%7.4f (gain=%7.4f) num good tiles = %5d"%(
min_disparity, max_disparity, max_offset_target, max_offset_result, strength_pow, rms0, rms1, rms0/rms1, good_tiles.sum() ),file=logfile)
rslt.append([rms0,rms1])
return rslt
def concentricSquares(radius):
side = 2 * radius + 1
return [[((i // side) >= var) and
((i // side) < (side - var)) and
((i % side) >= var) and
((i % side) < (side - var)) for i in range (side*side) ] for var in range(radius+1)]
qcstereo_losses.py 0 → 100644
View file @ b1111bca
#!/usr/bin/env python3
__copyright__ = "Copyright 2018, Elphel, Inc."
__license__ = "GPL-3.0+"
__email__ = "andrey@elphel.com"
#from numpy import float64
#import numpy as np
import tensorflow as tf
def smoothLoss(out_batch, # [batch_size,(1..2)] tf_result
target_disparity_batch, # [batch_size] tf placeholder
gt_ds_batch_clust, # [batch_size,25,2] tf placeholder
clip, # limit punishment for cutting corners (disparity pix)
absolute_disparity = False, #when false there should be no activation on disparity output !
cluster_radius = 2):
with tf.name_scope("SmoothLoss"):
center_tile_index = 2 * cluster_radius * (cluster_radius + 1)
cluster_side = 2 * cluster_radius + 1
cluster_size = cluster_side * cluster_side
w_corner = 0.7
w8 = [w_corner,1.0,w_corner,1.0,1.0,w_corner,1.0,w_corner]
w8 = [w/sum(w8) for w in w8]
tf_w8=tf.reshape(tf.constant(w8, dtype=tf.float32, name="w8_"), shape=[1,-1], name="w8")
i8 = []
for dy in [-1,0,1]:
for dx in [-1,0,1]:
if (dy != 0) or (dx != 0):
i8.append(center_tile_index+(dy*cluster_side)+dx)
tf_clip = tf.constant(clip, dtype=tf.float32, name = "clip")
tf_gt_ds_all = tf.reshape(gt_ds_batch_clust,[-1,cluster_size,gt_ds_batch_clust.shape[1]//cluster_size], name = "gt_ds_all")
tf_neibs8 = tf.gather(tf_gt_ds_all, indices = i8, axis = 1, name = "neibs8")
tf_gt_disparity8 = tf.reshape(tf_neibs8[:,:,0], [-1,8], name = "gt8_disparity") # (?,8)
tf_gt_strength8 = tf.reshape(tf_neibs8[:,:,1], [-1,8], name = "gt8_strength") # (?,8)
tf_w = tf.multiply(tf_gt_strength8, tf_w8, name = "w")
tf_dw = tf.multiply(tf_gt_disparity8, tf_w, name = "dw")
tf_sum_w = tf.reduce_sum(tf_w, axis = 1, name = "sum_w")
tf_sum_dw = tf.reduce_sum(tf_dw, axis = 1, name = "sum_dw")
tf_avg_disparity = tf.divide(tf_sum_dw, tf_sum_w, name = "avg_disparity") # (?,)
tf_gt_disparity = tf.reshape(tf_gt_ds_all[:,center_tile_index,0], [-1], name = "gt_disparity") # (?,)
"""
It is good to limit tf_gt_disparityby min/max (+margin) tf.reduce_min(tf_gt_disparity8, axis=1,...) but there could be zeros caused by undefined GT for the tile
"""
tf_gt_strength = tf.reshape(tf_gt_ds_all[:,center_tile_index,1], [-1], name = "gt_strength") # (?,)
tf_d0 = tf.abs(tf_gt_disparity - tf_avg_disparity, name = "tf_d0")
tf_d = tf.maximum(tf_d0, 0.001, name = "tf_d")
## tf_d2 = tf.multiply(tf_d, tf_d, name = "tf_d2")
tf_out = tf.reshape(out_batch[:,0],[-1], name = "tf_out")
if absolute_disparity:
tf_out_disparity = tf_out
else:
tf_out_disparity = tf.add(tf_out, tf.reshape(target_disparity_batch,[-1]),name = "out_disparity")
tf_offs = tf.subtract(tf_out_disparity, tf_avg_disparity, name = "offs")
tf_offs2 = tf.multiply(tf_offs, tf_offs, name = "offs2")
tf_offs2_d = tf.divide(tf_offs2, tf_d, name = "offs2_d")
tf_cost0 = tf.maximum(tf_d - tf_offs2_d, 0.0, name = "cost0")
tf_cost_nw = tf.minimum(tf_cost0, tf_clip, name = "cost_nw")
## tf_cost_nw = tf.maximum(tf_d2 - tf_offs2, 0.0, name = "cost_nw")
tf_cost_w = tf.multiply(tf_cost_nw, tf_gt_strength, name = "cost_w")
tf_sum_wc = tf.reduce_sum(tf_gt_strength, name = "sum_wc")
tf_sum_costw = tf.reduce_sum(tf_cost_w, name = "sum_costw")
tf_cost = tf.divide(tf_sum_costw, tf_sum_wc, name = "cost")
return tf_cost, tf_cost_nw, tf_cost_w, tf_d , tf_avg_disparity, tf_gt_disparity, tf_offs
def batchLoss(out_batch, # [batch_size,(1..2)] tf_result
target_disparity_batch, # [batch_size] tf placeholder
gt_ds_batch, # [batch_size,2] tf placeholder
batch_weights, # [batch_size] now batch index % 4 - different sources, even - low variance, odd - high variance
disp_diff_cap = 10.0, # cap disparity difference to this value (give up on large errors)
disp_diff_slope= 0.0, #allow squared error to grow above disp_diff_cap
absolute_disparity = False, #when false there should be no activation on disparity output !
use_confidence = False,
lambda_conf_avg = 0.01,
## lambda_conf_pwr = 0.1,
conf_pwr = 2.0,
gt_conf_offset = 0.08,
gt_conf_pwr = 1.0,
error2_offset = 0.0025, # 0.0, # 0.0025, # (0.05^2) ~= coring
disp_wmin = 1.0, # minimal disparity to apply weight boosting for small disparities
disp_wmax = 8.0, # maximal disparity to apply weight boosting for small disparities
use_out = False): # use calculated disparity for disparity weight boosting (False - use target disparity)
with tf.name_scope("BatchLoss"):
"""
Here confidence should be after relU. Disparity - may be also if absolute, but no activation if output is residual disparity
"""
tf_lambda_conf_avg = tf.constant(lambda_conf_avg, dtype=tf.float32, name="tf_lambda_conf_avg")
## tf_lambda_conf_pwr = tf.constant(lambda_conf_pwr, dtype=tf.float32, name="tf_lambda_conf_pwr")
## tf_conf_pwr = tf.constant(conf_pwr, dtype=tf.float32, name="tf_conf_pwr")
tf_gt_conf_offset = tf.constant(gt_conf_offset, dtype=tf.float32, name="tf_gt_conf_offset")
tf_gt_conf_pwr = tf.constant(gt_conf_pwr, dtype=tf.float32, name="tf_gt_conf_pwr")
tf_num_tiles = tf.shape(gt_ds_batch)[0]
tf_0f = tf.constant(0.0, dtype=tf.float32, name="tf_0f")
tf_1f = tf.constant(1.0, dtype=tf.float32, name="tf_1f")
## tf_maxw = tf.constant(1.0, dtype=tf.float32, name="tf_maxw")
tf_disp_diff_cap2= tf.constant(disp_diff_cap*disp_diff_cap, dtype=tf.float32, name="disp_diff_cap2")
tf_disp_diff_slope= tf.constant(disp_diff_slope, dtype=tf.float32, name="disp_diff_slope")
if gt_conf_pwr == 0:
w = tf.ones((out_batch.shape[0]), dtype=tf.float32,name="w_ones")
else:
w_slice = tf.reshape(gt_ds_batch[:,1],[-1], name = "w_gt_slice")
w_sub = tf.subtract (w_slice, tf_gt_conf_offset, name = "w_sub")
w_clip = tf.maximum(w_sub, tf_0f, name = "w_clip")
if gt_conf_pwr == 1.0:
w = w_clip
else:
w=tf.pow(w_clip, tf_gt_conf_pwr, name = "w_pow")
if use_confidence:
tf_num_tilesf = tf.cast(tf_num_tiles, dtype=tf.float32, name="tf_num_tilesf")
conf_slice = tf.reshape(out_batch[:,1],[-1], name = "conf_slice")
conf_sum = tf.reduce_sum(conf_slice, name = "conf_sum")
conf_avg = tf.divide(conf_sum, tf_num_tilesf, name = "conf_avg")
conf_avg1 = tf.subtract(conf_avg, tf_1f, name = "conf_avg1")
conf_avg2 = tf.square(conf_avg1, name = "conf_avg2")
cost2 = tf.multiply (conf_avg2, tf_lambda_conf_avg, name = "cost2")
iconf_avg = tf.divide(tf_1f, conf_avg, name = "iconf_avg")
nconf = tf.multiply (conf_slice, iconf_avg, name = "nconf") #normalized confidence
nconf_pwr = tf.pow(nconf, conf_pwr, name = "nconf_pwr")
nconf_pwr_sum = tf.reduce_sum(nconf_pwr, name = "nconf_pwr_sum")
nconf_pwr_offs = tf.subtract(nconf_pwr_sum, tf_1f, name = "nconf_pwr_offs")
cost3 = tf.multiply (conf_avg2, nconf_pwr_offs, name = "cost3")
w_all = tf.multiply (w, nconf, name = "w_all")
else:
w_all = w
# cost2 = 0.0
# cost3 = 0.0
# normalize weights
w_sum = tf.reduce_sum(w_all, name = "w_sum")
iw_sum = tf.divide(tf_1f, w_sum, name = "iw_sum")
w_norm = tf.multiply (w_all, iw_sum, name = "w_norm")
disp_slice = tf.reshape(out_batch[:,0],[-1], name = "disp_slice")
d_gt_slice = tf.reshape(gt_ds_batch[:,0],[-1], name = "d_gt_slice")
td_flat = tf.reshape(target_disparity_batch,[-1], name = "td_flat")
if absolute_disparity:
adisp = disp_slice
else:
adisp = tf.add(disp_slice, td_flat, name = "adisp")
out_diff = tf.subtract(adisp, d_gt_slice, name = "out_diff")
out_diff2 = tf.square(out_diff, name = "out_diff2")
pre_cap0 = tf.abs(out_diff, name = "pre_cap0")
pre_cap = tf.multiply(pre_cap0, tf_disp_diff_slope, name = "pre_cap")
diff_cap = tf.add(pre_cap, tf_disp_diff_cap2, name = "diff_cap")
out_diff2_capped = tf.minimum(out_diff2, diff_cap, name = "out_diff2_capped")
out_wdiff2 = tf.multiply (out_diff2_capped, w_norm, name = "out_wdiff2")
cost1 = tf.reduce_sum(out_wdiff2, name = "cost1")
out_diff2_offset = tf.subtract(out_diff2, error2_offset, name = "out_diff2_offset")
out_diff2_biased = tf.maximum(out_diff2_offset, 0.0, name = "out_diff2_biased")
# calculate disparity-based weight boost
if use_out:
dispw = tf.clip_by_value(adisp, disp_wmin, disp_wmax, name = "dispw")
else:
dispw = tf.clip_by_value(td_flat, disp_wmin, disp_wmax, name = "dispw")
dispw_boost = tf.divide(disp_wmax, dispw, name = "dispw_boost")
dispw_comp = tf.multiply (dispw_boost, w_norm, name = "dispw_comp") #HERE??
if batch_weights.shape[0] > 1:
dispw_batch = tf.multiply (dispw_comp, batch_weights, name = "dispw_batch")# apply weights for high/low variance and sources
else:
dispw_batch = tf.multiply (dispw_comp, tf_1f, name = "dispw_batch")# apply weights for high/low variance and sources
dispw_sum = tf.reduce_sum(dispw_batch, name = "dispw_sum")
idispw_sum = tf.divide(tf_1f, dispw_sum, name = "idispw_sum")
dispw_norm = tf.multiply (dispw_batch, idispw_sum, name = "dispw_norm")
out_diff2_wbiased = tf.multiply(out_diff2_biased, dispw_norm, name = "out_diff2_wbiased")
# out_diff2_wbiased = tf.multiply(out_diff2_biased, w_norm, name = "out_diff2_wbiased")
cost1b = tf.reduce_sum(out_diff2_wbiased, name = "cost1b")
if use_confidence:
cost12 = tf.add(cost1b, cost2, name = "cost12")
cost123 = tf.add(cost12, cost3, name = "cost123")
return cost123, disp_slice, d_gt_slice, out_diff,out_diff2, w_norm, out_wdiff2, cost1
else:
return cost1b, disp_slice, d_gt_slice, out_diff,out_diff2, w_norm, out_wdiff2, cost1
def weightsLoss(inp_weights,
tile_layers,
tile_side,
wborders_zero):
# [batch_size,(1..2)] tf_result
# weights_lambdas): # single lambda or same length as inp_weights.shape[1]
"""
Enforcing 'smooth' weights for the input 2d correlation tiles
@return mean squared difference for each weight and average of 8 neighbors divided by mean squared weights
"""
weight_ortho = 1.0
weight_diag = 0.7
sw = 4.0 * (weight_ortho + weight_diag)
weight_ortho /= sw
weight_diag /= sw
# w_neib = tf.const([[weight_diag, weight_ortho, weight_diag],
# [weight_ortho, -1.0, weight_ortho],
# [weight_diag, weight_ortho, weight_diag]])
#WBORDERS_ZERO
with tf.name_scope("WeightsLoss"):
# Adding 1 tile border
# tf_inp = tf.reshape(inp_weights[:TILE_LAYERS * TILE_SIZE,:], [TILE_LAYERS, FILE_TILE_SIDE, FILE_TILE_SIDE, inp_weights.shape[1]], name = "tf_inp")
tf_inp = tf.reshape(inp_weights[:tile_layers * tile_side * tile_side,:], [tile_layers, tile_side, tile_side, inp_weights.shape[1]], name = "tf_inp")
if wborders_zero:
tf_zero_col = tf.constant(0.0, dtype=tf.float32, shape=[tf_inp.shape[0], tf_inp.shape[1], 1, tf_inp.shape[3]], name = "tf_zero_col")
tf_zero_row = tf.constant(0.0, dtype=tf.float32, shape=[tf_inp.shape[0], 1 , tf_inp.shape[2] + 2, tf_inp.shape[3]], name = "tf_zero_row")
tf_inp_ext_h = tf.concat([tf_zero_col, tf_inp, tf_zero_col ], axis = 2, name ="tf_inp_ext_h")
tf_inp_ext = tf.concat([tf_zero_row, tf_inp_ext_h, tf_zero_row ], axis = 1, name ="tf_inp_ext")
else:
tf_inp_ext_h = tf.concat([tf_inp [:, :, :1, :], tf_inp, tf_inp [:, :, -1:, :]], axis = 2, name ="tf_inp_ext_h")
tf_inp_ext = tf.concat([tf_inp_ext_h [:, :1, :, :], tf_inp_ext_h, tf_inp_ext_h[:, -1:, :, :]], axis = 1, name ="tf_inp_ext")
s_ortho = tf_inp_ext[:,1:-1,:-2,:] + tf_inp_ext[:,1:-1, 2:,:] + tf_inp_ext[:,1:-1,:-2,:] + tf_inp_ext[:,1:-1, 2:, :]
s_corn = tf_inp_ext[:, :-2,:-2,:] + tf_inp_ext[:, :-2, 2:,:] + tf_inp_ext[:,2:, :-2,:] + tf_inp_ext[:,2: , 2:, :]
w_diff = tf.subtract(tf_inp, s_ortho * weight_ortho + s_corn * weight_diag, name="w_diff")
w_diff2 = tf.multiply(w_diff, w_diff, name="w_diff2")
w_var = tf.reduce_mean(w_diff2, name="w_var")
w2_mean = tf.reduce_mean(inp_weights * inp_weights, name="w2_mean")
w_rel = tf.divide(w_var, w2_mean, name= "w_rel")
return w_rel # scalar, cost for weights non-smoothness in 2d
qcstereo_network.py 0 → 100644
View file @ b1111bca
#!/usr/bin/env python3
__copyright__ = "Copyright 2018, Elphel, Inc."
__license__ = "GPL-3.0+"
__email__ = "andrey@elphel.com"
#from numpy import float64
#import numpy as np
import tensorflow as tf
import tensorflow.contrib.slim as slim
NN_LAYOUTS = {0:[0, 0, 0, 32, 20, 16],
1:[0, 0, 0, 256, 128, 64],
2:[0, 128, 32, 32, 32, 16],
3:[0, 0, 40, 32, 20, 16],
4:[0, 0, 0, 0, 16, 16],
5:[0, 0, 64, 32, 32, 16],
6:[0, 0, 32, 16, 16, 16],
7:[0, 0, 64, 16, 16, 16],
8:[0, 0, 0, 64, 20, 16],
9:[0, 0, 256, 64, 32, 16],
10:[0, 256, 128, 64, 32, 16],
11:[0, 0, 0, 0, 64, 32],
12:[0, 0, 256, 128, 64, 32],
13:[0, 0, 0, 256, 128, 32],
}
def lrelu(x):
return tf.maximum(x*0.2,x)
# return tf.nn.relu(x)
def sym_inputs8(inp, cluster_radius = 2):
"""
get input vector [?:4*9*9+1] (last being target_disparity) and reorder for horizontal flip,
vertical flip and transpose (8 variants, mode + 1 - hor, +2 - vert, +4 - transpose)
return same lengh, reordered
"""
tile_side = 2 * cluster_radius + 1
with tf.name_scope("sym_inputs8"):
td = inp[:,-1:] # tf.reshape(inp,[-1], name = "td")[-1]
inp_corr = tf.reshape(inp[:,:-1],[-1,4,tile_side,tile_side], name = "inp_corr")
inp_corr_h = tf.stack([-inp_corr [:,0,:,-1::-1], inp_corr [:,1,:,-1::-1], -inp_corr [:,3,:,-1::-1], -inp_corr [:,2,:,-1::-1]], axis=1, name = "inp_corr_h")
inp_corr_v = tf.stack([ inp_corr [:,0,-1::-1,:],-inp_corr [:,1,-1::-1,:], inp_corr [:,3,-1::-1,:], inp_corr [:,2,-1::-1,:]], axis=1, name = "inp_corr_v")
inp_corr_hv = tf.stack([ inp_corr_h[:,0,-1::-1,:],-inp_corr_h[:,1,-1::-1,:], inp_corr_h[:,3,-1::-1,:], inp_corr_h[:,2,-1::-1,:]], axis=1, name = "inp_corr_hv")
inp_corr_t = tf.stack([tf.transpose(inp_corr [:,1], perm=[0,2,1]),
tf.transpose(inp_corr [:,0], perm=[0,2,1]),
tf.transpose(inp_corr [:,2], perm=[0,2,1]),
-tf.transpose(inp_corr [:,3], perm=[0,2,1])], axis=1, name = "inp_corr_t")
inp_corr_ht = tf.stack([tf.transpose(inp_corr_h [:,1], perm=[0,2,1]),
tf.transpose(inp_corr_h [:,0], perm=[0,2,1]),
tf.transpose(inp_corr_h [:,2], perm=[0,2,1]),
-tf.transpose(inp_corr_h [:,3], perm=[0,2,1])], axis=1, name = "inp_corr_ht")
inp_corr_vt = tf.stack([tf.transpose(inp_corr_v [:,1], perm=[0,2,1]),
tf.transpose(inp_corr_v [:,0], perm=[0,2,1]),
tf.transpose(inp_corr_v [:,2], perm=[0,2,1]),
-tf.transpose(inp_corr_v [:,3], perm=[0,2,1])], axis=1, name = "inp_corr_vt")
inp_corr_hvt = tf.stack([tf.transpose(inp_corr_hv[:,1], perm=[0,2,1]),
tf.transpose(inp_corr_hv[:,0], perm=[0,2,1]),
tf.transpose(inp_corr_hv[:,2], perm=[0,2,1]),
-tf.transpose(inp_corr_hv[:,3], perm=[0,2,1])], axis=1, name = "inp_corr_hvt")
# return td, [inp_corr, inp_corr_h, inp_corr_v, inp_corr_hv, inp_corr_t, inp_corr_ht, inp_corr_vt, inp_corr_hvt]
"""
return [tf.concat([tf.reshape(inp_corr, [inp_corr.shape[0],-1]),td], axis=1,name = "out_corr"),
tf.concat([tf.reshape(inp_corr_h, [inp_corr.shape[0],-1]),td], axis=1,name = "out_corr_h"),
tf.concat([tf.reshape(inp_corr_v, [inp_corr.shape[0],-1]),td], axis=1,name = "out_corr_v"),
tf.concat([tf.reshape(inp_corr_hv, [inp_corr.shape[0],-1]),td], axis=1,name = "out_corr_hv"),
tf.concat([tf.reshape(inp_corr_t, [inp_corr.shape[0],-1]),td], axis=1,name = "out_corr_t"),
tf.concat([tf.reshape(inp_corr_ht, [inp_corr.shape[0],-1]),td], axis=1,name = "out_corr_ht"),
tf.concat([tf.reshape(inp_corr_vt, [inp_corr.shape[0],-1]),td], axis=1,name = "out_corr_vt"),
tf.concat([tf.reshape(inp_corr_hvt,[inp_corr.shape[0],-1]),td], axis=1,name = "out_corr_hvt")]
"""
cl = 4 * tile_side * tile_side
return [tf.concat([tf.reshape(inp_corr, [-1,cl]),td], axis=1,name = "out_corr"),
tf.concat([tf.reshape(inp_corr_h, [-1,cl]),td], axis=1,name = "out_corr_h"),
tf.concat([tf.reshape(inp_corr_v, [-1,cl]),td], axis=1,name = "out_corr_v"),
tf.concat([tf.reshape(inp_corr_hv, [-1,cl]),td], axis=1,name = "out_corr_hv"),
tf.concat([tf.reshape(inp_corr_t, [-1,cl]),td], axis=1,name = "out_corr_t"),
tf.concat([tf.reshape(inp_corr_ht, [-1,cl]),td], axis=1,name = "out_corr_ht"),
tf.concat([tf.reshape(inp_corr_vt, [-1,cl]),td], axis=1,name = "out_corr_vt"),
tf.concat([tf.reshape(inp_corr_hvt,[-1,cl]),td], axis=1,name = "out_corr_hvt")]
# inp_corr_h, inp_corr_v, inp_corr_hv, inp_corr_t, inp_corr_ht, inp_corr_vt, inp_corr_hvt]
def network_sub(input_tensor,
input_global, #add to all layers (but first) if not None
layout,
reuse,
sym8 = False,
cluster_radius = 2):
# last_indx = None;
fc = []
inp_weights = []
for i, num_outs in enumerate (layout):
if num_outs:
if fc:
if input_global is None:
inp = fc[-1]
else:
inp = tf.concat([fc[-1], input_global], axis = 1)
fc.append(slim.fully_connected(inp, num_outs, activation_fn=lrelu, scope='g_fc_sub'+str(i), reuse = reuse))
else:
inp = input_tensor
if sym8:
inp8 = sym_inputs8(inp, cluster_radius)
num_non_sum = num_outs % len(inp8) # if number of first layer outputs is not multiple of 8
num_sym8 = num_outs // len(inp8) # number of symmetrical groups
fc_sym = []
for j in range (len(inp8)): # ==8
reuse_this = reuse | (j > 0)
scp = 'g_fc_sub'+str(i)
fc_sym.append(slim.fully_connected(inp8[j], num_sym8, activation_fn=lrelu, scope= scp, reuse = reuse_this))
if not reuse_this:
with tf.compat.v1.variable_scope(scp,reuse=True) : # tf.AUTO_REUSE):
inp_weights.append(tf.get_variable('weights')) # ,shape=[inp.shape[1],num_outs]))
if num_non_sum > 0:
reuse_this = reuse
scp = 'g_fc_sub'+str(i)+"r"
fc_sym.append(slim.fully_connected(inp, num_non_sum, activation_fn=lrelu, scope=scp, reuse = reuse_this))
if not reuse_this:
with tf.compat.v1.variable_scope(scp,reuse=True) : # tf.AUTO_REUSE):
inp_weights.append(tf.get_variable('weights')) # ,shape=[inp.shape[1],num_outs]))
fc.append(tf.concat(fc_sym, 1, name='sym_input_layer'))
else:
scp = 'g_fc_sub'+str(i)
fc.append(slim.fully_connected(inp, num_outs, activation_fn=lrelu, scope= scp, reuse = reuse))
if not reuse:
with tf.compat.v1.variable_scope(scp, reuse=True) : # tf.AUTO_REUSE):
inp_weights.append(tf.get_variable('weights')) # ,shape=[inp.shape[1],num_outs]))
return fc[-1], inp_weights
def network_inter(input_tensor,
input_global, #add to all layers (but first) if not None
layout,
reuse=False,
use_confidence=False):
#last_indx = None;
fc = []
for i, num_outs in enumerate (layout):
if num_outs:
if fc:
if input_global is None:
inp = fc[-1]
else:
inp = tf.concat([fc[-1], input_global], axis = 1)
else:
inp = input_tensor
fc.append(slim.fully_connected(inp, num_outs, activation_fn=lrelu, scope='g_fc_inter'+str(i), reuse = reuse))
if use_confidence:
fc_out = slim.fully_connected(fc[-1], 2, activation_fn=lrelu, scope='g_fc_inter_out', reuse = reuse)
else:
fc_out = slim.fully_connected(fc[-1], 1, activation_fn=None, scope='g_fc_inter_out', reuse = reuse)
#If using residual disparity, split last layer into 2 or remove activation and add rectifier to confidence only
return fc_out
def networks_siam(input_tensor, # now [?,9,325]-> [?,25,325]
input_global, # add to all layers (but first) if not None
layout1,
layout2,
inter_convergence,
sym8 = False,
only_tile = None, # just for debugging - feed only data from the center sub-network
partials = None,
use_confidence=False,
cluster_radius = 2):
center_index = (input_tensor.shape[1] - 1) // 2
with tf.name_scope("Siam_net"):
inp_weights = []
num_legs = input_tensor.shape[1] # == 25
if partials is None:
partials = [[True] * num_legs]
inter_lists = [[] for _ in partials]
reuse = False
for i in range (num_legs):
if ((only_tile is None) or (i == only_tile)) and any([p[i] for p in partials]) :
if input_global is None:
ig = None
else:
ig =input_global[:,i,:]
ns, ns_weights = network_sub(input_tensor[:,i,:],
ig, # input_global[:,i,:],
layout= layout1,
reuse= reuse,
sym8 = sym8,
cluster_radius = cluster_radius)
for n, partial in enumerate(partials):
if partial[i]:
inter_lists[n].append(ns)
else:
inter_lists[n].append(tf.zeros_like(ns))
inp_weights += ns_weights
reuse = True
outs = []
for n, _ in enumerate(partials):
if input_global is None:
ig = None
else:
ig =input_global[:,center_index,:]
outs.append(network_inter (input_tensor = tf.concat(inter_lists[n],
axis=1,
name='inter_tensor'+str(n)),
input_global = [None, ig][inter_convergence], # optionally feed all convergence values (from each tile of a cluster)
layout = layout2,
reuse = (n > 0),
use_confidence = use_confidence))
return outs, inp_weights
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment