display weights

nn_ds_inmem4_tmp.py

 #!/usr/bin/env python3
 from numpy import float64
+from _stat import S_IEXEC
 
 __copyright__ = "Copyright 2018, Elphel, Inc."
 __license__   = "GPL-3.0+"
@@ -33,17 +34,17 @@ MAX_EPOCH =        500
 #LR =               1e-4 # learning rate
 LR =               1e-3 # learning rate
 USE_CONFIDENCE =     False
-ABSOLUTE_DISPARITY = False # True # False
+ABSOLUTE_DISPARITY = True # True # False
 DEBUG_PLT_LOSS =     True
 FEATURES_PER_TILE =  324
 EPOCHS_TO_RUN =     10000 #0
 RUN_TOT_AVG =       100 # last batches to average. Epoch is 307 training  batches
 BATCH_SIZE =       1000 # Each batch of tiles has balanced D/S tiles, shuffled batches but not inside batches
 SHUFFLE_EPOCH =    True
-NET_ARCH =           3 # overwrite with argv?
+NET_ARCH =           0 # overwrite with argv?
 #DEBUG_PACK_TILES = True
 SUFFIX=str(NET_ARCH)+ (["R","A"][ABSOLUTE_DISPARITY])
-MAX_TRAIN_FILES_TFR = 4
+MAX_TRAIN_FILES_TFR = 6
 #http://stackoverflow.com/questions/287871/print-in-terminal-with-colors-using-python
 class bcolors:
     HEADER = '\033[95m'
@@ -211,6 +212,9 @@ def lrelu(x):
 #    return tf.nn.relu(x)
 
 def network_fc_simple(input, arch = 0):
+
+    global image_summary_op1
+
     layouts = {0:[0,   0,   0,   32,  20,  16],
                1:[0,   0,   0,  256, 128,  64],
                2:[0, 128,  32,   32,  32,  16],
@@ -226,9 +230,86 @@ def network_fc_simple(input, arch = 0):
                inp = input
 
             fc.append(slim.fully_connected(inp, num_outs, activation_fn=lrelu,scope='g_fc'+str(i)))
-            with tf.variable_scope('g_fc'+str(i)+'/fully_connected',reuse=tf.AUTO_REUSE):
+            
+            #with tf.variable_scope('g_fc'+str(i)+'/fully_connected',reuse=tf.AUTO_REUSE):
+            with tf.variable_scope('g_fc'+str(i),reuse=tf.AUTO_REUSE):
+                
                 w = tf.get_variable('weights',shape=[inp.shape[1],num_outs])
-                b = tf.get_variable('weights',shape=[inp.shape[1],num_outs])
+                #image = tf.get_variable('w_images',shape=[1, inp.shape[1],num_outs,1])
+                if (i==3):
+                    
+                    # red border
+                    grid = tf.constant([255,100,100],dtype=tf.float32,name="GRID")
+                    
+                    # (325,32)
+                    wimg_1 = w
+                    # (32,325)
+                    wimg_2 = tf.transpose(wimg_1,[1,0])
+                    # (32,324)
+                    wimg_3 = wimg_2[:,:-1]
+                    
+                    # res?
+                    #wimg_res = tf.get_variable('wimg_res',shape=[32*(9+1),(9+1)*4, 3])
+                    
+                    # long list                    
+                    tmp1 = []
+                    for mi in range(32):
+                        tmp2 = []
+                        for mj in range(4):
+                            
+                            s_i = mj*81
+                            e_i = (mj+1)*81
+                            
+                            tile = tf.reshape(wimg_3[mi,s_i:e_i],shape=(9,9))                            
+                            tiles = tf.stack([tile]*3,axis=2)
+                            
+                            #gtiles1 = tf.concat([tiles, tf.reshape(9*[grid],shape=(1,9,3))],axis=0)
+                            gtiles1 = tf.concat([tiles,  tf.expand_dims(9*[grid],0)],axis=0)
+                            gtiles2 = tf.concat([gtiles1,tf.expand_dims(10*[grid],1)],axis=1)
+                            tmp2.append(gtiles2)
+                        
+                        ts = tf.concat(tmp2,axis=2)
+                        tmp1.append(ts)
+                    
+                    image_summary_op2 = tf.concat(tmp1,axis=0)
+                                                
+                    #image_summary_op1 = tf.assign(wimg_res,tf.zeros(shape=[32*(9+1),(9+1)*4, 3],dtype=tf.float32))
+                    
+                    #wimgo1 = tf.zeros(shape=[32*(9+1),(9+1)*4, 3],dtype=tf.float32)
+                                                            
+                    #tf.summary.image("wimg_res1",tf.reshape(wimg_res,[1,32*(9+1),(9+1)*4, 3]))
+                    
+                    #tf.summary.image("wimgo1",tf.reshape(wimgo1,[1,32*(9+1),(9+1)*4, 3]))
+                    #tf.summary.image("wimgo2",tf.reshape(wimgo2,[1,32*(9+1),(9+1)*4, 3]))
+                    tf.summary.image("SWEIGTS",tf.reshape(gtiles2,[1,10,10,3]))
+                    tf.summary.image("WEIGTS",tf.reshape(image_summary_op2,[1,320,40,3]))
+                    
+                    # borders
+                    #for mi in range(0,wimg_res.shape[0],10):
+                    #    for mj in range(wimg_res.shape[1]):
+                    #        wimg_res[mi,mj].assign([255,255,255])
+                    
+                    #wimg_res[9::(9+1),:].assign([255,0,0])
+                    #wimg_res[:,9::(9+1)].assign([255,0,0])
+                    
+                    #for mi in range(0,wimg_res.shape[0],10):
+                    #    print(mi)
+                        #wimg_res = tf.stack([wing_res,])
+                    
+                    #wimg_1  = tf.reshape(w,[1,inp.shape[1],num_outs,1])
+                    #wimg_1t = tf.transpose(wimg_1,[0,2,1,3])
+                    
+                    # w = w[a,b]
+                    # wt = w[b,a]
+                    # for i in range(b):
+                    #     tmp =  
+                    
+                    
+                    #tf.summary.image("wimg_1",wimg_1)
+                    #tf.summary.image("wimg_1t",wimg_1t)
+                    #tf.summary.image("wimg_res1",tf.reshape(wimg_res,[1,32*(9+1),(9+1)*4, 3]))
+                
+                b = tf.get_variable('biases',shape=[num_outs])
                 tf.summary.histogram("weights",w)
                 tf.summary.histogram("biases",b)
     """
@@ -247,7 +328,8 @@ def network_fc_simple(input, arch = 0):
 
         with tf.variable_scope('g_fc_out',reuse=tf.AUTO_REUSE):
             w = tf.get_variable('weights',shape=[fc[-1].shape[1],2])
-            b = tf.get_variable('biases',shape=[fc[-1].shape[1],2])
+            tf.summary.image("wimage",tf.reshape(w,[1,fc[-1].shape[1],2,1]))
+            b = tf.get_variable('biases',shape=[2])
             tf.summary.histogram("weights",w)
             tf.summary.histogram("biases",b)
 
@@ -256,6 +338,7 @@ def network_fc_simple(input, arch = 0):
 
         with tf.variable_scope('g_fc_out',reuse=tf.AUTO_REUSE):
             w = tf.get_variable('weights',shape=[fc[-1].shape[1],1])
+            tf.summary.image("wimage",tf.reshape(w,[1,fc[-1].shape[1],1,1]))
             b = tf.get_variable('biases',shape=[1])
             tf.summary.histogram("weights",w)
             tf.summary.histogram("biases",b)
@@ -504,6 +587,8 @@ with tf.Session()  as sess:
 #       if SHUFFLE_EPOCH:
 #        dataset_train = dataset_train.shuffle(buffer_size=10000)
 
+        # RUN TRAIN SESSION
+
         sess.run(iterator_train.initializer, feed_dict={corr2d_train_placeholder: corr2d_trains[train_file_index],
                                                         target_disparity_train_placeholder: target_disparity_trains[train_file_index],
                                                         gt_ds_train_placeholder: gt_ds_trains[train_file_index]})
@@ -530,6 +615,7 @@ with tf.Session()  as sess:
                 #train_writer.add_summary(train_summary, i)
                 loss_train_hist[i] =  G_loss_trained
                 loss2_train_hist[i] = out_cost1
+                                
             except tf.errors.OutOfRangeError:
                 print("train done at step %d"%(i))
                 break
@@ -537,10 +623,7 @@ with tf.Session()  as sess:
         train_avg = np.average(loss_train_hist).astype(np.float32)
         train2_avg = np.average(loss2_train_hist).astype(np.float32)
 
-        #_,_=sess.run([tf_ph_G_loss,tf_ph_sq_diff],feed_dict={tf_ph_G_loss:train_avg, tf_ph_sq_diff:train2_avg})
-
-#tf_ph_G_loss = tf.placeholder(tf.float32,shape=None,name='G_loss_avg')
-#tf_ph_sq_diff = tf.placeholder(tf.float32,shape=None,name='sq_diff_avg')
+        # RUN TEST SESSION
 
         sess.run(iterator_train.initializer, feed_dict={corr2d_train_placeholder: corr2d_test,
                                                         target_disparity_train_placeholder: target_disparity_test,
@@ -567,12 +650,10 @@ with tf.Session()  as sess:
                 print("test done at step %d"%(i))
                 break
 
-#            print_time("%d:%d -> %f"%(epoch,i,G_current))
         test_avg =  np.average(loss_test_hist).astype(np.float32)
         test2_avg = np.average(loss2_test_hist).astype(np.float32)
         
-#        _,_=sess.run([tf_ph_G_loss,tf_ph_sq_diff],feed_dict={tf_ph_G_loss:test_avg, tf_ph_sq_diff:test2_avg})
-        
+        # they include image summaries as well
         train_writer.add_summary(train_summary, epoch)
         test_writer.add_summary(test_summary, epoch)