Commit 21896135 authored by Clement Vachet's avatar Clement Vachet
Browse files

Update configuration filenames

parent fc5c34cb

AI_Inference_CSV.py

+4 −33
Original line number Diff line number Diff line
@@ -162,29 +162,18 @@ def main(args=None): 
			padding_mode = padding_mode,

		)

		len_grid_sampler = len(grid_sampler)

		#print('length grid_sampler', len(grid_sampler))



		patch_loader = torch.utils.data.DataLoader(grid_sampler, batch_size=bs)

		aggregator = tio.inference.GridAggregator(grid_sampler, overlap_mode = 'average')



		with torch.no_grad():

			for patch_idx, patches_batch in enumerate(patch_loader):

				# print('\n\t\t patch_idx: ', patch_idx)



				#print('\t\t Preparing data...')

				inputs = patches_batch['Combined'][tio.DATA]

				# print('\t\t inputs shape: ', inputs.shape)

				input1_tiles, input2_tiles_real, GroundTruth_real = dataset.prepare_data_withfiltering(inputs, nb_image_layers, nb_corr_layers, tile_size, adjacent_tiles_dim)

				#print('\t\t Preparing data - done -')



				input1_tiles = input1_tiles.to(device)

				input2_tiles_real = input2_tiles_real.to(device)

				#GroundTruth_real = GroundTruth_real.to(self.device)

				# Reducing last dimension to compute loss

				#GroundTruth_real = torch.squeeze(GroundTruth_real, dim=2)



				# print('\t\t input1_tiles shape: ', input1_tiles.shape)

				# print('\t\t input2_tiles_real shape:', input2_tiles_real.shape)



				if mc_dropout:

					# Perform multiple inference (mc_passes)
@@ -193,47 +182,37 @@ def main(args=None): 
						outputs = model(input1_tiles, input2_tiles_real)

						outputs_all[i] = torch.squeeze(outputs)



					# Compute mean, median, std, CV (coefficient of variation), SE (standard error)

					# Compute mean, median, std, CV (coefficient of variation)

					outputs_mean = torch.mean(outputs_all,0)

					outputs_median = torch.median(outputs_all,0)[0]

					outputs_std = torch.std(outputs_all,0)

					outputs_cv = torch.div(outputs_std, torch.abs(outputs_mean))

					# outputs_se = torch.div(outputs_std, math.sqrt(mc_passes))

					outputs_combined = torch.stack((outputs_mean, outputs_median, outputs_cv), dim=1)

				else:

					outputs_combined = model(input1_tiles, input2_tiles_real)

				# print('\t\t outputs_combined shape: ', outputs_combined.shape)

				# print('outputs_combined device', outputs_combined.device)



				# Reshape outputs

				outputs_combined_reshape = torch.reshape(outputs_combined,[outputs_combined.shape[0],outputs_combined.shape[1],1,1,1])

				print('\t\t outputs_combined_reshape shape: ', outputs_combined_reshape.shape)

				

				input_location = patches_batch[tio.LOCATION]

				# print('\t\t input_location shape: ', input_location.shape)

				# print('\t\t input_location: ', input_location)





				# Reshape input_location to prediction_location, to fit output image size (78,62,1)

				pred_location = dataset.prediction_patch_location(input_location, tile_size, adjacent_tiles_dim)

				# print('\t\t pred_location shape: ', pred_location.shape)

				# print('\t\t pred_location: ', pred_location)





				# Add batch with location to TorchIO aggregator

				aggregator.add_batch(outputs_combined_reshape, pred_location)



		# output_tensor shape [3, 1170, 930, 122]

		output_tensor_combined = aggregator.get_output_tensor()

		# print('output_tensor_combined type: ', output_tensor_combined.dtype)

		# print('output_tensor_combined device', output_tensor_combined.device)

		# print('output_tensor_combined shape: ', output_tensor_combined.shape)



		# Extract real information of interest [3,78,62]

		output_tensor_combined_real = output_tensor_combined[:,:NbTiles_H,:NbTiles_W,0]

		# print('output_tensor_combined_real shape: ', output_tensor_combined_real.shape)



		output_combined_np = output_tensor_combined_real.numpy().squeeze()

		# print('output_combined_np type', output_combined_np.dtype)

		# print('output_combined_np shape', output_combined_np.shape)





		if mc_dropout:

			output_mean_np = output_combined_np[0,...]
@@ -243,26 +222,18 @@ def main(args=None): 
			imageio_output_mean = np.moveaxis(output_mean_np, 0,1)

			imageio_output_median = np.moveaxis(output_median_np, 0,1)

			imageio_output_cv = np.moveaxis(output_cv_np, 0,1)

			# print('imageio_output_mean shape', imageio_output_mean.shape)

			# print('imageio_output_median shape', imageio_output_median.shape)

			# print('imageio_output_cv shape', imageio_output_cv.shape)

		else:

			output_np = output_combined_np

			imageio_output = np.moveaxis(output_np, 0,1)

			# print('imageio_output shape', imageio_output.shape)



		time_elapsed2 = time.time() - since2

		time_inference_list.append(time_elapsed2)



		if mc_dropout:

			# print('Writing output mean image via imageio...')

			imageio.imwrite(PredictionFile_mean, imageio_output_mean)

			# print('Writing output median image via imageio...')

			imageio.imwrite(PredictionFile_median, imageio_output_median)

			# print('Writing output CV image via imageio...')

			imageio.imwrite(PredictionFile_cv, imageio_output_cv)

		else:

			# print('Writing output image via imageio...')

			imageio.imwrite(PredictionFile, imageio_output)



		print('\t\t Inference in {:.2f}s---'.format(time_elapsed2))

AI_Inference_Direct.py

+3 −54
Original line number Diff line number Diff line
@@ -92,19 +92,11 @@ def main(args=None): 
	NbImageLayers = InputFile_Shape[3]

	NbCorrLayers = NbImageLayers -4

	InputDepth = NbCorrLayers

	print('InputFile_Shape: ', InputFile_Shape)

	print('NbTiles_H: ', NbTiles_H)

	print('NbTiles_W: ', NbTiles_W)

	print('NbImageLayers: ', NbImageLayers)

	print('InputDepth: ', InputDepth)





	# GridSampler

	print('\nGenerating Grid Sampler...')

	patch_size, patch_overlap, padding_mode = dataset.initialize_gridsampler_variables(NbImageLayers, TileSize, AdjacentTilesDim, padding_mode=None)

	print('patch_size: ',patch_size)

	print('patch_overlap: ',patch_overlap)

	print('padding_mode: ',padding_mode)

	



	grid_sampler = tio.data.GridSampler(
@@ -114,7 +106,6 @@ def main(args=None): 
		padding_mode = padding_mode,

	)

	len_grid_sampler = len(grid_sampler)

	print('length grid_sampler', len(grid_sampler))

	

	patch_loader = torch.utils.data.DataLoader(grid_sampler, batch_size=bs)

	aggregator = tio.data.GridAggregator(grid_sampler, overlap_mode = 'average')
@@ -134,22 +125,12 @@ def main(args=None): 
	#model = nn.Identity().eval()

	with torch.no_grad():

		for patch_idx, patches_batch in enumerate(patch_loader):

			print('\t patch_idx: ', patch_idx)



			#print('\t\t Preparing data...')

			inputs = patches_batch['Combined'][tio.DATA]

			print('\t\t inputs shape: ', inputs.shape)

			input1_tiles, input2_tiles_real, GroundTruth_real = dataset.prepare_data_withfiltering(inputs, NbImageLayers, NbCorrLayers, TileSize, AdjacentTilesDim)

			#print('\t\t Preparing data - done -')



			input1_tiles = input1_tiles.to(device)

			input2_tiles_real = input2_tiles_real.to(device)

			#GroundTruth_real = GroundTruth_real.to(device)

			# Reducing last dimension to compute loss

			#GroundTruth_real = torch.squeeze(GroundTruth_real, dim=2)



			print('\t\t input1_tiles shape: ', input1_tiles.shape)

			print('\t\t input2_tiles_real shape:', input2_tiles_real.shape)



			if mc_dropout:

				# Perform multiple inference (mc_passes)
@@ -158,62 +139,34 @@ def main(args=None): 
					outputs = model(input1_tiles, input2_tiles_real)

					outputs_all[i] = torch.squeeze(outputs)



				# Compute mean, std, CV (coefficient of variation), SE (standard error)

				# Compute mean, std, CV (coefficient of variation)

				outputs_mean = torch.mean(outputs_all,0)

				outputs_median = torch.median(outputs_all,0)[0]

				outputs_std = torch.std(outputs_all,0)

				outputs_cv = torch.div(outputs_std, torch.abs(outputs_mean))

				# outputs_se = torch.div(outputs_std, math.sqrt(mc_passes))

				outputs_combined = torch.stack((outputs_mean, outputs_median, outputs_cv), dim=1)

				

				print('\t\t outputs shape: ',outputs.shape)

				print('\t\t outputs device', outputs.device)

				print('\t\t outputs_all shape: ', outputs_all.shape)

				print('\t\t outputs_all device', outputs_all.device)

				print('\t\t outputs_mean shape: ', outputs_mean.shape)

				print('\t\t outputs_median shape: ', outputs_median.shape)

				print('\t\t outputs_median type: ', outputs_median.type())

				print('\t\t outputs_combined shape: ', outputs_combined.shape)



				print('\t\t outputs_mean[:20]',outputs_mean[:20])

				print('\t\t outputs_median[:20]',outputs_median[:20])

				print('\t\t outputs_std[:20]',outputs_std[:20])

				print('\t\t outputs_cv[:20]',outputs_cv[:20])

			else:

				outputs_combined = model(input1_tiles, input2_tiles_real)

			

			print('\t\t outputs_combined device', outputs_combined.device)

			print('\t\t outputs_combined shape: ', outputs_combined.shape)



			# Reshape outputs to match location dimensions

			outputs_combined_reshape = torch.reshape(outputs_combined,[outputs_combined.shape[0],outputs_combined.shape[1],1,1,1])

			print('\t\t outputs_combined_reshape shape: ', outputs_combined_reshape.shape)

			

			input_location = patches_batch[tio.LOCATION]

			print('\t\t input_location shape: ', input_location.shape)

			print('\t\t input_location type: ', input_location.dtype)

			print('\t\t input_location[:20]: ', input_location[:20])



			# Reshape input_location to prediction_location, to fit output image size (78,62,1)

			pred_location = dataset.prediction_patch_location(input_location, TileSize, AdjacentTilesDim)

			print('\t\t pred_location shape: ', pred_location.shape)

			print('\t\t pred_location[:20]: ', pred_location[:20])



			# Add batch with location to TorchIO aggregator

			aggregator.add_batch(outputs_combined_reshape, pred_location)



	# output_tensor shape [3, 1170, 930, 124]

	output_tensor_combined = aggregator.get_output_tensor()

	print('output_tensor_combined type: ', output_tensor_combined.dtype)

	print('output_tensor_combined shape: ', output_tensor_combined.shape)



	# Extract real information of interest [3, 78,62]

	output_tensor_combined_real = output_tensor_combined[:,:NbTiles_H,:NbTiles_W,0]

	print('output_tensor_combined_real shape: ', output_tensor_combined_real.shape)



	output_combined_np = output_tensor_combined_real.numpy().squeeze()

	print('output_combined_np type', output_combined_np.dtype)

	print('output_combined_np shape', output_combined_np.shape)



	if mc_dropout:

		output_mean_np = output_combined_np[0,...]
@@ -223,14 +176,10 @@ def main(args=None): 
		imageio_output_mean = np.moveaxis(output_mean_np, 0,1)

		imageio_output_median = np.moveaxis(output_median_np, 0,1)

		imageio_output_cv = np.moveaxis(output_cv_np, 0,1)

		print('imageio_output_mean shape', imageio_output_mean.shape)

		print('imageio_output_median shape', imageio_output_median.shape)

		print('imageio_output_cv shape', imageio_output_cv.shape)



	else:

		output_np = output_combined_np

		imageio_output = np.moveaxis(output_np, 0,1)

		print('imageio_output shape', imageio_output.shape)

		

	

	time_elapsed2 = time.time() - since2

README.md

+2 −2
Original line number Diff line number Diff line
@@ -29,12 +29,12 @@ Update your CSV files to point to your training and validation datasets. 


Command line:

```

python3 AI_Training.py --config ./Config_Files/AI_Training_Config.yaml

python3 AI_Training.py --config ./Config_Files/AI_Training_Config_Tiles3x3.yaml

``` 

 

###  3.2 DNN inference using configuration file



Command line:

```

python3 AI_Inference_CSV.py --config ./Config_Files/AI_Inference_Config.yaml --verbose

python3 AI_Inference_CSV.py --config ./Config_Files/AI_Inference_Config_Tiles3x3.yaml --verbose

```