![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10796179%2Fb3627d139c3ef770771ef54605f484d4%2F1.jpg?generation=1687875857220799&alt=media\")
Based on advice I tried to use Segmentation-models-pytorch for this task. Here is link:
\nhttps://github.com/qubvel/segmentation_models.pytorch/blob/master/examples/cars%20segmentation%20(camvid).ipynb
Below is an extract of useful code: model, loss, optimizer, TrainEpochs, Dataloaders. Only thing changed is CLASSES = ['blood_vessel', 'glomerulus', 'unsure'].
\nENCODER = \nENCODER_WEIGHTS = \nCLASSES = [, , ]\nACTIVATION = # could be None logits multiclass segmentation\nDEVICE = \n\nmodel = smp.FPN(\n =ENCODER, \n =ENCODER_WEIGHTS, \n =len(CLASSES), \n =ACTIVATION,\n)\n\npreprocessing_fn = smp.encoders.get_preprocessing_fn(ENCODER, ENCODER_WEIGHTS)\n\n segmentation_models_pytorch import utils\n\nloss = smp.utils.losses.DiceLoss()\nmetrics = [\n smp.utils.metrics.IoU(=0.5),\n]\n\noptimizer = torch.optim.Adam([ \n dict(=model.parameters(), =0.0001),\n])\ntrain_epoch = smp.utils.train.TrainEpoch(\n model, \n =loss, \n =metrics, \n =optimizer,\n =DEVICE,\n =,\n)\n\nvalid_epoch = smp.utils.train.ValidEpoch(\n model, \n =loss, \n =metrics, \n =DEVICE,\n =,\n)\n torch.utils.data DataLoader\n torch.utils.data Dataset BaseDataset\n\n ():\n\n CLASSES = [, , ]\n\n ():\n self.ids = os.listdir(images_dir)\n self.images_fps = [os.path.join(images_dir, image_id) image_id self.ids]\n self.masks_fps = [os.path.join(masks_dir, image_id) image_id self.ids]\n\n \n self.class_values = [self.CLASSES.index(cls.lower()) cls classes]\n (self.class_values)\n\n self.augmentation = augmentation\n self.preprocessing = preprocessing\n\n ():\n\n \n image = cv2.imread(self.images_fps[i])\n image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n mask = cv2.imread(self.masks_fps[i], )\n \n masks = [(mask == v) v self.class_values]\n mask = np.stack(masks, axis=-).astype()\n\n \n self.augmentation:\n sample = self.augmentation(image=image, mask=mask)\n image, mask = sample[], sample[]\n\n \n self.preprocessing:\n sample = self.preprocessing(image=image, mask=mask)\n image, mask = sample[], sample[]\n\n image, mask\n\n ():\n (self.ids)\n
x_valid_dir = \ny_valid_dir = \nx_train_dir = \ny_train_dir = \n\n\ntrain_dataset = Dataset(\n x_train_dir, \n y_train_dir, \n =get_training_augmentation(), \n =get_preprocessing(preprocessing_fn),\n =CLASSES,\n)\n\nvalid_dataset = Dataset(\n x_valid_dir, \n y_valid_dir, \n =get_validation_augmentation(), \n =get_preprocessing(preprocessing_fn),\n =CLASSES,\n)\n\ntrain_loader = DataLoader(train_dataset, =8, =, =12)\nvalid_loader = DataLoader(valid_dataset, =1, =, =4)\nmax_score = \n\n (, ):\n\n ((i))\n train_logs = train_epoch(train_loader)\n valid_logs = valid_epoch(valid_loader)\n\n max_score < valid_logs:\n max_score = valid_logs\n torch(model, )\n ()\n\n == :\n optimizer = e-\n ()\nbest_model = torch.load()\nvalid_dataset = Dataset(\n x_valid_dir, \n y_valid_dir, \n =get_validation_augmentation(), \n =get_preprocessing(preprocessing_fn),\n =CLASSES,\n)\n\nvalid_loader = DataLoader(valid_dataset, =1, =, =4)\npredictions = []\n image, _ valid_loader:\n image = image.(DEVICE)\n pr_mask = best_model.predict(image)\n pr_mask = (pr_mask.squeeze().cpu().numpy().round())\n predictions.append(pr_mask)\n i predictions:\n plt.imshow(.(i, (, , )))\n plt.(=False)\n\n # Wait user input\n = input()\n\n # Check the 'q' was pressed\n == 'q':\n \n\nplt.()\nExamples of images and masks are below. We can see that train and valid data is ok.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10796179%2Fd4c36abdf71808d1a37a0455cde7213d%2F2.jpg?generation=1687875869381267&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10796179%2Fff7dd5ff5ae506fe868eb717d351097a%2F3.jpg?generation=1687875931997715&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10796179%2F066022c384861e992b98d533953f5540%2F4.jpg?generation=1687875988408869&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10796179%2F4dcddce33eaa4c70a228096529fa2bb5%2F5.jpg?generation=1687876000422419&alt=media\")
you need to use argmax on the output, else it will just be a stacked masks
", "rawMarkdown": "you need to use argmax on the output, else it will just be a stacked masks", "votes": 1, "replies": [ { "id": 2321416, "postDate": "2023-06-28T14:52:57.163Z", "content": "Thank you, but this does not help. I have changed one line in output:
\nfor image, _ in valid_loader:\n image = image.(DEVICE)\n pr_mask = best_model.(image)\n pr_mask = torch.(pr_mask, dim=) # argmax here\n pr_mask = (pr_mask.().().().())\n predictions.(pr_mask) \nNow all squares are purple as background.
\nBefore argmax tensors are like
tensor([[[[9.9973e, 9.9984e, 9.9990e, ..., 9.9987e,\n 9.9976e, 9.9955e],\n [9.9988e, 9.9994e, 9.9997e, ..., 9.9995e,\n 9.9991e, 9.9981e],\n [9.9994e, 9.9997e, 9.9999e, ..., 9.9998e,\n 9.9996e, 9.9992e],\nAnd after it and squeeze all becomes 0.
\n\nhas_not_zero = (predictions != )()\n\nOutput is
\n(512, 512)
\nFalse
\nIf we do not argmax, there will be really small numbers
\npredictions\n\nindexes = np(predictions > )\n idx indexes:\n (idx)\nOutput is
\n(3, 512, 512)
\n[0 0 0]
\n[0 0 1]
\n[0 0 2]
\n[ 0 0 16]
\n[0 1 0]
\n[ 0 510 511]
\n[ 0 511 510]
Epoch: 0
\ntrain: 100%|██████████| 184/184 [01:59<00:00, 1.53it/s, dice_loss - 0.1374, iou_score - 0.7633]
\nvalid: 100%|██████████| 163/163 [00:07<00:00, 23.07it/s, dice_loss - 0.02045, iou_score - 0.9608]
\nModel saved!
\nEpoch: 1
\ntrain: 100%|██████████| 184/184 [01:52<00:00, 1.63it/s, dice_loss - 0.1279, iou_score - 0.777]
\nvalid: 100%|██████████| 163/163 [00:07<00:00, 20.97it/s, dice_loss - 0.02045, iou_score - 0.9608]
\nEpoch: 2
\ntrain: 100%|██████████| 184/184 [01:52<00:00, 1.63it/s, dice_loss - 0.1283, iou_score - 0.7765]
\nvalid: 100%|██████████| 163/163 [00:07<00:00, 22.92it/s, dice_loss - 0.02045, iou_score - 0.9608]
\nEpoch: 3
\ntrain: 100%|██████████| 184/184 [01:54<00:00, 1.61it/s, dice_loss - 0.1276, iou_score - 0.7771]
\nvalid: 100%|██████████| 163/163 [00:06<00:00, 23.72it/s, dice_loss - 0.02045, iou_score - 0.9608]
\nEpoch: 4
\ntrain: 100%|██████████| 184/184 [01:53<00:00, 1.62it/s, dice_loss - 0.1306, iou_score - 0.7726]
\nvalid: 100%|██████████| 163/163 [00:06<00:00, 23.71it/s, dice_loss - 0.02045, iou_score - 0.9608]
you have to choose the class that you want to output, i.e if your output has only one class then pred = pr_mask[:, 1], after this you can choose a threshold
", "rawMarkdown": "you have to choose the class that you want to output, i.e if your output has only one class then pred = pr_mask[:, 1], after this you can choose a threshold" }, { "id": 2325554, "postDate": "2023-07-01T12:54:27.293Z", "content": "Does not help at all…
\ntorch.Size([512])
\ntensor([0.9999, 0.9999, 0.9999, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000,
\n 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000,
may i know the shape of the output?
", "rawMarkdown": "may i know the shape of the output?" }, { "id": 2325829, "postDate": "2023-07-01T15:55:31.150Z", "content": "I decided to make it public, so you can test, here you are) Also changed output to be closer to original, but issue still there.
\nhttps://www.kaggle.com/code/kapedalex/segmentation-cars2hubmap
", "rawMarkdown": "I decided to make it public, so you can test, here you are) Also changed output to be closer to original, but issue still there.\n\nhttps://www.kaggle.com/code/kapedalex/segmentation-cars2hubmap" }, { "id": 2327763, "postDate": "2023-07-03T06:46:45.723Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2501864, "postDate": "2023-10-27T18:04:10.070Z", "content": "I'm experiencing the same problem in a different framework. I'm trying to do sem seg on an aerial dataset, I've tryied many networks. The loss converge but there is only 1 layer (class) of the ouput tensor where all the energy is and this leads to a black output. Did you eventually find out why?
", "rawMarkdown": "I'm experiencing the same problem in a different framework. I'm trying to do sem seg on an aerial dataset, I've tryied many networks. The loss converge but there is only 1 layer (class) of the ouput tensor where all the energy is and this leads to a black output. Did you eventually find out why? " } ] } ] }, { "id": 2320156, "postDate": "2023-06-27T14:30:38.997Z", "content": "Based on advice I tried to use Segmentation-models-pytorch for this task. Here is link:
\nhttps://github.com/qubvel/segmentation_models.pytorch/blob/master/examples/cars%20segmentation%20(camvid).ipynb
Below is an extract of useful code: model, loss, optimizer, TrainEpochs, Dataloaders. Only thing changed is CLASSES = ['blood_vessel', 'glomerulus', 'unsure'].
\nENCODER = \nENCODER_WEIGHTS = \nCLASSES = [, , ]\nACTIVATION = # could be None logits multiclass segmentation\nDEVICE = \n\nmodel = smp.FPN(\n =ENCODER, \n =ENCODER_WEIGHTS, \n =len(CLASSES), \n =ACTIVATION,\n)\n\npreprocessing_fn = smp.encoders.get_preprocessing_fn(ENCODER, ENCODER_WEIGHTS)\n\n segmentation_models_pytorch import utils\n\nloss = smp.utils.losses.DiceLoss()\nmetrics = [\n smp.utils.metrics.IoU(=0.5),\n]\n\noptimizer = torch.optim.Adam([ \n dict(=model.parameters(), =0.0001),\n])\ntrain_epoch = smp.utils.train.TrainEpoch(\n model, \n =loss, \n =metrics, \n =optimizer,\n =DEVICE,\n =,\n)\n\nvalid_epoch = smp.utils.train.ValidEpoch(\n model, \n =loss, \n =metrics, \n =DEVICE,\n =,\n)\n torch.utils.data DataLoader\n torch.utils.data Dataset BaseDataset\n\n ():\n\n CLASSES = [, , ]\n\n ():\n self.ids = os.listdir(images_dir)\n self.images_fps = [os.path.join(images_dir, image_id) image_id self.ids]\n self.masks_fps = [os.path.join(masks_dir, image_id) image_id self.ids]\n\n \n self.class_values = [self.CLASSES.index(cls.lower()) cls classes]\n (self.class_values)\n\n self.augmentation = augmentation\n self.preprocessing = preprocessing\n\n ():\n\n \n image = cv2.imread(self.images_fps[i])\n image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n mask = cv2.imread(self.masks_fps[i], )\n \n masks = [(mask == v) v self.class_values]\n mask = np.stack(masks, axis=-).astype()\n\n \n self.augmentation:\n sample = self.augmentation(image=image, mask=mask)\n image, mask = sample[], sample[]\n\n \n self.preprocessing:\n sample = self.preprocessing(image=image, mask=mask)\n image, mask = sample[], sample[]\n\n image, mask\n\n ():\n (self.ids)\n
x_valid_dir = \ny_valid_dir = \nx_train_dir = \ny_train_dir = \n\n\ntrain_dataset = Dataset(\n x_train_dir, \n y_train_dir, \n =get_training_augmentation(), \n =get_preprocessing(preprocessing_fn),\n =CLASSES,\n)\n\nvalid_dataset = Dataset(\n x_valid_dir, \n y_valid_dir, \n =get_validation_augmentation(), \n =get_preprocessing(preprocessing_fn),\n =CLASSES,\n)\n\ntrain_loader = DataLoader(train_dataset, =8, =, =12)\nvalid_loader = DataLoader(valid_dataset, =1, =, =4)\nmax_score = \n\n (, ):\n\n ((i))\n train_logs = train_epoch(train_loader)\n valid_logs = valid_epoch(valid_loader)\n\n max_score < valid_logs:\n max_score = valid_logs\n torch(model, )\n ()\n\n == :\n optimizer = e-\n ()\nbest_model = torch.load()\nvalid_dataset = Dataset(\n x_valid_dir, \n y_valid_dir, \n =get_validation_augmentation(), \n =get_preprocessing(preprocessing_fn),\n =CLASSES,\n)\n\nvalid_loader = DataLoader(valid_dataset, =1, =, =4)\npredictions = []\n image, _ valid_loader:\n image = image.(DEVICE)\n pr_mask = best_model.predict(image)\n pr_mask = (pr_mask.squeeze().cpu().numpy().round())\n predictions.append(pr_mask)\n i predictions:\n plt.imshow(.(i, (, , )))\n plt.(=False)\n\n # Wait user input\n = input()\n\n # Check the 'q' was pressed\n == 'q':\n \n\nplt.()\nExamples of images and masks are below. We can see that train and valid data is ok.
\n
you need to use argmax on the output, else it will just be a stacked masks
", "votes": 1, "replies": [ { "id": 2321416, "author_name": "kapedalex", "author_url": "", "post_date": "2023-06-28T14:52:57.163000", "content": "Thank you, but this does not help. I have changed one line in output:
\nfor image, _ in valid_loader:\n image = image.(DEVICE)\n pr_mask = best_model.(image)\n pr_mask = torch.(pr_mask, dim=) # argmax here\n pr_mask = (pr_mask.().().().())\n predictions.(pr_mask) \nNow all squares are purple as background.
\nBefore argmax tensors are like
tensor([[[[9.9973e, 9.9984e, 9.9990e, ..., 9.9987e,\n 9.9976e, 9.9955e],\n [9.9988e, 9.9994e, 9.9997e, ..., 9.9995e,\n 9.9991e, 9.9981e],\n [9.9994e, 9.9997e, 9.9999e, ..., 9.9998e,\n 9.9996e, 9.9992e],\nAnd after it and squeeze all becomes 0.
\n\nhas_not_zero = (predictions != )()\n\nOutput is
\n(512, 512)
\nFalse
\nIf we do not argmax, there will be really small numbers
\npredictions\n\nindexes = np(predictions > )\n idx indexes:\n (idx)\nOutput is
\n(3, 512, 512)
\n[0 0 0]
\n[0 0 1]
\n[0 0 2]
\n[ 0 0 16]
\n[0 1 0]
\n[ 0 510 511]
\n[ 0 511 510]
Epoch: 0
\ntrain: 100%|██████████| 184/184 [01:59<00:00, 1.53it/s, dice_loss - 0.1374, iou_score - 0.7633]
\nvalid: 100%|██████████| 163/163 [00:07<00:00, 23.07it/s, dice_loss - 0.02045, iou_score - 0.9608]
\nModel saved!
\nEpoch: 1
\ntrain: 100%|██████████| 184/184 [01:52<00:00, 1.63it/s, dice_loss - 0.1279, iou_score - 0.777]
\nvalid: 100%|██████████| 163/163 [00:07<00:00, 20.97it/s, dice_loss - 0.02045, iou_score - 0.9608]
\nEpoch: 2
\ntrain: 100%|██████████| 184/184 [01:52<00:00, 1.63it/s, dice_loss - 0.1283, iou_score - 0.7765]
\nvalid: 100%|██████████| 163/163 [00:07<00:00, 22.92it/s, dice_loss - 0.02045, iou_score - 0.9608]
\nEpoch: 3
\ntrain: 100%|██████████| 184/184 [01:54<00:00, 1.61it/s, dice_loss - 0.1276, iou_score - 0.7771]
\nvalid: 100%|██████████| 163/163 [00:06<00:00, 23.72it/s, dice_loss - 0.02045, iou_score - 0.9608]
\nEpoch: 4
\ntrain: 100%|██████████| 184/184 [01:53<00:00, 1.62it/s, dice_loss - 0.1306, iou_score - 0.7726]
\nvalid: 100%|██████████| 163/163 [00:06<00:00, 23.71it/s, dice_loss - 0.02045, iou_score - 0.9608]
you have to choose the class that you want to output, i.e if your output has only one class then pred = pr_mask[:, 1], after this you can choose a threshold
", "votes": 0, "replies": [] }, { "id": 2325554, "author_name": "kapedalex", "author_url": "", "post_date": "2023-07-01T12:54:27.293000", "content": "Does not help at all…
\ntorch.Size([512])
\ntensor([0.9999, 0.9999, 0.9999, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000,
\n 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000,
may i know the shape of the output?
", "votes": 0, "replies": [] }, { "id": 2325829, "author_name": "kapedalex", "author_url": "", "post_date": "2023-07-01T15:55:31.150000", "content": "I decided to make it public, so you can test, here you are) Also changed output to be closer to original, but issue still there.
\nhttps://www.kaggle.com/code/kapedalex/segmentation-cars2hubmap
", "votes": 0, "replies": [] }, { "id": 2327763, "author_name": "", "author_url": "", "post_date": "2023-07-03T06:46:45.723000", "content": "", "votes": 0, "replies": [] }, { "id": 2501864, "author_name": "PietroBrugnolo11", "author_url": "", "post_date": "2023-10-27T18:04:10.070000", "content": "I'm experiencing the same problem in a different framework. I'm trying to do sem seg on an aerial dataset, I've tryied many networks. The loss converge but there is only 1 layer (class) of the ouput tensor where all the energy is and this leads to a black output. Did you eventually find out why?
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2320687": "you need to use argmax on the output, else it will just be a stacked masks", "2320156": "Based on advice I tried to use Segmentation-models-pytorch for this task. Here is link:\nhttps://github.com/qubvel/segmentation_models.pytorch/blob/master/examples/cars%20segmentation%20(camvid).ipynb\n\nBelow is an extract of useful code: model, loss, optimizer, TrainEpochs, Dataloaders. Only thing changed is CLASSES = ['blood_vessel', 'glomerulus', 'unsure'].\n\n```\nENCODER = 'se_resnext50_32x4d'\nENCODER_WEIGHTS = 'imagenet'\nCLASSES = ['blood_vessel', 'glomerulus', 'unsure']\nACTIVATION = 'softmax2d' # could be None for logits or 'softmax2d' for multiclass segmentation\nDEVICE = 'cuda'\n\nmodel = smp.FPN(\n encoder_name=ENCODER, \n encoder_weights=ENCODER_WEIGHTS, \n classes=len(CLASSES), \n activation=ACTIVATION,\n)\n\npreprocessing_fn = smp.encoders.get_preprocessing_fn(ENCODER, ENCODER_WEIGHTS)\n\nfrom segmentation_models_pytorch import utils\n\nloss = smp.utils.losses.DiceLoss()\nmetrics = [\n smp.utils.metrics.IoU(threshold=0.5),\n]\n\noptimizer = torch.optim.Adam([ \n dict(params=model.parameters(), lr=0.0001),\n])\n```\n```\ntrain_epoch = smp.utils.train.TrainEpoch(\n model, \n loss=loss, \n metrics=metrics, \n optimizer=optimizer,\n device=DEVICE,\n verbose=True,\n)\n\nvalid_epoch = smp.utils.train.ValidEpoch(\n model, \n loss=loss, \n metrics=metrics, \n device=DEVICE,\n verbose=True,\n)\n```\n```\nfrom torch.utils.data import DataLoader\nfrom torch.utils.data import Dataset as BaseDataset\n\nclass Dataset(BaseDataset):\n \n CLASSES = ['blood_vessel', 'glomerulus', 'unsure']\n \n def __init__(\n self, \n images_dir, \n masks_dir, \n classes=None, \n augmentation=None, \n preprocessing=None,\n ):\n self.ids = os.listdir(images_dir)\n self.images_fps = [os.path.join(images_dir, image_id) for image_id in self.ids]\n self.masks_fps = [os.path.join(masks_dir, image_id) for image_id in self.ids]\n \n # convert str names to class values on masks\n self.class_values = [self.CLASSES.index(cls.lower()) for cls in classes]\n print(self.class_values)\n \n self.augmentation = augmentation\n self.preprocessing = preprocessing\n \n def __getitem__(self, i):\n \n # read data\n image = cv2.imread(self.images_fps[i])\n image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n mask = cv2.imread(self.masks_fps[i], 0)\n # extract certain classes from mask (e.g. cars)\n masks = [(mask == v) for v in self.class_values]\n mask = np.stack(masks, axis=-1).astype('float')\n \n # apply augmentations\n if self.augmentation:\n sample = self.augmentation(image=image, mask=mask)\n image, mask = sample['image'], sample['mask']\n \n # apply preprocessing\n if self.preprocessing:\n sample = self.preprocessing(image=image, mask=mask)\n image, mask = sample['image'], sample['mask']\n \n return image, mask\n \n def __len__(self):\n return len(self.ids)\n```\n#Then I try to teach it, using only annotated data. Problem is, that ALL predictions are just red squares. What I miss?\n\n\n\n```\nx_valid_dir = '/kaggle/working/validation_data/x_valid'\ny_valid_dir = '/kaggle/working/validation_data/y_valid'\nx_train_dir = '/kaggle/working/validation_data/x_train'\ny_train_dir = '/kaggle/working/validation_data/y_train'\n\n# Create the Dataset and DataLoader for training and validation\ntrain_dataset = Dataset(\n x_train_dir, \n y_train_dir, \n augmentation=get_training_augmentation(), \n preprocessing=get_preprocessing(preprocessing_fn),\n classes=CLASSES,\n)\n\nvalid_dataset = Dataset(\n x_valid_dir, \n y_valid_dir, \n augmentation=get_validation_augmentation(), \n preprocessing=get_preprocessing(preprocessing_fn),\n classes=CLASSES,\n)\n\ntrain_loader = DataLoader(train_dataset, batch_size=8, shuffle=True, num_workers=12)\nvalid_loader = DataLoader(valid_dataset, batch_size=1, shuffle=False, num_workers=4)\n```\n```\nmax_score = 0\n\nfor i in range(0, 3):\n \n print('\\nEpoch: {}'.format(i))\n train_logs = train_epoch.run(train_loader)\n valid_logs = valid_epoch.run(valid_loader)\n \n if max_score < valid_logs['iou_score']:\n max_score = valid_logs['iou_score']\n torch.save(model, './best_model.pth')\n print('Model saved!')\n \n if i == 25:\n optimizer.param_groups[0]['lr'] = 1e-5\n print('Decrease decoder learning rate to 1e-5!')\n```\n```\nbest_model = torch.load('./best_model.pth')\nvalid_dataset = Dataset(\n x_valid_dir, \n y_valid_dir, \n augmentation=get_validation_augmentation(), \n preprocessing=get_preprocessing(preprocessing_fn),\n classes=CLASSES,\n)\n\nvalid_loader = DataLoader(valid_dataset, batch_size=1, shuffle=False, num_workers=4)\npredictions = []\nfor image, _ in valid_loader:\n image = image.to(DEVICE)\n pr_mask = best_model.predict(image)\n pr_mask = (pr_mask.squeeze().cpu().numpy().round())\n predictions.append(pr_mask)\n```\n\n```\nfor i in predictions:\n plt.imshow(np.transpose(i, (1, 2, 0)))\n plt.show(block=False)\n \n # Wait for user input\n key = input(\"Press 'q' to quit or any other key to continue: \")\n \n # Check if the 'q' key was pressed\n if key == 'q':\n break\n\nplt.close()\n```\nExamples of images and masks are below. We can see that train and valid data is ok.\n\n\n\n\n\n\n\n" } }