Project update by Peter, minor path changes by Gudmundur

embeddings_and_difficulty/configs/config_augmentations/standard_augmenter.yaml

@@ -4,6 +4,6 @@ same_size: False
 mean: [0.0,0.0,0.0]
 std: [1.0,1.0,1.0]
 full_rot: 180
-scale: (0.8, 1.2)
+scale: [0.8, 1.2]
 shear: 10
 cutout: 16
\ No newline at end of file

embeddings_and_difficulty/configs/config_optimizers/config_optim.yaml

 ADAM:
     lr: 0.001
-    betas: (0.9, 0.999)
+    betas: [0.9, 0.999]
     eps: 1e-08
     weight_decay: 0
 

embeddings_and_difficulty/configs/general.yaml

@@ -6,10 +6,11 @@ TRAIN:
   CHECKPOINT_PERIOD: 2
   AUTO_RESUME: True
 DATA:
-  PATH_TO_DATA: r"C:\Users\ptrkm\PycharmProjects\BachelorDeeplearning\Embeddings\New_Embeddings"
-  PATH_TO_LABEL: r"C:\Users\ptrkm\PycharmProjects\BachelorDeeplearning\Embeddings\New_Embeddings"
-  PATH_TO_DIFFICULTIES: r"C:\Users\ptrkm\PycharmProjects\BachelorDeeplearning\Embeddings\New_Embeddings"
-  PATH_TO_SPLIT: r"C:\Users\ptrkm\PycharmProjects\BachelorDeeplearning\Embeddings\New_Embeddings"
+  PATH_TO_DATA: [data/processed/additional-dermoscopic-images,
+                 data/processed/main-dermoscopic-images]
+  PATH_TO_LABEL: data/processed/labels.csv
+  PATH_TO_DIFFICULTIES: data/processed/difficulties.pkl
+  PATH_TO_SPLIT: data/processed/splits.pkl
 NETWORK:
   PATH_TO_SAVED: None
   BACKBONE:
@@ -24,11 +25,11 @@ NETWORK:
 TRAINING:
   BACKBONE:
     MAX_EPOCH: 100
-    LOSS: contrastive
+    LOSS: Contrastive
     EARLY_STOP_PATIENCE: 3
   HEAD:
     MAX_EPOCH: 20
-    LOSS: least_squares
+    LOSS: LeastSquares
     EARLY_STOP_PATIENCE: 2
   COMBINED:
     MAX_EPOCH: 10
@@ -39,17 +40,21 @@ SOLVER:
   MOMENTUM: 0.9
   WEIGHT_DECAY: 1e-4
   WARMUP_START_LR: 0.01
-  OPTIMIZING_METHOD: ADAM
+  OPTIMIZER: ADAM
+  ALPHA: 0.5
 AUGMENTATION:
   NAME: ngessert
-  CONFIG: standard_augmenter.yaml
+  CONFIG: config_augmentations/standard_augmenter.yaml
 TEST:
   ENABLE: True
   BATCH_SIZE: 64
 DATA_LOADER:
-  NUM_WORKERS: 8
+  NUM_WORKERS: 0
   PIN_MEMORY: True
+EVAL_METRICS:
+  BACKBONE: knn
+  HEAD: MSE
 NUM_GPUS: 1 # Not set up to handle more currently
 NUM_SHARDS: 1
 RNG_SEED: 0
-OUTPUT_DIR: r"C:\Users\ptrkm\PycharmProjects\BachelorDeeplearning\Embeddings\New_Embeddings"
\ No newline at end of file
+OUTPUT_DIR: r"data/output"
\ No newline at end of file

embeddings_and_difficulty/data_augmentations/augmentations.py

@@ -6,7 +6,7 @@ from torchvision import transforms, utils
 import math
 from PIL import Image
 from numba import jit
-import color_constancy as cc
+# import color_constancy as cc
 import pickle
 from argparse import Namespace
 
@@ -42,10 +42,9 @@ class DataAugmentISIC_AISC:
         elif self.random_resize:
             all_transforms.append(transforms.RandomResizedCrop(self.input_size[0], scale=(0.08, 1.0)))
 
-        all_transforms.append(cc.general_color_constancy(gaussian_differentiation=0, minkowski_norm=6, sigma=0))
+        # all_transforms.append(cc.general_color_constancy(gaussian_differentiation=0, minkowski_norm=6, sigma=0))
         all_transforms.append(transforms.RandomHorizontalFlip())
         all_transforms.append(transforms.RandomVerticalFlip())
-
         all_transforms.append(transforms.RandomChoice([transforms.RandomAffine(model_params.get('full_rot',180),
                                                                                scale=model_params.get('scale', (0.8,1.2)),
                                                                                shear=model_params.get('shear', 10),
@@ -69,7 +68,7 @@ class DataAugmentISIC_AISC:
         self.composed_train = transforms.Compose(all_transforms)
 
         self.composed_eval = transforms.Compose([
-            cc.general_color_constancy(gaussian_differentiation=0, minkowski_norm=6, sigma = 0),
+            # cc.general_color_constancy(gaussian_differentiation=0, minkowski_norm=6, sigma = 0),
             transforms.Resize(self.input_size),
             transforms.ToTensor(),
             transforms.Normalize(np.float32(model_params.get('mean', np.array([0.0, 0.0, 0.0]))),
@@ -148,7 +147,7 @@ class Cutout_v0(object):
 DATA_AUGMENTERS = {'ngessert': DataAugmentISIC_AISC}
 
 def get_data_augmenter(augment_params):
-    return DATA_AUGMENTERS[augment_params.NAME](augment_params.vals)
+    return DATA_AUGMENTERS[augment_params.name](augment_params.vals)
 
 
 if __name__ == "__main__":

embeddings_and_difficulty/dataloaders/AISC.py

@@ -4,13 +4,14 @@ from PIL import Image
 import os
 import pickle
 import pandas as pd
-from Embeddings.New_Embeddings.data_augmentations import augmentations as aug
+from data_augmentations import augmentations as aug
 
 
 class AISC(Dataset):
     def __init__(self, dataset_params):
+
         self.path_to_data = dataset_params.PATH_TO_DATA
-        self.path_to_labels = dataset_params.PATH_TO_LABELS
+        self.path_to_labels = dataset_params.PATH_TO_LABEL
         self.path_to_difficulties = dataset_params.PATH_TO_DIFFICULTIES
         self.path_to_split = dataset_params.PATH_TO_SPLIT
         self.difficulties = None
@@ -76,8 +77,8 @@ class AISC(Dataset):
 
         labels = pd.read_csv(self.path_to_labels)
         label_names = list(labels['names'])
-        labels = labels.drop('names', axis=1).values()
-
+        labels = labels.drop('names', axis=1)
+        labels = labels.values
         return label_names, labels
 
     def read_difficulties(self):
@@ -86,6 +87,7 @@ class AISC(Dataset):
         :return: (dict) with image names as keys (not full path) and difficulty as value
         """
         if not os.path.isfile(self.path_to_difficulties):
+            breakpoint()
             raise ValueError("Chosen path to difficulties is not a file on this device")
 
         difficulties = pickle.load(open(self.path_to_difficulties, 'rb'))
@@ -100,12 +102,15 @@ class AISC(Dataset):
 
         temp = dict()
         loading_order = dict()
-        for mode, names in split.items():
+        for split, val in split.items():
+            for mode, names in val.items():
                 temp[mode] = {
                     name: self.name_to_file_label_difficulty[name]
                     for name in names
                 }
                 loading_order[mode] = names
+
+
         return temp, loading_order
 
     def __getitem__(self, item):
@@ -115,16 +120,18 @@ class AISC(Dataset):
         :param item: (int) conforming to the index of names
         :return: (tuple) of (torch.Tensor, torch.Tensor, torch.Tensor) of image, label and difficulty
         """
-        file, label, difficulty, has_diff = self.name_to_file_label_difficulty[
+
+        file, label, difficulty, has_diff = (self.name_to_file_label_difficulty[self.mode][
             self.loading_order[self.mode][item]
-        ]
+        ]).values()
+
 
         image = Image.open(file)
         image = self.data_augmenter(image, self.mode)
-        label = torch.tensor(label)
+        label = torch.tensor(label).reshape(-1)
         difficulty = torch.tensor(difficulty)
 
         if self.mode == 'train':
-            return image, label, difficulty, has_diff
+            return image, label, difficulty
         else:
             return image, label, difficulty, file, has_diff

embeddings_and_difficulty/losses/losses.py

@@ -3,12 +3,12 @@
 
 def create_loss(losses, args):
 
-    alpha = args.ALPHA
+    alpha = args.SOLVER.ALPHA
     loss_backbone = losses[0]
     loss_head = losses[1]
 
     def loss(embeddings, est_difficulties,labels, difficulties, score_keeper):
-
+        labels = labels.reshape(-1)
         if score_keeper.is_training == 'backbone':
             return loss_backbone(embeddings, labels)
         if score_keeper.is_training == 'head':

embeddings_and_difficulty/losses/losses_backbone.py

@@ -15,7 +15,8 @@ with the same name e.g. contrastive.yaml this should correspond to the string pu
 when added it should also be added to the dictionary in the bottom named all_losses
 """
 def get_contrastive(args):
-    return losses.ContrastiveLoss(args.pos_margin, args.neg_margin, **args.kwargs)
+
+    return losses.ContrastiveLoss(args.pos_margin, args.neg_margin)
 
 def get_triplet_margin(args):
     return losses.TripletMarginLoss(margin=args.margin,
@@ -27,9 +28,9 @@ def get_triplet_margin(args):
 
 
 all_losses = {
-    'contrastive': get_contrastive,
-    'triplet_marging': get_triplet_margin
+    'Contrastive': get_contrastive,
+    'TripletMarging': get_triplet_margin
 }
 
 def get_loss(loss, loss_args):
-    return all_losses[loss](loss_args)
+    return all_losses[loss](loss_args.__dict__[loss])

embeddings_and_difficulty/losses/losses_head.py

@@ -5,7 +5,8 @@ from pytorch_metric_learning import losses
 import torch.nn as nn
 
 def get_least_squares(args):
-    return nn.MSELoss(reduction=args.reductions)
+
+    return nn.MSELoss(reduction=args.reduction)
 
 def get_l1(args):
     return nn.L1Loss(reduction=args.reductions)
@@ -21,9 +22,10 @@ class KendallsTau(nn.modules.loss._Loss):
         tau = 2/(len(difficulty) * (len(difficulty)-1)) * torch.sum(torch)
 
 all_losses = {
-    'least_squares': get_least_squares,
+    'LeastSquares': get_least_squares,
     'L1': get_l1,
 }
 
 def get_loss(loss, loss_args):
-    return all_losses[loss](loss_args)
\ No newline at end of file
+
+    return all_losses[loss](loss_args.__dict__[loss])
\ No newline at end of file