impl. working version

configs/exp_resnet18.yaml

@@ -3,15 +3,16 @@
 # In Python, this is all read as a dict.
 
 # environment/computational parameters
-device: CUDA
+device: cuda
 num_workers: 4
 
 # dataset parameters
-data_root: /path/to/dataset
-num_classes: 32
+data_root: datasets/CaltechCT
+num_classes: 16
 
 # training hyperparameters
+image_size: [224, 224]
 num_epochs: 200
 batch_size: 128
-learning_rate: 1e-3
-weight_decay: 1e-3
+learning_rate: 0.001
+weight_decay: 0.001

ct_classifier/dataset.py

@@ -15,20 +15,23 @@
 import os
 import json
 from torch.utils.data import Dataset
-from torchvision.transforms import ToTensor
+from torchvision.transforms import Compose, Resize, ToTensor
 from PIL import Image
 
 
 class CTDataset(Dataset):
 
-    def __init__(self, data_root, split='train'):
+    def __init__(self, cfg, split='train'):
         '''
             Constructor. Here, we collect and index the dataset inputs and
             labels.
         '''
-        self.data_root = data_root
+        self.data_root = cfg['data_root']
         self.split = split
-        self.transform = ToTensor()
+        self.transform = Compose([              # Transforms. Here's where we could add data augmentation (see Björn's lecture on August 11).
+            Resize((cfg['image_size'])),        # For now, we just resize the images to the same dimensions...
+            ToTensor()                          # ...and convert them to torch.Tensor.
+        ])
 
         # index data into list
         self.data = []
@@ -40,6 +43,23 @@ def __init__(self, data_root, split='train'):
             'train_annotations.json' if self.split=='train' else 'cis_val_annotations.json'
         )
         meta = json.load(open(annoPath, 'r'))
+
+        images = dict([[i['id'], i['file_name']] for i in meta['images']])          # image id to filename lookup
+        labels = dict([[c['id'], idx] for idx, c in enumerate(meta['categories'])]) # custom labelclass indices that start at zero
+
+        # since we're doing classification, we're just taking the first annotation per image and drop the rest
+        images_covered = set()      # all those images for which we have already assigned a label
+        for anno in meta['annotations']:
+            imgID = anno['image_id']
+            if imgID in images_covered:
+                continue
+
+            # append image-label tuple to data
+            imgFileName = images[imgID]
+            label = anno['category_id']
+            labelIndex = labels[label]
+            self.data.append([imgFileName, labelIndex])
+            images_covered.add(imgID)       # make sure image is only added once to dataset
 
 
     def __len__(self):
@@ -54,17 +74,13 @@ def __getitem__(self, idx):
             Returns a single data point at given idx.
             Here's where we actually load the image.
         '''
-        image_name, label = self.data[idx]
+        image_name, label = self.data[idx]              # see line 57 above where we added these two items to the self.data list
 
         # load image
-        image_path = os.path.join(self.data_root, image_path)
-        img = Image.open(image_path)
+        image_path = os.path.join(self.data_root, 'eccv_18_all_images_sm', image_name)
+        img = Image.open(image_path).convert('RGB')     # the ".convert" makes sure we always get three bands in Red, Green, Blue order
 
-        # transform: convert to torch.Tensor
-        # here's where we could do data augmentation:
-        # https://pytorch.org/vision/stable/transforms.html
-        # see Björn's lecture on Thursday, August 11.
-        # For now, we only convert the image to torch.Tensor
+        # transform: see lines 31ff above where we define our transformations
         img_tensor = self.transform(img)
 
         return img_tensor, label

ct_classifier/model.py

@@ -23,10 +23,10 @@ def __init__(self, num_classes):
         # replace the very last layer from the original, 1000-class output
         # ImageNet to a new one that outputs num_classes
         last_layer = self.feature_extractor.fc                          # tip: print(self.feature_extractor) to get info on how model is set up
-        num_features = last_layer.num_features
+        in_features = last_layer.in_features                            # number of input dimensions to last (classifier) layer
         self.feature_extractor.fc = nn.Identity()                       # discard last layer...
 
-        self.classifier = nn.Linear(num_features, num_classes)          # ...and create a new one
+        self.classifier = nn.Linear(in_features, num_classes)           # ...and create a new one
 
 
     def forward(self, x):

ct_classifier/train.py

@@ -5,6 +5,7 @@
     2022 Benjamin Kellenberger
 '''
 
+import os
 import argparse
 import yaml
 import glob
@@ -16,8 +17,8 @@
 from torch.optim import SGD
 
 # let's import our own classes and functions!
-from ct_classifier.dataset import CTDataset
-from ct_classifier.model import CustomResNet18
+from dataset import CTDataset
+from model import CustomResNet18
 
 
 
@@ -26,7 +27,7 @@ def create_dataloader(cfg, split='train'):
         Loads a dataset according to the provided split and wraps it in a
         PyTorch DataLoader object.
     '''
-    dataset_instance = CTDataset(cfg['data_root'], split)        # create an object instance of our CTDataset class
+    dataset_instance = CTDataset(cfg, split)        # create an object instance of our CTDataset class
 
     dataLoader = DataLoader(
             dataset=dataset_instance,
@@ -66,6 +67,9 @@ def load_model(cfg):
 
 
 def save_model(epoch, model, stats):
+    # make sure save directory exists; create if not
+    os.makedirs('model_states', exist_ok=True)
+
     # get model parameters and add to stats...
     stats['model'] = model.state_dict()
 
@@ -100,11 +104,11 @@ def train(cfg, dataLoader, model, optimizer):
     criterion = nn.CrossEntropyLoss()
 
     # running averages
-    loss_total, oa_total = 0.0, 0.0     # for now, we just log the loss and overall accuracy (OA)
+    loss_total, oa_total = 0.0, 0.0                         # for now, we just log the loss and overall accuracy (OA)
 
     # iterate over dataLoader
     progressBar = trange(len(dataLoader))
-    for idx, (data, labels) in enumerate(dataLoader):
+    for idx, (data, labels) in enumerate(dataLoader):       # see the last line of file "dataset.py" where we return the image tensor (data) and label
 
         # put data and labels on device
         data, labels = data.to(device), labels.to(device)
@@ -125,18 +129,19 @@ def train(cfg, dataLoader, model, optimizer):
         optimizer.step()
 
         # log statistics
-        loss_total += loss.item()           # the .item() command retrieves the value of a single-valued tensor, regardless of its data type and device of tensor
+        loss_total += loss.item()                       # the .item() command retrieves the value of a single-valued tensor, regardless of its data type and device of tensor
 
-        pred_label = torch.argmax(prediction)           # the predicted label is the one at position (class index) with highest predicted value
+        pred_label = torch.argmax(prediction, dim=1)    # the predicted label is the one at position (class index) with highest predicted value
         oa = torch.mean((pred_label == labels).float()) # OA: number of correct predictions divided by batch size (i.e., average/mean)
         oa_total += oa.item()
 
         progressBar.set_description(
-            '[Train] Loss: {:.2f}; OA: {:.2f}'.format(
+            '[Train] Loss: {:.2f}; OA: {:.2f}%'.format(
                 loss_total/(idx+1),
-                oa_total/(idx+1)
+                100*oa_total/(idx+1)
             )
         )
+        progressBar.update(1)
 
     # end of epoch; finalize
     progressBar.close()
@@ -179,16 +184,17 @@ def validate(cfg, dataLoader, model):
             # log statistics
             loss_total += loss.item()
 
-            pred_label = torch.argmax(prediction)
+            pred_label = torch.argmax(prediction, dim=1)
             oa = torch.mean((pred_label == labels).float())
             oa_total += oa.item()
 
             progressBar.set_description(
-                '[Val ] Loss: {:.2f}; OA: {:.2f}'.format(
+                '[Val ] Loss: {:.2f}; OA: {:.2f}%'.format(
                     loss_total/(idx+1),
-                    oa_total/(idx+1)
+                    100*oa_total/(idx+1)
                 )
             )
+            progressBar.update(1)
 
     # end of epoch; finalize
     progressBar.close()

license

@@ -1,6 +1,6 @@
     MIT License
 
-    Copyright (c) Microsoft Corporation. All rights reserved.
+    Copyright (c) ECEO, EPFL. All rights reserved.
 
     Permission is hereby granted, free of charge, to any person obtaining a copy
     of this software and associated documentation files (the "Software"), to deal

readme.md

@@ -17,8 +17,10 @@ pip install -r requirements.txt
 
 3. Download dataset
 
+**NOTE:** Requires the [azcopy CLI](https://docs.microsoft.com/en-us/azure/storage/common/storage-use-azcopy-v10) to be installed and set up on your machine.
+
 ```bash
-./scripts/download_dataset.sh 
+sh scripts/download_dataset.sh 
 ```
 
 This downloads the [CCT20](https://lila.science/datasets/caltech-camera-traps) subset to the `datasets/CaltechCT` folder.