DataLoader __getitem__ using float indices?

[Riya_Tyagi]

I’m trying to train a CNN on retinal images. I’m using Torch ImageFolder to load in data. However, I’m running into a TypeError that suggests the Torch dataloader is iterating through my dataset using np.float64 indices rather than integer indices. I’ve uploaded my code with comments and the full stack trace. I haven’t been able to find an answer online and I’m quite confused as to how my code could have resulted in this error. Any help would be very very much appreciated.

Full Stack Trace:

TypeError                                 Traceback (most recent call last)
<ipython-input-46-94e0e5ea3e09> in <module>
     13     os.makedirs(os.path.join('outputs', 'optimizers', experiment_name))
     14 
---> 15 train(data_dir, experiment_name, X_train, y_train) # save X_train in csv

<ipython-input-44-153a42274efe> in train(data_dir, experiment_name, X_train, y_train, num_classes, batch_size, num_epochs, lr, image_size)
    125 
    126         print ("Model is fitting. Thank you for your patience.")
--> 127         net.fit(train_subset, y=None)
    128 
    129         print ("Model is performing inference. Results saved in probabilities folder.")

/usr/local/lib/python3.6/dist-packages/skorch/classifier.py in fit(self, X, y, **fit_params)
    139         # this is actually a pylint bug:
    140         # https://github.com/PyCQA/pylint/issues/1085
--> 141         return super(NeuralNetClassifier, self).fit(X, y, **fit_params)
    142 
    143     def predict_proba(self, X):

/usr/local/lib/python3.6/dist-packages/skorch/net.py in fit(self, X, y, **fit_params)
   1213             self.initialize()
   1214 
-> 1215         self.partial_fit(X, y, **fit_params)
   1216         return self
   1217 

/usr/local/lib/python3.6/dist-packages/skorch/net.py in partial_fit(self, X, y, classes, **fit_params)
   1172         self.notify('on_train_begin', X=X, y=y)
   1173         try:
-> 1174             self.fit_loop(X, y, **fit_params)
   1175         except KeyboardInterrupt:
   1176             pass

/usr/local/lib/python3.6/dist-packages/skorch/net.py in fit_loop(self, X, y, epochs, **fit_params)
   1089 
   1090             self.run_single_epoch(dataset_valid, training=False, prefix="valid",
-> 1091                                   step_fn=self.validation_step, **fit_params)
   1092 
   1093             self.notify("on_epoch_end", **on_epoch_kwargs)

/usr/local/lib/python3.6/dist-packages/skorch/net.py in run_single_epoch(self, dataset, training, prefix, step_fn, **fit_params)
   1118 
   1119         batch_count = 0
-> 1120         for batch in self.get_iterator(dataset, training=training):
   1121             self.notify("on_batch_begin", batch=batch, training=training)
   1122             step = step_fn(batch, **fit_params)

/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py in __next__(self)
    519             if self._sampler_iter is None:
    520                 self._reset()
--> 521             data = self._next_data()
    522             self._num_yielded += 1
    523             if self._dataset_kind == _DatasetKind.Iterable and \

/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py in _next_data(self)
   1201             else:
   1202                 del self._task_info[idx]
-> 1203                 return self._process_data(data)
   1204 
   1205     def _try_put_index(self):

/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py in _process_data(self, data)
   1227         self._try_put_index()
   1228         if isinstance(data, ExceptionWrapper):
-> 1229             data.reraise()
   1230         return data
   1231 

/usr/local/lib/python3.6/dist-packages/torch/_utils.py in reraise(self)
    423             # have message field
    424             raise self.exc_type(message=msg)
--> 425         raise self.exc_type(msg)
    426 
    427 

TypeError: Caught TypeError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/_utils/worker.py", line 287, in _worker_loop
    data = fetcher.fetch(index)
  File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/_utils/fetch.py", line 44, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataset.py", line 311, in __getitem__
    return self.dataset[self.indices[idx]]
  File "/usr/local/lib/python3.6/dist-packages/torchvision/datasets/folder.py", line 231, in __getitem__
    path, target = self.samples[index]
TypeError: list indices must be integers or slices, not numpy.float64

My Train Code

def train(data_dir, experiment_name, X_train, y_train,
          num_classes=2, batch_size=64, num_epochs=50, lr=0.001, image_size = (224, 224)):
    
    # USING GPU
    
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    if device == 'cuda:0':
        torch.cuda.empty_cache()
        
    # TRANSFORMING TRAIN/TEST DATA
    
    train_transforms = transforms.Compose([transforms.Resize(image_size), # since not applying any of my own funcs
                                          transforms.RandomHorizontalFlip(),
                                          transforms.RandomVerticalFlip(),
                                          transforms.RandomRotation(25),
                                          transforms.ToTensor(),
                                          transforms.Normalize([0.485, 0.456, 0.406],
                                                               [0.229, 0.224, 0.225])]) # why this normalizing?
    
    test_transforms = transforms.Compose([transforms.Resize(image_size),
                                         transforms.ToTensor(),
                                         transforms.Normalize([0.485, 0.456, 0.406],
                                                              [0.229, 0.224, 0.225])])
    
    # APPENDING DATA TO DATA FOLDERS. THE FOLDERS ARE SETUP TO EACH CONTAIN THE TWO CLASS FOLDERS
    # TRAIN/CLASS1 TRAIN/CLASS2 TEST/CLASS1 TEST/CLASS2
    train_folder = os.path.join(data_dir, 'train')
    test_folder = os.path.join(data_dir, 'test')
    
    train_dataset = datasets.ImageFolder(train_folder, train_transforms)
    test_dataset = datasets.ImageFolder(test_folder, test_transforms)
    
    # DEFINING VALIDATION SPLIT FOR 5-FOLD CROSS VALIDATION
    
    indxs = np.arange(0,len(train_dataset))
    
    skf = StratifiedKFold(n_splits=5, shuffle=True)  # different splits each time  
    for i, (train_split, val_split) in enumerate(skf.split(X_train, y_train)):     
        f_params = f'./outputs/checkpoints/{experiment_name}/model_fold{i+1}.pt'
        f_history = f'./outputs/histories/{experiment_name}/model_fold{i+1}.json'
        f_optimize = f'./outputs/optimizers/{experiment_name}/model_fold{i+1}.json'
        f_criterion = f'./outputs/criterions/{experiment_name}/model_fold{i+1}.json'
        csv_name = f'./outputs/probabilities/{experiment_name}/model_fold{i+1}.csv'
                
        val_patient = np.array(X_train.loc[val_split]['id'])
        train_patient = np.array(X_train.loc[train_split]['id'])
        
        val_indxs = np.negative(np.ones(len(train_dataset)))
                        
        for j in range(len(train_dataset.samples)):
            sample_id = re.sub('.*/', '',str(train_dataset.samples[j][0]))
            sample_id = sample_id.split("_")[0] # bc of new naming conv.

            if np.any(val_patient == int(sample_id)) == True:
                val_indxs[j] = j
        
        val_indxs = val_indxs[val_indxs >= 0]
                
        train_indxs = list(set(indxs) - set(val_indxs)) 
        
        print(len(val_indxs), len(train_indxs))
                     
        train_subset = torch.utils.data.Subset(train_dataset, train_indxs)
        val_subset = torch.utils.data.Subset(train_dataset, val_indxs)
        
        print ("Train/Test/Val datasets have been created.")
                
        # WEIGHTED SAMPLING}
        labels = [x[1] for x in train_subset]
        labels = np.array(labels)
        labels = labels.astype(int) 
        
        black_weight = 1 / len(labels[labels == 0]) 
        white_weight = 1 / len(labels[labels == 1])
        sample_weights = np.array([black_weight, white_weight])
        weights = sample_weights[labels]
        sampler = torch.utils.data.WeightedRandomSampler(weights, len(train_subset), replacement=True)

        print()
        print(f'Data Directory: {data_dir}')
        print(f'Number of black eyes: {len(labels[labels == 0])}')
        print(f'Number of white eyes: {len(labels[labels == 1])}')
        print(f'Device: {device}')
        print()

        # SETTING MODEL PARAMETERS WITH SKORCH

        checkpoint = Checkpoint(monitor='valid_loss_best',
                                f_params=f_params,
                                f_history=f_history,
                                f_optimizer=f_optimize,
                                f_criterion=f_criterion)

        train_acc = EpochScoring(scoring='accuracy',
                                 on_train=True,
                                 name='train_acc',
                                 lower_is_better=False)

        train_auc = EpochScoring(scoring='roc_auc', name ='train_auc', on_train=True, lower_is_better=False)
        valid_auc = EpochScoring(scoring='roc_auc', name ='valid_auc', lower_is_better=False)


        early_stopping = EarlyStopping()
        callbacks = [train_auc, valid_auc, checkpoint, train_acc, early_stopping]

        net = NeuralNetClassifier(PretrainedModel,
                                  criterion=nn.CrossEntropyLoss,
                                  lr=lr,
                                  batch_size=batch_size,
                                  max_epochs=num_epochs,
                                  module__output_features=num_classes,
                                  optimizer=optim.SGD,
                                  optimizer__momentum=0.9,
                                  iterator_train__num_workers=1,
                                  iterator_train__sampler=sampler,
                                  iterator_valid__shuffle=False,
                                  iterator_valid__num_workers=1,
                                  train_split=predefined_split(val_subset),
                                  callbacks=callbacks,
                                  device=device)
        # SKORCH MODEL FITTING
        print ("Model is fitting. Thank you for your patience.")
        net.fit(train_subset, y=None)

        print ("Model is performing inference. Results saved in probabilities folder.")
        # SAVING RESULTS
        img_locs = [loc for loc, _ in test_dataset.samples]
        test_probs = net.predict_proba(test_dataset)
        test_probs = [prob[0] for prob in test_probs]
        data = {'img_loc' : img_locs, 'probability' : test_probs}
        pd.DataFrame(data=data).to_csv(csv_name, index=False)
        
        print (f"Model {i} is done and saved.")

    print ("The code is done.")```

[Karthik_Ganesan]

Based on a quick look, it seems like you are passing numpy arrays instead of torch tensors. You can re-write everything you need using torch tensors which will avoid these types of errors.

Also, it seems like you are trying to implement a weighted sampler for your dataset. PyTorch provides code for doing this already, here. If at all possible, you should try to use existing PyTorch code instead of writing your own.

[Riya_Tyagi]

Thank you! It actually turns out that val_indxs which I was using to generate the data Subset val_subset contained floats. All I had to do was typecast it!

Additionally, I am already using WeightedRandomSampler provided by PyTorch. The code I have before that is just to calculate weights for the sampler.

Thanks again!