Revert "fixed image_size and training loss, accuracy for inception mo…

…dels"

This reverts commit fbbd351.

distiller/apputils/data_loaders.py

@@ -58,20 +58,20 @@ def classification_get_input_shape(dataset):
         raise ValueError("dataset %s is not supported" % dataset)
 
 
-def __dataset_factory(dataset, is_inception):
+def __dataset_factory(dataset):
     return {'cifar10': cifar10_get_datasets,
             'mnist': mnist_get_datasets,
             'imagenet': imagenet_get_datasets}.get(dataset, None)
 
 
 def load_data(dataset, data_dir, batch_size, workers, validation_split=0.1, deterministic=False,
               effective_train_size=1., effective_valid_size=1., effective_test_size=1.,
-              fixed_subset=False, sequential=False, is_inception=False):
+              fixed_subset=False, sequential=False):
     """Load a dataset.
 
     Args:
         dataset: a string with the name of the dataset to load (cifar10/imagenet)
-        data_dir: the directory where the datset resides
+        data_dir: the directory where the dataset resides
         batch_size: the batch size
         workers: the number of worker threads to use for loading the data
         validation_split: portion of training dataset to set aside for validation
@@ -86,8 +86,8 @@ def load_data(dataset, data_dir, batch_size, workers, validation_split=0.1, dete
     """
     if dataset not in DATASETS_NAMES:
         raise ValueError('load_data does not support dataset %s" % dataset')
-    datasets_fn = __dataset_factory(dataset, is_inception)
-    return get_data_loaders(datasets_fn, data_dir, batch_size, workers, is_inception,
+    datasets_fn = __dataset_factory(dataset)
+    return get_data_loaders(datasets_fn, data_dir, batch_size, workers, 
                             validation_split=validation_split,
                             deterministic=deterministic, 
                             effective_train_size=effective_train_size,
@@ -155,23 +155,17 @@ def cifar10_get_datasets(data_dir):
     return train_dataset, test_dataset
 
 
-
-def imagenet_get_datasets(data_dir, is_inception):
+def imagenet_get_datasets(data_dir):
     """
     Load the ImageNet dataset.
     """
-    # Inception Network accepts image of size 3, 299, 299
-    if is_inception == True:
-        resize, crop = 336, 299
-    else:
-        resize, crop = 256, 224
     train_dir = os.path.join(data_dir, 'train')
     test_dir = os.path.join(data_dir, 'val')
     normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                      std=[0.229, 0.224, 0.225])
 
     train_transform = transforms.Compose([
-        transforms.RandomResizedCrop(crop),
+        transforms.RandomResizedCrop(224),
         transforms.RandomHorizontalFlip(),
         transforms.ToTensor(),
         normalize,
@@ -180,8 +174,8 @@ def imagenet_get_datasets(data_dir, is_inception):
     train_dataset = datasets.ImageFolder(train_dir, train_transform)
 
     test_transform = transforms.Compose([
-        transforms.Resize(resize),
-        transforms.CenterCrop(crop),
+        transforms.Resize(256),
+        transforms.CenterCrop(224),
         transforms.ToTensor(),
         normalize,
     ])
@@ -190,6 +184,7 @@ def imagenet_get_datasets(data_dir, is_inception):
 
     return train_dataset, test_dataset
 
+
 def __image_size(dataset):
     # un-squeeze is used here to add the batch dimension (value=1), which is missing
     return dataset[0][0].unsqueeze(0).size()
@@ -266,10 +261,10 @@ def _get_sampler(data_source, effective_size, fixed_subset=False, sequential=Fal
     return SwitchingSubsetRandomSampler(data_source, effective_size)
 
 
-def get_data_loaders(datasets_fn, data_dir, batch_size, num_workers, is_inception, validation_split=0.1, deterministic=False,
+def get_data_loaders(datasets_fn, data_dir, batch_size, num_workers, validation_split=0.1, deterministic=False,
                      effective_train_size=1., effective_valid_size=1., effective_test_size=1., fixed_subset=False,
                      sequential=False):
-    train_dataset, test_dataset = datasets_fn(data_dir, is_inception) # load data according to model type
+    train_dataset, test_dataset = datasets_fn(data_dir)
 
     worker_init_fn = None
     if deterministic:

distiller/apputils/image_classifier.py

@@ -470,16 +470,11 @@ def save_collectors_data(collectors, directory):
 
 
 def load_data(args, fixed_subset=False, sequential=False, load_train=True, load_val=True, load_test=True):
-    # create flag to indicate if model is inception to use proper image size
-    if args.arch in ['inception_v3, inceptionv3, inceptionv4', 'inceptionresnetv2']:
-        is_inception = True
-    else:
-        is_inception = False
     train_loader, val_loader, test_loader, _ =  apputils.load_data(args.dataset, 
                               os.path.expanduser(args.data), args.batch_size,
                               args.workers, args.validation_split, args.deterministic,
                               args.effective_train_size, args.effective_valid_size, args.effective_test_size,
-                              fixed_subset, sequential, is_inception)
+                              fixed_subset, sequential)
     msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d',
                    len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler))
 
@@ -580,17 +575,9 @@ def _log_training_progress():
             output = args.kd_policy.forward(inputs)
 
         if not early_exit_mode(args):
-            # For inception, we need to add losses for both classifier outputs
-            if isinstance(output, tuple):
-                loss = sum((criterion(o,target) for o in output))
-            else:
-                loss = criterion(output, target)
+            loss = criterion(output, target)
             # Measure accuracy
-            # For inception, we only consider accuracy of main classifier
-            if isinstance(output, tuple):
-                classerr.add(output[0].detach(), target)
-            else:
-                classerr.add(output.detach(), target)
+            classerr.add(output.detach(), target)
             acc_stats.append([classerr.value(1), classerr.value(5)])
         else:
             # Measure accuracy and record loss