Merge pull request #3 from RaivoKoot/multiple_label_support

Multi-Label samples now supported

Author: RaivoKoot

README.md

@@ -31,6 +31,7 @@ def plot_video(rows, cols, frame_list, plot_width, plot_height):
     2. Minimal demo without sparse temporal sampling for single continuous frame clips, without image transforms
     3. Demo with image transforms
     4. Demo 3 continued with PyTorch dataloader
+    5. Demo of using a dataset where samples have multiple separate class labels
     
     """
     videos_root = os.path.join(os.getcwd(), 'demo_dataset')
@@ -145,7 +146,63 @@ def denormalize(video_tensor):
             """
             Insert Training Code Here
             """
-            print("Video Batch Tensor Size:", video_batch.size())
-            print("Labels Size:", labels.size())
+            print(labels)
+            print("\nVideo Batch Tensor Size:", video_batch.size())
+            print("Batch Labels Size:", labels.size())
             break
         break
+
+
+    """ DEMO 5: SAMPLES WITH MULTIPLE LABELS """
+    """
+    Apart from supporting just a single label per sample, VideoFrameDataset also supports multi-label samples,
+    where a sample can be associated with more than just one label. EPIC-KITCHENS, for example, associates a
+    noun, verb, and action with each video clip. To support this, instead of each row in annotations.txt
+    being (VIDEO_PATH, START_FRAME, END_FRAME, LABEL_ID), each row can also be
+    (VIDEO_PATH, START_FRAME, END_FRAME, LABEL_1_ID, ..., LABEL_N_ID). An example of this can be seen in the
+    directory `demo_dataset_multilabel`.
+    
+    Each sample returned by VideoFrameDataset is then ((FRAMESxCHANNELSxHEIGHTxWIDTH), (LABEL_1, ..., LABEL_N)).
+    When paired with the `torch.utils.data.DataLoader`, instead of yielding each batch as
+    ((BATCHxFRAMESxCHANNELSxHEIGHTxWIDTH), (BATCH)) where the second tuple item is the labels of the batch,
+    `torch.utils.data.DataLoader` returns a batch as ((BATCHxFRAMESxCHANNELSxHEIGHTxWIDTH), ((BATCH),...,(BATCH))
+    where the second tuple item is itself a tuple, with N BATCH-sized tensors of labels, where N is the 
+    number of labels assigned to each sample.
+    """
+    videos_root = os.path.join(os.getcwd(), 'demo_dataset_multilabel')
+    annotation_file = os.path.join(videos_root, 'annotations.txt')
+
+    dataset = VideoFrameDataset(
+        root_path=videos_root,
+        annotationfile_path=annotation_file,
+        num_segments=5,
+        frames_per_segment=1,
+        imagefile_template='img_{:05d}.jpg',
+        transform=preprocess,
+        random_shift=True,
+        test_mode=False
+    )
+
+    dataloader = torch.utils.data.DataLoader(
+        dataset=dataset,
+        batch_size=3,
+        shuffle=True,
+        num_workers=2,
+        pin_memory=True
+    )
+
+    print("\nMulti-Label Example")
+    for epoch in range(10):
+        for batch in dataloader:
+            """
+            Insert Training Code Here
+            """
+            video_batch, (labels1, labels2, labels3) = batch
+
+            print("Video Batch Tensor Size:", video_batch.size())
+            print("Labels1 Size:", labels1.size())  # == batch_size
+            print("Labels2 Size:", labels2.size())  # == batch_size
+            print("Labels3 Size:", labels3.size())  # == batch_size
+
+            break
+        break

demo.py

@@ -0,0 +1,4 @@
+jumping/0001 1 17 0 2 4
+jumping/0002 1 18 0 1 3
+running/0001 1 15 1 1 2
+running/0002 1 15 1 3 3