update app.py and resnet50.py

app.py

@@ -34,8 +34,8 @@ def predict(input_file, downsample_size):
 
     zip_path = os.path.join(input_file.split('/')[-1][:-4] + ".zip")
 
-    mean = np.asarray([0.3156024, 0.33569682, 0.34337464])
-    std = np.asarray([0.16568947, 0.17827448, 0.18925823])
+    mean = np.asarray([0.3156024, 0.33569682, 0.34337464], dtype=np.float32)
+    std = np.asarray([0.16568947, 0.17827448, 0.18925823], dtype=np.float32)
 
     img_vecs = []
     with torch.no_grad():
@@ -46,8 +46,9 @@ def predict(input_file, downsample_size):
                                                     to_rgb=True)):
                 in_img = (in_img.astype(np.float32) / 255.)
                 in_img = (in_img - mean) / std
+                in_img = np.expand_dims(in_img, 0)
                 in_img = np.transpose(in_img, (0, 3, 1, 2))
-                in_img = torch.from_numpy(in_img)
+                in_img = torch.from_numpy(in_img).float()
                 encoded = avg_pool(model(in_img))[0, :, 0, 0].cpu().numpy()
                 img_vecs += [encoded]
 
@@ -82,7 +83,8 @@ def predict(input_file, downsample_size):
 
 demo = gr.Interface(
     fn=predict,
-    inputs=[gr.inputs.Video(label="Upload Video File"), gr.inputs.Number(label="Downsample size")],
+    inputs=[gr.inputs.Video(label="Upload Video File"),
+            gr.inputs.Number(label="Downsample size")],
     outputs=gr.outputs.File(label="Zip"))
 
 demo.launch()

resnet50.py

@@ -314,7 +314,7 @@ class ResNet(nn.Module):
         )[1], 0)
         start_idx = 0
         for end_idx in idx_crops:
-            _out = self.forward_backbone(torch.cat(inputs[start_idx: end_idx]).cuda(non_blocking=True))
+            _out = self.forward_backbone(torch.cat(inputs[start_idx: end_idx])) # .cuda(non_blocking=True)
             if start_idx == 0:
                 output = _out
             else: