Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,92 +1,92 @@
|
|
1 |
-
import torch
|
2 |
-
import torch.nn as nn
|
3 |
-
import cv2
|
4 |
-
import os
|
5 |
-
import gradio as gr
|
6 |
-
from PIL import Image
|
7 |
-
from transformers import ViTImageProcessor, ViTModel
|
8 |
-
|
9 |
-
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
10 |
-
MODEL_PATH =
|
11 |
-
NUM_CLASSES = 5
|
12 |
-
MAX_FRAMES = 16
|
13 |
-
|
14 |
-
class ViT_LSTM(nn.Module):
|
15 |
-
def __init__(self, feature_dim=768, hidden_dim=512, num_classes=NUM_CLASSES):
|
16 |
-
super(ViT_LSTM, self).__init__()
|
17 |
-
self.lstm = nn.LSTM(feature_dim, hidden_dim, batch_first=True, num_layers=2, bidirectional=True)
|
18 |
-
self.fc = nn.Linear(hidden_dim * 2, num_classes)
|
19 |
-
self.dropout = nn.Dropout(0.3)
|
20 |
-
|
21 |
-
def forward(self, x):
|
22 |
-
lstm_out, _ = self.lstm(x)
|
23 |
-
lstm_out = lstm_out[:, -1, :]
|
24 |
-
out = self.dropout(lstm_out)
|
25 |
-
out = self.fc(out)
|
26 |
-
return out
|
27 |
-
|
28 |
-
vit_processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224")
|
29 |
-
vit_model = ViTModel.from_pretrained("google/vit-base-patch16-224").to(DEVICE)
|
30 |
-
|
31 |
-
model = ViT_LSTM()
|
32 |
-
model.load_state_dict(torch.load(MODEL_PATH, map_location=DEVICE))
|
33 |
-
model.to(DEVICE)
|
34 |
-
model.eval()
|
35 |
-
LABELS = ["BaseballPitch", "Basketball", "BenchPress", "Biking", "Billiards"]
|
36 |
-
|
37 |
-
def extract_vit_features(video_path, max_frames=MAX_FRAMES):
|
38 |
-
cap = cv2.VideoCapture(video_path)
|
39 |
-
frames = []
|
40 |
-
frame_count = 0
|
41 |
-
|
42 |
-
while cap.isOpened() and frame_count < max_frames:
|
43 |
-
ret, frame = cap.read()
|
44 |
-
if not ret:
|
45 |
-
break
|
46 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
47 |
-
frames.append(Image.fromarray(frame))
|
48 |
-
frame_count += 1
|
49 |
-
|
50 |
-
cap.release()
|
51 |
-
|
52 |
-
if not frames:
|
53 |
-
return None
|
54 |
-
|
55 |
-
print(f"Extracted {len(frames)} frames from video.")
|
56 |
-
|
57 |
-
inputs = vit_processor(images=frames, return_tensors="pt")["pixel_values"].to(DEVICE)
|
58 |
-
|
59 |
-
with torch.no_grad():
|
60 |
-
features = vit_model(inputs).last_hidden_state.mean(dim=1)
|
61 |
-
|
62 |
-
return features
|
63 |
-
|
64 |
-
def predict_action(video_file):
|
65 |
-
video_path = video_file.name
|
66 |
-
print(f"Received video path: {video_path}")
|
67 |
-
|
68 |
-
features = extract_vit_features(video_path)
|
69 |
-
|
70 |
-
if features is None:
|
71 |
-
return "No frames extracted, please upload a valid video."
|
72 |
-
|
73 |
-
features = features.unsqueeze(0)
|
74 |
-
|
75 |
-
with torch.no_grad():
|
76 |
-
output = model(features)
|
77 |
-
predicted_class = torch.argmax(output, dim=1).item()
|
78 |
-
|
79 |
-
return f"Predicted Action: {LABELS[predicted_class]}"
|
80 |
-
|
81 |
-
# Gradio Interface
|
82 |
-
with gr.Blocks() as demo:
|
83 |
-
gr.Markdown("# Action Recognition with ViT-LSTM")
|
84 |
-
gr.Markdown("Upload a short video to predict the action.")
|
85 |
-
|
86 |
-
video_input = gr.File(label="Upload a video")
|
87 |
-
output_text = gr.Textbox(label="Prediction")
|
88 |
-
|
89 |
-
predict_btn = gr.Button("Predict Action")
|
90 |
-
predict_btn.click(fn=predict_action, inputs=video_input, outputs=output_text)
|
91 |
-
|
92 |
-
demo.launch()
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import cv2
|
4 |
+
import os
|
5 |
+
import gradio as gr
|
6 |
+
from PIL import Image
|
7 |
+
from transformers import ViTImageProcessor, ViTModel
|
8 |
+
|
9 |
+
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
10 |
+
MODEL_PATH = "Vit_LSTM.pth"
|
11 |
+
NUM_CLASSES = 5
|
12 |
+
MAX_FRAMES = 16
|
13 |
+
|
14 |
+
class ViT_LSTM(nn.Module):
|
15 |
+
def __init__(self, feature_dim=768, hidden_dim=512, num_classes=NUM_CLASSES):
|
16 |
+
super(ViT_LSTM, self).__init__()
|
17 |
+
self.lstm = nn.LSTM(feature_dim, hidden_dim, batch_first=True, num_layers=2, bidirectional=True)
|
18 |
+
self.fc = nn.Linear(hidden_dim * 2, num_classes)
|
19 |
+
self.dropout = nn.Dropout(0.3)
|
20 |
+
|
21 |
+
def forward(self, x):
|
22 |
+
lstm_out, _ = self.lstm(x)
|
23 |
+
lstm_out = lstm_out[:, -1, :]
|
24 |
+
out = self.dropout(lstm_out)
|
25 |
+
out = self.fc(out)
|
26 |
+
return out
|
27 |
+
|
28 |
+
vit_processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224")
|
29 |
+
vit_model = ViTModel.from_pretrained("google/vit-base-patch16-224").to(DEVICE)
|
30 |
+
|
31 |
+
model = ViT_LSTM()
|
32 |
+
model.load_state_dict(torch.load(MODEL_PATH, map_location=DEVICE))
|
33 |
+
model.to(DEVICE)
|
34 |
+
model.eval()
|
35 |
+
LABELS = ["BaseballPitch", "Basketball", "BenchPress", "Biking", "Billiards"]
|
36 |
+
|
37 |
+
def extract_vit_features(video_path, max_frames=MAX_FRAMES):
|
38 |
+
cap = cv2.VideoCapture(video_path)
|
39 |
+
frames = []
|
40 |
+
frame_count = 0
|
41 |
+
|
42 |
+
while cap.isOpened() and frame_count < max_frames:
|
43 |
+
ret, frame = cap.read()
|
44 |
+
if not ret:
|
45 |
+
break
|
46 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
47 |
+
frames.append(Image.fromarray(frame))
|
48 |
+
frame_count += 1
|
49 |
+
|
50 |
+
cap.release()
|
51 |
+
|
52 |
+
if not frames:
|
53 |
+
return None
|
54 |
+
|
55 |
+
print(f"Extracted {len(frames)} frames from video.")
|
56 |
+
|
57 |
+
inputs = vit_processor(images=frames, return_tensors="pt")["pixel_values"].to(DEVICE)
|
58 |
+
|
59 |
+
with torch.no_grad():
|
60 |
+
features = vit_model(inputs).last_hidden_state.mean(dim=1)
|
61 |
+
|
62 |
+
return features
|
63 |
+
|
64 |
+
def predict_action(video_file):
|
65 |
+
video_path = video_file.name
|
66 |
+
print(f"Received video path: {video_path}")
|
67 |
+
|
68 |
+
features = extract_vit_features(video_path)
|
69 |
+
|
70 |
+
if features is None:
|
71 |
+
return "No frames extracted, please upload a valid video."
|
72 |
+
|
73 |
+
features = features.unsqueeze(0)
|
74 |
+
|
75 |
+
with torch.no_grad():
|
76 |
+
output = model(features)
|
77 |
+
predicted_class = torch.argmax(output, dim=1).item()
|
78 |
+
|
79 |
+
return f"Predicted Action: {LABELS[predicted_class]}"
|
80 |
+
|
81 |
+
# Gradio Interface
|
82 |
+
with gr.Blocks() as demo:
|
83 |
+
gr.Markdown("# Action Recognition with ViT-LSTM")
|
84 |
+
gr.Markdown("Upload a short video to predict the action.")
|
85 |
+
|
86 |
+
video_input = gr.File(label="Upload a video")
|
87 |
+
output_text = gr.Textbox(label="Prediction")
|
88 |
+
|
89 |
+
predict_btn = gr.Button("Predict Action")
|
90 |
+
predict_btn.click(fn=predict_action, inputs=video_input, outputs=output_text)
|
91 |
+
|
92 |
+
demo.launch()
|