Spaces:
Runtime error
Runtime error
Commit
Β·
fd37619
1
Parent(s):
793723d
- app.py +0 -11
- app/app_utils.py +0 -333
- app/face_utils.py +0 -68
- app/model.py +0 -78
- app/model_architectures.py +0 -46
app.py
CHANGED
@@ -4,13 +4,6 @@ from tabs.FACS_analysis import create_facs_analysis_tab
|
|
4 |
from ui_components import CUSTOM_CSS, HEADER_HTML, DISCLAIMER_HTML
|
5 |
import spaces # Importing spaces to utilize Zero GPU
|
6 |
|
7 |
-
# Initialize Zero GPU
|
8 |
-
if torch.cuda.is_available():
|
9 |
-
zero = torch.Tensor([0]).cuda()
|
10 |
-
print(f"Initial device: {zero.device}")
|
11 |
-
else:
|
12 |
-
zero = torch.Tensor([0])
|
13 |
-
print("CUDA is not available. Using CPU.")
|
14 |
|
15 |
# Define the tab structure
|
16 |
TAB_STRUCTURE = [
|
@@ -22,10 +15,6 @@ TAB_STRUCTURE = [
|
|
22 |
# Decorate GPU-dependent function with Zero GPU
|
23 |
@spaces.GPU(duration=120) # Allocates GPU for 120 seconds when needed
|
24 |
def create_demo():
|
25 |
-
if torch.cuda.is_available():
|
26 |
-
print(f"Device inside create_demo: {zero.device}")
|
27 |
-
else:
|
28 |
-
print("CUDA is not available inside create_demo.")
|
29 |
|
30 |
# Gradio blocks to create the interface
|
31 |
with gr.Blocks(css=CUSTOM_CSS) as demo:
|
|
|
4 |
from ui_components import CUSTOM_CSS, HEADER_HTML, DISCLAIMER_HTML
|
5 |
import spaces # Importing spaces to utilize Zero GPU
|
6 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
# Define the tab structure
|
9 |
TAB_STRUCTURE = [
|
|
|
15 |
# Decorate GPU-dependent function with Zero GPU
|
16 |
@spaces.GPU(duration=120) # Allocates GPU for 120 seconds when needed
|
17 |
def create_demo():
|
|
|
|
|
|
|
|
|
18 |
|
19 |
# Gradio blocks to create the interface
|
20 |
with gr.Blocks(css=CUSTOM_CSS) as demo:
|
app/app_utils.py
DELETED
@@ -1,333 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
import numpy as np
|
3 |
-
import mediapipe as mp
|
4 |
-
from PIL import Image
|
5 |
-
import cv2
|
6 |
-
from pytorch_grad_cam.utils.image import show_cam_on_image
|
7 |
-
import matplotlib.pyplot as plt
|
8 |
-
|
9 |
-
# Importing necessary components for the Gradio app
|
10 |
-
from app.model import pth_model_static, pth_model_dynamic, cam, pth_processing
|
11 |
-
from app.face_utils import get_box, display_info
|
12 |
-
from app.config import DICT_EMO, config_data
|
13 |
-
from app.plot import statistics_plot
|
14 |
-
|
15 |
-
mp_face_mesh = mp.solutions.face_mesh
|
16 |
-
|
17 |
-
def get_device():
|
18 |
-
if torch.backends.mps.is_available():
|
19 |
-
return torch.device("mps")
|
20 |
-
elif torch.cuda.is_available():
|
21 |
-
return torch.device("cuda")
|
22 |
-
else:
|
23 |
-
return torch.device("cpu")
|
24 |
-
|
25 |
-
device = get_device()
|
26 |
-
print(f"Using device: {device}")
|
27 |
-
|
28 |
-
# Move models to the selected device
|
29 |
-
pth_model_static = pth_model_static.to(device)
|
30 |
-
pth_model_dynamic = pth_model_dynamic.to(device)
|
31 |
-
|
32 |
-
def preprocess_image_and_predict(inp):
|
33 |
-
inp = np.array(inp)
|
34 |
-
|
35 |
-
if inp is None:
|
36 |
-
return None, None, None
|
37 |
-
|
38 |
-
try:
|
39 |
-
h, w = inp.shape[:2]
|
40 |
-
except Exception:
|
41 |
-
return None, None, None
|
42 |
-
|
43 |
-
with mp_face_mesh.FaceMesh(
|
44 |
-
max_num_faces=1,
|
45 |
-
refine_landmarks=False,
|
46 |
-
min_detection_confidence=0.5,
|
47 |
-
min_tracking_confidence=0.5,
|
48 |
-
) as face_mesh:
|
49 |
-
results = face_mesh.process(inp)
|
50 |
-
if results.multi_face_landmarks:
|
51 |
-
for fl in results.multi_face_landmarks:
|
52 |
-
startX, startY, endX, endY = get_box(fl, w, h)
|
53 |
-
cur_face = inp[startY:endY, startX:endX]
|
54 |
-
cur_face_n = pth_processing(Image.fromarray(cur_face)).to(device)
|
55 |
-
with torch.no_grad():
|
56 |
-
prediction = (
|
57 |
-
torch.nn.functional.softmax(pth_model_static(cur_face_n), dim=1)
|
58 |
-
.detach()
|
59 |
-
.cpu()
|
60 |
-
.numpy()[0]
|
61 |
-
)
|
62 |
-
confidences = {DICT_EMO[i]: float(prediction[i]) for i in range(7)}
|
63 |
-
grayscale_cam = cam(input_tensor=cur_face_n)
|
64 |
-
grayscale_cam = grayscale_cam[0, :]
|
65 |
-
cur_face_hm = cv2.resize(cur_face,(224,224))
|
66 |
-
cur_face_hm = np.float32(cur_face_hm) / 255
|
67 |
-
heatmap = show_cam_on_image(cur_face_hm, grayscale_cam, use_rgb=True)
|
68 |
-
|
69 |
-
return cur_face, heatmap, confidences
|
70 |
-
|
71 |
-
def preprocess_frame_and_predict_aus(frame):
|
72 |
-
if len(frame.shape) == 2:
|
73 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_GRAY2RGB)
|
74 |
-
elif frame.shape[2] == 4:
|
75 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_RGBA2RGB)
|
76 |
-
|
77 |
-
with mp_face_mesh.FaceMesh(
|
78 |
-
max_num_faces=1,
|
79 |
-
refine_landmarks=False,
|
80 |
-
min_detection_confidence=0.5,
|
81 |
-
min_tracking_confidence=0.5
|
82 |
-
) as face_mesh:
|
83 |
-
results = face_mesh.process(frame)
|
84 |
-
|
85 |
-
if results.multi_face_landmarks:
|
86 |
-
h, w = frame.shape[:2]
|
87 |
-
for fl in results.multi_face_landmarks:
|
88 |
-
startX, startY, endX, endY = get_box(fl, w, h)
|
89 |
-
cur_face = frame[startY:endY, startX:endX]
|
90 |
-
cur_face_n = pth_processing(Image.fromarray(cur_face)).to(device)
|
91 |
-
|
92 |
-
with torch.no_grad():
|
93 |
-
features = pth_model_static(cur_face_n)
|
94 |
-
au_intensities = features_to_au_intensities(features)
|
95 |
-
|
96 |
-
grayscale_cam = cam(input_tensor=cur_face_n)
|
97 |
-
grayscale_cam = grayscale_cam[0, :]
|
98 |
-
cur_face_hm = cv2.resize(cur_face, (224, 224))
|
99 |
-
cur_face_hm = np.float32(cur_face_hm) / 255
|
100 |
-
heatmap = show_cam_on_image(cur_face_hm, grayscale_cam, use_rgb=True)
|
101 |
-
|
102 |
-
return cur_face, au_intensities, heatmap
|
103 |
-
|
104 |
-
return None, None, None
|
105 |
-
|
106 |
-
def features_to_au_intensities(features):
|
107 |
-
features_np = features.detach().cpu().numpy()[0]
|
108 |
-
au_intensities = (features_np - features_np.min()) / (features_np.max() - features_np.min())
|
109 |
-
return au_intensities[:24] # Assuming we want 24 AUs
|
110 |
-
|
111 |
-
def preprocess_video_and_predict(video):
|
112 |
-
cap = cv2.VideoCapture(video)
|
113 |
-
w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
114 |
-
h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
115 |
-
fps = np.round(cap.get(cv2.CAP_PROP_FPS))
|
116 |
-
|
117 |
-
path_save_video_face = 'result_face.mp4'
|
118 |
-
vid_writer_face = cv2.VideoWriter(path_save_video_face, cv2.VideoWriter_fourcc(*'mp4v'), fps, (224, 224))
|
119 |
-
|
120 |
-
path_save_video_hm = 'result_hm.mp4'
|
121 |
-
vid_writer_hm = cv2.VideoWriter(path_save_video_hm, cv2.VideoWriter_fourcc(*'mp4v'), fps, (224, 224))
|
122 |
-
|
123 |
-
lstm_features = []
|
124 |
-
count_frame = 1
|
125 |
-
count_face = 0
|
126 |
-
probs = []
|
127 |
-
frames = []
|
128 |
-
au_intensities_list = []
|
129 |
-
last_output = None
|
130 |
-
last_heatmap = None
|
131 |
-
last_au_intensities = None
|
132 |
-
cur_face = None
|
133 |
-
|
134 |
-
with mp_face_mesh.FaceMesh(
|
135 |
-
max_num_faces=1,
|
136 |
-
refine_landmarks=False,
|
137 |
-
min_detection_confidence=0.5,
|
138 |
-
min_tracking_confidence=0.5) as face_mesh:
|
139 |
-
|
140 |
-
while cap.isOpened():
|
141 |
-
_, frame = cap.read()
|
142 |
-
if frame is None: break
|
143 |
-
|
144 |
-
frame_copy = frame.copy()
|
145 |
-
frame_copy.flags.writeable = False
|
146 |
-
frame_copy = cv2.cvtColor(frame_copy, cv2.COLOR_BGR2RGB)
|
147 |
-
results = face_mesh.process(frame_copy)
|
148 |
-
frame_copy.flags.writeable = True
|
149 |
-
|
150 |
-
if results.multi_face_landmarks:
|
151 |
-
for fl in results.multi_face_landmarks:
|
152 |
-
startX, startY, endX, endY = get_box(fl, w, h)
|
153 |
-
cur_face = frame_copy[startY:endY, startX: endX]
|
154 |
-
|
155 |
-
if count_face%config_data.FRAME_DOWNSAMPLING == 0:
|
156 |
-
cur_face_copy = pth_processing(Image.fromarray(cur_face)).to(device)
|
157 |
-
with torch.no_grad():
|
158 |
-
features = torch.nn.functional.relu(pth_model_static.extract_features(cur_face_copy)).detach().cpu().numpy()
|
159 |
-
au_intensities = features_to_au_intensities(pth_model_static(cur_face_copy))
|
160 |
-
|
161 |
-
grayscale_cam = cam(input_tensor=cur_face_copy)
|
162 |
-
grayscale_cam = grayscale_cam[0, :]
|
163 |
-
cur_face_hm = cv2.resize(cur_face,(224,224), interpolation = cv2.INTER_AREA)
|
164 |
-
cur_face_hm = np.float32(cur_face_hm) / 255
|
165 |
-
heatmap = show_cam_on_image(cur_face_hm, grayscale_cam, use_rgb=False)
|
166 |
-
last_heatmap = heatmap
|
167 |
-
last_au_intensities = au_intensities
|
168 |
-
|
169 |
-
if len(lstm_features) == 0:
|
170 |
-
lstm_features = [features]*10
|
171 |
-
else:
|
172 |
-
lstm_features = lstm_features[1:] + [features]
|
173 |
-
|
174 |
-
lstm_f = torch.from_numpy(np.vstack(lstm_features)).to(device)
|
175 |
-
lstm_f = torch.unsqueeze(lstm_f, 0)
|
176 |
-
with torch.no_grad():
|
177 |
-
output = pth_model_dynamic(lstm_f).detach().cpu().numpy()
|
178 |
-
last_output = output
|
179 |
-
|
180 |
-
if count_face == 0:
|
181 |
-
count_face += 1
|
182 |
-
|
183 |
-
else:
|
184 |
-
if last_output is not None:
|
185 |
-
output = last_output
|
186 |
-
heatmap = last_heatmap
|
187 |
-
au_intensities = last_au_intensities
|
188 |
-
|
189 |
-
elif last_output is None:
|
190 |
-
output = np.empty((1, 7))
|
191 |
-
output[:] = np.nan
|
192 |
-
au_intensities = np.empty(24)
|
193 |
-
au_intensities[:] = np.nan
|
194 |
-
|
195 |
-
probs.append(output[0])
|
196 |
-
frames.append(count_frame)
|
197 |
-
au_intensities_list.append(au_intensities)
|
198 |
-
else:
|
199 |
-
if last_output is not None:
|
200 |
-
lstm_features = []
|
201 |
-
empty = np.empty((7))
|
202 |
-
empty[:] = np.nan
|
203 |
-
probs.append(empty)
|
204 |
-
frames.append(count_frame)
|
205 |
-
au_intensities_list.append(np.full(24, np.nan))
|
206 |
-
|
207 |
-
if cur_face is not None:
|
208 |
-
heatmap_f = display_info(heatmap, 'Frame: {}'.format(count_frame), box_scale=.3)
|
209 |
-
|
210 |
-
cur_face = cv2.cvtColor(cur_face, cv2.COLOR_RGB2BGR)
|
211 |
-
cur_face = cv2.resize(cur_face, (224,224), interpolation = cv2.INTER_AREA)
|
212 |
-
cur_face = display_info(cur_face, 'Frame: {}'.format(count_frame), box_scale=.3)
|
213 |
-
vid_writer_face.write(cur_face)
|
214 |
-
vid_writer_hm.write(heatmap_f)
|
215 |
-
|
216 |
-
count_frame += 1
|
217 |
-
if count_face != 0:
|
218 |
-
count_face += 1
|
219 |
-
|
220 |
-
vid_writer_face.release()
|
221 |
-
vid_writer_hm.release()
|
222 |
-
|
223 |
-
stat = statistics_plot(frames, probs)
|
224 |
-
au_stat = au_statistics_plot(frames, au_intensities_list)
|
225 |
-
|
226 |
-
if not stat or not au_stat:
|
227 |
-
return None, None, None, None, None
|
228 |
-
|
229 |
-
return video, path_save_video_face, path_save_video_hm, stat, au_stat
|
230 |
-
|
231 |
-
# The rest of the functions remain the same
|
232 |
-
# ...
|
233 |
-
|
234 |
-
def au_statistics_plot(frames, au_intensities_list):
|
235 |
-
fig, ax = plt.subplots(figsize=(12, 6))
|
236 |
-
au_intensities_array = np.array(au_intensities_list)
|
237 |
-
|
238 |
-
for i in range(au_intensities_array.shape[1]):
|
239 |
-
ax.plot(frames, au_intensities_array[:, i], label=f'AU{i+1}')
|
240 |
-
|
241 |
-
ax.set_xlabel('Frame')
|
242 |
-
ax.set_ylabel('AU Intensity')
|
243 |
-
ax.set_title('Action Unit Intensities Over Time')
|
244 |
-
ax.legend(bbox_to_anchor=(1.05, 1), loc='upper left')
|
245 |
-
plt.tight_layout()
|
246 |
-
return fig
|
247 |
-
|
248 |
-
def preprocess_video_and_predict_sleep_quality(video):
|
249 |
-
cap = cv2.VideoCapture(video)
|
250 |
-
w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
251 |
-
h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
252 |
-
fps = np.round(cap.get(cv2.CAP_PROP_FPS))
|
253 |
-
|
254 |
-
path_save_video_original = 'result_original.mp4'
|
255 |
-
path_save_video_face = 'result_face.mp4'
|
256 |
-
path_save_video_sleep = 'result_sleep.mp4'
|
257 |
-
|
258 |
-
vid_writer_original = cv2.VideoWriter(path_save_video_original, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
|
259 |
-
vid_writer_face = cv2.VideoWriter(path_save_video_face, cv2.VideoWriter_fourcc(*'mp4v'), fps, (224, 224))
|
260 |
-
vid_writer_sleep = cv2.VideoWriter(path_save_video_sleep, cv2.VideoWriter_fourcc(*'mp4v'), fps, (224, 224))
|
261 |
-
|
262 |
-
frames = []
|
263 |
-
sleep_quality_scores = []
|
264 |
-
eye_bags_images = []
|
265 |
-
|
266 |
-
with mp_face_mesh.FaceMesh(
|
267 |
-
max_num_faces=1,
|
268 |
-
refine_landmarks=False,
|
269 |
-
min_detection_confidence=0.5,
|
270 |
-
min_tracking_confidence=0.5) as face_mesh:
|
271 |
-
|
272 |
-
while cap.isOpened():
|
273 |
-
ret, frame = cap.read()
|
274 |
-
if not ret:
|
275 |
-
break
|
276 |
-
|
277 |
-
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
278 |
-
results = face_mesh.process(frame_rgb)
|
279 |
-
|
280 |
-
if results.multi_face_landmarks:
|
281 |
-
for fl in results.multi_face_landmarks:
|
282 |
-
startX, startY, endX, endY = get_box(fl, w, h)
|
283 |
-
cur_face = frame_rgb[startY:endY, startX:endX]
|
284 |
-
|
285 |
-
sleep_quality_score, eye_bags_image = analyze_sleep_quality(cur_face)
|
286 |
-
sleep_quality_scores.append(sleep_quality_score)
|
287 |
-
eye_bags_images.append(cv2.resize(eye_bags_image, (224, 224)))
|
288 |
-
|
289 |
-
sleep_quality_viz = create_sleep_quality_visualization(cur_face, sleep_quality_score)
|
290 |
-
|
291 |
-
cur_face = cv2.resize(cur_face, (224, 224))
|
292 |
-
|
293 |
-
vid_writer_face.write(cv2.cvtColor(cur_face, cv2.COLOR_RGB2BGR))
|
294 |
-
vid_writer_sleep.write(sleep_quality_viz)
|
295 |
-
|
296 |
-
vid_writer_original.write(frame)
|
297 |
-
frames.append(len(frames) + 1)
|
298 |
-
|
299 |
-
cap.release()
|
300 |
-
vid_writer_original.release()
|
301 |
-
vid_writer_face.release()
|
302 |
-
vid_writer_sleep.release()
|
303 |
-
|
304 |
-
sleep_stat = sleep_quality_statistics_plot(frames, sleep_quality_scores)
|
305 |
-
|
306 |
-
if eye_bags_images:
|
307 |
-
average_eye_bags_image = np.mean(np.array(eye_bags_images), axis=0).astype(np.uint8)
|
308 |
-
else:
|
309 |
-
average_eye_bags_image = np.zeros((224, 224, 3), dtype=np.uint8)
|
310 |
-
|
311 |
-
return (path_save_video_original, path_save_video_face, path_save_video_sleep,
|
312 |
-
average_eye_bags_image, sleep_stat)
|
313 |
-
|
314 |
-
def analyze_sleep_quality(face_image):
|
315 |
-
# Placeholder function - implement your sleep quality analysis here
|
316 |
-
sleep_quality_score = np.random.random()
|
317 |
-
eye_bags_image = cv2.resize(face_image, (224, 224))
|
318 |
-
return sleep_quality_score, eye_bags_image
|
319 |
-
|
320 |
-
def create_sleep_quality_visualization(face_image, sleep_quality_score):
|
321 |
-
viz = face_image.copy()
|
322 |
-
cv2.putText(viz, f"Sleep Quality: {sleep_quality_score:.2f}", (10, 30),
|
323 |
-
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
|
324 |
-
return cv2.cvtColor(viz, cv2.COLOR_RGB2BGR)
|
325 |
-
|
326 |
-
def sleep_quality_statistics_plot(frames, sleep_quality_scores):
|
327 |
-
# Placeholder function - implement your statistics plotting here
|
328 |
-
fig, ax = plt.subplots()
|
329 |
-
ax.plot(frames, sleep_quality_scores)
|
330 |
-
ax.set_xlabel('Frame')
|
331 |
-
ax.set_ylabel('Sleep Quality Score')
|
332 |
-
ax.set_title('Sleep Quality Over Time')
|
333 |
-
return fig
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app/face_utils.py
DELETED
@@ -1,68 +0,0 @@
|
|
1 |
-
"""
|
2 |
-
File: face_utils.py
|
3 |
-
Author: Elena Ryumina and Dmitry Ryumin
|
4 |
-
Description: This module contains utility functions related to facial landmarks and image processing.
|
5 |
-
License: MIT License
|
6 |
-
"""
|
7 |
-
|
8 |
-
import numpy as np
|
9 |
-
import math
|
10 |
-
import cv2
|
11 |
-
|
12 |
-
|
13 |
-
def norm_coordinates(normalized_x, normalized_y, image_width, image_height):
|
14 |
-
x_px = min(math.floor(normalized_x * image_width), image_width - 1)
|
15 |
-
y_px = min(math.floor(normalized_y * image_height), image_height - 1)
|
16 |
-
return x_px, y_px
|
17 |
-
|
18 |
-
|
19 |
-
def get_box(fl, w, h):
|
20 |
-
idx_to_coors = {}
|
21 |
-
for idx, landmark in enumerate(fl.landmark):
|
22 |
-
landmark_px = norm_coordinates(landmark.x, landmark.y, w, h)
|
23 |
-
if landmark_px:
|
24 |
-
idx_to_coors[idx] = landmark_px
|
25 |
-
|
26 |
-
x_min = np.min(np.asarray(list(idx_to_coors.values()))[:, 0])
|
27 |
-
y_min = np.min(np.asarray(list(idx_to_coors.values()))[:, 1])
|
28 |
-
endX = np.max(np.asarray(list(idx_to_coors.values()))[:, 0])
|
29 |
-
endY = np.max(np.asarray(list(idx_to_coors.values()))[:, 1])
|
30 |
-
|
31 |
-
(startX, startY) = (max(0, x_min), max(0, y_min))
|
32 |
-
(endX, endY) = (min(w - 1, endX), min(h - 1, endY))
|
33 |
-
|
34 |
-
return startX, startY, endX, endY
|
35 |
-
|
36 |
-
def display_info(img, text, margin=1.0, box_scale=1.0):
|
37 |
-
img_copy = img.copy()
|
38 |
-
img_h, img_w, _ = img_copy.shape
|
39 |
-
line_width = int(min(img_h, img_w) * 0.001)
|
40 |
-
thickness = max(int(line_width / 3), 1)
|
41 |
-
|
42 |
-
font_face = cv2.FONT_HERSHEY_SIMPLEX
|
43 |
-
font_color = (0, 0, 0)
|
44 |
-
font_scale = thickness / 1.5
|
45 |
-
|
46 |
-
t_w, t_h = cv2.getTextSize(text, font_face, font_scale, None)[0]
|
47 |
-
|
48 |
-
margin_n = int(t_h * margin)
|
49 |
-
sub_img = img_copy[0 + margin_n: 0 + margin_n + t_h + int(2 * t_h * box_scale),
|
50 |
-
img_w - t_w - margin_n - int(2 * t_h * box_scale): img_w - margin_n]
|
51 |
-
|
52 |
-
white_rect = np.ones(sub_img.shape, dtype=np.uint8) * 255
|
53 |
-
|
54 |
-
img_copy[0 + margin_n: 0 + margin_n + t_h + int(2 * t_h * box_scale),
|
55 |
-
img_w - t_w - margin_n - int(2 * t_h * box_scale):img_w - margin_n] = cv2.addWeighted(sub_img, 0.5, white_rect, .5, 1.0)
|
56 |
-
|
57 |
-
cv2.putText(img=img_copy,
|
58 |
-
text=text,
|
59 |
-
org=(img_w - t_w - margin_n - int(2 * t_h * box_scale) // 2,
|
60 |
-
0 + margin_n + t_h + int(2 * t_h * box_scale) // 2),
|
61 |
-
fontFace=font_face,
|
62 |
-
fontScale=font_scale,
|
63 |
-
color=font_color,
|
64 |
-
thickness=thickness,
|
65 |
-
lineType=cv2.LINE_AA,
|
66 |
-
bottomLeftOrigin=False)
|
67 |
-
|
68 |
-
return img_copy
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app/model.py
DELETED
@@ -1,78 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
import torch
|
3 |
-
import torch.nn as nn
|
4 |
-
import torchvision.transforms as transforms
|
5 |
-
from pytorch_grad_cam import GradCAM
|
6 |
-
from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
|
7 |
-
import logging
|
8 |
-
from app.model_architectures import ResNet50, LSTMPyTorch
|
9 |
-
|
10 |
-
# Set up logging
|
11 |
-
logging.basicConfig(level=logging.INFO)
|
12 |
-
logger = logging.getLogger(__name__)
|
13 |
-
|
14 |
-
# Determine the device
|
15 |
-
device = torch.device('mps' if torch.backends.mps.is_available() else 'cuda' if torch.cuda.is_available() else 'cpu')
|
16 |
-
logger.info(f"Using device: {device}")
|
17 |
-
|
18 |
-
# Define paths
|
19 |
-
STATIC_MODEL_PATH = 'assets/models/FER_static_ResNet50_AffectNet.pt'
|
20 |
-
DYNAMIC_MODEL_PATH = 'assets/models/FER_dynamic_LSTM.pt'
|
21 |
-
|
22 |
-
def load_model(model_class, model_path, *args, **kwargs):
|
23 |
-
model = model_class(*args, **kwargs).to(device)
|
24 |
-
if os.path.exists(model_path):
|
25 |
-
try:
|
26 |
-
model.load_state_dict(torch.load(model_path, map_location=device))
|
27 |
-
model.eval()
|
28 |
-
logger.info(f"Model loaded successfully from {model_path}")
|
29 |
-
except Exception as e:
|
30 |
-
logger.error(f"Error loading model from {model_path}: {str(e)}")
|
31 |
-
logger.info("Initializing with random weights.")
|
32 |
-
else:
|
33 |
-
logger.warning(f"Model file not found at {model_path}. Initializing with random weights.")
|
34 |
-
return model
|
35 |
-
|
36 |
-
# Load the static model
|
37 |
-
pth_model_static = load_model(ResNet50, STATIC_MODEL_PATH, num_classes=7, channels=3)
|
38 |
-
|
39 |
-
# Load the dynamic model
|
40 |
-
pth_model_dynamic = load_model(LSTMPyTorch, DYNAMIC_MODEL_PATH, input_size=2048, hidden_size=256, num_layers=2, num_classes=7)
|
41 |
-
|
42 |
-
# Set up GradCAM
|
43 |
-
target_layers = [pth_model_static.resnet.layer4[-1]]
|
44 |
-
cam = GradCAM(model=pth_model_static, target_layers=target_layers)
|
45 |
-
|
46 |
-
# Define image preprocessing
|
47 |
-
pth_transform = transforms.Compose([
|
48 |
-
transforms.Resize((224, 224)),
|
49 |
-
transforms.ToTensor(),
|
50 |
-
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
|
51 |
-
])
|
52 |
-
|
53 |
-
def pth_processing(img):
|
54 |
-
img = pth_transform(img).unsqueeze(0).to(device)
|
55 |
-
return img
|
56 |
-
|
57 |
-
def predict_emotion(img):
|
58 |
-
with torch.no_grad():
|
59 |
-
output = pth_model_static(pth_processing(img))
|
60 |
-
_, predicted = torch.max(output, 1)
|
61 |
-
return predicted.item()
|
62 |
-
|
63 |
-
def get_emotion_probabilities(img):
|
64 |
-
with torch.no_grad():
|
65 |
-
output = nn.functional.softmax(pth_model_static(pth_processing(img)), dim=1)
|
66 |
-
return output.squeeze().cpu().numpy()
|
67 |
-
|
68 |
-
def generate_cam(img):
|
69 |
-
input_tensor = pth_processing(img)
|
70 |
-
targets = [ClassifierOutputTarget(predict_emotion(img))]
|
71 |
-
grayscale_cam = cam(input_tensor=input_tensor, targets=targets)
|
72 |
-
return grayscale_cam[0, :]
|
73 |
-
|
74 |
-
# Add any other necessary functions or variables here
|
75 |
-
|
76 |
-
if __name__ == "__main__":
|
77 |
-
logger.info("Model initialization complete.")
|
78 |
-
# You can add some test code here to verify everything is working correctly
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app/model_architectures.py
DELETED
@@ -1,46 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
import torch.nn as nn
|
3 |
-
import torchvision.models as models
|
4 |
-
|
5 |
-
class ResNet50(nn.Module):
|
6 |
-
def __init__(self, num_classes=7, channels=3):
|
7 |
-
super(ResNet50, self).__init__()
|
8 |
-
self.resnet = models.resnet50(pretrained=True)
|
9 |
-
# Modify the first convolutional layer if channels != 3
|
10 |
-
if channels != 3:
|
11 |
-
self.resnet.conv1 = nn.Conv2d(channels, 64, kernel_size=7, stride=2, padding=3, bias=False)
|
12 |
-
num_features = self.resnet.fc.in_features
|
13 |
-
self.resnet.fc = nn.Linear(num_features, num_classes)
|
14 |
-
|
15 |
-
def forward(self, x):
|
16 |
-
return self.resnet(x)
|
17 |
-
|
18 |
-
def extract_features(self, x):
|
19 |
-
x = self.resnet.conv1(x)
|
20 |
-
x = self.resnet.bn1(x)
|
21 |
-
x = self.resnet.relu(x)
|
22 |
-
x = self.resnet.maxpool(x)
|
23 |
-
|
24 |
-
x = self.resnet.layer1(x)
|
25 |
-
x = self.resnet.layer2(x)
|
26 |
-
x = self.resnet.layer3(x)
|
27 |
-
x = self.resnet.layer4(x)
|
28 |
-
|
29 |
-
x = self.resnet.avgpool(x)
|
30 |
-
x = torch.flatten(x, 1)
|
31 |
-
return x
|
32 |
-
|
33 |
-
class LSTMPyTorch(nn.Module):
|
34 |
-
def __init__(self, input_size, hidden_size, num_layers, num_classes):
|
35 |
-
super(LSTMPyTorch, self).__init__()
|
36 |
-
self.hidden_size = hidden_size
|
37 |
-
self.num_layers = num_layers
|
38 |
-
self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True)
|
39 |
-
self.fc = nn.Linear(hidden_size, num_classes)
|
40 |
-
|
41 |
-
def forward(self, x):
|
42 |
-
h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(x.device)
|
43 |
-
c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(x.device)
|
44 |
-
out, _ = self.lstm(x, (h0, c0))
|
45 |
-
out = self.fc(out[:, -1, :])
|
46 |
-
return out
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|