Spaces:
Running
on
Zero
Running
on
Zero
fix vae nan bug
Browse files
app.py
CHANGED
@@ -217,21 +217,7 @@ if NEW_MODEL:
|
|
217 |
model.eval()
|
218 |
print(missing_keys, extra_keys)
|
219 |
assert len(missing_keys) == 0
|
220 |
-
|
221 |
-
print(f"vae_state_dict encoder dtype: {vae_state_dict['encoder.conv_in.weight'].dtype}")
|
222 |
-
autoencoder = vqvae.create_model(3, 3, opts.latent_dim).eval().requires_grad_(False)
|
223 |
-
print(f"autoencoder encoder dtype: {next(autoencoder.encoder.parameters()).dtype}")
|
224 |
-
print(f"encoder before load_state_dict parameters min: {min([p.min() for p in autoencoder.encoder.parameters()])}")
|
225 |
-
print(f"encoder before load_state_dict parameters max: {max([p.max() for p in autoencoder.encoder.parameters()])}")
|
226 |
-
missing_keys, extra_keys = autoencoder.load_state_dict(vae_state_dict, strict=False)
|
227 |
-
print(f"encoder after load_state_dict parameters min: {min([p.min() for p in autoencoder.encoder.parameters()])}")
|
228 |
-
print(f"encoder after load_state_dict parameters max: {max([p.max() for p in autoencoder.encoder.parameters()])}")
|
229 |
-
autoencoder = autoencoder.to(device)
|
230 |
-
autoencoder.eval()
|
231 |
-
print(f"encoder after eval() min: {min([p.min() for p in autoencoder.encoder.parameters()])}")
|
232 |
-
print(f"encoder after eval() max: {max([p.max() for p in autoencoder.encoder.parameters()])}")
|
233 |
-
print(f"autoencoder encoder after eval() dtype: {next(autoencoder.encoder.parameters()).dtype}")
|
234 |
-
assert len(missing_keys) == 0
|
235 |
# else:
|
236 |
# opts = HandDiffOpts()
|
237 |
# model_path = './finetune_epoch=5-step=130000.ckpt'
|
@@ -266,127 +252,6 @@ hands = mp_hands.Hands(
|
|
266 |
min_detection_confidence=0.1,
|
267 |
)
|
268 |
|
269 |
-
# def make_ref_cond(
|
270 |
-
# image
|
271 |
-
# ):
|
272 |
-
# print("ready to run autoencoder")
|
273 |
-
# # print(f"image.device: {image.device}, type(image): {type(image)}")
|
274 |
-
# # image = image.to("cuda")
|
275 |
-
# print(f"autoencoder device: {next(autoencoder.parameters()).device}")
|
276 |
-
# latent = opts.latent_scaling_factor * autoencoder.encode(image[None, ...]).sample()
|
277 |
-
# return image[None, ...], latent
|
278 |
-
|
279 |
-
|
280 |
-
# def get_ref_anno(ref):
|
281 |
-
# print("inside get_ref_anno")
|
282 |
-
# if ref is None:
|
283 |
-
# return (
|
284 |
-
# None,
|
285 |
-
# None,
|
286 |
-
# None,
|
287 |
-
# None,
|
288 |
-
# None,
|
289 |
-
# )
|
290 |
-
# img = ref["composite"][..., :3]
|
291 |
-
# img = cv2.resize(img, opts.image_size, interpolation=cv2.INTER_AREA)
|
292 |
-
# keypts = np.zeros((42, 2))
|
293 |
-
# print("ready to run mediapipe")
|
294 |
-
# if REF_POSE_MASK:
|
295 |
-
# print(f"type(img): {type(img)}, img.shape: {img.shape}, img.dtype: {img.dtype}")
|
296 |
-
# mp_pose = hands.process(img)
|
297 |
-
# print("processed mediapipe")
|
298 |
-
# detected = np.array([0, 0])
|
299 |
-
# start_idx = 0
|
300 |
-
# if mp_pose.multi_hand_landmarks:
|
301 |
-
# # handedness is flipped assuming the input image is mirrored in MediaPipe
|
302 |
-
# for hand_landmarks, handedness in zip(
|
303 |
-
# mp_pose.multi_hand_landmarks, mp_pose.multi_handedness
|
304 |
-
# ):
|
305 |
-
# # actually right hand
|
306 |
-
# if handedness.classification[0].label == "Left":
|
307 |
-
# start_idx = 0
|
308 |
-
# detected[0] = 1
|
309 |
-
# # actually left hand
|
310 |
-
# elif handedness.classification[0].label == "Right":
|
311 |
-
# start_idx = 21
|
312 |
-
# detected[1] = 1
|
313 |
-
# for i, landmark in enumerate(hand_landmarks.landmark):
|
314 |
-
# keypts[start_idx + i] = [
|
315 |
-
# landmark.x * opts.image_size[1],
|
316 |
-
# landmark.y * opts.image_size[0],
|
317 |
-
# ]
|
318 |
-
|
319 |
-
# sam_predictor.set_image(img)
|
320 |
-
# l = keypts[:21].shape[0]
|
321 |
-
# if keypts[0].sum() != 0 and keypts[21].sum() != 0:
|
322 |
-
# input_point = np.array([keypts[0], keypts[21]])
|
323 |
-
# input_label = np.array([1, 1])
|
324 |
-
# elif keypts[0].sum() != 0:
|
325 |
-
# input_point = np.array(keypts[:1])
|
326 |
-
# input_label = np.array([1])
|
327 |
-
# elif keypts[21].sum() != 0:
|
328 |
-
# input_point = np.array(keypts[21:22])
|
329 |
-
# input_label = np.array([1])
|
330 |
-
# print("ready to run SAM")
|
331 |
-
# masks, _, _ = sam_predictor.predict(
|
332 |
-
# point_coords=input_point,
|
333 |
-
# point_labels=input_label,
|
334 |
-
# multimask_output=False,
|
335 |
-
# )
|
336 |
-
# print("finished SAM")
|
337 |
-
# hand_mask = masks[0]
|
338 |
-
# masked_img = img * hand_mask[..., None] + 255 * (1 - hand_mask[..., None])
|
339 |
-
# ref_pose = visualize_hand(keypts, masked_img)
|
340 |
-
# else:
|
341 |
-
# raise gr.Error("No hands detected in the reference image.")
|
342 |
-
# else:
|
343 |
-
# hand_mask = np.zeros_like(img[:,:, 0])
|
344 |
-
# ref_pose = np.zeros_like(img)
|
345 |
-
|
346 |
-
# image_transform = Compose(
|
347 |
-
# [
|
348 |
-
# ToTensor(),
|
349 |
-
# Resize(opts.image_size),
|
350 |
-
# Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
|
351 |
-
# ]
|
352 |
-
# )
|
353 |
-
# image = image_transform(img)
|
354 |
-
# kpts_valid = check_keypoints_validity(keypts, opts.image_size)
|
355 |
-
# heatmaps = torch.tensor(
|
356 |
-
# keypoint_heatmap(
|
357 |
-
# scale_keypoint(keypts, opts.image_size, opts.latent_size), opts.latent_size, var=1.0
|
358 |
-
# )
|
359 |
-
# * kpts_valid[:, None, None],
|
360 |
-
# dtype=torch.float,
|
361 |
-
# # device=device,
|
362 |
-
# )[None, ...]
|
363 |
-
# mask = torch.tensor(
|
364 |
-
# cv2.resize(
|
365 |
-
# hand_mask.astype(int),
|
366 |
-
# dsize=opts.latent_size,
|
367 |
-
# interpolation=cv2.INTER_NEAREST,
|
368 |
-
# ),
|
369 |
-
# dtype=torch.float,
|
370 |
-
# # device=device,
|
371 |
-
# ).unsqueeze(0)[None, ...]
|
372 |
-
# image, latent = make_ref_cond(
|
373 |
-
# image,
|
374 |
-
# # keypts,
|
375 |
-
# # hand_mask,
|
376 |
-
# # device=device,
|
377 |
-
# # target_size=opts.image_size,
|
378 |
-
# # latent_size=opts.latent_size,
|
379 |
-
# )
|
380 |
-
# print("finished autoencoder")
|
381 |
-
|
382 |
-
# if not REF_POSE_MASK:
|
383 |
-
# heatmaps = torch.zeros_like(heatmaps)
|
384 |
-
# mask = torch.zeros_like(mask)
|
385 |
-
# ref_cond = torch.cat([latent, heatmaps, mask], 1)
|
386 |
-
|
387 |
-
# return img, ref_pose, ref_cond
|
388 |
-
|
389 |
-
|
390 |
def get_ref_anno(ref):
|
391 |
if ref is None:
|
392 |
return (
|
@@ -396,6 +261,24 @@ def get_ref_anno(ref):
|
|
396 |
None,
|
397 |
None,
|
398 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
399 |
img = ref["composite"][..., :3]
|
400 |
img = cv2.resize(img, opts.image_size, interpolation=cv2.INTER_AREA)
|
401 |
keypts = np.zeros((42, 2))
|
|
|
217 |
model.eval()
|
218 |
print(missing_keys, extra_keys)
|
219 |
assert len(missing_keys) == 0
|
220 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
221 |
# else:
|
222 |
# opts = HandDiffOpts()
|
223 |
# model_path = './finetune_epoch=5-step=130000.ckpt'
|
|
|
252 |
min_detection_confidence=0.1,
|
253 |
)
|
254 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
255 |
def get_ref_anno(ref):
|
256 |
if ref is None:
|
257 |
return (
|
|
|
261 |
None,
|
262 |
None,
|
263 |
)
|
264 |
+
|
265 |
+
vae_state_dict = torch.load(vae_path, map_location='cpu')['state_dict']
|
266 |
+
print(f"vae_state_dict encoder dtype: {vae_state_dict['encoder.conv_in.weight'].dtype}")
|
267 |
+
autoencoder = vqvae.create_model(3, 3, opts.latent_dim).eval().requires_grad_(False)
|
268 |
+
print(f"autoencoder encoder dtype: {next(autoencoder.encoder.parameters()).dtype}")
|
269 |
+
print(f"encoder before load_state_dict parameters min: {min([p.min() for p in autoencoder.encoder.parameters()])}")
|
270 |
+
print(f"encoder before load_state_dict parameters max: {max([p.max() for p in autoencoder.encoder.parameters()])}")
|
271 |
+
missing_keys, extra_keys = autoencoder.load_state_dict(vae_state_dict, strict=False)
|
272 |
+
print(f"encoder after load_state_dict parameters min: {min([p.min() for p in autoencoder.encoder.parameters()])}")
|
273 |
+
print(f"encoder after load_state_dict parameters max: {max([p.max() for p in autoencoder.encoder.parameters()])}")
|
274 |
+
autoencoder = autoencoder.to(device)
|
275 |
+
autoencoder.eval()
|
276 |
+
print(f"encoder after eval() min: {min([p.min() for p in autoencoder.encoder.parameters()])}")
|
277 |
+
print(f"encoder after eval() max: {max([p.max() for p in autoencoder.encoder.parameters()])}")
|
278 |
+
print(f"autoencoder encoder after eval() dtype: {next(autoencoder.encoder.parameters()).dtype}")
|
279 |
+
assert len(missing_keys) == 0
|
280 |
+
|
281 |
+
|
282 |
img = ref["composite"][..., :3]
|
283 |
img = cv2.resize(img, opts.image_size, interpolation=cv2.INTER_AREA)
|
284 |
keypts = np.zeros((42, 2))
|