Spaces:
Runtime error
Runtime error
File size: 14,300 Bytes
9f83ce9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 | import torch
import logging
import onnxruntime as ort
from time import time
from typing import Union
from configs import ModelConfig, InferenceConfig
from utils import (
POSE_BASED_MODELS,
RGB_BASED_MODELS,
HUGGINGFACE_RGB_BASED_MODELS,
TORCHHUB_RGB_BASED_MODELS,
)
from transformers import (
ImageProcessingMixin,
FeatureExtractionMixin,
AutoModelForVideoClassification,
AutoModel,
Pipeline,
pipeline,
)
from transformers.pipelines import PIPELINE_REGISTRY
from visualization import draw_text_on_image
from utils import exists_on_hf
from models import (
Swin3DConfig, Swin3DImageProcessor, Swin3DForVideoClassification,
S3DConfig, S3DImageProcessor, S3DForVideoClassification,
VideoResNetConfig, VideoResNetImageProcessor, VideoResNetForVideoClassification,
MViTConfig, MViTImageProcessor, MViTForVideoClassification,
SLGCNConfig, SLGCNFeatureExtractor, SLGCNForGraphClassification,
SPOTERConfig, SPOTERFeatureExtractor, SPOTERForGraphClassification,
DSTASLRConfig, DSTASLRFeatureExtractor, DSTASLRForGraphClassification,
VideoMAEConfig, VideoMAEImageProcessor, VideoMAEForVideoClassification
)
from pipelines import (
VideoClassificationPipeline,
SLGCNGraphClassificationPipeline,
SPOTERGraphClassificationPipeline,
)
def load_model(
model_config: ModelConfig,
label2id: dict = None,
id2label: dict = None,
do_train: bool = False,
) -> tuple:
'''
'''
if do_train:
if model_config.arch in POSE_BASED_MODELS:
return load_pose_model_for_training(model_config, label2id, id2label)
return load_rgb_model_for_training(model_config, label2id, id2label)
if model_config.arch in POSE_BASED_MODELS:
processor = FeatureExtractionMixin.from_pretrained(
model_config.pretrained,
trust_remote_code=True,
cache_dir="models/huggingface",
)
model = AutoModel.from_pretrained(
model_config.pretrained,
trust_remote_code=True,
cache_dir="models/huggingface",
)
else:
processor = ImageProcessingMixin.from_pretrained(
model_config.pretrained,
trust_remote_code=True,
cache_dir="models/huggingface",
)
model = AutoModelForVideoClassification.from_pretrained(
model_config.pretrained,
trust_remote_code=True,
cache_dir="models/huggingface",
)
model.eval()
return model.config, processor, model
def load_rgb_model_for_training(
model_config: ModelConfig,
label2id: dict = None,
id2label: dict = None,
) -> tuple:
'''
'''
if model_config.arch in HUGGINGFACE_RGB_BASED_MODELS:
if model_config.arch == "videomae":
config_class = VideoMAEConfig
processor_class = VideoMAEImageProcessor
model_class = VideoMAEForVideoClassification
elif exists_on_hf(model_config.pretrained):
processor = ImageProcessingMixin.from_pretrained(
model_config.pretrained,
trust_remote_code=True,
cache_dir="models/huggingface",
)
model = AutoModelForVideoClassification.from_pretrained(
model_config.pretrained,
label2id,
id2label,
ignore_mismatched_sizes=True,
trust_remote_code=True,
cache_dir="models/huggingface",
)
return model.config, processor, model
elif model_config.arch in TORCHHUB_RGB_BASED_MODELS:
if model_config.arch in ['swin3d_t', 'swin3d_s', 'swin3d_b']:
config_class = Swin3DConfig
processor_class = Swin3DImageProcessor
model_class = Swin3DForVideoClassification
elif model_config.arch in ['r3d_18', 'mc3_18', 'r2plus1d_18']:
config_class = VideoResNetConfig
processor_class = VideoResNetImageProcessor
model_class = VideoResNetForVideoClassification
elif model_config.arch in ['s3d']:
config_class = S3DConfig
processor_class = S3DImageProcessor
model_class = S3DForVideoClassification
elif model_config.arch in ['mvit_v1_b', 'mvit_v2_s']:
config_class = MViTConfig
processor_class = MViTImageProcessor
model_class = MViTForVideoClassification
else:
logging.error(f"Model {model_config.arch} is not supported")
exit(1)
config_class.register_for_auto_class()
processor_class.register_for_auto_class("AutoImageProcessor")
model_class.register_for_auto_class("AutoModel")
model_class.register_for_auto_class("AutoModelForVideoClassification")
logging.info(f"{model_config.arch} classes registered")
config = config_class(**vars(model_config))
processor = processor_class(config=config)
model = model_class(config=config, label2id=label2id, id2label=id2label)
return config, processor, model
def load_pose_model_for_training(
model_config: ModelConfig,
label2id: dict = None,
id2label: dict = None,
) -> tuple:
'''
'''
if exists_on_hf(model_config.pretrained):
processor = FeatureExtractionMixin.from_pretrained(
model_config.pretrained,
trust_remote_code=True,
cache_dir="models/huggingface",
)
model = AutoModel.from_pretrained(
model_config.pretrained,
label2id=label2id,
id2label=id2label,
ignore_mismatched_sizes=True,
trust_remote_code=True,
cache_dir="models/huggingface",
)
return model.config, processor, model
elif model_config.arch in POSE_BASED_MODELS:
if model_config.arch == "spoter":
config_class = SPOTERConfig
processor_class = SPOTERFeatureExtractor
model_class = SPOTERForGraphClassification
elif model_config.arch == "sl_gcn":
config_class = SLGCNConfig
processor_class = SLGCNFeatureExtractor
model_class = SLGCNForGraphClassification
elif model_config.arch == "dsta_slr":
config_class = DSTASLRConfig
processor_class = DSTASLRFeatureExtractor
model_class = DSTASLRForGraphClassification
else:
logging.error(f"Model {model_config.arch} is not supported")
exit(1)
config_class.register_for_auto_class()
processor_class.register_for_auto_class("AutoFeatureExtractor")
model_class.register_for_auto_class("AutoModel")
logging.info(F"Registering {model_config.arch} classes")
config = config_class(**vars(model_config))
processor = processor_class(config=config)
model = model_class(config=config, label2id=label2id, id2label=id2label)
return config, processor, model
class Predictions:
def __init__(
self,
predictions: list[dict] = None,
inference_time: float = 0,
start_time: float = 0,
end_time: float = 0,
) -> None:
self.predictions = predictions
self.inference_time = inference_time
self.start_time = start_time
self.end_time = end_time
def visualize(
self,
frame: torch.Tensor,
position: tuple = (20, 100),
prefix: str = "Predictions",
color: tuple = (0, 0, 255),
) -> None:
text = prefix + ": " + self.get_pred_message()
return draw_text_on_image(
image=frame,
text=text,
position=position,
color=color,
font_size=20,
)
def get_pred_message(self) -> str:
if not any((
self.start_time,
self.end_time,
self.inference_time,
self.predictions
)):
return ""
return ', '.join(
[
f"{pred['gloss']} ({pred['score']*100:.2f}%)"
for pred in self.predictions
]
)
def __str__(self) -> str:
if not any((
self.start_time,
self.end_time,
self.inference_time,
self.predictions
)):
return ""
predictions = self.get_pred_message()
message = "Sample start: {:.2f}s - end: {:.2f}s | Runtime: {:.2f}s | Predictions: {}"
return message.format(self.start_time, self.end_time, self.inference_time, predictions)
def merge_results(self, results: dict = None) -> dict:
if results is None:
results = {
"start_time": [],
"end_time": [],
"inference_time": [],
"prediction": [],
}
results["start_time"].append(self.start_time)
results["end_time"].append(self.end_time)
results["inference_time"].append(self.inference_time)
results["prediction"].append(self.predictions)
return results
def get_predictions(
inputs: torch.Tensor,
model: Union[ort.InferenceSession, AutoModel],
id2gloss: dict,
k: int = 3,
) -> Predictions:
'''
Get the top-k predictions.
Parameters
----------
inputs : torch.Tensor
Model inputs (Time, Height, Width, Channels).
model : Union[ort.InferenceSession, AutoModel]
Model to get predictions from.
id2gloss : dict
Mapping of class indices to glosses.
k : int, optional
Number of predictions to return, by default 3.
Returns
-------
tuple
List of top-k predictions and inference time.
'''
if inputs is None:
return Predictions()
# Get logits
start_time = time()
if isinstance(model, ort.InferenceSession):
inputs = inputs.cpu().numpy()
logits = torch.from_numpy(model.run(None, {"pixel_values": inputs})[0])
else:
logits = model(inputs.to(model.device)).logits
inference_time = time() - start_time
# Get top-3 predictions
topk_scores, topk_indices = torch.topk(logits, k, dim=1)
topk_scores = torch.nn.functional.softmax(topk_scores, dim=1).squeeze().detach().numpy()
topk_indices = topk_indices.squeeze().detach().numpy()
predictions = [
{
'gloss': id2gloss[str(topk_indices[i])],
'score': topk_scores[i],
}
for i in range(k)
]
return Predictions(predictions=predictions, inference_time=inference_time)
def register_pipeline(model_config: ModelConfig) -> Pipeline:
'''
'''
_, processor, model = load_model(model_config)
if model_config.arch == "spoter":
PIPELINE_REGISTRY.register_pipeline(
"video-classification",
pipeline_class=SPOTERGraphClassificationPipeline,
pt_model=AutoModel,
default={"pt": ("vsltranslation/spoter_v3.0", "main")},
type="multimodal",
)
return SPOTERGraphClassificationPipeline(
model=model,
feature_extractor=processor,
)
elif model_config.arch in ["sl_gcn", "dsta_slr"]:
PIPELINE_REGISTRY.register_pipeline(
"video-classification",
pipeline_class=SLGCNGraphClassificationPipeline,
pt_model=AutoModel,
default={"pt": ("vsltranslation/sl_gcn_joint_v1.0", "main")},
type="multimodal",
)
return SLGCNGraphClassificationPipeline(
model=model,
feature_extractor=processor,
)
PIPELINE_REGISTRY.register_pipeline(
"video-classification",
pipeline_class=VideoClassificationPipeline,
pt_model=AutoModelForVideoClassification,
default={"pt": ("vsltranslation/swin3d_t_v1.0", "main")},
type="multimodal",
)
return VideoClassificationPipeline(
model=model,
image_processor=processor,
)
def load_pipeline(
model_config: ModelConfig,
inference_config: InferenceConfig,
) -> Pipeline:
'''
'''
if model_config.arch in POSE_BASED_MODELS:
return pipeline(
"video-classification",
model=model_config.pretrained,
feature_extractor=model_config.pretrained,
device=inference_config.device,
model_kwargs={
"cache_dir": inference_config.cache_dir,
},
trust_remote_code=True,
use_onnx=inference_config.use_onnx,
top_k=inference_config.top_k,
bone_stream=inference_config.bone_stream,
motion_stream=inference_config.motion_stream,
)
return pipeline(
"video-classification",
model=model_config.pretrained,
image_processor=model_config.pretrained,
device=inference_config.device,
model_kwargs={
"cache_dir": inference_config.cache_dir,
},
trust_remote_code=True,
use_onnx=inference_config.use_onnx,
top_k=inference_config.top_k,
)
def get_input_shape(
arch: str,
processor: Union[ImageProcessingMixin, FeatureExtractionMixin],
batch_size: int = 1,
) -> tuple:
'''
Get the input shape for the model.
Parameters
----------
processor : Union[ImageProcessingMixin, FeatureExtractionMixin]
Model processor.
batch_size : int, optional
Batch size, by default 1.
Returns
-------
tuple
Input shape.
'''
if arch in RGB_BASED_MODELS:
return (
batch_size,
processor.num_frames,
3,
processor.size["height"],
processor.size["width"]
)
elif arch in POSE_BASED_MODELS:
if arch == "spoter":
return (
batch_size,
processor.num_frames,
processor.num_points,
processor.in_channels,
)
elif arch in ["sl_gcn", "dsta_slr"]:
return (
batch_size,
processor.in_channels,
processor.window_size,
processor.num_points,
processor.num_people,
)
else:
logging.error(f"Model {arch} is not supported")
exit(1)
else:
logging.error(f"Model {arch} is not supported")
exit(1)
|