Spaces:
Running
Running
feat: allow latest version only
Browse files- dev/inference/wandb-backend.ipynb +154 -358
dev/inference/wandb-backend.ipynb
CHANGED
|
@@ -32,9 +32,25 @@
|
|
| 32 |
"metadata": {},
|
| 33 |
"outputs": [],
|
| 34 |
"source": [
|
| 35 |
-
"
|
|
|
|
| 36 |
"VQGAN_REPO, VQGAN_COMMIT_ID = 'dalle-mini/vqgan_imagenet_f16_16384', None\n",
|
| 37 |
-
"normalize_text = True"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
]
|
| 39 |
},
|
| 40 |
{
|
|
@@ -104,18 +120,6 @@
|
|
| 104 |
" samples = [samples[i:i+batch_size] for i in range(0, len(samples), batch_size)]"
|
| 105 |
]
|
| 106 |
},
|
| 107 |
-
{
|
| 108 |
-
"cell_type": "code",
|
| 109 |
-
"execution_count": null,
|
| 110 |
-
"id": "3ffb1d09-bd1c-4f57-9ae5-3eda6f7d3a08",
|
| 111 |
-
"metadata": {},
|
| 112 |
-
"outputs": [],
|
| 113 |
-
"source": [
|
| 114 |
-
"# TODO: iterate on runs\n",
|
| 115 |
-
"wandb_run = wandb_runs[0]\n",
|
| 116 |
-
"model_pmapped = False"
|
| 117 |
-
]
|
| 118 |
-
},
|
| 119 |
{
|
| 120 |
"cell_type": "code",
|
| 121 |
"execution_count": null,
|
|
@@ -123,12 +127,14 @@
|
|
| 123 |
"metadata": {},
|
| 124 |
"outputs": [],
|
| 125 |
"source": [
|
| 126 |
-
"def get_artifact_versions(run_id):\n",
|
| 127 |
" try:\n",
|
| 128 |
-
"
|
|
|
|
|
|
|
|
|
|
| 129 |
" except:\n",
|
| 130 |
-
"
|
| 131 |
-
" return versions"
|
| 132 |
]
|
| 133 |
},
|
| 134 |
{
|
|
@@ -139,7 +145,7 @@
|
|
| 139 |
"outputs": [],
|
| 140 |
"source": [
|
| 141 |
"def get_training_config(run_id):\n",
|
| 142 |
-
" training_run = api.run(f'
|
| 143 |
" config = training_run.config\n",
|
| 144 |
" return config"
|
| 145 |
]
|
|
@@ -155,7 +161,7 @@
|
|
| 155 |
"def get_last_inference_version(run_id):\n",
|
| 156 |
" try:\n",
|
| 157 |
" inference_run = api.run(f'dalle-mini/dalle-mini/inference-{run_id}')\n",
|
| 158 |
-
" return inference_run.summary.get('
|
| 159 |
" except:\n",
|
| 160 |
" return None"
|
| 161 |
]
|
|
@@ -186,68 +192,142 @@
|
|
| 186 |
{
|
| 187 |
"cell_type": "code",
|
| 188 |
"execution_count": null,
|
| 189 |
-
"id": "
|
| 190 |
"metadata": {},
|
| 191 |
"outputs": [],
|
| 192 |
"source": [
|
| 193 |
-
"
|
| 194 |
-
"
|
| 195 |
-
" last_inference_version = get_last_inference_version(run_id)\n",
|
| 196 |
-
" training_config = get_training_config(run_id)\n",
|
| 197 |
-
" run = None\n",
|
| 198 |
-
" p_generate = None\n",
|
| 199 |
-
" model_files = ['config.json', 'flax_model.msgpack', 'merges.txt', 'special_tokens_map.json', 'tokenizer.json', 'tokenizer_config.json', 'vocab.json']\n",
|
| 200 |
-
" for artifact in artifact_versions:\n",
|
| 201 |
-
" print(f'Processing artifact: {artifact.name}')\n",
|
| 202 |
-
" version = int(artifact.version[1:])\n",
|
| 203 |
-
" if last_version_inference is None:\n",
|
| 204 |
-
" # we should start from v0\n",
|
| 205 |
-
" assert version == 0\n",
|
| 206 |
-
" elif version <= last_version_inference:\n",
|
| 207 |
-
" print(f'v{version} has already been logged (versions logged up to v{last_version_inference}')\n",
|
| 208 |
-
" else:\n",
|
| 209 |
-
" # check we are logging the correct version\n",
|
| 210 |
-
" assert version == last_version_inference + 1\n",
|
| 211 |
-
" \n",
|
| 212 |
-
" # start/resume corresponding run\n",
|
| 213 |
-
" if run is None:\n",
|
| 214 |
-
" run = wandb.init(job_type='inference', config=config, id=f'inference-{wandb_run}', resume='allow')\n",
|
| 215 |
-
" \n",
|
| 216 |
-
" # work in temporary directory\n",
|
| 217 |
-
" with tempfile.TemporaryDirectory() as tmp:\n",
|
| 218 |
-
" \n",
|
| 219 |
-
" # download model files\n",
|
| 220 |
-
" artifact = run.use_artifact(artifact)\n",
|
| 221 |
-
" for f in model_files:\n",
|
| 222 |
-
" artifact.get_path(f).download(tmp)\n",
|
| 223 |
-
" \n",
|
| 224 |
-
" # load tokenizer and model\n",
|
| 225 |
-
" tokenizer = BartTokenizer.from_pretrained(tmp)\n",
|
| 226 |
-
" model = CustomFlaxBartForConditionalGeneration.from_pretrained(tmp)\n",
|
| 227 |
-
" model_params = replicate(model.params)\n",
|
| 228 |
-
" \n",
|
| 229 |
-
" # pmap model function needs to happen only once per model config\n",
|
| 230 |
-
" if p_generate is None:\n",
|
| 231 |
-
" p_generate = pmap_model_function(model)\n",
|
| 232 |
-
" \n",
|
| 233 |
-
" for batch in tqdm(samples):\n",
|
| 234 |
-
" prompts = [x['Caption'] for x in batch]\n",
|
| 235 |
-
" processed_prompts = [text_normalizer(x) for x in prompts] if normalize_text else prompts\n",
|
| 236 |
-
" \n",
|
| 237 |
-
"\n",
|
| 238 |
-
" \n",
|
| 239 |
-
" \n",
|
| 240 |
-
" "
|
| 241 |
]
|
| 242 |
},
|
| 243 |
{
|
| 244 |
"cell_type": "code",
|
| 245 |
"execution_count": null,
|
| 246 |
-
"id": "
|
| 247 |
"metadata": {},
|
| 248 |
-
"outputs": [
|
| 249 |
-
|
| 250 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 251 |
]
|
| 252 |
},
|
| 253 |
{
|
|
@@ -257,296 +337,12 @@
|
|
| 257 |
"metadata": {},
|
| 258 |
"outputs": [],
|
| 259 |
"source": [
|
|
|
|
| 260 |
"def log_runs(runs):\n",
|
| 261 |
" for run in tqdm(runs):\n",
|
| 262 |
" log_run(run)"
|
| 263 |
]
|
| 264 |
},
|
| 265 |
-
{
|
| 266 |
-
"cell_type": "code",
|
| 267 |
-
"execution_count": null,
|
| 268 |
-
"id": "7a24b903-777b-4e3d-817c-00ed613a7021",
|
| 269 |
-
"metadata": {},
|
| 270 |
-
"outputs": [],
|
| 271 |
-
"source": [
|
| 272 |
-
"# TODO: loop over samples\n",
|
| 273 |
-
"batch = samples[0]\n",
|
| 274 |
-
"prompts = [x['Caption'] for x in batch]\n",
|
| 275 |
-
"processed_prompts = [text_normalizer(x) for x in prompts] if normalize_text else prompts"
|
| 276 |
-
]
|
| 277 |
-
},
|
| 278 |
-
{
|
| 279 |
-
"cell_type": "code",
|
| 280 |
-
"execution_count": null,
|
| 281 |
-
"id": "d77aa785-dc05-4070-aba2-aa007524d20b",
|
| 282 |
-
"metadata": {},
|
| 283 |
-
"outputs": [],
|
| 284 |
-
"source": [
|
| 285 |
-
"processed_prompts"
|
| 286 |
-
]
|
| 287 |
-
},
|
| 288 |
-
{
|
| 289 |
-
"cell_type": "code",
|
| 290 |
-
"execution_count": null,
|
| 291 |
-
"id": "95db38fb-8948-4814-98ae-c172ca7c6d0a",
|
| 292 |
-
"metadata": {},
|
| 293 |
-
"outputs": [],
|
| 294 |
-
"source": [
|
| 295 |
-
"repeated_prompts = processed_prompts * jax.device_count()"
|
| 296 |
-
]
|
| 297 |
-
},
|
| 298 |
-
{
|
| 299 |
-
"cell_type": "code",
|
| 300 |
-
"execution_count": null,
|
| 301 |
-
"id": "e948ba9e-3700-4e87-926f-580a10f3e5cd",
|
| 302 |
-
"metadata": {},
|
| 303 |
-
"outputs": [],
|
| 304 |
-
"source": [
|
| 305 |
-
"tokenized_prompt = tokenizer(repeated_prompts, return_tensors='jax', padding='max_length', truncation=True, max_length=128).data\n",
|
| 306 |
-
"tokenized_prompt = shard(tokenized_prompt)"
|
| 307 |
-
]
|
| 308 |
-
},
|
| 309 |
-
{
|
| 310 |
-
"cell_type": "code",
|
| 311 |
-
"execution_count": null,
|
| 312 |
-
"id": "30d96812-fc17-4acf-bb64-5fdb8d0cd313",
|
| 313 |
-
"metadata": {},
|
| 314 |
-
"outputs": [],
|
| 315 |
-
"source": [
|
| 316 |
-
"tokenized_prompt['input_ids'].shape"
|
| 317 |
-
]
|
| 318 |
-
},
|
| 319 |
-
{
|
| 320 |
-
"cell_type": "code",
|
| 321 |
-
"execution_count": null,
|
| 322 |
-
"id": "92ea034b-2649-4d18-ab6d-877ed04ae5c4",
|
| 323 |
-
"metadata": {},
|
| 324 |
-
"outputs": [],
|
| 325 |
-
"source": [
|
| 326 |
-
"images = []\n",
|
| 327 |
-
"for i in range(num_images // jax.device_count()):\n",
|
| 328 |
-
" key, subkey = jax.random.split(key, 2)\n",
|
| 329 |
-
" \n",
|
| 330 |
-
" encoded_images = p_generate(tokenized_prompt, shard_prng_key(subkey), model_params)\n",
|
| 331 |
-
" encoded_images = encoded_images.sequences[..., 1:]\n",
|
| 332 |
-
" \n",
|
| 333 |
-
" decoded_images = p_decode(encoded_images, vqgan_params)\n",
|
| 334 |
-
" decoded_images = decoded_images.clip(0., 1.).reshape((-1, 256, 256, 3))\n",
|
| 335 |
-
" \n",
|
| 336 |
-
" for img in decoded_images:\n",
|
| 337 |
-
" images.append(Image.fromarray(np.asarray(img * 255, dtype=np.uint8)))\n",
|
| 338 |
-
" "
|
| 339 |
-
]
|
| 340 |
-
},
|
| 341 |
-
{
|
| 342 |
-
"cell_type": "code",
|
| 343 |
-
"execution_count": null,
|
| 344 |
-
"id": "84d52f30-44c9-4a74-9992-fb2578f19b90",
|
| 345 |
-
"metadata": {},
|
| 346 |
-
"outputs": [],
|
| 347 |
-
"source": [
|
| 348 |
-
"len(images)"
|
| 349 |
-
]
|
| 350 |
-
},
|
| 351 |
-
{
|
| 352 |
-
"cell_type": "code",
|
| 353 |
-
"execution_count": null,
|
| 354 |
-
"id": "beb594f9-5b91-47fe-98bd-41e68c6b1d73",
|
| 355 |
-
"metadata": {},
|
| 356 |
-
"outputs": [],
|
| 357 |
-
"source": [
|
| 358 |
-
"images[0]"
|
| 359 |
-
]
|
| 360 |
-
},
|
| 361 |
-
{
|
| 362 |
-
"cell_type": "code",
|
| 363 |
-
"execution_count": null,
|
| 364 |
-
"id": "bb135190-64e5-44af-b416-e688b034da44",
|
| 365 |
-
"metadata": {},
|
| 366 |
-
"outputs": [],
|
| 367 |
-
"source": [
|
| 368 |
-
"images[1]"
|
| 369 |
-
]
|
| 370 |
-
},
|
| 371 |
-
{
|
| 372 |
-
"cell_type": "code",
|
| 373 |
-
"execution_count": null,
|
| 374 |
-
"id": "d78a0d92-72c2-4f82-a6ab-b3f5865dd863",
|
| 375 |
-
"metadata": {},
|
| 376 |
-
"outputs": [],
|
| 377 |
-
"source": [
|
| 378 |
-
"clip_inputs = processor(text=prompts, images=images, return_tensors='np', padding='max_length', max_length=77, truncation=True).data"
|
| 379 |
-
]
|
| 380 |
-
},
|
| 381 |
-
{
|
| 382 |
-
"cell_type": "code",
|
| 383 |
-
"execution_count": null,
|
| 384 |
-
"id": "89ff78a6-bfa4-44d9-ad66-07a4a68b4352",
|
| 385 |
-
"metadata": {},
|
| 386 |
-
"outputs": [],
|
| 387 |
-
"source": [
|
| 388 |
-
"# each shard will have one prompt\n",
|
| 389 |
-
"clip_inputs['input_ids'].shape"
|
| 390 |
-
]
|
| 391 |
-
},
|
| 392 |
-
{
|
| 393 |
-
"cell_type": "code",
|
| 394 |
-
"execution_count": null,
|
| 395 |
-
"id": "2cda8984-049c-4c87-96ad-7b0412750656",
|
| 396 |
-
"metadata": {},
|
| 397 |
-
"outputs": [],
|
| 398 |
-
"source": [
|
| 399 |
-
"# each shard needs to have the images corresponding to a specific prompt\n",
|
| 400 |
-
"clip_inputs['pixel_values'].shape"
|
| 401 |
-
]
|
| 402 |
-
},
|
| 403 |
-
{
|
| 404 |
-
"cell_type": "code",
|
| 405 |
-
"execution_count": null,
|
| 406 |
-
"id": "0a044e8f-be29-404b-b6c7-8f2395c5efc6",
|
| 407 |
-
"metadata": {},
|
| 408 |
-
"outputs": [],
|
| 409 |
-
"source": [
|
| 410 |
-
"images_per_prompt_indices = np.asarray(range(0, len(images), batch_size))\n",
|
| 411 |
-
"images_per_prompt_indices"
|
| 412 |
-
]
|
| 413 |
-
},
|
| 414 |
-
{
|
| 415 |
-
"cell_type": "code",
|
| 416 |
-
"execution_count": null,
|
| 417 |
-
"id": "7a6c61b3-12e0-45d8-b39a-830288324d3d",
|
| 418 |
-
"metadata": {},
|
| 419 |
-
"outputs": [],
|
| 420 |
-
"source": []
|
| 421 |
-
},
|
| 422 |
-
{
|
| 423 |
-
"cell_type": "code",
|
| 424 |
-
"execution_count": null,
|
| 425 |
-
"id": "7318e67e-4214-46f9-bf60-6d139d4bd00f",
|
| 426 |
-
"metadata": {},
|
| 427 |
-
"outputs": [],
|
| 428 |
-
"source": [
|
| 429 |
-
"# reorder so each shard will have correct images\n",
|
| 430 |
-
"clip_inputs['pixel_values'] = jnp.concatenate(list(clip_inputs['pixel_values'][images_per_prompt_indices + i] for i in range(batch_size)))"
|
| 431 |
-
]
|
| 432 |
-
},
|
| 433 |
-
{
|
| 434 |
-
"cell_type": "code",
|
| 435 |
-
"execution_count": null,
|
| 436 |
-
"id": "90c949a2-8e2a-4905-b6d4-92038f1704b8",
|
| 437 |
-
"metadata": {},
|
| 438 |
-
"outputs": [],
|
| 439 |
-
"source": [
|
| 440 |
-
"clip_inputs = shard(clip_inputs)"
|
| 441 |
-
]
|
| 442 |
-
},
|
| 443 |
-
{
|
| 444 |
-
"cell_type": "code",
|
| 445 |
-
"execution_count": null,
|
| 446 |
-
"id": "58fa836e-5ebb-45e7-af77-ab10646dfbfb",
|
| 447 |
-
"metadata": {},
|
| 448 |
-
"outputs": [],
|
| 449 |
-
"source": [
|
| 450 |
-
"logits = p_clip(clip_inputs)"
|
| 451 |
-
]
|
| 452 |
-
},
|
| 453 |
-
{
|
| 454 |
-
"cell_type": "code",
|
| 455 |
-
"execution_count": null,
|
| 456 |
-
"id": "fd7a3f91-3a1f-4a0a-8b3e-3c926cd367fb",
|
| 457 |
-
"metadata": {},
|
| 458 |
-
"outputs": [],
|
| 459 |
-
"source": [
|
| 460 |
-
"logits.shape"
|
| 461 |
-
]
|
| 462 |
-
},
|
| 463 |
-
{
|
| 464 |
-
"cell_type": "code",
|
| 465 |
-
"execution_count": null,
|
| 466 |
-
"id": "fa406db7-0a21-4e4b-9890-4c7aece4280c",
|
| 467 |
-
"metadata": {},
|
| 468 |
-
"outputs": [],
|
| 469 |
-
"source": [
|
| 470 |
-
"logits = logits.reshape(-1, num_images)"
|
| 471 |
-
]
|
| 472 |
-
},
|
| 473 |
-
{
|
| 474 |
-
"cell_type": "code",
|
| 475 |
-
"execution_count": null,
|
| 476 |
-
"id": "9c359a8c-2c27-4e68-8775-371857397723",
|
| 477 |
-
"metadata": {},
|
| 478 |
-
"outputs": [],
|
| 479 |
-
"source": [
|
| 480 |
-
"logits.shape"
|
| 481 |
-
]
|
| 482 |
-
},
|
| 483 |
-
{
|
| 484 |
-
"cell_type": "code",
|
| 485 |
-
"execution_count": null,
|
| 486 |
-
"id": "a56b9f28-dd91-4382-bc47-11e89fda1254",
|
| 487 |
-
"metadata": {},
|
| 488 |
-
"outputs": [],
|
| 489 |
-
"source": [
|
| 490 |
-
"logits"
|
| 491 |
-
]
|
| 492 |
-
},
|
| 493 |
-
{
|
| 494 |
-
"cell_type": "code",
|
| 495 |
-
"execution_count": null,
|
| 496 |
-
"id": "0bed8167-0a6d-46c1-badf-8bdc20b93c31",
|
| 497 |
-
"metadata": {},
|
| 498 |
-
"outputs": [],
|
| 499 |
-
"source": [
|
| 500 |
-
"top_idx = logits.argsort()[:, -top_k:][..., ::-1]"
|
| 501 |
-
]
|
| 502 |
-
},
|
| 503 |
-
{
|
| 504 |
-
"cell_type": "code",
|
| 505 |
-
"execution_count": null,
|
| 506 |
-
"id": "188c5333-6b8c-4a17-8cc8-15651c77ef99",
|
| 507 |
-
"metadata": {},
|
| 508 |
-
"outputs": [],
|
| 509 |
-
"source": [
|
| 510 |
-
"len(images)"
|
| 511 |
-
]
|
| 512 |
-
},
|
| 513 |
-
{
|
| 514 |
-
"cell_type": "code",
|
| 515 |
-
"execution_count": null,
|
| 516 |
-
"id": "babd22b3-e773-467d-8bbb-f0323f57a44b",
|
| 517 |
-
"metadata": {},
|
| 518 |
-
"outputs": [],
|
| 519 |
-
"source": [
|
| 520 |
-
"results = []\n",
|
| 521 |
-
"columns = ['Caption', 'Theme'] + [f'Image {i+1}' for i in range(top_k)] + [f'Score {i+1}' for i in range(top_k)]\n",
|
| 522 |
-
"logits = jax.device_get(logits)"
|
| 523 |
-
]
|
| 524 |
-
},
|
| 525 |
-
{
|
| 526 |
-
"cell_type": "code",
|
| 527 |
-
"execution_count": null,
|
| 528 |
-
"id": "75976c9f-dea5-48e3-8920-55a1bbfd91c2",
|
| 529 |
-
"metadata": {},
|
| 530 |
-
"outputs": [],
|
| 531 |
-
"source": [
|
| 532 |
-
"for i, (idx, scores, sample) in enumerate(zip(top_idx, logits, batch)):\n",
|
| 533 |
-
" if sample['Caption'] == padding_item: continue\n",
|
| 534 |
-
" cur_images = [images[x] for x in images_per_prompt_indices + i]\n",
|
| 535 |
-
" top_images = [wandb.Image(cur_images[x]) for x in idx]\n",
|
| 536 |
-
" top_scores = [scores[x] for x in idx]\n",
|
| 537 |
-
" results.append([sample['Caption'], sample['Theme']] + top_images + top_scores)"
|
| 538 |
-
]
|
| 539 |
-
},
|
| 540 |
-
{
|
| 541 |
-
"cell_type": "code",
|
| 542 |
-
"execution_count": null,
|
| 543 |
-
"id": "4bf40461-99d3-4d36-b7cc-e0129a3c9053",
|
| 544 |
-
"metadata": {},
|
| 545 |
-
"outputs": [],
|
| 546 |
-
"source": [
|
| 547 |
-
"table = wandb.Table(columns=columns, data=results)"
|
| 548 |
-
]
|
| 549 |
-
},
|
| 550 |
{
|
| 551 |
"cell_type": "code",
|
| 552 |
"execution_count": null,
|
|
|
|
| 32 |
"metadata": {},
|
| 33 |
"outputs": [],
|
| 34 |
"source": [
|
| 35 |
+
"run_ids = ['rjf3rycy']\n",
|
| 36 |
+
"ENTITY, PROJECT = 'dalle-mini', 'dalle-mini' # used only for training run\n",
|
| 37 |
"VQGAN_REPO, VQGAN_COMMIT_ID = 'dalle-mini/vqgan_imagenet_f16_16384', None\n",
|
| 38 |
+
"normalize_text = True\n",
|
| 39 |
+
"latest_only = False # log only latest or all versions"
|
| 40 |
+
]
|
| 41 |
+
},
|
| 42 |
+
{
|
| 43 |
+
"cell_type": "code",
|
| 44 |
+
"execution_count": null,
|
| 45 |
+
"id": "23e00271-941c-4e1b-b6a9-107a1b77324d",
|
| 46 |
+
"metadata": {},
|
| 47 |
+
"outputs": [],
|
| 48 |
+
"source": [
|
| 49 |
+
"run_ids = ['4oh3u7ca']\n",
|
| 50 |
+
"ENTITY, PROJECT = 'wandb', 'hf-flax-dalle-mini'\n",
|
| 51 |
+
"VQGAN_REPO, VQGAN_COMMIT_ID = 'dalle-mini/vqgan_imagenet_f16_16384', None\n",
|
| 52 |
+
"normalize_text = False\n",
|
| 53 |
+
"latest_only = True # log only latest or all versions"
|
| 54 |
]
|
| 55 |
},
|
| 56 |
{
|
|
|
|
| 120 |
" samples = [samples[i:i+batch_size] for i in range(0, len(samples), batch_size)]"
|
| 121 |
]
|
| 122 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 123 |
{
|
| 124 |
"cell_type": "code",
|
| 125 |
"execution_count": null,
|
|
|
|
| 127 |
"metadata": {},
|
| 128 |
"outputs": [],
|
| 129 |
"source": [
|
| 130 |
+
"def get_artifact_versions(run_id, latest_only=False):\n",
|
| 131 |
" try:\n",
|
| 132 |
+
" if latest_only:\n",
|
| 133 |
+
" return [api.artifact(type='bart_model', name=f'{ENTITY}/{PROJECT}/model-{run_id}:latest')]\n",
|
| 134 |
+
" else:\n",
|
| 135 |
+
" return api.artifact_versions(type_name='bart_model', name=f'{ENTITY}/{PROJECT}/model-{run_id}', per_page=10000)\n",
|
| 136 |
" except:\n",
|
| 137 |
+
" return []"
|
|
|
|
| 138 |
]
|
| 139 |
},
|
| 140 |
{
|
|
|
|
| 145 |
"outputs": [],
|
| 146 |
"source": [
|
| 147 |
"def get_training_config(run_id):\n",
|
| 148 |
+
" training_run = api.run(f'{ENTITY}/{PROJECT}/{run_id}')\n",
|
| 149 |
" config = training_run.config\n",
|
| 150 |
" return config"
|
| 151 |
]
|
|
|
|
| 161 |
"def get_last_inference_version(run_id):\n",
|
| 162 |
" try:\n",
|
| 163 |
" inference_run = api.run(f'dalle-mini/dalle-mini/inference-{run_id}')\n",
|
| 164 |
+
" return inference_run.summary.get('version', None)\n",
|
| 165 |
" except:\n",
|
| 166 |
" return None"
|
| 167 |
]
|
|
|
|
| 192 |
{
|
| 193 |
"cell_type": "code",
|
| 194 |
"execution_count": null,
|
| 195 |
+
"id": "23b2444c-67a9-44d7-abd1-187ed83a9431",
|
| 196 |
"metadata": {},
|
| 197 |
"outputs": [],
|
| 198 |
"source": [
|
| 199 |
+
"run_id = run_ids[0]\n",
|
| 200 |
+
"# TODO: turn everything into a class"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 201 |
]
|
| 202 |
},
|
| 203 |
{
|
| 204 |
"cell_type": "code",
|
| 205 |
"execution_count": null,
|
| 206 |
+
"id": "bba70f33-af8b-4eb3-9973-7be672301a0b",
|
| 207 |
"metadata": {},
|
| 208 |
+
"outputs": [
|
| 209 |
+
{
|
| 210 |
+
"name": "stdout",
|
| 211 |
+
"output_type": "stream",
|
| 212 |
+
"text": [
|
| 213 |
+
"Processing artifact: model-4oh3u7ca:v54\n"
|
| 214 |
+
]
|
| 215 |
+
},
|
| 216 |
+
{
|
| 217 |
+
"name": "stderr",
|
| 218 |
+
"output_type": "stream",
|
| 219 |
+
"text": [
|
| 220 |
+
"\u001b[34m\u001b[1mwandb\u001b[0m: Currently logged in as: \u001b[33mborisd13\u001b[0m (use `wandb login --relogin` to force relogin)\n"
|
| 221 |
+
]
|
| 222 |
+
},
|
| 223 |
+
{
|
| 224 |
+
"data": {
|
| 225 |
+
"text/html": [
|
| 226 |
+
"\n",
|
| 227 |
+
" Syncing run <strong><a href=\"https://wandb.ai/dalle-mini/dalle-mini/runs/inference-4oh3u7ca\" target=\"_blank\">inference-4oh3u7ca</a></strong> to <a href=\"https://wandb.ai/dalle-mini/dalle-mini\" target=\"_blank\">Weights & Biases</a> (<a href=\"https://docs.wandb.com/integrations/jupyter.html\" target=\"_blank\">docs</a>).<br/>\n",
|
| 228 |
+
"\n",
|
| 229 |
+
" "
|
| 230 |
+
],
|
| 231 |
+
"text/plain": [
|
| 232 |
+
"<IPython.core.display.HTML object>"
|
| 233 |
+
]
|
| 234 |
+
},
|
| 235 |
+
"metadata": {},
|
| 236 |
+
"output_type": "display_data"
|
| 237 |
+
}
|
| 238 |
+
],
|
| 239 |
+
"source": [
|
| 240 |
+
"artifact_versions = get_artifact_versions(run_id, latest_only)\n",
|
| 241 |
+
"last_inference_version = get_last_inference_version(run_id)\n",
|
| 242 |
+
"training_config = get_training_config(run_id)\n",
|
| 243 |
+
"run = None\n",
|
| 244 |
+
"p_generate = None\n",
|
| 245 |
+
"model_files = ['config.json', 'flax_model.msgpack', 'merges.txt', 'special_tokens_map.json', 'tokenizer.json', 'tokenizer_config.json', 'vocab.json']\n",
|
| 246 |
+
"for artifact in artifact_versions:\n",
|
| 247 |
+
" print(f'Processing artifact: {artifact.name}')\n",
|
| 248 |
+
" version = int(artifact.version[1:])\n",
|
| 249 |
+
" results = []\n",
|
| 250 |
+
" columns = ['Caption', 'Theme'] + [f'Image {i+1}' for i in range(top_k)] + [f'Score {i+1}' for i in range(top_k)]\n",
|
| 251 |
+
" \n",
|
| 252 |
+
" if latest_only:\n",
|
| 253 |
+
" assert last_inference_version is None or version > last_inference_version\n",
|
| 254 |
+
" else:\n",
|
| 255 |
+
" if last_inference_version is None:\n",
|
| 256 |
+
" # we should start from v0\n",
|
| 257 |
+
" assert version == 0\n",
|
| 258 |
+
" elif version <= last_inference_version:\n",
|
| 259 |
+
" print(f'v{version} has already been logged (versions logged up to v{last_inference_version}')\n",
|
| 260 |
+
" else:\n",
|
| 261 |
+
" # check we are logging the correct version\n",
|
| 262 |
+
" assert version == last_inference_version + 1\n",
|
| 263 |
+
"\n",
|
| 264 |
+
" # start/resume corresponding run\n",
|
| 265 |
+
" if run is None:\n",
|
| 266 |
+
" run = wandb.init(job_type='inference', entity='dalle-mini', project='dalle-mini', config=training_config, id=f'inference-{run_id}', resume='allow')\n",
|
| 267 |
+
"\n",
|
| 268 |
+
" # work in temporary directory\n",
|
| 269 |
+
" with tempfile.TemporaryDirectory() as tmp:\n",
|
| 270 |
+
"\n",
|
| 271 |
+
" # download model files\n",
|
| 272 |
+
" artifact = run.use_artifact(artifact)\n",
|
| 273 |
+
" for f in model_files:\n",
|
| 274 |
+
" artifact.get_path(f).download(tmp)\n",
|
| 275 |
+
"\n",
|
| 276 |
+
" # load tokenizer and model\n",
|
| 277 |
+
" tokenizer = BartTokenizer.from_pretrained(tmp)\n",
|
| 278 |
+
" model = CustomFlaxBartForConditionalGeneration.from_pretrained(tmp)\n",
|
| 279 |
+
" model_params = replicate(model.params)\n",
|
| 280 |
+
"\n",
|
| 281 |
+
" # pmap model function needs to happen only once per model config\n",
|
| 282 |
+
" if p_generate is None:\n",
|
| 283 |
+
" p_generate = pmap_model_function(model)\n",
|
| 284 |
+
"\n",
|
| 285 |
+
" # process one batch of captions\n",
|
| 286 |
+
" for batch in tqdm(samples):\n",
|
| 287 |
+
" prompts = [x['Caption'] for x in batch]\n",
|
| 288 |
+
" processed_prompts = [text_normalizer(x) for x in prompts] if normalize_text else prompts\n",
|
| 289 |
+
"\n",
|
| 290 |
+
" # repeat the prompts to distribute over each device and tokenize\n",
|
| 291 |
+
" processed_prompts = processed_prompts * jax.device_count()\n",
|
| 292 |
+
" tokenized_prompt = tokenizer(processed_prompts, return_tensors='jax', padding='max_length', truncation=True, max_length=128).data\n",
|
| 293 |
+
" tokenized_prompt = shard(tokenized_prompt)\n",
|
| 294 |
+
"\n",
|
| 295 |
+
" # generate images\n",
|
| 296 |
+
" print('Generating images')\n",
|
| 297 |
+
" images = []\n",
|
| 298 |
+
" for i in tqdm(range(num_images // jax.device_count())):\n",
|
| 299 |
+
" key, subkey = jax.random.split(key)\n",
|
| 300 |
+
" encoded_images = p_generate(tokenized_prompt, shard_prng_key(subkey), model_params)\n",
|
| 301 |
+
" encoded_images = encoded_images.sequences[..., 1:]\n",
|
| 302 |
+
" decoded_images = p_decode(encoded_images, vqgan_params)\n",
|
| 303 |
+
" decoded_images = decoded_images.clip(0., 1.).reshape((-1, 256, 256, 3))\n",
|
| 304 |
+
" for img in decoded_images:\n",
|
| 305 |
+
" images.append(Image.fromarray(np.asarray(img * 255, dtype=np.uint8)))\n",
|
| 306 |
+
"\n",
|
| 307 |
+
" # get clip scores\n",
|
| 308 |
+
" print('Calculating CLIP scores')\n",
|
| 309 |
+
" clip_inputs = processor(text=prompts, images=images, return_tensors='np', padding='max_length', max_length=77, truncation=True).data\n",
|
| 310 |
+
" # each shard will have one prompt, images need to be reorganized to be associated to the correct shard\n",
|
| 311 |
+
" images_per_prompt_indices = np.asarray(range(0, len(images), batch_size))\n",
|
| 312 |
+
" clip_inputs['pixel_values'] = jnp.concatenate(list(clip_inputs['pixel_values'][images_per_prompt_indices + i] for i in range(batch_size)))\n",
|
| 313 |
+
" clip_inputs = shard(clip_inputs)\n",
|
| 314 |
+
" logits = p_clip(clip_inputs)\n",
|
| 315 |
+
" logits = logits.reshape(-1, num_images)\n",
|
| 316 |
+
" top_scores = logits.argsort()[:, -top_k:][..., ::-1]\n",
|
| 317 |
+
" logits = jax.device_get(logits)\n",
|
| 318 |
+
"\n",
|
| 319 |
+
" # add to results table\n",
|
| 320 |
+
" for i, (idx, scores, sample) in enumerate(zip(top_scores, logits, batch)):\n",
|
| 321 |
+
" if sample['Caption'] == padding_item: continue\n",
|
| 322 |
+
" cur_images = [images[x] for x in images_per_prompt_indices + i]\n",
|
| 323 |
+
" top_images = [wandb.Image(cur_images[x]) for x in idx]\n",
|
| 324 |
+
" top_scores = [scores[x] for x in idx]\n",
|
| 325 |
+
" results.append([sample['Caption'], sample['Theme']] + top_images + top_scores)\n",
|
| 326 |
+
"\n",
|
| 327 |
+
" # log results\n",
|
| 328 |
+
" table = wandb.Table(columns=columns, data=results)\n",
|
| 329 |
+
" run.log({'Samples': table, 'version': version})\n",
|
| 330 |
+
" wandb.finish()"
|
| 331 |
]
|
| 332 |
},
|
| 333 |
{
|
|
|
|
| 337 |
"metadata": {},
|
| 338 |
"outputs": [],
|
| 339 |
"source": [
|
| 340 |
+
"# TODO: not implemented\n",
|
| 341 |
"def log_runs(runs):\n",
|
| 342 |
" for run in tqdm(runs):\n",
|
| 343 |
" log_run(run)"
|
| 344 |
]
|
| 345 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 346 |
{
|
| 347 |
"cell_type": "code",
|
| 348 |
"execution_count": null,
|