File size: 60,437 Bytes
d6e8aa8 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 |
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true,
"id": "2s48Vmoo9EB5"
},
"outputs": [],
"source": [
"!pip install -q torchmetrics sacrebleu x-transformers"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "Lz8buKsjvA_w"
},
"source": [
"## CONFIG"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "df355sdDrNSb"
},
"outputs": [],
"source": [
"!pip install -q torchmetrics sacrebleu x-transformers\n",
"\n",
"## CONFIG\n",
"\n",
"# --- Data & Task Size ---\n",
"MAX_LENGTH = 128\n",
"\n",
"MODEL_CHOICE = \"Name_Your_Model\" # Renamed for clarity\n",
"\n",
"# --- Model Architecture Config ---\n",
"D_MODEL = 512\n",
"NUM_HEADS = 8\n",
"D_FF = 2048\n",
"DROPOUT = 0.1\n",
"\n",
"# --- Layer counts ---\n",
"NUM_ENCODER_LAYERS = 7\n",
"NUM_DECODER_LAYERS = 6\n",
"\n",
"# --- Training Config (ADJUSTED FOR FAIR COMPARISON) ---\n",
"\n",
"TARGET_TRAINING_STEPS = 100000\n",
"GRAD_ACCUMULATION_STEPS = 2\n",
"\n",
"\n",
"VALIDATION_SCHEDULE = [\n",
" 2000, 4000, 5000, 7500, 10000, 15000, 20000,\n",
" 25000, 30000, 35000, 42500, 50000, 57500, 65000, 72500, 90000, 100000\n",
"]\n",
"PEAK_LEARNING_RATE = 6e-4\n",
"WARMUP_STEPS = 600 # Warmup can stay similar or scale slightly, 600 is fine\n",
"WEIGHT_DECAY = 0.01\n",
"\n",
"# --- Regularization Config ---\n",
"LABEL_SMOOTHING_EPSILON = 0.1\n",
"\n",
"# --- Other Constants ---\n",
"DRIVE_BASE_PATH = \"/content/drive/MyDrive/AIAYN\"\n",
"ORIGINAL_BUCKETED_REPO_ID = \"prism-lab/wmt14-de-en-bucketed-w4\" # Use the bucketed one (we will ignore buckets)\n",
"MODEL_CHECKPOINT = \"Helsinki-NLP/opus-mt-de-en\""
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "W5l1HHRFXxPA"
},
"source": [
"## DATALOADERS"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true,
"id": "FA5SqFzeMrpK"
},
"outputs": [],
"source": [
"import torch\n",
"import torch.nn as nn\n",
"from torch.utils.data import DataLoader\n",
"from transformers import AutoTokenizer\n",
"from datasets import load_dataset\n",
"import math\n",
"import os\n",
"from tqdm.auto import tqdm\n",
"from torch.utils.tensorboard import SummaryWriter\n",
"import random\n",
"import numpy as np\n",
"import torch\n",
"from transformers import get_cosine_schedule_with_warmup\n",
"from typing import List\n",
"from transformers import AutoModel\n",
"from transformers import DataCollatorForSeq2Seq\n",
"\n",
"\n",
"def set_seed(seed_value=5):\n",
" \"\"\"Sets the seed for reproducibility.\"\"\"\n",
" random.seed(seed_value)\n",
" np.random.seed(seed_value)\n",
" torch.manual_seed(seed_value)\n",
" torch.cuda.manual_seed_all(seed_value)\n",
" torch.backends.cudnn.deterministic = True\n",
" torch.backends.cudnn.benchmark = False\n",
"\n",
"SEED = 117\n",
"set_seed(SEED)\n",
"print(f\"Reproducibility seed set to {SEED}\")\n",
"os.environ[\"CUBLAS_WORKSPACE_CONFIG\"] = \":4096:8\"\n",
"\n",
"#torch.use_deterministic_algorithms(True)\n",
"\n",
"print(\"--- Loading Modernized Configuration ---\")\n",
"def seed_worker(worker_id):\n",
" worker_seed = torch.initial_seed() % 2**32\n",
" np.random.seed(worker_seed)\n",
" random.seed(worker_seed)\n",
"\n",
"torch.set_float32_matmul_precision('high')\n",
"print(\"✅ PyTorch matmul precision set to 'high'\")\n",
"\n",
"# --- Device Setup ---\n",
"device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
"print(f\"Using device: {device}\")\n",
"\n",
"tokenizer = AutoTokenizer.from_pretrained(MODEL_CHECKPOINT)\n",
"\n",
"VOCAB_SIZE = len(tokenizer)\n",
"print(f\"Vocab size: {VOCAB_SIZE}\")\n",
"\n",
"\n",
"# DATA LOADING & PREPARATION\n",
"\n",
"# --- 1. DEFINE THE FNET COLLATOR (FORCE FIXED LENGTH) ---\n",
"# This is crucial. It forces every sentence to be exactly 128 tokens.\n",
"fnet_collator = DataCollatorForSeq2Seq(\n",
" tokenizer=tokenizer,\n",
" padding=\"max_length\", # <--- FORCE PADDING\n",
" max_length=MAX_LENGTH, # <--- 128 (defined in your config)\n",
" pad_to_multiple_of=None\n",
")\n",
"\n",
"# --- 2. LOAD DATASET ---\n",
"print(f\"Loading original bucketed samples from: {ORIGINAL_BUCKETED_REPO_ID}\")\n",
"original_datasets = load_dataset(ORIGINAL_BUCKETED_REPO_ID)\n",
"\n",
"# --- 3. CREATE DATALOADERS (STANDARD FIXED SIZE) ---\n",
"FNET_PHYSICAL_BATCH_SIZE = 320\n",
"\n",
"g = torch.Generator()\n",
"g.manual_seed(SEED)\n",
"\n",
"train_dataloader = DataLoader(\n",
" original_datasets[\"train\"],\n",
" batch_size=FNET_PHYSICAL_BATCH_SIZE, # <--- FIXED BATCH SIZE (Safe from OOM)\n",
" shuffle=True, # <--- GLOBAL SHUFFLE\n",
" num_workers=8,\n",
" collate_fn=fnet_collator,\n",
" pin_memory=True,\n",
" worker_init_fn=seed_worker,\n",
" generator=g,\n",
")\n",
"\n",
"val_dataloader = DataLoader(\n",
" original_datasets[\"validation\"],\n",
" batch_size=FNET_PHYSICAL_BATCH_SIZE,\n",
" collate_fn=fnet_collator,\n",
" num_workers=8,\n",
" pin_memory=True,\n",
" worker_init_fn=seed_worker,\n",
" generator=g,\n",
")\n",
"\n",
"print(f\"Train Dataloader is now a STANDARD iterator.\")\n",
"print(f\"Physical Batch Size: {FNET_PHYSICAL_BATCH_SIZE}\")\n",
"print(f\"Gradient Accumulation: {GRAD_ACCUMULATION_STEPS}\")\n",
"print(f\"Effective Batch Size: {FNET_PHYSICAL_BATCH_SIZE * GRAD_ACCUMULATION_STEPS}\")\n",
"\n",
"# --- SANITY CHECK ---\n",
"print(\"\\n--- Running Sanity Check on new FNet DataLoader ---\")\n",
"train_dataloader.generator.manual_seed(SEED)\n",
"temp_iterator = iter(train_dataloader)\n",
"print(\"Shapes of first 3 batches (Should all be [64, 128]):\")\n",
"for i in range(3):\n",
" batch = next(temp_iterator)\n",
" print(f\" Batch {i+1}: input_ids shape = {batch['input_ids'].shape}\")\n",
"print(\"--- Sanity Check Complete ---\\n\")\n",
"# --- VERIFY SHUFFLE IS WORKING ---\n",
"print(\"🕵️ INSPECTING ONE BATCH 🕵️\")\n",
"\n",
"# Get one batch from your active train_dataloader\n",
"batch = next(iter(train_dataloader))\n",
"input_ids = batch['input_ids']\n",
"\n",
"# Calculate real lengths (ignoring padding)\n",
"# We count how many tokens are NOT the pad token (usually 0 or 58100)\n",
"real_lengths = (input_ids != tokenizer.pad_token_id).sum(dim=1)\n",
"\n",
"print(f\"Batch Shape: {input_ids.shape}\")\n",
"print(\"Random Sample of 20 lengths in this batch:\")\n",
"print(real_lengths[:20].tolist())\n",
"\n",
"# Check diversity\n",
"if real_lengths.float().std() < 5:\n",
" print(\"\\n⚠️ WARNING: LENGTHS LOOK CLUSTERED! (Bad shuffling)\")\n",
"else:\n",
" print(f\"\\n✅ PASSED: Lengths are highly variable (Std Dev: {real_lengths.float().std():.2f}). Shuffling is working.\")"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "cS4JvJGRhClv"
},
"source": [
"## Models"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "SMhlM0YvO1A7"
},
"outputs": [],
"source": [
"import torch\n",
"import torch.nn as nn\n",
"import torch.nn.functional as F\n",
"import math\n",
"from x_transformers import Encoder, Decoder\n",
"\n",
"class RoPETransformer(nn.Module):\n",
" def __init__(self, num_encoder_layers, num_decoder_layers, num_heads, d_model, dff, vocab_size, max_length, dropout):\n",
" super().__init__()\n",
" self.d_model = d_model\n",
" self.embedding = nn.Embedding(vocab_size, d_model)\n",
"\n",
" # We REMOVE self.pos_encoder (RoPE handles position internally)\n",
" self.dropout_layer = nn.Dropout(dropout)\n",
"\n",
" # --- x-transformers Encoder ---\n",
" self.encoder = Encoder(\n",
" dim = d_model,\n",
" depth = num_encoder_layers,\n",
" heads = num_heads,\n",
" attn_dim_head = d_model // num_heads,\n",
" ff_mult = dff / d_model,\n",
" rotary_pos_emb = True,\n",
" attn_flash = True,\n",
" attn_dropout = dropout,\n",
" ff_dropout = dropout,\n",
" use_rmsnorm = True\n",
" )\n",
"\n",
" # --- x-transformers Decoder ---\n",
" self.decoder = Decoder(\n",
" dim = d_model,\n",
" depth = num_decoder_layers,\n",
" heads = num_heads,\n",
" attn_dim_head = d_model // num_heads,\n",
" ff_mult = dff / d_model,\n",
" rotary_pos_emb = True,\n",
" cross_attend = True,\n",
" attn_flash = True,\n",
" attn_dropout = dropout,\n",
" ff_dropout = dropout,\n",
" use_rmsnorm = True\n",
" )\n",
"\n",
" self.final_linear = nn.Linear(d_model, vocab_size)\n",
" self.final_linear.weight = self.embedding.weight\n",
"\n",
" def forward(self, src, tgt, src_padding_mask, tgt_padding_mask, memory_key_padding_mask, tgt_mask):\n",
" # 1. Embeddings (No Absolute Positional Encoding added!)\n",
" src_emb = self.embedding(src) * math.sqrt(self.d_model)\n",
" src_emb = self.dropout_layer(src_emb)\n",
"\n",
" tgt_emb = self.embedding(tgt) * math.sqrt(self.d_model)\n",
" tgt_emb = self.dropout_layer(tgt_emb)\n",
"\n",
" # 2. Mask Conversion\n",
" # User provides True=PAD. x-transformers wants True=KEEP.\n",
" # We invert the boolean mask using ~\n",
" enc_mask = ~src_padding_mask if src_padding_mask is not None else None\n",
" dec_mask = ~tgt_padding_mask if tgt_padding_mask is not None else None\n",
"\n",
" # Note: 'tgt_mask' (causal mask) is handled automatically by x-transformers Decoder!\n",
" # We do NOT pass the square causal mask manually.\n",
"\n",
" # 3. Encoder\n",
" # x-transformers takes embeddings directly\n",
" memory = self.encoder(src_emb, mask=enc_mask)\n",
"\n",
" # 4. Decoder\n",
" # context = memory (from encoder)\n",
" # context_mask = mask for memory (encoder mask)\n",
" decoder_output = self.decoder(\n",
" tgt_emb,\n",
" context=memory,\n",
" mask=dec_mask,\n",
" context_mask=enc_mask\n",
" )\n",
"\n",
" return self.final_linear(decoder_output)\n",
"\n",
" # Keep your existing create_masks (used for Data Processing mostly)\n",
" def create_masks(self, src, tgt):\n",
" src_padding_mask = (src == tokenizer.pad_token_id)\n",
" tgt_padding_mask = (tgt == tokenizer.pad_token_id)\n",
" # We still generate this for compatibility, though x-transformers handles causality internally\n",
" tgt_mask = nn.Transformer.generate_square_subsequent_mask(\n",
" sz=tgt.size(1), device=src.device, dtype=torch.bool\n",
" )\n",
" return src_padding_mask, tgt_padding_mask, src_padding_mask, tgt_mask\n",
"\n",
" @torch.no_grad()\n",
" def generate(self, src: torch.Tensor, max_length: int, num_beams: int = 5) -> torch.Tensor:\n",
" self.eval()\n",
" # Create Mask (True=PAD)\n",
" src_padding_mask = (src == tokenizer.pad_token_id)\n",
" # Invert for x-transformers (True=KEEP)\n",
" enc_mask = ~src_padding_mask\n",
"\n",
" # Encode\n",
" src_emb = self.embedding(src) * math.sqrt(self.d_model)\n",
" # No Pos Encoder\n",
" memory = self.encoder(self.dropout_layer(src_emb), mask=enc_mask)\n",
"\n",
" batch_size = src.shape[0]\n",
" # Expand for beams\n",
" memory = memory.repeat_interleave(num_beams, dim=0)\n",
" enc_mask = enc_mask.repeat_interleave(num_beams, dim=0)\n",
"\n",
" initial_token = tokenizer.pad_token_id\n",
" beams = torch.full((batch_size * num_beams, 1), initial_token, dtype=torch.long, device=src.device)\n",
" beam_scores = torch.zeros(batch_size * num_beams, device=src.device)\n",
" finished_beams = torch.zeros(batch_size * num_beams, dtype=torch.bool, device=src.device)\n",
"\n",
" for _ in range(max_length - 1):\n",
" if finished_beams.all(): break\n",
"\n",
" # Embed beams\n",
" tgt_emb = self.embedding(beams) * math.sqrt(self.d_model)\n",
" # No Pos Encoder\n",
"\n",
" # Decode\n",
" # x-transformers automatically handles the causal masking for the sequence length of tgt_emb\n",
" decoder_output = self.decoder(\n",
" self.dropout_layer(tgt_emb),\n",
" context=memory,\n",
" context_mask=enc_mask\n",
" )\n",
"\n",
" logits = self.final_linear(decoder_output[:, -1, :])\n",
" log_probs = F.log_softmax(logits, dim=-1)\n",
"\n",
" # ... (Rest of your Beam Search Logic remains identical) ...\n",
" log_probs[:, tokenizer.pad_token_id] = -torch.inf\n",
" if finished_beams.any(): log_probs[finished_beams, tokenizer.eos_token_id] = 0\n",
"\n",
" total_scores = beam_scores.unsqueeze(1) + log_probs\n",
" if _ == 0:\n",
" total_scores = total_scores.view(batch_size, num_beams, -1)\n",
" total_scores[:, 1:, :] = -torch.inf\n",
" total_scores = total_scores.view(batch_size * num_beams, -1)\n",
" else:\n",
" total_scores = beam_scores.unsqueeze(1) + log_probs\n",
"\n",
" total_scores = total_scores.view(batch_size, -1)\n",
" top_scores, top_indices = torch.topk(total_scores, k=num_beams, dim=1)\n",
"\n",
" beam_indices = top_indices // log_probs.shape[-1]\n",
" token_indices = top_indices % log_probs.shape[-1]\n",
"\n",
" batch_indices = torch.arange(batch_size, device=src.device).unsqueeze(1)\n",
" effective_indices = (batch_indices * num_beams + beam_indices).view(-1)\n",
"\n",
" beams = beams[effective_indices]\n",
" beams = torch.cat([beams, token_indices.view(-1, 1)], dim=1)\n",
" beam_scores = top_scores.view(-1)\n",
" finished_beams = finished_beams | (beams[:, -1] == tokenizer.eos_token_id)\n",
"\n",
" final_beams = beams.view(batch_size, num_beams, -1)\n",
" final_scores = beam_scores.view(batch_size, num_beams)\n",
" normalized_scores = final_scores / (final_beams != tokenizer.pad_token_id).sum(-1).float().clamp(min=1)\n",
" best_beams = final_beams[torch.arange(batch_size), normalized_scores.argmax(1), :]\n",
" self.train()\n",
" return best_beams\n",
"\n",
"class RMSNorm(nn.Module):\n",
" def __init__(self, dim, eps=1e-8):\n",
" super().__init__()\n",
" self.eps = eps\n",
" self.gamma = nn.Parameter(torch.ones(dim))\n",
"\n",
" def forward(self, x):\n",
" # 1. Calculate the mean of the squares\n",
" mean_square = x.pow(2).mean(dim=-1, keepdim=True)\n",
"\n",
" # 2. Calculate the inverse square root (1 / RMS)\n",
" # We add eps before the sqrt for stability\n",
" inv_rms = torch.rsqrt(mean_square + self.eps)\n",
"\n",
" # 3. Normalize and scale\n",
" return x * inv_rms * self.gamma\n",
"\n",
"\n",
"class FNetBlock(nn.Module):\n",
" def __init__(self, d_model, d_ff, dropout):\n",
" super().__init__()\n",
" self.norm_mix = nn.LayerNorm(d_model) # LayerNorm is safer for FNet than RMSNorm\n",
" self.norm_ff = nn.LayerNorm(d_model)\n",
"\n",
" self.ff = nn.Sequential(\n",
" nn.Linear(d_model, d_ff),\n",
" nn.GELU(),\n",
" nn.Dropout(dropout),\n",
" nn.Linear(d_ff, d_model),\n",
" nn.Dropout(dropout)\n",
" )\n",
"\n",
" def forward(self, x):\n",
" # 1. Fourier Mixing Branch\n",
" residual = x\n",
" x = self.norm_mix(x)\n",
"\n",
" # --- THE FIX ---\n",
" with torch.cuda.amp.autocast(enabled=False):\n",
" x = x.float()\n",
" # norm='ortho' makes the FFT energy-preserving.\n",
" # Output magnitude will match input magnitude (~1).\n",
" x = torch.fft.fftn(x, dim=(-2, -1), norm='ortho').real\n",
" x = x.to(dtype=residual.dtype)\n",
" # ---------------\n",
"\n",
" # Now 'x' and 'residual' have roughly same magnitude.\n",
" # The skip connection works again.\n",
" x = x + residual\n",
"\n",
" # 2. Feed Forward Branch\n",
" residual = x\n",
" x = self.norm_ff(x)\n",
" x = self.ff(x)\n",
" return x + residual\n",
"\n",
"\n",
"class FNetEncoder(nn.Module):\n",
" def __init__(self, depth, d_model, d_ff, dropout):\n",
" super().__init__()\n",
" self.layers = nn.ModuleList([\n",
" FNetBlock(d_model, d_ff, dropout) for _ in range(depth)\n",
" ])\n",
" # [FIX] Use LayerNorm here to match the blocks\n",
" self.norm_out = nn.LayerNorm(d_model)\n",
"\n",
" def forward(self, x):\n",
" for layer in self.layers:\n",
" x = layer(x)\n",
" return self.norm_out(x)\n",
"\n",
"# --- Main Hybrid Model ---\n",
"\n",
"class FNetHybridTransformer(nn.Module):\n",
" def __init__(self, num_encoder_layers, num_decoder_layers, num_heads, d_model, dff, vocab_size, max_length, dropout):\n",
" super().__init__()\n",
" self.d_model = d_model\n",
"\n",
" # Shared Embeddings\n",
" # padding_idx=tokenizer.pad_token_id forces the vector at this index to be strict ZEROS.\n",
" # It does not have gradients, it stays zero forever.\n",
" self.embedding = nn.Embedding(vocab_size, d_model, padding_idx=tokenizer.pad_token_id)\n",
"\n",
" # FNet REQUIRES Absolute Positional Embeddings because FFT mixes information\n",
" # but doesn't inherently understand sequence order like RoPE/RNNs do initially.\n",
" self.pos_embedding = nn.Embedding(max_length, d_model)\n",
"\n",
" self.dropout_layer = nn.Dropout(dropout)\n",
"\n",
" # --- Custom FNet Encoder ---\n",
" self.encoder = FNetEncoder(\n",
" depth=num_encoder_layers,\n",
" d_model=d_model,\n",
" d_ff=dff,\n",
" dropout=dropout\n",
" )\n",
"\n",
" # --- x-transformers Decoder (Retains RoPE) ---\n",
" self.decoder = Decoder(\n",
" dim=d_model,\n",
" depth=num_decoder_layers,\n",
" heads=num_heads,\n",
" attn_dim_head=d_model // num_heads,\n",
" ff_mult=dff / d_model,\n",
" rotary_pos_emb=True, # Decoder still uses RoPE\n",
" cross_attend=True,\n",
" attn_flash=True,\n",
" attn_dropout=dropout,\n",
" ff_dropout=dropout,\n",
" use_rmsnorm=True\n",
" )\n",
"\n",
" self.final_linear = nn.Linear(d_model, vocab_size)\n",
" self.final_linear.weight = self.embedding.weight\n",
"\n",
" def forward(self, src, tgt, src_padding_mask, tgt_padding_mask, memory_key_padding_mask, tgt_mask):\n",
" # 1. Embeddings\n",
" # Source (Encoder) gets Absolute Positional Embeddings\n",
" B, L_src = src.shape\n",
" pos_ids = torch.arange(L_src, device=src.device).unsqueeze(0)\n",
" src_emb = self.embedding(src) * math.sqrt(self.d_model)\n",
" src_emb = src_emb + self.pos_embedding(pos_ids)\n",
" src_emb = self.dropout_layer(src_emb)\n",
"\n",
" # Target (Decoder) gets NO Positional Embeddings here (RoPE handles it inside Decoder)\n",
" tgt_emb = self.embedding(tgt) * math.sqrt(self.d_model)\n",
" tgt_emb = self.dropout_layer(tgt_emb)\n",
"\n",
" # 2. Prepare Masks\n",
" # x-transformers requires True = Keep, False = Mask\n",
" # Your dataloader provides True = Pad\n",
" enc_mask = ~src_padding_mask if src_padding_mask is not None else None\n",
" dec_mask = ~tgt_padding_mask if tgt_padding_mask is not None else None\n",
"\n",
" # 3. FNet Encoder\n",
" # Note: FNet mixes ALL tokens (including padding).\n",
" memory = self.encoder(src_emb)\n",
"\n",
" # CRITICAL: Zero out padding positions in encoder output so Decoder doesn't attend to them.\n",
" if src_padding_mask is not None:\n",
" memory = memory.masked_fill(src_padding_mask.unsqueeze(-1), 0.0)\n",
"\n",
" # 4. RoPE Decoder\n",
" # The decoder uses RoPE for self-attention on 'tgt',\n",
" # and standard cross-attention to 'memory' (FNet output).\n",
" decoder_output = self.decoder(\n",
" tgt_emb,\n",
" context=memory,\n",
" mask=dec_mask,\n",
" context_mask=enc_mask\n",
" )\n",
"\n",
" return self.final_linear(decoder_output)\n",
"\n",
" def create_masks(self, src, tgt):\n",
" # Standard mask creation (Same as your original)\n",
" src_padding_mask = (src == tokenizer.pad_token_id)\n",
" tgt_padding_mask = (tgt == tokenizer.pad_token_id)\n",
" tgt_mask = nn.Transformer.generate_square_subsequent_mask(\n",
" sz=tgt.size(1), device=src.device, dtype=torch.bool\n",
" )\n",
" return src_padding_mask, tgt_padding_mask, src_padding_mask, tgt_mask\n",
"\n",
" @torch.no_grad()\n",
" def generate(self, src: torch.Tensor, max_length: int, num_beams: int = 5) -> torch.Tensor:\n",
" self.eval()\n",
" B, L_src = src.shape\n",
"\n",
" # 1. Encode with FNet\n",
" pos_ids = torch.arange(L_src, device=src.device).unsqueeze(0)\n",
" src_emb = self.embedding(src) * math.sqrt(self.d_model)\n",
" src_emb = src_emb + self.pos_embedding(pos_ids)\n",
"\n",
" memory = self.encoder(self.dropout_layer(src_emb))\n",
"\n",
" # Masking padding in memory\n",
" src_padding_mask = (src == tokenizer.pad_token_id)\n",
" memory = memory.masked_fill(src_padding_mask.unsqueeze(-1), 0.0)\n",
"\n",
" # Prepare for Decoder (x-transformers style mask: True=Keep)\n",
" enc_mask = ~src_padding_mask\n",
"\n",
" # --- BEAM SEARCH SETUP ---\n",
" # Expand memory for beams\n",
" memory = memory.repeat_interleave(num_beams, dim=0)\n",
" enc_mask = enc_mask.repeat_interleave(num_beams, dim=0)\n",
"\n",
" initial_token = tokenizer.pad_token_id\n",
" beams = torch.full((B * num_beams, 1), initial_token, dtype=torch.long, device=src.device)\n",
" beam_scores = torch.zeros(B * num_beams, device=src.device)\n",
" finished_beams = torch.zeros(B * num_beams, dtype=torch.bool, device=src.device)\n",
"\n",
" for _ in range(max_length - 1):\n",
" if finished_beams.all(): break\n",
"\n",
" # Decoder Step (RoPE handled internally)\n",
" tgt_emb = self.embedding(beams) * math.sqrt(self.d_model)\n",
"\n",
" decoder_output = self.decoder(\n",
" self.dropout_layer(tgt_emb),\n",
" context=memory,\n",
" context_mask=enc_mask\n",
" )\n",
"\n",
" logits = self.final_linear(decoder_output[:, -1, :])\n",
" log_probs = F.log_softmax(logits, dim=-1)\n",
"\n",
" # --- STANDARD BEAM LOGIC (No changes needed here) ---\n",
" log_probs[:, tokenizer.pad_token_id] = -torch.inf\n",
" if finished_beams.any(): log_probs[finished_beams, tokenizer.eos_token_id] = 0\n",
"\n",
" total_scores = beam_scores.unsqueeze(1) + log_probs\n",
" if _ == 0:\n",
" total_scores = total_scores.view(B, num_beams, -1)\n",
" total_scores[:, 1:, :] = -torch.inf\n",
" total_scores = total_scores.view(B * num_beams, -1)\n",
" else:\n",
" total_scores = beam_scores.unsqueeze(1) + log_probs\n",
"\n",
" total_scores = total_scores.view(B, -1)\n",
" top_scores, top_indices = torch.topk(total_scores, k=num_beams, dim=1)\n",
"\n",
" beam_indices = top_indices // log_probs.shape[-1]\n",
" token_indices = top_indices % log_probs.shape[-1]\n",
"\n",
" batch_indices = torch.arange(B, device=src.device).unsqueeze(1)\n",
" effective_indices = (batch_indices * num_beams + beam_indices).view(-1)\n",
"\n",
" beams = beams[effective_indices]\n",
" beams = torch.cat([beams, token_indices.view(-1, 1)], dim=1)\n",
" beam_scores = top_scores.view(-1)\n",
" finished_beams = finished_beams | (beams[:, -1] == tokenizer.eos_token_id)\n",
"\n",
" final_beams = beams.view(B, num_beams, -1)\n",
" final_scores = beam_scores.view(B, num_beams)\n",
" normalized_scores = final_scores / (final_beams != tokenizer.pad_token_id).sum(-1).float().clamp(min=1)\n",
" best_beams = final_beams[torch.arange(B), normalized_scores.argmax(1), :]\n",
" self.train()\n",
" return best_beams"
]
},
{
"cell_type": "code",
"source": [
"def count_parameters(model):\n",
" table_data = []\n",
" total_params = 0\n",
" trainable_params = 0\n",
"\n",
" # 1. Global Counts\n",
" for p in model.parameters():\n",
" total_params += p.numel()\n",
" if p.requires_grad:\n",
" trainable_params += p.numel()\n",
"\n",
" print(\"=\"*40)\n",
" print(f\"📊 MODEL STATISTICS\")\n",
" print(\"=\"*40)\n",
" print(f\"Total Parameters: {total_params:,} ({total_params/1e6:.2f}M)\")\n",
" print(f\"Trainable Parameters: {trainable_params:,} ({trainable_params/1e6:.2f}M)\")\n",
" print(\"-\" * 40)\n",
"\n",
" # 2. Section Breakdown\n",
" def get_params(module):\n",
" return sum(p.numel() for p in module.parameters())\n",
"\n",
" if hasattr(model, 'encoder'):\n",
" enc_p = get_params(model.encoder)\n",
" print(f\" • Encoder (FNet): {enc_p:,} ({enc_p/1e6:.2f}M)\")\n",
"\n",
" if hasattr(model, 'decoder'):\n",
" dec_p = get_params(model.decoder)\n",
" print(f\" • Decoder (RoPE): {dec_p:,} ({dec_p/1e6:.2f}M)\")\n",
"\n",
" if hasattr(model, 'embedding'):\n",
" emb_p = get_params(model.embedding)\n",
" print(f\" • Embeddings: {emb_p:,} ({emb_p/1e6:.2f}M)\")\n",
"\n",
" print(\"=\"*40)\n",
"\n"
],
"metadata": {
"id": "wpmz-H9Slko1"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "Zd3AFTmhrCJq"
},
"source": [
"## Functions (Loss, Eval etc)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "Te1qTyUKrDEd"
},
"outputs": [],
"source": [
"\n",
"translation_loss_fn = nn.CrossEntropyLoss(\n",
" ignore_index=-100, # We don't calculate loss for pad tokens. Pad tokens are replaced with -100 by DataCollatorForSeq2Seq.\n",
" label_smoothing=LABEL_SMOOTHING_EPSILON\n",
")\n",
"def calculate_combined_loss(model_outputs, target_labels):\n",
" \"\"\"Calculates the loss based on the model's output structure.\"\"\"\n",
" logits = model_outputs\n",
" translation_loss = translation_loss_fn(logits.reshape(-1, logits.shape[-1]), target_labels.reshape(-1))\n",
" loss_dict = {'total': translation_loss.item()}\n",
" return translation_loss, loss_dict\n",
"\n",
"from torchmetrics.text import SacreBLEUScore\n",
"\n",
"def evaluate(model, dataloader, device):\n",
" # Use SacreBLEUScore (defaults to '13a' tokenizer, the WMT standard)\n",
" metric = SacreBLEUScore().to(device)\n",
"\n",
" model.eval()\n",
"\n",
" # Use no_grad to save memory and speed up validation\n",
" with torch.no_grad():\n",
" for batch in tqdm(dataloader, desc=\"Evaluating\", leave=False):\n",
" input_ids = batch['input_ids'].to(device)\n",
" labels = batch['labels']\n",
"\n",
" # Generate predictions\n",
" generated_ids = model.generate(input_ids, max_length=MAX_LENGTH, num_beams=5)\n",
"\n",
" # Decode predictions\n",
" pred_texts = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)\n",
"\n",
" # Decode labels (Fixing -100 padding)\n",
" labels[labels == -100] = tokenizer.pad_token_id\n",
" ref_texts = tokenizer.batch_decode(labels, skip_special_tokens=True)\n",
"\n",
" # Update Metric\n",
" # SacreBLEU expects references as a list of lists: [[ref1], [ref2], ...]\n",
" formatted_refs = [[ref] for ref in ref_texts]\n",
" metric.update(pred_texts, formatted_refs)\n",
"\n",
" model.train()\n",
"\n",
" # Compute returns a tensor, .item() converts it to a standard python float\n",
" return metric.compute().item()\n",
"\n",
"\n",
"\n",
"## WARNING! THIS CAN'T BE USED FOR FNET\n",
"def generate_sample_translations(model, device, sentences_de):\n",
" \"\"\"Generates and prints sample translations using beam search.\"\"\"\n",
" print(\"\\n--- Generating Sample Translations (with Beam Search) ---\")\n",
" orig_model = getattr(model, '_orig_mod', model)\n",
" orig_model.eval()\n",
"\n",
" inputs = tokenizer(sentences_de, return_tensors=\"pt\", padding=True, truncation=True, max_length=MAX_LENGTH)\n",
" input_ids = inputs.input_ids.to(device)\n",
" generated_ids = orig_model.generate(input_ids, max_length=MAX_LENGTH, num_beams=5)\n",
"\n",
" translations = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)\n",
" for src, out in zip(sentences_de, translations):\n",
" print(f\" DE Source: {src}\")\n",
" print(f\" EN Output: {out}\")\n",
" print(\"-\" * 20)\n",
" orig_model.train()\n",
"\n",
"sample_sentences_de_for_tracking = [\n",
" \"Eine Katze sitzt auf der Matte.\",\n",
" \"Ein Mann in einem roten Hemd liest ein Buch.\",\n",
" \"Was ist die Hauptstadt von Deutschland?\",\n",
" \"Ich gehe ins Kino, weil der Film sehr gut ist.\",\n",
"]\n",
"\n",
"def init_other_linear_weights(m):\n",
" if isinstance(m, nn.Linear):\n",
" # The 'is not' check correctly skips the final_linear layer,\n",
" # leaving its weights tied to the correctly initialized embeddings.\n",
" if m is not getattr(model, '_orig_mod', model).final_linear:\n",
" nn.init.xavier_uniform_(m.weight)\n",
" if m.bias is not None:\n",
" nn.init.zeros_(m.bias)\n",
"\n",
"\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "YwPXbSwR50I2"
},
"outputs": [],
"source": [
"import json\n",
"import os\n",
"import subprocess\n",
"import torch\n",
"import hashlib\n",
"import sys\n",
"import shutil\n",
"\n",
"# This logger will be configured and used in the main training script\n",
"import logging\n",
"logger = logging.getLogger(__name__)\n",
"\n",
"\n",
"def log_to_run_specific_file(run_dir):\n",
" run_log_path = os.path.join(run_dir, \"run_log.txt\")\n",
" file_handler = logging.FileHandler(run_log_path)\n",
" file_handler.setFormatter(logging.Formatter('%(asctime)s [%(levelname)s] %(message)s'))\n",
" logger.addHandler(file_handler)\n",
" return file_handler\n",
"\n",
"def log_configurations(log_dir, config_vars):\n",
" # (Same as your provided function)\n",
" config_path = os.path.join(log_dir, \"config.json\")\n",
" try:\n",
" with open(config_path, 'w') as f:\n",
" serializable_configs = {k: v for k, v in config_vars.items() if isinstance(v, (int, float, str, bool, list, dict, type(None)))}\n",
" json.dump(serializable_configs, f, indent=4)\n",
" logger.info(f\"Configurations saved to {config_path}\")\n",
" except Exception as e:\n",
" logger.error(f\"Could not save configurations: {e}\")\n",
"\n",
"def log_environment(log_dir):\n",
" # (Same as your provided function)\n",
" env_path = os.path.join(log_dir, \"environment.txt\")\n",
" try:\n",
" with open(env_path, 'w') as f:\n",
" f.write(f\"--- Timestamp (UTC): {datetime.datetime.utcnow().isoformat()} ---\\n\")\n",
" f.write(f\"Python Version: {sys.version}\\n\")\n",
" f.write(f\"PyTorch Version: {torch.__version__}\\n\")\n",
" f.write(f\"CUDA Available: {torch.cuda.is_available()}\\n\")\n",
" if torch.cuda.is_available():\n",
" f.write(f\"CUDA Version: {torch.version.cuda}\\n\")\n",
" f.write(f\"CuDNN Version: {torch.backends.cudnn.version()}\\n\")\n",
" f.write(f\"Number of GPUs: {torch.cuda.device_count()}\\n\")\n",
" f.write(f\"GPU Name: {torch.cuda.get_device_name(0)}\\n\")\n",
" f.write(\"\\n--- Full pip freeze ---\\n\")\n",
" result = subprocess.run([sys.executable, '-m', 'pip', 'freeze'], stdout=subprocess.PIPE, text=True, check=True)\n",
" f.write(result.stdout)\n",
" logger.info(f\"Environment info saved to {env_path}\")\n",
" except Exception as e:\n",
" logger.error(f\"Could not save environment info: {e}\")\n",
"\n",
"def log_code_snapshot(log_dir, script_path):\n",
" # NOTE: In Colab, you must save your notebook as a .py file for this to work.\n",
" # For example, file -> \"Save a copy as .py\"\n",
" code_dir = os.path.join(log_dir, \"code_snapshot\")\n",
" os.makedirs(code_dir, exist_ok=True)\n",
" if script_path and os.path.exists(script_path):\n",
" try:\n",
" shutil.copy(script_path, os.path.join(code_dir, os.path.basename(script_path)))\n",
" logger.info(f\"Copied script '{script_path}' to snapshot directory for verification.\")\n",
" except Exception as e:\n",
" logger.error(f\"Could not copy script for snapshot: {e}\")\n",
" else:\n",
" logger.warning(f\"Code Snapshot: Script path '{script_path}' not found. SKIPPING.\")\n",
"\n",
"def get_file_hash(filepath):\n",
" # (Same as your provided function)\n",
" sha256_hash = hashlib.sha256()\n",
" try:\n",
" with open(filepath, \"rb\") as f:\n",
" for byte_block in iter(lambda: f.read(4096), b\"\"):\n",
" sha256_hash.update(byte_block)\n",
" return sha256_hash.hexdigest()\n",
" except Exception as e:\n",
" logger.error(f\"Could not generate hash for {filepath}: {e}\")\n",
" return None\n",
"\n",
"def create_checksum_file(run_dir, artifacts_dict):\n",
" checksum_file_path = os.path.join(run_dir, \"checksums.sha256\")\n",
" logger.info(f\"--- Creating digital fingerprints for key artifacts ---\")\n",
" with open(checksum_file_path, \"w\") as f:\n",
" f.write(f\"SHA256 Checksums for run: {os.path.basename(run_dir)}\\n\")\n",
" for name, path in artifacts_dict.items():\n",
" if path and os.path.exists(path):\n",
" file_hash = get_file_hash(path)\n",
" if file_hash:\n",
" log_message = f\" - {name} ({os.path.basename(path)}): {file_hash}\"\n",
" logger.info(log_message)\n",
" f.write(f\"{file_hash} {os.path.basename(path)}\\n\")\n",
" else:\n",
" logger.warning(f\" - Skipped hashing '{name}', file not found: {path}\")\n",
" logger.info(f\"Checksums saved to {checksum_file_path}\")\n",
"\n",
"def init_weights_kaiming(m):\n",
" \"\"\"\n",
" Applies Kaiming He initialization to Linear layers.\n",
" This is the standard, superior way to initialize deep Transformers.\n",
" NOTE: We will handle the Embedding layer separately.\n",
" \"\"\"\n",
"\n",
" if isinstance(m, nn.Linear):\n",
" nn.init.kaiming_uniform_(m.weight, a=math.sqrt(5)) # a=sqrt(5) mimics default PyTorch for LeakyReLU\n",
" if m.bias is not None:\n",
" fan_in, _ = nn.init._calculate_fan_in_and_fan_out(m.weight)\n",
" bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0\n",
" nn.init.uniform_(m.bias, -bound, bound)\n",
"\n",
"\n",
"def init_weights_fnet(m):\n",
" \"\"\"\n",
" Specific initialization for FNet Hybrid.\n",
" FNet is essentially a BERT-like encoder, so we use BERT-style initialization\n",
" (Truncated Normal or Xavier) rather than Kaiming.\n",
" \"\"\"\n",
" if isinstance(m, nn.Linear):\n",
" # Xavier (Glorot) Uniform is the standard for Transformer/FNet attention/FFN layers\n",
" nn.init.xavier_uniform_(m.weight)\n",
" if m.bias is not None:\n",
" nn.init.zeros_(m.bias)\n",
"\n",
" elif isinstance(m, nn.Embedding):\n",
" # Critical: Keep embedding variance low (0.02)\n",
" nn.init.normal_(m.weight, mean=0.0, std=0.02)\n",
"\n",
" # Handle the RMSNorms if they have learnable parameters\n",
" elif isinstance(m, (nn.LayerNorm, RMSNorm)):\n",
" if hasattr(m, 'weight') and m.weight is not None:\n",
" nn.init.ones_(m.weight)\n",
" if hasattr(m, 'bias') and m.bias is not None:\n",
" nn.init.zeros_(m.bias)\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "ijTUk5dHu494"
},
"source": [
"## Training Loop"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "pyHZ1moluyA2"
},
"outputs": [],
"source": [
"if __name__ == '__main__':\n",
"\n",
" experiment_name = f\"{MODEL_CHOICE}\"\n",
" CURRENT_RUN_DIR = os.path.join(DRIVE_BASE_PATH, experiment_name)\n",
" SAVE_DIR = os.path.join(CURRENT_RUN_DIR, \"models\")\n",
" LOG_DIR_TENSORBOARD = os.path.join(CURRENT_RUN_DIR, \"tensorboard_logs\")\n",
" LOG_FILE_TXT = os.path.join(CURRENT_RUN_DIR, \"run_log.txt\")\n",
"\n",
" os.makedirs(SAVE_DIR, exist_ok=True)\n",
" os.makedirs(LOG_DIR_TENSORBOARD, exist_ok=True)\n",
"\n",
" logging.basicConfig(\n",
" level=logging.INFO,\n",
" format='%(asctime)s [%(levelname)s] %(message)s',\n",
" handlers=[logging.FileHandler(LOG_FILE_TXT), logging.StreamHandler(sys.stdout)],\n",
" force=True\n",
" )\n",
" logger = logging.getLogger(__name__)\n",
" writer = SummaryWriter(LOG_DIR_TENSORBOARD)\n",
"\n",
" logger.info(f\"--- LAUNCHING EXPERIMENT: {experiment_name} ---\")\n",
"\n",
" all_configs = {k: v for k, v in globals().items() if k.isupper()}\n",
" log_configurations(CURRENT_RUN_DIR, all_configs)\n",
" log_environment(CURRENT_RUN_DIR)\n",
"\n",
" logger.info(f\"--- Initializing FNetHybridTransformer ---\")\n",
" model = FNetHybridTransformer(\n",
" num_encoder_layers=NUM_ENCODER_LAYERS,\n",
" num_decoder_layers=NUM_DECODER_LAYERS,\n",
" num_heads=NUM_HEADS,\n",
" d_model=D_MODEL,\n",
" dff=D_FF,\n",
" vocab_size=VOCAB_SIZE,\n",
" max_length=MAX_LENGTH,\n",
" dropout=DROPOUT\n",
" )\n",
"\n",
" model.apply(init_weights_fnet)\n",
" nn.init.normal_(model.pos_embedding.weight, mean=0.0, std=0.02)\n",
" model.final_linear.weight = model.embedding.weight\n",
"\n",
" model.to(device)\n",
" count_parameters(model)\n",
"\n",
" # 4. SETUP OPTIMIZER\n",
" optimizer = torch.optim.AdamW(model.parameters(), lr=PEAK_LEARNING_RATE, betas=(0.9, 0.98),\n",
" eps=1e-9, weight_decay=WEIGHT_DECAY)\n",
"\n",
" # Scheduler\n",
" scheduler = get_cosine_schedule_with_warmup(optimizer=optimizer, num_warmup_steps=WARMUP_STEPS,\n",
" num_training_steps=TARGET_TRAINING_STEPS)\n",
" scaler = torch.cuda.amp.GradScaler()\n",
"\n",
"# --- AUTO-RESUME LOGIC (SMARTER VERSION) ---\n",
" global_step = 0\n",
" best_bleu = 0.0\n",
" LAST_CHECKPOINT_PATH = os.path.join(SAVE_DIR, \"last.pt\")\n",
" BEST_CHECKPOINT_PATH = os.path.join(SAVE_DIR, \"best.pt\")\n",
"\n",
" # 1. Try to find the latest checkpoint (if it exists)\n",
" if os.path.exists(LAST_CHECKPOINT_PATH):\n",
" logger.info(f\"🔄 Found checkpoint at {LAST_CHECKPOINT_PATH}. Resuming...\")\n",
" checkpoint = torch.load(LAST_CHECKPOINT_PATH, map_location=device)\n",
"\n",
" model.load_state_dict(checkpoint['model_state_dict'])\n",
" optimizer.load_state_dict(checkpoint['optimizer_state_dict'])\n",
" scheduler.load_state_dict(checkpoint['scheduler_state_dict'])\n",
" scaler.load_state_dict(checkpoint['scaler_state_dict'])\n",
"\n",
" global_step = checkpoint['global_step']\n",
" best_bleu = checkpoint.get('best_bleu', 0.0)\n",
" logger.info(f\" ✅ Resumed from Step {global_step} (LAST)\")\n",
"\n",
" # 2. If no LAST, try to find the BEST checkpoint (Fall back to this!)\n",
" elif os.path.exists(BEST_CHECKPOINT_PATH):\n",
" logger.info(f\"🔙 'last.pt' not found. Falling back to BEST checkpoint: {BEST_CHECKPOINT_PATH}\")\n",
" checkpoint = torch.load(BEST_CHECKPOINT_PATH, map_location=device)\n",
"\n",
" model.load_state_dict(checkpoint['model_state_dict'])\n",
" optimizer.load_state_dict(checkpoint['optimizer_state_dict'])\n",
" scheduler.load_state_dict(checkpoint['scheduler_state_dict'])\n",
" scaler.load_state_dict(checkpoint['scaler_state_dict'])\n",
"\n",
" global_step = checkpoint['global_step']\n",
" best_bleu = checkpoint.get('best_bleu', 0.0)\n",
" logger.info(f\" ✅ Resumed from Step {global_step} (BEST)\")\n",
"\n",
" # 3. Start Fresh\n",
" else:\n",
" logger.info(\"🆕 No checkpoint found. Starting fresh training.\")\n",
" # 5. TRAINING LOOP\n",
" model.train()\n",
"\n",
" # Resume progress bar from global_step\n",
" progress_bar = tqdm(total=TARGET_TRAINING_STEPS, initial=global_step, desc=\"Training Steps\")\n",
" training_complete = False\n",
"\n",
" # Initialize gradients\n",
" optimizer.zero_grad(set_to_none=True)\n",
"\n",
" # We iterate until global_step reaches the target\n",
" epoch = 0\n",
" while not training_complete:\n",
" train_dataloader.generator.manual_seed(SEED + epoch)\n",
" epoch += 1\n",
"\n",
" for batch_idx, batch in enumerate(train_dataloader):\n",
" if global_step >= TARGET_TRAINING_STEPS:\n",
" training_complete = True\n",
" break\n",
"\n",
" input_ids = batch['input_ids'].to(device, non_blocking=True)\n",
" labels = batch['labels'].to(device, non_blocking=True)\n",
"\n",
" decoder_start_token = torch.full((labels.shape[0], 1), tokenizer.pad_token_id, dtype=torch.long, device=device)\n",
" decoder_input_ids = torch.cat([decoder_start_token, labels[:, :-1]], dim=1)\n",
" decoder_input_ids[decoder_input_ids == -100] = tokenizer.pad_token_id\n",
" target_labels = labels\n",
"\n",
" src_padding_mask, tgt_padding_mask, mem_key_padding_mask, tgt_mask = model.create_masks(input_ids, decoder_input_ids)\n",
" tgt_padding_mask[:, 0] = False\n",
"\n",
" with torch.autocast(device_type=\"cuda\", dtype=torch.float16):\n",
" model_outputs = model(src=input_ids, tgt=decoder_input_ids, src_padding_mask=src_padding_mask,\n",
" tgt_padding_mask=tgt_padding_mask, memory_key_padding_mask=mem_key_padding_mask,\n",
" tgt_mask=tgt_mask)\n",
" loss, loss_components = calculate_combined_loss(model_outputs, target_labels)\n",
"\n",
" # --- GRADIENT ACCUMULATION SCALING ---\n",
" loss = loss / GRAD_ACCUMULATION_STEPS\n",
"\n",
" # Accumulate gradients (no optimizer step yet)\n",
" scaler.scale(loss).backward()\n",
"\n",
" # --- OPTIMIZER STEP (Conditional) ---\n",
" if (batch_idx + 1) % GRAD_ACCUMULATION_STEPS == 0:\n",
" scaler.unscale_(optimizer)\n",
" total_grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)\n",
"\n",
" scaler.step(optimizer)\n",
" scaler.update()\n",
" scheduler.step()\n",
"\n",
" # Reset gradients\n",
" optimizer.zero_grad(set_to_none=True)\n",
"\n",
" global_step += 1\n",
" progress_bar.update(1)\n",
" lr = scheduler.get_last_lr()[0]\n",
"\n",
" if global_step % 20 == 0:\n",
" # Scale loss back up for logging purposes\n",
" logged_loss = loss.item() * GRAD_ACCUMULATION_STEPS\n",
" writer.add_scalar('train/loss', logged_loss, global_step)\n",
" writer.add_scalar('train/learning_rate', lr, global_step)\n",
" writer.add_scalar('train/gradient_norm', total_grad_norm.item(), global_step)\n",
" progress_bar.set_postfix(\n",
" loss=f\"{logged_loss:.2f}\",\n",
" lr=f\"{lr:.2e}\",\n",
" grad=f\"{total_grad_norm.item():.2f}\" # Showing Gradients\n",
" )\n",
"\n",
" # --- PERIODIC SAVING (Every 500 Steps) ---\n",
" # Saves you if Colab crashes mid-epoch\n",
" if global_step % 500 == 0:\n",
" torch.save({\n",
" 'global_step': global_step,\n",
" 'model_state_dict': model.state_dict(),\n",
" 'optimizer_state_dict': optimizer.state_dict(),\n",
" 'scheduler_state_dict': scheduler.state_dict(),\n",
" 'scaler_state_dict': scaler.state_dict(),\n",
" 'best_bleu': best_bleu\n",
" }, LAST_CHECKPOINT_PATH)\n",
"\n",
" # --- VALIDATION CHECK ---\n",
" if global_step in VALIDATION_SCHEDULE:\n",
" logger.info(f\"\\n--- Validation at Step {global_step} ---\")\n",
" bleu_score = evaluate(model, val_dataloader, device)\n",
" writer.add_scalar('validation/bleu', bleu_score, global_step)\n",
" logger.info(f\"Validation BLEU: {bleu_score:.4f} (Best: {best_bleu:.4f})\")\n",
" #generate_sample_translations(model, device, sample_sentences_de_for_tracking)\n",
"\n",
" if bleu_score > best_bleu:\n",
" best_bleu = bleu_score\n",
" logger.info(f\" New best BLEU! Saving best model...\")\n",
" # Save EVERYTHING so you can resume even from best model\n",
" torch.save({\n",
" 'global_step': global_step,\n",
" 'model_state_dict': model.state_dict(),\n",
" 'optimizer_state_dict': optimizer.state_dict(),\n",
" 'scheduler_state_dict': scheduler.state_dict(),\n",
" 'scaler_state_dict': scaler.state_dict(),\n",
" 'best_bleu': best_bleu\n",
" }, BEST_CHECKPOINT_PATH)\n",
"\n",
" model.train()\n",
"\n",
" progress_bar.close()\n",
" writer.close()\n",
"\n",
" # Save Final (With States)\n",
" torch.save({\n",
" 'global_step': global_step,\n",
" 'model_state_dict': model.state_dict(),\n",
" 'optimizer_state_dict': optimizer.state_dict(),\n",
" 'scheduler_state_dict': scheduler.state_dict(),\n",
" 'scaler_state_dict': scaler.state_dict(),\n",
" 'best_bleu': best_bleu\n",
" }, LAST_CHECKPOINT_PATH)\n",
"\n",
" print(\"\\n\" + \"*\"*80)\n",
" print(\" EXPERIMENT COMPLETE \")\n",
" print(\"*\"*80)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "UsS6qhLtJaMF"
},
"outputs": [],
"source": [
"import os\n",
"import sys\n",
"import torch\n",
"import transformers\n",
"import datasets\n",
"import torchmetrics\n",
"import numpy\n",
"import pkg_resources\n",
"\n",
"def log_environment_separate(log_dir):\n",
" # Define the separate file path\n",
" meta_file = os.path.join(log_dir, \"system_metadata.txt\")\n",
"\n",
" with open(meta_file, \"w\") as f:\n",
" # --- PART 1: SUMMARY ---\n",
" f.write(\"=\"*40 + \"\\n\")\n",
" f.write(\"CORE ENVIRONMENT SUMMARY\\n\")\n",
" f.write(\"=\"*40 + \"\\n\")\n",
" f.write(f\"Python: {sys.version.split()[0]}\\n\")\n",
" f.write(f\"PyTorch: {torch.__version__}\\n\")\n",
" f.write(f\"Transformers: {transformers.__version__}\\n\")\n",
" f.write(f\"Datasets: {datasets.__version__}\\n\")\n",
" f.write(f\"TorchMetrics: {torchmetrics.__version__}\\n\")\n",
" f.write(f\"NumPy: {numpy.__version__}\\n\")\n",
"\n",
" try:\n",
" import sacrebleu\n",
" f.write(f\"SacreBLEU: {sacrebleu.__version__}\\n\")\n",
" except ImportError:\n",
" f.write(\"SacreBLEU: Not Installed\\n\")\n",
"\n",
" if torch.cuda.is_available():\n",
" f.write(f\"GPU Name: {torch.cuda.get_device_name(0)}\\n\")\n",
" f.write(f\"CUDA Ver: {torch.version.cuda}\\n\")\n",
" f.write(f\"Capability: {torch.cuda.get_device_capability(0)}\\n\")\n",
" else:\n",
" f.write(\"GPU: None (CPU Only)\\n\")\n",
"\n",
" # --- PART 2: FULL FREEZE ---\n",
" f.write(\"\\n\" + \"=\"*40 + \"\\n\")\n",
" f.write(\"FULL LIBRARY DEPENDENCIES (PIP FREEZE)\\n\")\n",
" f.write(\"=\"*40 + \"\\n\")\n",
"\n",
" installed_packages = {d.project_name: d.version for d in pkg_resources.working_set}\n",
" for package, version in sorted(installed_packages.items()):\n",
" f.write(f\"{package}=={version}\\n\")\n",
"\n",
" print(f\"✅ Environment details saved SEPARATELY to: {meta_file}\")\n",
"\n",
"# Execute\n",
"# Assumes CURRENT_RUN_DIR is defined from your config\n",
"log_environment_separate(CURRENT_RUN_DIR)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "tqDiOyy18clU"
},
"outputs": [],
"source": [
"# TENSORBOARD VISUALIZATION\n",
"\n",
"%load_ext tensorboard\n",
"\n",
"TENSORBOARD_BASE_DIR = os.path.join(DRIVE_BASE_PATH)\n",
"\n",
"%tensorboard --logdir \"{TENSORBOARD_BASE_DIR}\""
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "AmOcgwNnJqOj"
},
"outputs": [],
"source": [
"from google.colab import runtime\n",
"runtime.unassign()"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "eI0-qVlWVVpx"
},
"source": [
"## End"
]
}
],
"metadata": {
"accelerator": "GPU",
"colab": {
"gpuType": "A100",
"provenance": [],
"machine_shape": "hm"
},
"kernelspec": {
"display_name": "Python 3",
"name": "python3"
},
"language_info": {
"name": "python"
}
},
"nbformat": 4,
"nbformat_minor": 0
} |