File size: 19,839 Bytes
f6fb767
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
ELECTRA-style discriminative pre-training for ModernProteinLM.

Generator (small): ~25% of discriminator size, trained with MLM.
Discriminator (main model): Trained to detect replaced tokens (RTD objective).

Key improvements over standard ELECTRA:
1. Curriculum masking: start at 30%, decay to 5%
2. Span masking: mask contiguous regions (protein structural motifs)
3. Generator-distillation: generator temperature annealing
4. No NSP, no dropout (following ESM-2)
"""

import os
import math
import random
from typing import Dict, List, Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset
from transformers import (
    PreTrainedTokenizerFast,
    get_cosine_schedule_with_warmup,
    get_linear_schedule_with_warmup,
)
from datasets import load_dataset, concatenate_datasets
import numpy as np
from tqdm import tqdm

from modeling_modern_protein import ModernProteinLM, ModernProteinLMConfig


class ProteinTokenizer:
    """Simple protein tokenizer matching ESM-2 vocab."""
    
    ALL_AA = "LAGVSERTIDPQKNFYWMHCXBUZO"
    
    def __init__(self):
        # ESM-2 vocab
        # 0: <cls>, 1: <pad>, 2: <eos>, 3: <unk>
        # 4-29: amino acids
        # 30: <mask>, 31: <sep>, 32: <mask> (duplicate for compatibility)
        self.vocab = {
            "<cls>": 0, "<pad>": 1, "<eos>": 2, "<unk>": 3,
            "L": 4, "A": 5, "G": 6, "V": 7, "S": 8, "E": 9, "R": 10,
            "T": 11, "I": 12, "D": 13, "P": 14, "Q": 15, "K": 16, "N": 17,
            "F": 18, "Y": 19, "W": 20, "M": 21, "H": 22, "C": 23, "X": 24,
            "B": 25, "U": 26, "Z": 27, "O": 28, "<mask>": 29,
            "<sep>": 30,  # additional sep
        }
        # Pad to 33 for ESM compatibility
        while len(self.vocab) < 33:
            self.vocab[f"<special_{len(self.vocab)}>"] = len(self.vocab)
        
        self.id_to_token = {v: k for k, v in self.vocab.items()}
        self.mask_token_id = 29
        self.pad_token_id = 1
        self.cls_token_id = 0
        self.eos_token_id = 2
        
    def encode(self, sequence: str, max_length: int = 1024, add_special_tokens: bool = True):
        tokens = []
        if add_special_tokens:
            tokens.append(self.cls_token_id)
        
        for aa in sequence.upper():
            if aa in self.vocab:
                tokens.append(self.vocab[aa])
            else:
                tokens.append(self.vocab["<unk>"])
        
        if add_special_tokens:
            tokens.append(self.eos_token_id)
        
        # Truncate or pad
        if len(tokens) > max_length:
            tokens = tokens[:max_length]
        
        attention_mask = [1] * len(tokens)
        while len(tokens) < max_length:
            tokens.append(self.pad_token_id)
            attention_mask.append(0)
        
        return {
            "input_ids": tokens,
            "attention_mask": attention_mask,
        }
    
    def batch_encode(self, sequences: List[str], max_length: int = 1024):
        results = [self.encode(seq, max_length) for seq in sequences]
        return {
            "input_ids": torch.tensor([r["input_ids"] for r in results], dtype=torch.long),
            "attention_mask": torch.tensor([r["attention_mask"] for r in results], dtype=torch.long),
        }
    
    def decode(self, token_ids):
        if isinstance(token_ids, torch.Tensor):
            token_ids = token_ids.tolist()
        return "".join([self.id_to_token.get(t, "<unk>") for t in token_ids])


def create_span_mask(length, mask_ratio=0.30, mean_span_length=3, min_span_length=1):
    """Create span mask for protein sequences."""
    num_to_mask = max(1, int(length * mask_ratio))
    mask = [False] * length
    
    attempts = 0
    masked = 0
    while masked < num_to_mask and attempts < num_to_mask * 10:
        span_len = max(min_span_length, min(mean_span_length + random.randint(-1, 1), num_to_mask - masked))
        start = random.randint(0, max(0, length - span_len - 1))
        
        # Don't mask if already masked
        if any(mask[start:start+span_len]):
            attempts += 1
            continue
        
        for i in range(start, min(start + span_len, length)):
            mask[i] = True
            masked += 1
    
    return mask


class ProteinDataset(Dataset):
    def __init__(self, sequences, tokenizer, max_length=1024, mask_ratio=0.30, 
                 mean_span_length=3, curriculum_start_ratio=0.30, curriculum_end_ratio=0.05,
                 total_steps=100000, current_step=0):
        self.sequences = sequences
        self.tokenizer = tokenizer
        self.max_length = max_length
        self.mean_span_length = mean_span_length
        self.curriculum_start_ratio = curriculum_start_ratio
        self.curriculum_end_ratio = curriculum_end_ratio
        self.total_steps = total_steps
        self.current_step = current_step
        
    def get_current_mask_ratio(self):
        """Linear decay from start to end ratio."""
        progress = min(1.0, self.current_step / self.total_steps)
        return self.curriculum_start_ratio + (self.curriculum_end_ratio - self.curriculum_start_ratio) * progress
    
    def __len__(self):
        return len(self.sequences)
    
    def __getitem__(self, idx):
        seq = self.sequences[idx]
        encoded = self.tokenizer.encode(seq, max_length=self.max_length)
        input_ids = encoded["input_ids"]
        attention_mask = encoded["attention_mask"]
        
        # Find actual sequence length (before padding)
        seq_len = sum(attention_mask)
        # Exclude special tokens from masking
        effective_len = seq_len - 2 if seq_len > 2 else seq_len
        
        # Apply span masking
        mask_ratio = self.get_current_mask_ratio()
        span_mask = create_span_mask(effective_len, mask_ratio, self.mean_span_length)
        
        # Create masked input and labels
        masked_input = input_ids.copy()
        labels = [-100] * len(input_ids)  # -100 = ignore in loss
        replaced = [False] * len(input_ids)  # For discriminator
        
        for i in range(1, 1 + effective_len):  # Skip CLS
            if span_mask[i - 1]:
                labels[i] = input_ids[i]
                replaced[i] = True
                # 80% mask, 10% random, 10% keep
                r = random.random()
                if r < 0.8:
                    masked_input[i] = self.tokenizer.mask_token_id
                elif r < 0.9:
                    masked_input[i] = random.randint(4, 28)  # Random AA
                # else: keep original
        
        return {
            "input_ids": torch.tensor(masked_input, dtype=torch.long),
            "attention_mask": torch.tensor(attention_mask, dtype=torch.long),
            "labels": torch.tensor(labels, dtype=torch.long),
            "replaced": torch.tensor(replaced, dtype=torch.bool),
            "original_ids": torch.tensor(input_ids, dtype=torch.long),
        }


class GeneratorModel(nn.Module):
    """Small generator model for ELECTRA."""
    
    def __init__(self, vocab_size, hidden_size=256, num_layers=4, num_heads=4, intermediate_size=1024):
        super().__init__()
        config = ModernProteinLMConfig(
            vocab_size=vocab_size,
            hidden_size=hidden_size,
            num_hidden_layers=num_layers,
            num_attention_heads=num_heads,
            intermediate_size=intermediate_size,
            tie_word_embeddings=True,
        )
        self.model = ModernProteinLM(config)
    
    def forward(self, input_ids, attention_mask, labels):
        return self.model(input_ids, attention_mask, labels=labels)


class DiscriminatorModel(ModernProteinLM):
    """Discriminator with additional classification head for RTD."""
    
    def __init__(self, config):
        super().__init__(config)
        self.discriminator_head = nn.Linear(config.hidden_size, 1)
    
    def forward(self, input_ids, attention_mask, labels=None):
        outputs = super().forward(input_ids, attention_mask, return_dict=True)
        hidden = outputs.hidden_states[-1]  # (B, T, H)
        
        # Discriminator logits: real vs fake
        disc_logits = self.discriminator_head(hidden).squeeze(-1)  # (B, T)
        
        disc_loss = None
        if labels is not None:
            # labels: 1 = real, 0 = fake (replaced)
            loss_fct = nn.BCEWithLogitsLoss()
            active_loss = labels != -100
            active_logits = disc_logits[active_loss]
            active_labels = labels[active_loss].float()
            disc_loss = loss_fct(active_logits, active_labels)
        
        return {
            "loss": disc_loss,
            "logits": disc_logits,
            "hidden_states": outputs.hidden_states,
        }


class ELECTRAProteinTrainer:
    def __init__(
        self,
        generator: GeneratorModel,
        discriminator: DiscriminatorModel,
        tokenizer,
        train_dataset,
        eval_dataset,
        output_dir="./electra_protein",
        lr=5e-4,
        batch_size=32,
        max_steps=100000,
        warmup_steps=10000,
        weight_decay=0.01,
        grad_clip=1.0,
        generator_weight=1.0,
        discriminator_weight=50.0,
        device="cuda",
    ):
        self.generator = generator.to(device)
        self.discriminator = discriminator.to(device)
        self.tokenizer = tokenizer
        self.train_dataset = train_dataset
        self.eval_dataset = eval_dataset
        self.output_dir = output_dir
        self.device = device
        self.max_steps = max_steps
        self.grad_clip = grad_clip
        self.generator_weight = generator_weight
        self.discriminator_weight = discriminator_weight
        
        os.makedirs(output_dir, exist_ok=True)
        
        # Optimizers
        self.gen_optimizer = torch.optim.AdamW(
            generator.parameters(), lr=lr, betas=(0.9, 0.98), eps=1e-6, weight_decay=weight_decay
        )
        self.disc_optimizer = torch.optim.AdamW(
            discriminator.parameters(), lr=lr, betas=(0.9, 0.98), eps=1e-6, weight_decay=weight_decay
        )
        
        # Schedulers
        self.gen_scheduler = get_cosine_schedule_with_warmup(
            self.gen_optimizer, num_warmup_steps=warmup_steps, num_training_steps=max_steps
        )
        self.disc_scheduler = get_cosine_schedule_with_warmup(
            self.disc_optimizer, num_warmup_steps=warmup_steps, num_training_steps=max_steps
        )
        
        self.train_loader = DataLoader(
            train_dataset, batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True
        )
        self.eval_loader = DataLoader(
            eval_dataset, batch_size=batch_size, shuffle=False, num_workers=4, pin_memory=True
        )
        
        self.global_step = 0
        self.best_eval_loss = float("inf")
    
    def train(self):
        self.generator.train()
        self.discriminator.train()
        
        pbar = tqdm(total=self.max_steps, desc="Training")
        
        for batch in self.train_loader:
            if self.global_step >= self.max_steps:
                break
            
            self._train_step(batch)
            self.global_step += 1
            pbar.update(1)
            
            if self.global_step % 1000 == 0:
                eval_loss = self.evaluate()
                if eval_loss < self.best_eval_loss:
                    self.best_eval_loss = eval_loss
                    self.save_checkpoint("best")
                self.generator.train()
                self.discriminator.train()
            
            if self.global_step % 5000 == 0:
                self.save_checkpoint(f"step_{self.global_step}")
        
        pbar.close()
        self.save_checkpoint("final")
    
    def _train_step(self, batch):
        input_ids = batch["input_ids"].to(self.device)
        attention_mask = batch["attention_mask"].to(self.device)
        mlm_labels = batch["labels"].to(self.device)
        replaced_positions = batch["replaced"].to(self.device)
        original_ids = batch["original_ids"].to(self.device)
        
        # ====== GENERATOR STEP ======
        gen_outputs = self.generator(input_ids, attention_mask, mlm_labels)
        gen_loss = gen_outputs.loss
        
        # Sample from generator to create corrupted input for discriminator
        with torch.no_grad():
            gen_logits = gen_outputs.logits  # (B, T, V)
            gen_probs = F.softmax(gen_logits, dim=-1)
            sampled_ids = torch.multinomial(
                gen_probs.view(-1, gen_probs.size(-1)), 1
            ).view(gen_probs.shape[:-1])
            
            # Replace masked positions with generator samples
            corrupted_input = original_ids.clone()
            mask_positions = mlm_labels != -100
            corrupted_input[mask_positions] = sampled_ids[mask_positions]
        
        # ====== DISCRIMINATOR STEP ======
        # Create discriminator labels: 1 = original, 0 = replaced
        disc_labels = torch.ones_like(original_ids, dtype=torch.float)  # (B, T)
        disc_labels[replaced_positions] = 0.0
        # Ignore padding
        disc_labels[attention_mask == 0] = -100
        
        disc_outputs = self.discriminator(corrupted_input, attention_mask, disc_labels)
        disc_loss = disc_outputs["loss"]
        
        # ====== BACKWARD ======
        # Combined loss with weighting
        total_loss = self.generator_weight * gen_loss + self.discriminator_weight * disc_loss
        
        total_loss.backward()
        
        torch.nn.utils.clip_grad_norm_(self.generator.parameters(), self.grad_clip)
        torch.nn.utils.clip_grad_norm_(self.discriminator.parameters(), self.grad_clip)
        
        self.gen_optimizer.step()
        self.disc_optimizer.step()
        self.gen_scheduler.step()
        self.disc_scheduler.step()
        
        self.gen_optimizer.zero_grad()
        self.disc_optimizer.zero_grad()
        
        if self.global_step % 100 == 0:
            pbar = tqdm.get_tqdm()
            pbar.set_postfix({
                "gen_loss": f"{gen_loss.item():.4f}",
                "disc_loss": f"{disc_loss.item():.4f}",
                "lr": f"{self.gen_scheduler.get_last_lr()[0]:.2e}",
            })
    
    def evaluate(self):
        self.generator.eval()
        self.discriminator.eval()
        
        total_gen_loss = 0
        total_disc_loss = 0
        total_samples = 0
        
        with torch.no_grad():
            for batch in self.eval_loader:
                input_ids = batch["input_ids"].to(self.device)
                attention_mask = batch["attention_mask"].to(self.device)
                mlm_labels = batch["labels"].to(self.device)
                replaced_positions = batch["replaced"].to(self.device)
                original_ids = batch["original_ids"].to(self.device)
                
                gen_outputs = self.generator(input_ids, attention_mask, mlm_labels)
                total_gen_loss += gen_outputs.loss.item() * input_ids.size(0)
                
                disc_labels = torch.ones_like(original_ids, dtype=torch.float)
                disc_labels[replaced_positions] = 0.0
                disc_labels[attention_mask == 0] = -100
                
                disc_outputs = self.discriminator(input_ids, attention_mask, disc_labels)
                total_disc_loss += disc_outputs["loss"].item() * input_ids.size(0)
                total_samples += input_ids.size(0)
        
        avg_gen = total_gen_loss / total_samples
        avg_disc = total_disc_loss / total_samples
        
        print(f"Eval - Gen Loss: {avg_gen:.4f}, Disc Loss: {avg_disc:.4f}")
        return avg_gen + avg_disc
    
    def save_checkpoint(self, name):
        path = os.path.join(self.output_dir, name)
        os.makedirs(path, exist_ok=True)
        
        torch.save({
            "generator": self.generator.state_dict(),
            "discriminator": self.discriminator.state_dict(),
            "gen_optimizer": self.gen_optimizer.state_dict(),
            "disc_optimizer": self.disc_optimizer.state_dict(),
            "step": self.global_step,
        }, os.path.join(path, "checkpoint.pt"))
        
        # Save discriminator config (main model)
        self.discriminator.config.save_pretrained(path)
        
        print(f"Saved checkpoint to {path}")


def load_protein_sequences(dataset_name="lamm-mit/protein_secondary_structure_from_PDB", split="train", max_seqs=None):
    """Load protein sequences from HF dataset."""
    ds = load_dataset(dataset_name, split=split, streaming=True)
    sequences = []
    
    for i, example in enumerate(ds):
        if max_seqs and i >= max_seqs:
            break
        # Try common column names
        seq = None
        for key in ["input", "primary", "sequences", "sequence", "protein", "text"]:
            if key in example:
                seq = example[key]
                break
        if seq and len(seq) > 10:
            sequences.append(seq)
    
    return sequences


def main():
    # Config
    DISC_CONFIG = ModernProteinLMConfig(
        vocab_size=33,
        hidden_size=576,
        num_hidden_layers=28,
        num_attention_heads=9,
        intermediate_size=2304,
        use_geglu=True,
        tie_word_embeddings=True,
        max_position_embeddings=1026,
        position_embedding_type="rotary",
        rope_theta=10000.0,
    )
    
    # Generator: ~25% of discriminator size
    GEN_CONFIG = ModernProteinLMConfig(
        vocab_size=33,
        hidden_size=320,
        num_hidden_layers=8,
        num_attention_heads=8,
        intermediate_size=1280,
        use_geglu=True,
        tie_word_embeddings=True,
    )
    
    tokenizer = ProteinTokenizer()
    
    # Load data
    print("Loading protein sequences...")
    train_seqs = load_protein_sequences("lamm-mit/protein_secondary_structure_from_PDB", "train", max_seqs=50000)
    eval_seqs = load_protein_sequences("lamm-mit/protein_secondary_structure_from_PDB", "train", max_seqs=5000)
    
    print(f"Loaded {len(train_seqs)} train, {len(eval_seqs)} eval sequences")
    
    train_dataset = ProteinDataset(
        train_seqs, tokenizer, max_length=1024,
        curriculum_start_ratio=0.30, curriculum_end_ratio=0.05,
        total_steps=100000,
    )
    eval_dataset = ProteinDataset(
        eval_seqs, tokenizer, max_length=1024,
        curriculum_start_ratio=0.30, curriculum_end_ratio=0.05,
        total_steps=100000, current_step=100000,  # Fixed at end ratio for eval
    )
    
    # Models
    generator = GeneratorModel(
        vocab_size=33,
        hidden_size=GEN_CONFIG.hidden_size,
        num_layers=GEN_CONFIG.num_hidden_layers,
        num_heads=GEN_CONFIG.num_attention_heads,
        intermediate_size=GEN_CONFIG.intermediate_size,
    )
    discriminator = DiscriminatorModel(DISC_CONFIG)
    
    # Count parameters
    gen_params = sum(p.numel() for p in generator.parameters())
    disc_params = sum(p.numel() for p in discriminator.parameters())
    print(f"Generator params: {gen_params/1e6:.1f}M")
    print(f"Discriminator params: {disc_params/1e6:.1f}M")
    
    trainer = ELECTRAProteinTrainer(
        generator=generator,
        discriminator=discriminator,
        tokenizer=tokenizer,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        output_dir="./modern_protein_electra",
        lr=5e-4,
        batch_size=16,
        max_steps=100000,
        warmup_steps=10000,
        weight_decay=0.01,
        grad_clip=1.0,
        generator_weight=1.0,
        discriminator_weight=50.0,
        device="cuda" if torch.cuda.is_available() else "cpu",
    )
    
    print("Starting ELECTRA pre-training...")
    trainer.train()


if __name__ == "__main__":
    main()