train_openfold.py 22.8 KB
Newer Older
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
1
2
3
import argparse
import logging
import os
4
import random
5
import sys
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
6
7
import time

8
import numpy as np
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
9
import pytorch_lightning as pl
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
10
from pytorch_lightning.callbacks.lr_monitor import LearningRateMonitor
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
11
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
12
from pytorch_lightning.loggers import WandbLogger
13
from pytorch_lightning.plugins.training_type import DeepSpeedPlugin, DDPPlugin
14
from pytorch_lightning.plugins.environments import SLURMEnvironment
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
15
16
17
import torch

from openfold.config import model_config
18
from openfold.data.data_modules import (
19
    OpenFoldDataModule,OpenFoldMultimerDataModule,
20
    DummyDataLoader,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
21
)
22
from openfold.model.model import AlphaFold
23
from openfold.model.torchscript import script_preset_
24
from openfold.np import residue_constants
25
from openfold.utils.argparse_utils import remove_arguments
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
26
27
28
from openfold.utils.callbacks import (
    EarlyStoppingVerbose,
)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
29
from openfold.utils.exponential_moving_average import ExponentialMovingAverage
30
from openfold.utils.loss import AlphaFoldLoss, AlphaFoldMultimerLoss,lddt_ca
31
from openfold.utils.lr_schedulers import AlphaFoldLRScheduler
32
from openfold.utils.seed import seed_everything
33
from openfold.utils.superimposition import superimpose
34
from openfold.utils.tensor_utils import tensor_tree_map
35
36
37
38
39
from openfold.utils.validation_metrics import (
    drmsd,
    gdt_ts,
    gdt_ha,
)
40
41
42
from openfold.utils.import_weights import (
    import_jax_weights_,
)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
43
from scripts.zero_to_fp32 import (
44
45
    get_fp32_state_dict_from_zero_checkpoint,
    get_global_step_from_zero_checkpoint
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
46
)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
47

Marta's avatar
Marta committed
48
49
from openfold.utils.logger import PerformanceLoggingCallback

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
50
51
52
53
54

class OpenFoldWrapper(pl.LightningModule):
    def __init__(self, config):
        super(OpenFoldWrapper, self).__init__()
        self.config = config
55
        self.model = AlphaFold(config)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
56
        self.loss = AlphaFoldLoss(config.loss)
57
58
59
        self.ema = ExponentialMovingAverage(
            model=self.model, decay=config.ema.decay
        )
60
61
        
        self.cached_weights = None
62
        self.last_lr_step = -1
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
63
64
65
66

    def forward(self, batch):
        return self.model(batch)

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
    def _log(self, loss_breakdown, batch, outputs, train=True):
        phase = "train" if train else "val"
        for loss_name, indiv_loss in loss_breakdown.items():
            self.log(
                f"{phase}/{loss_name}", 
                indiv_loss, 
                on_step=train, on_epoch=(not train), logger=True,
            )

            if(train):
                self.log(
                    f"{phase}/{loss_name}_epoch",
                    indiv_loss,
                    on_step=False, on_epoch=True, logger=True,
                )

        with torch.no_grad():
            other_metrics = self._compute_validation_metrics(
                batch, 
                outputs,
                superimposition_metrics=(not train)
            )

        for k,v in other_metrics.items():
            self.log(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
92
93
                f"{phase}/{k}",
                torch.mean(v),
94
95
96
                on_step=False, on_epoch=True, logger=True
            )

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
97
    def training_step(self, batch, batch_idx):
98
99
100
        if(self.ema.device != batch["aatype"].device):
            self.ema.to(batch["aatype"].device)

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
101
102
        # Run the model
        outputs = self(batch)
103

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
104
105
106
107
        # Remove the recycling dimension
        batch = tensor_tree_map(lambda t: t[..., -1], batch)

        # Compute loss
108
109
110
        loss, loss_breakdown = self.loss(
            outputs, batch, _return_breakdown=True
        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
111

112
113
        # Log it
        self._log(loss_breakdown, batch, outputs)
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
114

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
115
        return loss
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
116

117
118
    def on_before_zero_grad(self, *args, **kwargs):
        self.ema.update(self.model)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
119

120
121
122
    def validation_step(self, batch, batch_idx):
        # At the start of validation, load the EMA weights
        if(self.cached_weights is None):
123
124
125
126
127
            # model.state_dict() contains references to model weights rather
            # than copies. Therefore, we need to clone them before calling 
            # load_state_dict().
            clone_param = lambda t: t.detach().clone()
            self.cached_weights = tensor_tree_map(clone_param, self.model.state_dict())
128
            self.model.load_state_dict(self.ema.state_dict()["params"])
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
129
       
130
        # Run the model
131
132
        outputs = self(batch)
        batch = tensor_tree_map(lambda t: t[..., -1], batch)
133
134
135
136
137

        # Compute loss and other metrics
        batch["use_clamped_fape"] = 0.
        _, loss_breakdown = self.loss(
            outputs, batch, _return_breakdown=True
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
138
        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
139

140
141
        self._log(loss_breakdown, batch, outputs, train=False)
        
142
143
144
145
    def validation_epoch_end(self, _):
        # Restore the model weights to normal
        self.model.load_state_dict(self.cached_weights)
        self.cached_weights = None
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
    def _compute_validation_metrics(self, 
        batch, 
        outputs, 
        superimposition_metrics=False
    ):
        metrics = {}
        
        gt_coords = batch["all_atom_positions"]
        pred_coords = outputs["final_atom_positions"]
        all_atom_mask = batch["all_atom_mask"]
    
        # This is super janky for superimposition. Fix later
        gt_coords_masked = gt_coords * all_atom_mask[..., None]
        pred_coords_masked = pred_coords * all_atom_mask[..., None]
        ca_pos = residue_constants.atom_order["CA"]
        gt_coords_masked_ca = gt_coords_masked[..., ca_pos, :]
        pred_coords_masked_ca = pred_coords_masked[..., ca_pos, :]
        all_atom_mask_ca = all_atom_mask[..., ca_pos]
    
        lddt_ca_score = lddt_ca(
            pred_coords,
            gt_coords,
            all_atom_mask,
            eps=self.config.globals.eps,
            per_residue=False,
        )
   
        metrics["lddt_ca"] = lddt_ca_score
   
        drmsd_ca_score = drmsd(
            pred_coords_masked_ca,
            gt_coords_masked_ca,
            mask=all_atom_mask_ca, # still required here to compute n
        )
   
        metrics["drmsd_ca"] = drmsd_ca_score
    
        if(superimposition_metrics):
            superimposed_pred, alignment_rmsd = superimpose(
                gt_coords_masked_ca, pred_coords_masked_ca, all_atom_mask_ca,
            )
            gdt_ts_score = gdt_ts(
                superimposed_pred, gt_coords_masked_ca, all_atom_mask_ca
            )
            gdt_ha_score = gdt_ha(
                superimposed_pred, gt_coords_masked_ca, all_atom_mask_ca
            )

            metrics["alignment_rmsd"] = alignment_rmsd
            metrics["gdt_ts"] = gdt_ts_score
            metrics["gdt_ha"] = gdt_ha_score
    
        return metrics

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
201
202
    def configure_optimizers(self, 
        learning_rate: float = 1e-3,
203
        eps: float = 1e-5,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
204
    ) -> torch.optim.Adam:
205
206
207
208
209
#        return torch.optim.Adam(
#            self.model.parameters(),
#            lr=learning_rate,
#            eps=eps
#        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
210
        # Ignored as long as a DeepSpeed optimizer is configured
211
        optimizer = torch.optim.Adam(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
212
213
214
215
            self.model.parameters(), 
            lr=learning_rate, 
            eps=eps
        )
216
217
218
219
220
221

        if self.last_lr_step != -1:
            for group in optimizer.param_groups:
                if 'initial_lr' not in group:
                    group['initial_lr'] = learning_rate

222
223
224
        lr_scheduler = AlphaFoldLRScheduler(
            optimizer,
        )
225

226
227
228
229
230
231
232
233
        return {
            "optimizer": optimizer,
            "lr_scheduler": {
                "scheduler": lr_scheduler,
                "interval": "step",
                "name": "AlphaFoldLRScheduler",
            }
        }
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
234

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
235
    def on_load_checkpoint(self, checkpoint):
236
237
238
239
        ema = checkpoint["ema"]
        if(not self.model.template_config.enabled):
            ema["params"] = {k:v for k,v in ema["params"].items() if not "template" in k}
        self.ema.load_state_dict(ema)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
240

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
241
242
243
    def on_save_checkpoint(self, checkpoint):
        checkpoint["ema"] = self.ema.state_dict()

244
245
246
    def resume_last_lr_step(self, lr_step):
        self.last_lr_step = lr_step

247
248
249
250
251
252
253
254
255
256
257
    def load_from_jax(self, jax_path):
        model_basename = os.path.splitext(
                os.path.basename(
                    os.path.normpath(jax_path)
                )
        )[0]
        model_version = "_".join(model_basename.split("_")[1:])
        import_jax_weights_(
                self.model, jax_path, version=model_version
        )

258

259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
class OpenFoldMultimerWrapper(OpenFoldWrapper):
    def __init__(self, config):
        super(OpenFoldMultimerWrapper, self).__init__(config)
        self.config = config
        self.model = AlphaFold(config)
        self.loss = AlphaFoldMultimerLoss(config.loss)
        self.ema = ExponentialMovingAverage(
            model=self.model, decay=config.ema.decay
        )
        
        self.cached_weights = None
        self.last_lr_step = -1

    def forward(self, batch):
        return self.model(batch)

    def training_step(self, batch, batch_idx):
276
277
278
        # Log it
        if(self.ema.device != batch["aatype"].device):
            self.ema.to(batch["aatype"].device)
279
280

        # Run the model
281
282
283
284
285
        outputs = self(batch)

        # Remove the recycling dimension
        batch = tensor_tree_map(lambda t: t[..., -1], batch)

286
        # Compute loss
287
288
        loss, loss_breakdown = self.loss(
            outputs, batch, _return_breakdown=True
289
290
291
        )

        # Log it
292
        self._log(loss_breakdown, batch, outputs)
293
294
295
296
297
298
299
300
301
302
303
304
305
306

        return loss
    
    def validation_step(self, batch, batch_idx):
        # At the start of validation, load the EMA weights
        if(self.cached_weights is None):
            # model.state_dict() contains references to model weights rather
            # than copies. Therefore, we need to clone them before calling 
            # load_state_dict().
            clone_param = lambda t: t.detach().clone()
            self.cached_weights = tensor_tree_map(clone_param, self.model.state_dict())
            self.model.load_state_dict(self.ema.state_dict()["params"])
       
        # Run the model
307
        outputs = self(batch)
308
309

        # Compute loss and other metrics
310
        batch["use_clamped_fape"] = 0.
311
        _, loss_breakdown = self.loss(
312
            outputs, batch, _return_breakdown=True
313
314
        )

315
        self._log(loss_breakdown, batch, outputs, train=False)
316
317
318
319
320
321
        
    def validation_epoch_end(self, _):
        # Restore the model weights to normal
        self.model.load_state_dict(self.cached_weights)
        self.cached_weights = None

322

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
323
def main(args):
324
325
326
    if(args.seed is not None):
        seed_everything(args.seed) 

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
327
    config = model_config(
328
        args.config_preset, 
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
329
        train=True, 
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
330
        low_prec=(str(args.precision) == "16")
331
    ) 
332
333
334
335
    if "multimer" in args.config_preset:
        model_module = OpenFoldMultimerWrapper(config)   
    else: 
        model_module = OpenFoldWrapper(config)
336
    if(args.resume_from_ckpt):
337
338
339
340
341
        if(os.path.isdir(args.resume_from_ckpt)):  
            last_global_step = get_global_step_from_zero_checkpoint(args.resume_from_ckpt)
        else:
            sd = torch.load(args.resume_from_ckpt)
            last_global_step = int(sd['global_step'])
342
343
        model_module.resume_last_lr_step(last_global_step)
        logging.info("Successfully loaded last lr step...")
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
344
    if(args.resume_from_ckpt and args.resume_model_weights_only):
345
346
347
348
        if(os.path.isdir(args.resume_from_ckpt)):
            sd = get_fp32_state_dict_from_zero_checkpoint(args.resume_from_ckpt)
        else:
            sd = torch.load(args.resume_from_ckpt)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
349
350
351
        sd = {k[len("module."):]:v for k,v in sd.items()}
        model_module.load_state_dict(sd)
        logging.info("Successfully loaded model weights...")
Lucas Bickmann's avatar
Lucas Bickmann committed
352
353
354
    if(args.resume_from_jax_params):
        model_module.load_from_jax(args.resume_from_jax_params)
        logging.info(f"Successfully loaded JAX parameters at {args.resume_from_jax_params}...")
355
 
356
    # TorchScript components of the model
357
358
    if(args.script_modules):
        script_preset_(model_module)
359

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
360
    #data_module = DummyDataLoader("new_batch.pickle")
361
362
    if "multimer" in args.config_preset:
        data_module = OpenFoldMultimerDataModule(
363
364
365
366
        config=config.data, 
        batch_seed=args.seed,
        **vars(args)
    )
367
368
369
370
371
372
    else:
        data_module = OpenFoldDataModule(
            config=config.data, 
            batch_seed=args.seed,
            **vars(args)
        )
373

374
375
    data_module.prepare_data()
    data_module.setup()
376
    
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
377
    callbacks = []
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
378
    if(args.checkpoint_every_epoch):
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
379
        mc = ModelCheckpoint(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
380
            every_n_epochs=1,
381
382
            auto_insert_metric_name=False,
            save_top_k=-1,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
383
384
385
386
387
        )
        callbacks.append(mc)

    if(args.early_stopping):
        es = EarlyStoppingVerbose(
388
            monitor="val/lddt_ca",
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
389
390
391
            min_delta=args.min_delta,
            patience=args.patience,
            verbose=False,
392
            mode="max",
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
393
394
395
396
            check_finite=True,
            strict=True,
        )
        callbacks.append(es)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
397

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
398
    if(args.log_performance):
Marta's avatar
Marta committed
399
400
        global_batch_size = args.num_nodes * args.gpus
        perf = PerformanceLoggingCallback(
Marta's avatar
Marta committed
401
            log_file=os.path.join(args.output_dir, "performance_log.json"),
Marta's avatar
Marta committed
402
403
404
            global_batch_size=global_batch_size,
        )
        callbacks.append(perf)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
405

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
406
407
408
409
    if(args.log_lr):
        lr_monitor = LearningRateMonitor(logging_interval="step")
        callbacks.append(lr_monitor)

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
410
411
412
413
414
415
416
417
418
419
420
    loggers = []
    if(args.wandb):
        wdb_logger = WandbLogger(
            name=args.experiment_name,
            save_dir=args.output_dir,
            id=args.wandb_id,
            project=args.wandb_project,
            **{"entity": args.wandb_entity}
        )
        loggers.append(wdb_logger)

421
    if(args.deepspeed_config_path is not None):
422
423
424
        strategy = DeepSpeedPlugin(
            config=args.deepspeed_config_path,
        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
425
426
        if(args.wandb):
            wdb_logger.experiment.save(args.deepspeed_config_path)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
427
            wdb_logger.experiment.save("openfold/config.py")
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
428
    elif (args.gpus is not None and args.gpus > 1) or args.num_nodes > 1:
429
        strategy = DDPPlugin(find_unused_parameters=False)
430
431
    else:
        strategy = None
432
433
434
435
436
437
 
    if(args.wandb):
        freeze_path = f"{wdb_logger.experiment.dir}/package_versions.txt"
        os.system(f"{sys.executable} -m pip freeze > {freeze_path}")
        wdb_logger.experiment.save(f"{freeze_path}")

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
438
439
    trainer = pl.Trainer.from_argparse_args(
        args,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
440
        default_root_dir=args.output_dir,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
441
        strategy=strategy,
Marta's avatar
Marta committed
442
        callbacks=callbacks,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
443
        logger=loggers,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
444
445
446
447
448
449
450
451
452
453
454
    )

    if(args.resume_model_weights_only):
        ckpt_path = None
    else:
        ckpt_path = args.resume_from_ckpt

    trainer.fit(
        model_module, 
        datamodule=data_module,
        ckpt_path=ckpt_path,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
455
456
457
    )


Marta's avatar
Marta committed
458
459
460
461
462
463
464
465
466
467
def bool_type(bool_str: str):
    bool_str_lower = bool_str.lower()
    if bool_str_lower in ('false', 'f', 'no', 'n', '0'):
        return False
    elif bool_str_lower in ('true', 't', 'yes', 'y', '1'):
        return True
    else:
        raise ValueError(f'Cannot interpret {bool_str} as bool')


Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
468
469
470
471
472
473
474
475
476
477
478
479
480
481
if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "train_data_dir", type=str,
        help="Directory containing training mmCIF files"
    )
    parser.add_argument(
        "train_alignment_dir", type=str,
        help="Directory containing precomputed training alignments"
    )
    parser.add_argument(
        "template_mmcif_dir", type=str,
        help="Directory containing mmCIF files to search for templates"
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
482
483
484
485
486
    parser.add_argument(
        "output_dir", type=str,
        help='''Directory in which to output checkpoints, logs, etc. Ignored
                if not on rank 0'''
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
487
488
    parser.add_argument(
        "max_template_date", type=str,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
489
490
        help='''Cutoff for all templates. In training mode, templates are also 
                filtered by the release date of the target'''
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
491
    )
492
493
494
495
    parser.add_argument(
        "--train_mmcif_data_cache_path", type=str, default=None,
        help="path to the json file which records all the information of mmcif structures used during training"
    )
496
497
498
499
500
501
502
503
    parser.add_argument(
        "--distillation_data_dir", type=str, default=None,
        help="Directory containing training PDB files"
    )
    parser.add_argument(
        "--distillation_alignment_dir", type=str, default=None,
        help="Directory containing precomputed distillation alignments"
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
504
505
506
507
508
509
510
511
    parser.add_argument(
        "--val_data_dir", type=str, default=None,
        help="Directory containing validation mmCIF files"
    )
    parser.add_argument(
        "--val_alignment_dir", type=str, default=None,
        help="Directory containing precomputed validation alignments"
    )
512
513
514
515
    parser.add_argument(
        "--val_mmcif_data_cache_path", type=str, default=None,
        help="path to the json file which records all the information of mmcif structures used during training"
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
516
517
518
519
520
    parser.add_argument(
        "--kalign_binary_path", type=str, default='/usr/bin/kalign',
        help="Path to the kalign binary"
    )
    parser.add_argument(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
521
522
523
524
        "--train_filter_path", type=str, default=None,
        help='''Optional path to a text file containing names of training
                examples to include, one per line. Used to filter the training 
                set'''
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
525
526
    )
    parser.add_argument(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
527
528
        "--distillation_filter_path", type=str, default=None,
        help="""See --train_filter_path"""
529
    )
530
531
532
533
534
    parser.add_argument(
        "--obsolete_pdbs_file_path", type=str, default=None,
        help="""Path to obsolete.dat file containing list of obsolete PDBs and 
             their replacements."""
    )
535
536
    parser.add_argument(
        "--template_release_dates_cache_path", type=str, default=None,
537
538
        help="""Output of scripts/generate_mmcif_cache.py run on template mmCIF
                files."""
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
539
540
    )
    parser.add_argument(
Marta's avatar
Marta committed
541
        "--use_small_bfd", type=bool_type, default=False,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
542
543
544
        help="Whether to use a reduced version of the BFD database"
    )
    parser.add_argument(
545
546
        "--seed", type=int, default=None,
        help="Random seed"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
547
    )
548
549
550
551
    parser.add_argument(
        "--deepspeed_config_path", type=str, default=None,
        help="Path to DeepSpeed config. If not provided, DeepSpeed is disabled"
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
552
    parser.add_argument(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
553
554
        "--checkpoint_every_epoch", action="store_true", default=False,
        help="""Whether to checkpoint at the end of every training epoch"""
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
555
556
    )
    parser.add_argument(
Marta's avatar
Marta committed
557
        "--early_stopping", type=bool_type, default=False,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
558
559
560
561
562
563
564
565
566
567
568
        help="Whether to stop training when validation loss fails to decrease"
    )
    parser.add_argument(
        "--min_delta", type=float, default=0,
        help="""The smallest decrease in validation loss that counts as an 
                improvement for the purposes of early stopping"""
    )
    parser.add_argument(
        "--patience", type=int, default=3,
        help="Early stopping patience"
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
569
570
571
572
573
    parser.add_argument(
        "--resume_from_ckpt", type=str, default=None,
        help="Path to a model checkpoint from which to restore training state"
    )
    parser.add_argument(
Marta's avatar
Marta committed
574
        "--resume_model_weights_only", type=bool_type, default=False,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
575
576
        help="Whether to load just model weights as opposed to training state"
    )
Lucas Bickmann's avatar
Lucas Bickmann committed
577
    parser.add_argument(
578
579
        "--resume_from_jax_params", type=str, default=None,
        help="""Path to an .npz JAX parameter file with which to initialize the model"""
Lucas Bickmann's avatar
Lucas Bickmann committed
580
    )
Marta's avatar
Marta committed
581
    parser.add_argument(
582
        "--log_performance", type=bool_type, default=False,
Marta's avatar
Marta committed
583
584
        help="Measure performance"
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
585
586
    parser.add_argument(
        "--wandb", action="store_true", default=False,
587
        help="Whether to log metrics to Weights & Biases"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
588
589
590
    )
    parser.add_argument(
        "--experiment_name", type=str, default=None,
591
        help="Name of the current experiment. Used for wandb logging"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
592
593
594
    )
    parser.add_argument(
        "--wandb_id", type=str, default=None,
595
        help="ID of a previous run to be resumed"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
596
597
598
    )
    parser.add_argument(
        "--wandb_project", type=str, default=None,
599
        help="Name of the wandb project to which this run will belong"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
600
601
602
    )
    parser.add_argument(
        "--wandb_entity", type=str, default=None,
603
        help="wandb username or team name to which runs are attributed"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
604
    )
605
606
607
608
    parser.add_argument(
        "--script_modules", type=bool_type, default=False,
        help="Whether to TorchScript eligible components of them model"
    )
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
609
    parser.add_argument(
610
        "--train_chain_data_cache_path", type=str, default=None,
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
611
612
    )
    parser.add_argument(
613
        "--distillation_chain_data_cache_path", type=str, default=None,
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
614
615
616
    )
    parser.add_argument(
        "--train_epoch_len", type=int, default=10000,
617
618
619
620
621
622
        help=(
            "The virtual length of each training epoch. Stochastic filtering "
            "of training data means that training datasets have no "
            "well-defined length. This virtual length affects frequency of "
            "validation & checkpointing (by default, one of each per epoch)."
        )
623
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
624
    parser.add_argument(
625
626
        "--log_lr", action="store_true", default=False,
        help="Whether to log the actual learning rate"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
627
    )
628
    parser.add_argument(
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
        "--config_preset", type=str, default="initial_training",
        help=(
            'Config setting. Choose e.g. "initial_training", "finetuning", '
            '"model_1", etc. By default, the actual values in the config are '
            'used.'
        )
    )
    parser.add_argument(
        "--_distillation_structure_index_path", type=str, default=None,
    )
    parser.add_argument(
        "--alignment_index_path", type=str, default=None,
        help="Training alignment index. See the README for instructions."
    )
    parser.add_argument(
        "--distillation_alignment_index_path", type=str, default=None,
        help="Distillation alignment index. See the README for instructions."
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
646
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
647
    parser = pl.Trainer.add_argparse_args(parser)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
648
649
   
    # Disable the initial validation pass
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
650
651
652
653
    parser.set_defaults(
        num_sanity_val_steps=0,
    )

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
654
    # Remove some buggy/redundant arguments introduced by the Trainer
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
655
656
657
658
659
    remove_arguments(
        parser, 
        [
            "--accelerator", 
            "--resume_from_checkpoint",
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
660
661
            "--reload_dataloaders_every_epoch",
            "--reload_dataloaders_every_n_epochs",
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
662
663
        ]
    ) 
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
664

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
665
666
    args = parser.parse_args()

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
667
668
669
670
671
    if(args.seed is None and 
        ((args.gpus is not None and args.gpus > 1) or 
         (args.num_nodes is not None and args.num_nodes > 1))):
        raise ValueError("For distributed training, --seed must be specified")

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
672
    if(str(args.precision) == "16" and args.deepspeed_config_path is not None):
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
673
674
        raise ValueError("DeepSpeed and FP16 training are not compatible")

Lucas Bickmann's avatar
Lucas Bickmann committed
675
    if(args.resume_from_jax_params is not None and args.resume_from_ckpt is not None):
676
677
        raise ValueError("Choose between loading pretrained Jax-weights and a checkpoint-path")

Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
678
    # This re-applies the training-time filters at the beginning of every epoch
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
679
    args.reload_dataloaders_every_n_epochs = 1
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
680

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
681
    main(args)