train_openfold.py 21.3 KB
Newer Older
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
1
2
3
import argparse
import logging
import os
4
import sys
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
5
6

import pytorch_lightning as pl
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
7
from pytorch_lightning.callbacks.lr_monitor import LearningRateMonitor
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
8
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
9
from pytorch_lightning.loggers import WandbLogger
10
from pytorch_lightning.plugins.training_type import DeepSpeedPlugin, DDPPlugin
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
11
12
13
import torch

from openfold.config import model_config
14
from openfold.data.data_modules import OpenFoldDataModule, OpenFoldMultimerDataModule
15
from openfold.model.model import AlphaFold
16
from openfold.model.torchscript import script_preset_
17
from openfold.np import residue_constants
18
from openfold.utils.argparse_utils import remove_arguments
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
19
20
21
from openfold.utils.callbacks import (
    EarlyStoppingVerbose,
)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
22
from openfold.utils.exponential_moving_average import ExponentialMovingAverage
23
from openfold.utils.loss import AlphaFoldLoss, lddt_ca
24
from openfold.utils.lr_schedulers import AlphaFoldLRScheduler
25
from openfold.utils.multi_chain_permutation import multi_chain_permutation_align
26
from openfold.utils.seed import seed_everything
27
from openfold.utils.superimposition import superimpose
28
from openfold.utils.tensor_utils import tensor_tree_map
29
30
31
32
33
from openfold.utils.validation_metrics import (
    drmsd,
    gdt_ts,
    gdt_ha,
)
34
35
from openfold.utils.import_weights import (
    import_jax_weights_,
36
    import_openfold_weights_
37
)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
38
from scripts.zero_to_fp32 import (
39
40
    get_fp32_state_dict_from_zero_checkpoint,
    get_global_step_from_zero_checkpoint
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
41
)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
42

Marta's avatar
Marta committed
43
44
from openfold.utils.logger import PerformanceLoggingCallback

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
45
46
47
48
49

class OpenFoldWrapper(pl.LightningModule):
    def __init__(self, config):
        super(OpenFoldWrapper, self).__init__()
        self.config = config
50
        self.model = AlphaFold(config)
51
        self.is_multimer = self.config.globals.is_multimer
52

53
        self.loss = AlphaFoldLoss(config.loss)
54

55
56
57
        self.ema = ExponentialMovingAverage(
            model=self.model, decay=config.ema.decay
        )
58
59
        
        self.cached_weights = None
60
        self.last_lr_step = -1
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
61
62
63
64

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

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
    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
90
91
                f"{phase}/{k}",
                torch.mean(v),
92
93
94
                on_step=False, on_epoch=True, logger=True
            )

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
95
    def training_step(self, batch, batch_idx):
96
97
98
        if(self.ema.device != batch["aatype"].device):
            self.ema.to(batch["aatype"].device)

99
100
        ground_truth = batch.pop('gt_features', None)

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

107
108
109
110
111
        if self.is_multimer:
            batch = multi_chain_permutation_align(out=outputs,
                                                  features=batch,
                                                  ground_truth=ground_truth)

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
112
        # Compute loss
113
114
115
        loss, loss_breakdown = self.loss(
            outputs, batch, _return_breakdown=True
        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
116

117
118
        # Log it
        self._log(loss_breakdown, batch, outputs)
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
119

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
120
        return loss
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
121

122
123
    def on_before_zero_grad(self, *args, **kwargs):
        self.ema.update(self.model)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
124

125
126
127
    def validation_step(self, batch, batch_idx):
        # At the start of validation, load the EMA weights
        if(self.cached_weights is None):
128
129
130
131
132
            # 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())
133
            self.model.load_state_dict(self.ema.state_dict()["params"])
134
135
136

        ground_truth = batch.pop('gt_features', None)

137
        # Run the model
138
139
        outputs = self(batch)
        batch = tensor_tree_map(lambda t: t[..., -1], batch)
140
141

        batch["use_clamped_fape"] = 0.
142
143
144
145
146
147
148

        if self.is_multimer:
            batch = multi_chain_permutation_align(out=outputs,
                                                  features=batch,
                                                  ground_truth=ground_truth)

        # Compute loss and other metrics
149
150
        _, loss_breakdown = self.loss(
            outputs, batch, _return_breakdown=True
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
151
        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
152

153
154
        self._log(loss_breakdown, batch, outputs, train=False)
        
155
156
157
158
    def validation_epoch_end(self, _):
        # Restore the model weights to normal
        self.model.load_state_dict(self.cached_weights)
        self.cached_weights = None
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
    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
214
215
    def configure_optimizers(self, 
        learning_rate: float = 1e-3,
216
        eps: float = 1e-5,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
217
    ) -> torch.optim.Adam:
218
219
220
221
222
#        return torch.optim.Adam(
#            self.model.parameters(),
#            lr=learning_rate,
#            eps=eps
#        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
223
        # Ignored as long as a DeepSpeed optimizer is configured
224
        optimizer = torch.optim.Adam(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
225
226
227
228
            self.model.parameters(), 
            lr=learning_rate, 
            eps=eps
        )
229
230
231
232
233
234

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

235
236
        lr_scheduler = AlphaFoldLRScheduler(
            optimizer,
237
            last_epoch=self.last_lr_step
238
        )
239

240
241
242
243
244
245
246
247
        return {
            "optimizer": optimizer,
            "lr_scheduler": {
                "scheduler": lr_scheduler,
                "interval": "step",
                "name": "AlphaFoldLRScheduler",
            }
        }
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
248

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
249
    def on_load_checkpoint(self, checkpoint):
250
251
252
253
        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
254

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
255
256
257
    def on_save_checkpoint(self, checkpoint):
        checkpoint["ema"] = self.ema.state_dict()

258
259
260
    def resume_last_lr_step(self, lr_step):
        self.last_lr_step = lr_step

261
262
263
264
265
266
267
268
269
270
271
    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
        )

272

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
273
def main(args):
274
275
276
    if(args.seed is not None):
        seed_everything(args.seed) 

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
277
    config = model_config(
278
        args.config_preset, 
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
279
        train=True, 
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
280
        low_prec=(str(args.precision) == "16")
281
    ) 
282
283
    model_module = OpenFoldWrapper(config)

284
    if(args.resume_from_ckpt):
285
286
287
288
289
        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'])
290
291
        model_module.resume_last_lr_step(last_global_step)
        logging.info("Successfully loaded last lr step...")
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
292
    if(args.resume_from_ckpt and args.resume_model_weights_only):
293
294
295
296
        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
297
        sd = {k[len("module."):]:v for k,v in sd.items()}
298
        import_openfold_weights_(model=model_module, state_dict=sd)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
299
        logging.info("Successfully loaded model weights...")
Lucas Bickmann's avatar
Lucas Bickmann committed
300
301
302
    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}...")
303
 
304
    # TorchScript components of the model
305
306
    if(args.script_modules):
        script_preset_(model_module)
307

308
309
    if "multimer" in args.config_preset:
        data_module = OpenFoldMultimerDataModule(
310
311
312
313
        config=config.data, 
        batch_seed=args.seed,
        **vars(args)
    )
314
315
316
317
318
319
    else:
        data_module = OpenFoldDataModule(
            config=config.data, 
            batch_seed=args.seed,
            **vars(args)
        )
320

321
322
    data_module.prepare_data()
    data_module.setup()
323
    
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
324
    callbacks = []
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
325
    if(args.checkpoint_every_epoch):
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
326
        mc = ModelCheckpoint(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
327
            every_n_epochs=1,
328
329
            auto_insert_metric_name=False,
            save_top_k=-1,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
330
331
332
333
334
        )
        callbacks.append(mc)

    if(args.early_stopping):
        es = EarlyStoppingVerbose(
335
            monitor="val/lddt_ca",
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
336
337
338
            min_delta=args.min_delta,
            patience=args.patience,
            verbose=False,
339
            mode="max",
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
340
341
342
343
            check_finite=True,
            strict=True,
        )
        callbacks.append(es)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
344

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
345
    if(args.log_performance):
Marta's avatar
Marta committed
346
347
        global_batch_size = args.num_nodes * args.gpus
        perf = PerformanceLoggingCallback(
Marta's avatar
Marta committed
348
            log_file=os.path.join(args.output_dir, "performance_log.json"),
Marta's avatar
Marta committed
349
350
351
            global_batch_size=global_batch_size,
        )
        callbacks.append(perf)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
352

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
353
354
355
356
    if(args.log_lr):
        lr_monitor = LearningRateMonitor(logging_interval="step")
        callbacks.append(lr_monitor)

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
357
358
359
360
361
362
363
364
365
366
367
    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)

368
    if(args.deepspeed_config_path is not None):
369
370
371
        strategy = DeepSpeedPlugin(
            config=args.deepspeed_config_path,
        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
372
373
        if(args.wandb):
            wdb_logger.experiment.save(args.deepspeed_config_path)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
374
            wdb_logger.experiment.save("openfold/config.py")
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
375
    elif (args.gpus is not None and args.gpus > 1) or args.num_nodes > 1:
376
        strategy = DDPPlugin(find_unused_parameters=False)
377
378
    else:
        strategy = None
379
380
381
382
383
384
 
    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
385
386
    trainer = pl.Trainer.from_argparse_args(
        args,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
387
        default_root_dir=args.output_dir,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
388
        strategy=strategy,
Marta's avatar
Marta committed
389
        callbacks=callbacks,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
390
        logger=loggers,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
391
392
393
394
395
396
397
398
399
400
401
    )

    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
402
403
404
    )


Marta's avatar
Marta committed
405
406
407
408
409
410
411
412
413
414
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
415
416
417
418
419
420
421
422
423
424
425
426
427
428
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
429
430
431
432
433
    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
434
435
    parser.add_argument(
        "max_template_date", type=str,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
436
437
        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
438
    )
439
440
    parser.add_argument(
        "--train_mmcif_data_cache_path", type=str, default=None,
441
442
        help="Path to the json file which records all the information of mmcif structures used during training"
    )
443
    parser.add_argument(
444
        "--use_single_seq_mode", type=str, default=False,
445
        help="Use single sequence embeddings instead of MSAs."
446
    )
447
448
449
450
451
452
453
454
    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
455
456
457
458
459
460
461
462
    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"
    )
463
464
    parser.add_argument(
        "--val_mmcif_data_cache_path", type=str, default=None,
Dingquan Yu's avatar
Dingquan Yu committed
465
        help="path to the json file which records all the information of mmcif structures used during validation"
466
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
467
468
469
470
471
    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
472
473
474
475
        "--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
476
477
    )
    parser.add_argument(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
478
479
        "--distillation_filter_path", type=str, default=None,
        help="""See --train_filter_path"""
480
    )
481
482
483
484
485
    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."""
    )
486
487
    parser.add_argument(
        "--template_release_dates_cache_path", type=str, default=None,
488
489
        help="""Output of scripts/generate_mmcif_cache.py run on template mmCIF
                files."""
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
490
491
    )
    parser.add_argument(
Marta's avatar
Marta committed
492
        "--use_small_bfd", type=bool_type, default=False,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
493
494
495
        help="Whether to use a reduced version of the BFD database"
    )
    parser.add_argument(
496
497
        "--seed", type=int, default=None,
        help="Random seed"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
498
    )
499
500
501
502
    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
503
    parser.add_argument(
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
504
505
        "--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
506
507
    )
    parser.add_argument(
Marta's avatar
Marta committed
508
        "--early_stopping", type=bool_type, default=False,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
509
510
511
512
513
514
515
516
517
518
519
        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
520
521
522
523
524
    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
525
        "--resume_model_weights_only", type=bool_type, default=False,
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
526
527
        help="Whether to load just model weights as opposed to training state"
    )
Lucas Bickmann's avatar
Lucas Bickmann committed
528
    parser.add_argument(
529
530
        "--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
531
    )
Marta's avatar
Marta committed
532
    parser.add_argument(
533
        "--log_performance", type=bool_type, default=False,
Marta's avatar
Marta committed
534
535
        help="Measure performance"
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
536
537
    parser.add_argument(
        "--wandb", action="store_true", default=False,
538
        help="Whether to log metrics to Weights & Biases"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
539
540
541
    )
    parser.add_argument(
        "--experiment_name", type=str, default=None,
542
        help="Name of the current experiment. Used for wandb logging"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
543
544
545
    )
    parser.add_argument(
        "--wandb_id", type=str, default=None,
546
        help="ID of a previous run to be resumed"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
547
548
549
    )
    parser.add_argument(
        "--wandb_project", type=str, default=None,
550
        help="Name of the wandb project to which this run will belong"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
551
552
553
    )
    parser.add_argument(
        "--wandb_entity", type=str, default=None,
554
        help="wandb username or team name to which runs are attributed"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
555
    )
556
557
558
559
    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
560
    parser.add_argument(
561
        "--train_chain_data_cache_path", type=str, default=None,
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
562
563
    )
    parser.add_argument(
564
        "--distillation_chain_data_cache_path", type=str, default=None,
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
565
566
567
    )
    parser.add_argument(
        "--train_epoch_len", type=int, default=10000,
568
569
570
571
572
573
        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)."
        )
574
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
575
    parser.add_argument(
576
577
        "--log_lr", action="store_true", default=False,
        help="Whether to log the actual learning rate"
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
578
    )
579
    parser.add_argument(
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
        "--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
597
    )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
598
    parser = pl.Trainer.add_argparse_args(parser)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
599
600
   
    # Disable the initial validation pass
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
601
602
603
604
    parser.set_defaults(
        num_sanity_val_steps=0,
    )

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
605
    # Remove some buggy/redundant arguments introduced by the Trainer
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
606
607
608
609
610
    remove_arguments(
        parser, 
        [
            "--accelerator", 
            "--resume_from_checkpoint",
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
611
612
            "--reload_dataloaders_every_epoch",
            "--reload_dataloaders_every_n_epochs",
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
613
614
        ]
    ) 
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
615

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
616
617
    args = parser.parse_args()

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
618
619
620
621
622
    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
623
    if(str(args.precision) == "16" and args.deepspeed_config_path is not None):
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
624
625
        raise ValueError("DeepSpeed and FP16 training are not compatible")

Lucas Bickmann's avatar
Lucas Bickmann committed
626
    if(args.resume_from_jax_params is not None and args.resume_from_ckpt is not None):
627
628
        raise ValueError("Choose between loading pretrained Jax-weights and a checkpoint-path")

Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
629
    # This re-applies the training-time filters at the beginning of every epoch
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
630
    args.reload_dataloaders_every_n_epochs = 1
Gustaf Ahdritz's avatar
Fixes  
Gustaf Ahdritz committed
631

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
632
    main(args)