-
Notifications
You must be signed in to change notification settings - Fork 65
/
train.py
454 lines (438 loc) · 20.2 KB
/
train.py
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
import argparse
import json
import os
from datetime import datetime, timedelta
import pathlib
import numpy as np
import torch
import torch.multiprocessing as mp
from torch.optim.lr_scheduler import LambdaLR
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.optim import Adam
from torch.utils.data import DataLoader
import torch.distributed as dist
from torch.cuda.amp import GradScaler
from evodiff.model import ByteNetLMTime
from evodiff.utils import Tokenizer
from torch.utils.data import Subset
from sequence_models.samplers import SortishSampler, ApproxBatchSampler
from sequence_models.datasets import UniRefDataset
from sequence_models.constants import MSA_ALPHABET
from evodiff.collaters import OAMaskCollater, D3PMCollater
from evodiff.losses import OAMaskedCrossEntropyLoss, D3PMCELoss, D3PMLVBLoss
from sequence_models.metrics import MaskedAccuracy
from sequence_models.utils import warmup
import sys
sys.setrecursionlimit(1000) # must be as large as diffusion timesteps for Q_bar calculation
### SET RANDOM SEEDS ###
torch.cuda.empty_cache() # empty caches
home = str(pathlib.Path.home())
def main():
parser = argparse.ArgumentParser()
parser.add_argument('config_fpath')
parser.add_argument('out_fpath', type=str, nargs='?', default=os.getenv('PT_OUTPUT_DIR', '/tmp') + '/')
parser.add_argument('-n', '--nodes', default=1, type=int, metavar='N')
parser.add_argument('-g', '--gpus', default=1, type=int,
help='number of gpus per node')
parser.add_argument('-nr', '--nr', default=0, type=int,
help='ranking within the nodes')
parser.add_argument('-off', '--offset', default=0, type=int,
help='Number of GPU devices to skip.')
parser.add_argument('--dropout', type=float, default=0.0)
parser.add_argument('--weight_decay', type=float, default=0.0)
parser.add_argument('--tie_weights', action='store_true')
parser.add_argument('--task', default=None)
parser.add_argument('--dataset', default=None)
parser.add_argument('--aml', action='store_true') # Set true to do multi-node training on amlk8s
parser.add_argument('-sd', '--state_dict', default=None)
parser.add_argument('--decay', action='store_true')
parser.add_argument('--final_norm', action='store_true')
parser.add_argument('--norm_first', action='store_true') # turns norm_first on in transformer model
parser.add_argument('--mini_run', action='store_true') # Set to True if running on subset of data
parser.add_argument('--mask', type=str, default='oadm') # Set to True if running on subset of data
parser.add_argument('--warmup', action='store_true') # Set to True if running on subset of data
parser.add_argument('--checkpoint_freq', type=float, default=1) # in minutes
parser.add_argument('--log_freq', type=float, default=10) # in steps
parser.add_argument('--reweighting_term', type=float, default=0) # lambda reweighting term from Austin D3PM
parser.add_argument('--random_seed', type=int, default=0) # lambda reweighting term from Austin D3PM
parser.add_argument('--pretrained', action='store_true') # ONLY USE THIS FLAG FOR FIRST RUN OF PRETRAIN
args = parser.parse_args()
args.world_size = args.gpus * args.nodes
if args.aml:
pass
else:
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '8889'
#print(args.world_size, args.gpus, args.nodes)
mp.spawn(train, nprocs=args.gpus, args=(args,))
def train(gpu, args):
rs = torch.random.manual_seed(args.random_seed)
rs = np.random.seed(int(args.random_seed))
if args.aml:
args.nr = int(os.environ['RANK'])
rank = args.nr * args.gpus + gpu
print("nr", args.nr, "gpus", args.gpus, "gpu", gpu, "rank", rank)
dist.init_process_group(
backend='nccl',
init_method='env://',
world_size=args.world_size,
rank=rank)
torch.cuda.set_device(gpu + args.offset)
device = torch.device('cuda:' + str(gpu + args.offset))
with open(args.config_fpath, 'r') as f:
config = json.load(f)
n_tokens = len(MSA_ALPHABET)
d_embed = config['d_embed']
d_model = config['d_model']
n_layers = config['n_layers']
kernel_size = config['kernel_size']
r = config['r']
if 'slim' in config:
slim = config['slim']
else:
slim = True
if 'activation' in config:
activation = config['activation']
else:
activation = 'relu'
if 'accumulate' in config:
iters_to_accumulate = config['accumulate']
else:
iters_to_accumulate = 1 # dont accumulate
bucket_size = config['bucket_size']
max_tokens = config['max_tokens']
max_batch_size = config['max_batch_size']
epochs = config['epochs']
lr = config['lr']
opt_level = config['opt_level']
warmup_steps = config['warmup_steps']
if 'rank' in config:
weight_rank = config['rank']
else:
weight_rank = None
if args.task is not None:
config['task'] = args.task
if args.dataset is not None:
config['dataset'] = args.dataset
try:
data_top_dir = os.getenv('PT_DATA_DIR') + '/'
ptjob = True
except:
data_top_dir = home + '/Desktop/DMs/data/'
ptjob = False
data_dir = data_top_dir + config['dataset'] + '/'
if args.mini_run:
mini_size = 100 # For troubleshooting
# ----------------------------------------------------------
### COLLATORS ###
# ----------------------------------------------------------
if args.mask == 'oadm':
tokenizer = Tokenizer()
collater = OAMaskCollater(tokenizer=tokenizer)
diffusion_timesteps = None # Not input to model
# elif args.mask == 'so':
# tokenizer = Tokenizer()
# raise Exception("Autoreg in other script")
# collater = BertMaskCollater(tokenizer=tokenizer)
# diffusion_timesteps = None # Not input to model
elif args.mask == 'blosum' or args.mask == 'random':
diffusion_timesteps = config['diffusion_timesteps']
tokenizer = Tokenizer(path_to_blosum=data_top_dir+"blosum62-special-MSA.mat", sequences=True)
if args.mask == 'random':
Q_prod, Q_t = tokenizer.q_random_schedule(timesteps=diffusion_timesteps)
if args.mask == 'blosum':
Q_prod, Q_t = tokenizer.q_blosum_schedule(timesteps=diffusion_timesteps)
collater = D3PMCollater(tokenizer=tokenizer, num_timesteps=diffusion_timesteps, Q=Q_t, Q_bar=Q_prod)
else:
print("mask must be: 'oadm', 'blosum', or 'random'")
causal = False
if args.mask == 'so':
causal = True
# ----------------------------------------------------------
### DATALOADER ###
# ----------------------------------------------------------
metadata = np.load(data_dir + 'lengths_and_offsets.npz')
ds_train = UniRefDataset(data_dir, 'train', structure=False)
train_idx = ds_train.indices
if args.mini_run:
tindices = np.arange(0,1000) # np.arange(21546293,31546293,1)#(1000000,21546293, 1)
train_indices = np.sort(np.random.choice(tindices, mini_size, replace=False))
train_sampler = Subset(ds_train,train_indices)
len_train = train_indices
dl_train = DataLoader(dataset=train_sampler,
shuffle=True,
batch_size=1,
num_workers=4,
collate_fn=collater)
else:
len_train = metadata['ells'][train_idx]
train_sortish_sampler = SortishSampler(len_train, bucket_size, num_replicas=args.world_size, rank=rank)
train_sampler = ApproxBatchSampler(train_sortish_sampler, max_tokens, max_batch_size, len_train)
dl_train = DataLoader(dataset=ds_train,
batch_sampler=train_sampler,
num_workers=16,
collate_fn=collater)
if rank == 0:
ds_valid = UniRefDataset(data_dir, 'valid', structure=False)
valid_idx = ds_valid.indices
if args.mini_run:
vindices = np.arange(1, 80000, 1)
valid_indices = np.random.choice(vindices, mini_size)
len_valid = valid_indices
valid_sampler = Subset(ds_valid, valid_indices)
len_valid = valid_sampler
dl_valid = DataLoader(dataset=valid_sampler,
shuffle=True,
batch_size=1,
num_workers=4,
collate_fn=collater)
else:
len_valid = metadata['ells'][valid_idx]
valid_sortish_sampler = SortishSampler(len_valid, 1000, num_replicas=1, rank=0)
valid_sampler = ApproxBatchSampler(valid_sortish_sampler, max_tokens // 2, max_batch_size, len_valid)
dl_valid = DataLoader(dataset=ds_valid,
batch_sampler=valid_sampler,
num_workers=8,
collate_fn=collater)
# ----------------------------------------------------------
# Initiate model
# ----------------------------------------------------------
padding_idx = tokenizer.pad_id # PROTEIN_ALPHABET.index(PAD)
masking_idx = tokenizer.mask_id
print('Using {} as padding index'.format(padding_idx))
print('Using {} as masking index'.format(masking_idx))
#if args.model_type == 'ByteNet':
model = ByteNetLMTime(n_tokens, d_embed, d_model, n_layers, kernel_size, r,
causal=causal, padding_idx=masking_idx, rank=weight_rank, dropout=args.dropout,
tie_weights=args.tie_weights, final_ln=args.final_norm, slim=slim, activation=activation,
timesteps=diffusion_timesteps)
optimizer = Adam(model.parameters(), lr=lr, weight_decay=args.weight_decay)
outputs = os.listdir(args.out_fpath)
if len(outputs) > 0:
last_epoch = 0
for output in outputs:
if 'checkpoint' in output:
epoch = int(output.split('checkpoint')[-1][:-4])
if epoch > last_epoch:
args.state_dict = args.out_fpath + output
last_epoch = epoch
model = model.to(device)
if args.pretrained: # testing something w/ pretraining
args.state_dict = 'data/pretrained/checkpoint538468.tar'
if args.state_dict is not None:
print('Loading weights from ' + args.state_dict + '...')
sd = torch.load(args.state_dict, map_location=torch.device('cpu'))
msd = sd['model_state_dict']
msd = {k.split('module.')[1]: v for k,v in msd.items()}
model.load_state_dict(msd)
optimizer.load_state_dict(sd['optimizer_state_dict'])
initial_epoch = sd['epoch'] + 1
total_steps = sd['step']
total_tokens = sd['tokens']
else:
initial_epoch = 0
total_steps = 0
total_tokens = 0
scaler = GradScaler()
model = DDP(model)
# ----------------------------------------------------------
# Loss Function
# ----------------------------------------------------------
if args.warmup:
scheduler = LambdaLR(optimizer, warmup(warmup_steps), verbose=False)
else:
raise Exception("add --warmup flag to runtime")
if args.mask == 'oadm' or args.mask == 'so':
loss_func = OAMaskedCrossEntropyLoss(reweight=True)
elif args.mask == 'blosum' or args.mask == 'random':
# Austin = LVB + lambda * CE
loss_func1 = D3PMLVBLoss(tmax=diffusion_timesteps, tokenizer=tokenizer)
loss_func2 = D3PMCELoss(tokenizer=tokenizer)
_lambda = args.reweighting_term
accu_func = MaskedAccuracy()
# ----------------------------------------------------------
# Run
# ----------------------------------------------------------
def epoch(model, train, current_step=0, current_tokens=0):
start_time = datetime.now()
if train:
model = model.train()
loader = dl_train
t = 'Training:'
else:
model = model.eval()
loader = dl_valid
t = 'Validating:'
losses = []
nll_losses = []
accus = []
ns = []
num_seqs = []
chunk_time = datetime.now()
n_seen = 0
tokens_trained = current_tokens
if train:
if args.mini_run:
n_total = len(len_train)
else:
n_total = len(ds_train)
else:
if args.mini_run:
n_total = len(len_valid)
else:
n_total = len(ds_valid)
for i, batch in enumerate(loader):
# restarting from a checkpoint
if train and i == 1 and e == initial_epoch and args.state_dict is not None and not args.pretrained:
print("Restarting from checkpoint")
optimizer.load_state_dict(sd['optimizer_state_dict'])
scheduler.load_state_dict(sd['scheduler_state_dict'])
new_loss, new_nll_loss, new_accu, new_n, new_seqs, new_processed = step(model, batch, train)
if train:
dist.reduce(new_loss, 0, op=dist.ReduceOp.SUM)
dist.reduce(new_nll_loss, 0, op=dist.ReduceOp.SUM)
dist.reduce(new_accu, 0, op=dist.ReduceOp.SUM)
dist.reduce(new_n, 0, op=dist.ReduceOp.SUM)
dist.reduce(new_seqs, 0, op=dist.ReduceOp.SUM)
losses.append(new_loss.item())
nll_losses.append(new_nll_loss.item())
accus.append(new_accu.item())
ns.append(new_n.item())
num_seqs.append(new_seqs.item())
n_seen += new_seqs.item()
total_n = sum(ns)
r_loss = sum(losses) / total_n
r_nll_loss = sum(nll_losses) / total_n
raccu = sum(accus) / total_n
if train:
nsteps = current_step + i + 1
tokens_trained += new_processed.item()
else:
nsteps = i
if rank == 0:
if ptjob:
end = '\n'
start = ''
else:
start = ''
end = '\n'
print(start + '%s Epoch %d of %d Step %d ntokens %d Example %d of %d loss = %.4f nll loss = %.4f accu = %.4f'
% (t, e + 1, epochs, nsteps, tokens_trained, n_seen, n_total, r_loss, r_nll_loss, raccu),
end=end)
if train:
losses = losses[-999:]
accus = accus[-999:]
ns = ns[-999:]
num_seqs = num_seqs[-999:]
nll_losses = nll_losses[-999:]
if nsteps % args.log_freq == 0: # write to checkpoint frequency
if rank == 0:
with open(args.out_fpath + 'train-metrics.csv', 'a') as f:
f.write(','.join([str(r_loss), str(r_nll_loss), str(raccu), str(int(current_tokens)), str(nsteps), str(e)]))
f.write('\n')
if ((datetime.now() - chunk_time) > timedelta(minutes=args.checkpoint_freq)) or (n_seen == n_total):
if rank == 0:
print('Writing to checkpoint at', chunk_time)
with torch.no_grad():
if rank == 0:
ckpt_fpath = args.out_fpath + 'checkpoint%d.tar' % nsteps
torch.save({
'step': nsteps,
'tokens': tokens_trained,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': scheduler.state_dict(),
'epoch': e
}, ckpt_fpath)
_ = epoch(model, False, current_step=nsteps, current_tokens=tokens_trained)
chunk_time = datetime.now()
if not train:
if rank == 0:
with open(args.out_fpath + 'valid-metrics.csv', 'a') as f:
f.write(','.join([str(r_loss), str(r_nll_loss), str(raccu), str(int(current_tokens)), str(current_step), str(e)]))
f.write('\n')
print('Validation complete in ' + str(datetime.now() - start_time))
elif rank == 0:
print('Epoch complete in ' + str(datetime.now() - start_time))
return i, tokens_trained
def step(model, batch, train):
if args.mask == 'blosum' or args.mask == 'random':
src, src_onehot, timestep, tgt, tgt_onehot, Q, Q_bar, q = batch
q = q.to(device)
Q = Q.to(device)
Q_bar = Q_bar.to(device)
src_onehot = src_onehot.to(device)
tgt_onehot = tgt_onehot.to(device)
else:
src, timestep, tgt, mask = batch
mask = mask.to(device)
timestep = timestep.to(device)
src = src.to(device)
tgt = tgt.to(device)
input_mask = (src != padding_idx).float()
if args.mask == 'blosum' or args.mask == 'random':
n_tokens = input_mask.sum()
else:
n_tokens = mask.sum()
n_processed = input_mask.sum()
n_seqs = torch.tensor(len(src), device=device)
# step through model
if train:
optimizer.zero_grad() # reset gradients of model parameters
# Enables autocasting for the forward pass (model + loss)
with torch.cuda.amp.autocast(dtype=torch.float32):
outputs = model(src, timestep, input_mask=input_mask.unsqueeze(-1))
if args.mask == 'blosum' or args.mask == 'random':
lvb_loss = loss_func1(src_onehot, q, outputs, tgt, tgt_onehot, input_mask, timestep, Q, Q_bar)
ce_loss = loss_func2(outputs, tgt, input_mask)
lvb_loss = lvb_loss.to(torch.float32)
ce_loss = ce_loss.to(torch.float32)
loss = (lvb_loss + (_lambda * ce_loss)) * n_tokens
nll_loss = ce_loss * n_tokens
accu = accu_func(outputs, tgt, input_mask) * n_tokens
elif args.mask == 'oadm' or args.mask=='so':
ce_loss, nll_loss = loss_func(outputs, tgt, mask, timestep, input_mask) # sum(loss per token)
loss = ce_loss
accu = accu_func(outputs, tgt, mask) * n_tokens
if train:
# Exit the context manager before backward()
scaler.scale(loss).backward()
scaler.step(optimizer)
scale = scaler.get_scale()
scaler.update()
skip_scheduler = (scale > scaler.get_scale())
if not skip_scheduler:
scheduler.step()
# Gradient accumulation
#print("batch", i)
# if (i + 1) % iters_to_accumulate == 0: # If not accumulating gradients iters_to_accumulate = 1
# #print("accumulating every", iters_to_accumulate)
# #print("updating gradients at batch", i)
# scaler.step(optimizer)
# scale = scaler.get_scale()
# scaler.update()
#
# skip_scheduler = (scale > scaler.get_scale())
# if not skip_scheduler:
# scheduler.step()
if loss <= 0 or loss >= 1000000:
print(loss, lvb_loss, ce_loss, nll_loss, n_tokens, _lambda)
print(timestep)
print([tokenizer.untokenize(t) for t in tgt])
print([tokenizer.untokenize(s) for s in src])
import pdb; pdb.set_trace()
#print("lvb", lvb_loss, "ce", ce_loss, "loss", loss, "tokens", n_tokens, "timestep", timestep)
return loss, nll_loss, accu, n_tokens, n_seqs, n_processed
n_parameters = sum(p.numel() for p in model.parameters())
if rank == 0:
print('%d model parameters' %n_parameters)
print('%d training sequences' %len(len_train))
print('%d validation sequences' %len(len_valid))
for e in range(initial_epoch, epochs):
if not args.mini_run:
train_sortish_sampler.set_epoch(e + 1)
s, t = epoch(model, True, current_step=total_steps, current_tokens=total_tokens)
total_steps += s
total_tokens += t
if __name__ == '__main__':
main()