-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtrain_msa.py
More file actions
161 lines (131 loc) · 5.71 KB
/
train_msa.py
File metadata and controls
161 lines (131 loc) · 5.71 KB
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
from torch.utils.data import DataLoader
from dataset.dataset_msa import Dataset_msa, msa_dataset_collate
import torch
from network.msa.model import Model
import numpy as np
from tqdm import tqdm
import os, time
import wandb
import warnings
warnings.filterwarnings("ignore")
import random
from torch.optim import AdamW
from torch.optim.lr_scheduler import OneCycleLR
import torch.nn as nn
import sys
import esm
from eval.eval_msa import eval_msa
def train(device='cpu', resume=False, checkpoint=None, **kwargs):
msa_transformer, msa_transformer_alphabet = esm.pretrained.esm_msa1b_t12_100M_UR50S()
msa_transformer = msa_transformer.eval().to(device)
msa_transformer_batch_converter = msa_transformer_alphabet.get_batch_converter()
dataset = Dataset_msa(train_Datapath, data_num="all")
train_dataloader = DataLoader(dataset, batch_size=8, shuffle=True, num_workers=0, pin_memory=True, collate_fn=msa_dataset_collate, drop_last=True)
testset = Dataset_msa(val_Datapath, data_num=50)
test_dataloader = DataLoader(testset, batch_size=1, shuffle=False, num_workers=0, pin_memory=True, collate_fn=msa_dataset_collate)
timesteps = 1000
model=Model(
device=device,
timesteps=timesteps
).to(device)
print('model construct completed')
start_epoch = 0
total_epoch = 10000
train_total_step = 0
val_total_step = 0
best_auc = 0
auc = 0
min_loss = 1000
optimizer=AdamW(model.parameters(),lr=0.0001)
lr_scheduler=OneCycleLR(optimizer,0.0001,total_steps=total_epoch*len(train_dataloader),pct_start=0.25,anneal_strategy='cos')
loss_fn=nn.CrossEntropyLoss(ignore_index=-1)
t_loss_fn = nn.MSELoss()
print('optimizer construct completed')
current_time = time.strftime("%m_%d_%H_%M", time.localtime())
Log_dir = f'log/esm-{current_time}/'
wandb.init(project='esm', name=f'esm-{current_time}')
if checkpoint is not None and checkpoint != '':
model.load_state_dict(torch.load(checkpoint), strict=False)
for epoch in range(start_epoch, total_epoch):
if epoch == start_epoch:
auc = evaluate_msa(
model, test_dataloader, device, device, epoch,
msa_transformer, msa_transformer_batch_converter, debug=False
)
torch.cuda.empty_cache()
train_loss = train_msa(
model, train_dataloader, optimizer, lr_scheduler, loss_fn,
t_loss_fn, device, device, epoch, debug=False
)
os.makedirs(Log_dir,exist_ok=True)
torch.save(model.state_dict(), Log_dir + f'lastest.pth')
torch.cuda.empty_cache()
if epoch % 1 == 0:
auc = evaluate_msa(
model, test_dataloader, device, device, epoch,
msa_transformer, msa_transformer_batch_converter, debug=False
)
if auc > best_auc:
best_auc = auc
os.makedirs(Log_dir,exist_ok=True)
torch.save(model.state_dict(), Log_dir + f'best_model.pth')
torch.cuda.empty_cache()
def train_msa(model, dataloader, optimizer, lr_scheduler, loss_fn, t_loss_fn, device, epoch, debug=False):
model.train()
train_loss = 0.0
progress_bar = tqdm(enumerate(dataloader), total=len(dataloader), desc=f"Epoch {epoch}")
for i, (good_tokens, bad_tokens, full_tokens, msa, name, msa_array) in progress_bar:
good_tokens = [match.to(device) for match in good_tokens]
bad_tokens = [match.to(device) for match in bad_tokens]
full_tokens = [match.to(device) for match in full_tokens]
pred_label, gt_label, pred_t, gt_t = model(
good_tokens, bad_tokens, full_tokens, msa, name, msa_array, training=True
)
label_loss = loss_fn(pred_label, gt_label)
t_loss = t_loss_fn(pred_t, gt_t)
loss = label_loss + t_loss
train_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
lr_scheduler.step()
if i % 10 == 0:
progress_bar.set_postfix(loss=loss.item(), loss_avg=train_loss / (i + 1))
if not debug:
wandb.log({"loss": loss.item(), "epoch": epoch, "batch_idx": i})
train_loss /= len(dataloader)
if not debug:
wandb.log({"train_loss": train_loss})
return train_loss
def evaluate_msa(model, dataloader, device, epoch, msa_transformer, msa_transformer_batch_converter, debug=False):
model.eval()
AUC_list = []
progress_bar = tqdm(enumerate(dataloader), total=len(dataloader), desc="Validation")
with torch.no_grad():
for step, (good_tokens, bad_tokens, full_tokens, msa, name, msa_array) in progress_bar:
good_tokens = [match.to(device) for match in good_tokens]
bad_tokens = [match.to(device) for match in bad_tokens]
full_tokens = [match.to(device) for match in full_tokens]
denoised_msa_dic = model(
good_tokens, bad_tokens, full_tokens, msa, name, msa_array, training=False
)
auc = eval_msa(denoised_msa_dic, device, msa_transformer, msa_transformer_batch_converter)
AUC_list.append(auc)
progress_bar.set_postfix(PL=auc)
if not debug:
wandb.log({"Step_PL": auc})
MAUC = np.mean(AUC_list) if len(AUC_list) > 0 else 0.0
if not debug:
wandb.log({"PL": MAUC})
print("PL:", MAUC)
return MAUC
if __name__ == '__main__':
seed = 0
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
np.random.seed(seed)
device = torch.device("cuda:0" if torch.cuda.is_available() else 'cpu')
train(device=device, device=device, resume=False, checkpoint='')