Repository navigation
Expand file tree
/
Copy pathtrain.py
More file actions
195 lines (169 loc) · 7.28 KB
/
Copy pathtrain.py
File metadata and controls
195 lines (169 loc) · 7.28 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
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
import argparse
import os
import warnings
import pytorch_lightning as pl
import pytorch_lightning.callbacks as plc
import torch
import torch.multiprocessing as mp
from pytorch_lightning import Trainer, strategies
from pytorch_lightning.loggers import CSVLogger
from llamo.configuration_llamo import GraphConfig
from llamo.conformer_encoders.unimol import SimpleUniMolModel
from pipeline.config import load_config
from pipeline.data_utils.llamo_dm import LLaMoDM
from pipeline.dictionary import Dictionary
from pipeline.trainer.stage import LLaMoStage
# torch.set_default_dtype(torch.float16)
os.environ["OPENBLAS_NUM_THREADS"] = "1"
## for pyg bug
warnings.filterwarnings(
"ignore", category=UserWarning, message="TypedStorage is deprecated"
)
## for A5000 gpus
torch.set_float32_matmul_precision(
"medium"
) # can be medium (bfloat16), high (tensorfloat32), highest (float32)
class MyDDPStrategy(strategies.DDPStrategy):
def load_model_state_dict(self, checkpoint):
assert self.lightning_module is not None
self.lightning_module.load_state_dict(checkpoint["state_dict"], strict=False)
def main(args):
mp.set_sharing_strategy("file_system")
pl.seed_everything(args.seed)
config = load_config(args.config_file)
config = GraphConfig.from_dict(config)
# if args.init_checkpoint:
# model = LLaMoStage.load_from_checkpoint(args.init_checkpoint, strict=False, map_location="cpu", args=args, config=config)
# print(f"loaded init checkpoint from {args.init_checkpoint}")
if args.mode == "eval" and args.stage_path:
model = LLaMoStage(args, config)
ckpt = torch.load(args.stage_path, map_location="cpu")
if config.encode_mode == "3d":
new_dict = dict()
for k, v in ckpt["state_dict"].items():
if ".graph_model." in k:
new_dict[k.replace(".graph_model.", ".graph3d_model.")] = v
elif ".abstractor." in k:
new_dict[k.replace(".abstractor.", ".abstractor3d.")] = v
else:
new_dict[k] = v
ckpt["state_dict"] = new_dict
if config.llm_tune == "lora":
model.mllm.init_peft()
model.load_state_dict(ckpt["state_dict"], strict=True)
print(f"loaded model from {args.stage_path}")
elif args.stage_path:
model = LLaMoStage(args, config)
# model.mllm.init_peft()
ckpt = torch.load(args.stage_path, map_location="cpu")
model.load_state_dict(ckpt["state_dict"], strict=True)
print(f"loaded model from {args.stage_path}")
else:
model = LLaMoStage(args, config)
if config.llm_tune == "lora" and args.mode != "eval":
model.mllm.init_peft()
print("total params:", sum(p.numel() for p in model.parameters()))
tokenizer = model.tokenizer
if model.uni_dict is None:
uni_dict = Dictionary.load("./llamo/conformer_encoders/unimol_dict.txt")
uni_dict.add_symbol("[MASK]", is_special=True)
else:
uni_dict = model.uni_dict
# data
dm = LLaMoDM(
args.mode,
args.num_workers,
args.batch_size,
args.root_train,
args.root_eval,
config.text_max_len,
tokenizer,
uni_dict,
args,
config,
)
callbacks = []
## fixme save only used parameters
# callbacks.append(plc.ModelCheckpoint(dirpath="all_checkpoints/"+args.filename+"/", every_n_epochs=10, save_top_k=-1))
callbacks.append(
plc.ModelCheckpoint(
dirpath="/hub_data2/jinyoungp/checkpoints/" + args.filename + "/",
filename="{epoch:02d}",
every_n_epochs=args.save_every_n_epochs,
save_last=True,
save_top_k=-1,
save_on_train_epoch_end=True,
)
)
if len(args.devices.split(",")) > 1:
if args.strategy_name == "fsdp":
strategy = strategies.DDPFullyShardedNativeStrategy()
elif args.strategy_name == "deepspeed":
strategy = strategies.DeepSpeedStrategy(stage=3)
else:
strategy = MyDDPStrategy(find_unused_parameters=True, start_method="spawn")
else:
strategy = "ddp_find_unused_parameters_true"
args.devices = eval(args.devices)
logger = CSVLogger(save_dir=f"/hub_data2/jinyoungp/checkpoints/{args.filename}/")
# trainer = Trainer.from_argparse_args(args,
# callbacks=callbacks,
# strategy=strategy,
# logger=logger,
# # limit_train_batches=100,
# )
trainer = Trainer(
accelerator=args.accelerator,
devices=args.devices,
precision=args.precision,
max_epochs=args.max_epochs,
check_val_every_n_epoch=args.check_val_every_n_epoch,
callbacks=callbacks,
strategy=strategy,
logger=logger,
)
if args.mode in {"pretrain", "ft"}:
trainer.fit(model, datamodule=dm, ckpt_path=args.ckpt_path)
elif args.mode == "eval":
trainer.fit_loop.epoch_progress.current.completed = args.caption_eval_epoch - 1
trainer.test(model, datamodule=dm)
else:
raise NotImplementedError()
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--config_file", type=str, default="test.yaml")
parser.add_argument("--filename", type=str, default="stage2_test")
parser.add_argument("--seed", type=int, default=42, help="random seed")
# MM settings
parser.add_argument("--mode", type=str, default="pretrain")
parser.add_argument("--strategy_name", type=str, default=None)
parser.add_argument("--iupac_prediction", action="store_true", default=False)
parser.add_argument("--ckpt_path", type=str, default=None)
# parser = Trainer.add_argparse_args(parser)
parser = LLaMoStage.add_model_specific_args(parser)
parser = LLaMoDM.add_model_specific_args(parser)
parser = SimpleUniMolModel.add_args(parser)
parser.add_argument("--accelerator", type=str, default="gpu")
parser.add_argument("--devices", type=str, default="0,1,2,3")
parser.add_argument("--precision", type=str, default="bf16-mixed")
parser.add_argument("--max_epochs", type=int, default=10)
parser.add_argument("--accumulate_grad_batches", type=int, default=1)
parser.add_argument("--check_val_every_n_epoch", type=int, default=1)
#######
parser.add_argument("--not_use_kg", action="store_true", default=False)
parser.add_argument("--noise_step", type=int, default=999)
parser.add_argument("--noise_type", type=str, choices=["cd"], default="cd")
parser.add_argument("--use_vcd", action="store_true", default=False)
parser.add_argument("--changed", action="store_true", default=False)
parser.add_argument("--cd_alpha", type=float, default=0.5)
parser.add_argument("--cd_beta", type=float, default=0.1)
parser.add_argument("--project", type=str, default="linear")
parser.add_argument("--mlp_depth", type=int, default=2)
args = parser.parse_args()
print("=========================================")
for k, v in sorted(vars(args).items()):
print(k, "=", v)
print("=========================================")
return args
if __name__ == "__main__":
main(get_args())