Spaces:
Sleeping
Sleeping
Zhaohan Meng
commited on
Delete src/finetune/.ipynb_checkpoints
Browse files
src/finetune/.ipynb_checkpoints/finetune-checkpoint.py
DELETED
|
@@ -1,416 +0,0 @@
|
|
| 1 |
-
import argparse
|
| 2 |
-
import os
|
| 3 |
-
import random
|
| 4 |
-
import string
|
| 5 |
-
import sys
|
| 6 |
-
import pandas as pd
|
| 7 |
-
from datetime import datetime
|
| 8 |
-
|
| 9 |
-
sys.path.append("../")
|
| 10 |
-
import numpy as np
|
| 11 |
-
import torch
|
| 12 |
-
import lightgbm as lgb
|
| 13 |
-
import sklearn.metrics as metrics
|
| 14 |
-
from sklearn.utils import class_weight
|
| 15 |
-
from sklearn.model_selection import train_test_split
|
| 16 |
-
from sklearn.metrics import accuracy_score, precision_recall_curve, f1_score, precision_recall_fscore_support,roc_auc_score
|
| 17 |
-
from torch.utils.data import DataLoader
|
| 18 |
-
from tqdm.auto import tqdm
|
| 19 |
-
from transformers import EsmTokenizer, EsmForMaskedLM, BertModel, BertTokenizer, AutoTokenizer, EsmModel
|
| 20 |
-
from utils.downstream_disgenet import DisGeNETProcessor
|
| 21 |
-
from utils.metric_learning_models import GDA_Metric_Learning
|
| 22 |
-
|
| 23 |
-
def parse_config():
|
| 24 |
-
parser = argparse.ArgumentParser()
|
| 25 |
-
parser.add_argument('-f')
|
| 26 |
-
parser.add_argument("--step", type=int, default=0)
|
| 27 |
-
parser.add_argument(
|
| 28 |
-
"--save_model_path",
|
| 29 |
-
type=str,
|
| 30 |
-
default=None,
|
| 31 |
-
help="path of the pretrained disease model located",
|
| 32 |
-
)
|
| 33 |
-
parser.add_argument(
|
| 34 |
-
"--prot_encoder_path",
|
| 35 |
-
type=str,
|
| 36 |
-
default="facebook/esm2_t33_650M_UR50D",
|
| 37 |
-
#"facebook/galactica-6.7b", "Rostlab/prot_bert" “facebook/esm2_t33_650M_UR50D”
|
| 38 |
-
help="path/name of protein encoder model located",
|
| 39 |
-
)
|
| 40 |
-
parser.add_argument(
|
| 41 |
-
"--disease_encoder_path",
|
| 42 |
-
type=str,
|
| 43 |
-
default="microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext",
|
| 44 |
-
help="path/name of textual pre-trained language model",
|
| 45 |
-
)
|
| 46 |
-
parser.add_argument("--reduction_factor", type=int, default=8)
|
| 47 |
-
parser.add_argument(
|
| 48 |
-
"--loss",
|
| 49 |
-
help="{ms_loss|infoNCE|cosine_loss|circle_loss|triplet_loss}}",
|
| 50 |
-
default="infoNCE",
|
| 51 |
-
)
|
| 52 |
-
parser.add_argument(
|
| 53 |
-
"--input_feature_save_path",
|
| 54 |
-
type=str,
|
| 55 |
-
default="../../data/processed_disease",
|
| 56 |
-
help="path of tokenized training data",
|
| 57 |
-
)
|
| 58 |
-
parser.add_argument(
|
| 59 |
-
"--agg_mode", default="mean_all_tok", type=str, help="{cls|mean|mean_all_tok}"
|
| 60 |
-
)
|
| 61 |
-
parser.add_argument("--batch_size", type=int, default=256)
|
| 62 |
-
parser.add_argument("--patience", type=int, default=5)
|
| 63 |
-
parser.add_argument("--num_leaves", type=int, default=5)
|
| 64 |
-
parser.add_argument("--max_depth", type=int, default=5)
|
| 65 |
-
parser.add_argument("--lr", type=float, default=0.35)
|
| 66 |
-
parser.add_argument("--dropout", type=float, default=0.1)
|
| 67 |
-
parser.add_argument("--test", type=int, default=0)
|
| 68 |
-
parser.add_argument("--use_miner", action="store_true")
|
| 69 |
-
parser.add_argument("--miner_margin", default=0.2, type=float)
|
| 70 |
-
parser.add_argument("--freeze_prot_encoder", action="store_true")
|
| 71 |
-
parser.add_argument("--freeze_disease_encoder", action="store_true")
|
| 72 |
-
parser.add_argument("--use_adapter", action="store_true")
|
| 73 |
-
parser.add_argument("--use_pooled", action="store_true")
|
| 74 |
-
parser.add_argument("--device", type=str, default="cpu")
|
| 75 |
-
parser.add_argument(
|
| 76 |
-
"--use_both_feature",
|
| 77 |
-
help="use the both features of gnn_feature_v1_samples and pretrained models",
|
| 78 |
-
action="store_true",
|
| 79 |
-
)
|
| 80 |
-
parser.add_argument(
|
| 81 |
-
"--use_v1_feature_only",
|
| 82 |
-
help="use the features of gnn_feature_v1_samples only",
|
| 83 |
-
action="store_true",
|
| 84 |
-
)
|
| 85 |
-
parser.add_argument(
|
| 86 |
-
"--save_path_prefix",
|
| 87 |
-
type=str,
|
| 88 |
-
default="../../save_model_ckp/finetune/",
|
| 89 |
-
help="save the result in which directory",
|
| 90 |
-
)
|
| 91 |
-
parser.add_argument(
|
| 92 |
-
"--save_name", default="fine_tune", type=str, help="the name of the saved file"
|
| 93 |
-
)
|
| 94 |
-
# Add argument for input CSV file path
|
| 95 |
-
parser.add_argument("--input_csv_path", type=str, required=True, help="Path to the input CSV file.")
|
| 96 |
-
|
| 97 |
-
# Add argument for output CSV file path
|
| 98 |
-
parser.add_argument("--output_csv_path", type=str, required=True, help="Path to the output CSV file.")
|
| 99 |
-
return parser.parse_args()
|
| 100 |
-
|
| 101 |
-
def get_feature(model, dataloader, args):
|
| 102 |
-
x = list()
|
| 103 |
-
y = list()
|
| 104 |
-
with torch.no_grad():
|
| 105 |
-
for step, batch in tqdm(enumerate(dataloader)):
|
| 106 |
-
prot_input_ids, prot_attention_mask, dis_input_ids, dis_attention_mask, y1 = batch
|
| 107 |
-
prot_input = {
|
| 108 |
-
'input_ids': prot_input_ids.to(args.device),
|
| 109 |
-
'attention_mask': prot_attention_mask.to(args.device)
|
| 110 |
-
}
|
| 111 |
-
dis_input = {
|
| 112 |
-
'input_ids': dis_input_ids.to(args.device),
|
| 113 |
-
'attention_mask': dis_attention_mask.to(args.device)
|
| 114 |
-
}
|
| 115 |
-
feature_output = model.predict(prot_input, dis_input)
|
| 116 |
-
x1 = feature_output.cpu().numpy()
|
| 117 |
-
x.append(x1)
|
| 118 |
-
y.append(y1.cpu().numpy())
|
| 119 |
-
x = np.concatenate(x, axis=0)
|
| 120 |
-
y = np.concatenate(y, axis=0)
|
| 121 |
-
return x, y
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
def encode_pretrained_feature(args, disGeNET):
|
| 125 |
-
input_feat_file = os.path.join(
|
| 126 |
-
args.input_feature_save_path,
|
| 127 |
-
f"{args.model_short}_{args.step}_use_{'pooled' if args.use_pooled else 'cls'}_feat.npz",
|
| 128 |
-
)
|
| 129 |
-
|
| 130 |
-
if os.path.exists(input_feat_file):
|
| 131 |
-
print(f"load prior feature data from {input_feat_file}.")
|
| 132 |
-
loaded = np.load(input_feat_file)
|
| 133 |
-
x_train, y_train = loaded["x_train"], loaded["y_train"]
|
| 134 |
-
x_valid, y_valid = loaded["x_valid"], loaded["y_valid"]
|
| 135 |
-
# x_test, y_test = loaded["x_test"], loaded["y_test"]
|
| 136 |
-
|
| 137 |
-
prot_tokenizer = EsmTokenizer.from_pretrained(args.prot_encoder_path, do_lower_case=False)
|
| 138 |
-
# prot_tokenizer = BertTokenizer.from_pretrained(args.prot_encoder_path, do_lower_case=False)
|
| 139 |
-
print("prot_tokenizer", len(prot_tokenizer))
|
| 140 |
-
disease_tokenizer = BertTokenizer.from_pretrained(args.disease_encoder_path)
|
| 141 |
-
print("disease_tokenizer", len(disease_tokenizer))
|
| 142 |
-
|
| 143 |
-
prot_model = EsmModel.from_pretrained(args.prot_encoder_path)
|
| 144 |
-
# prot_model = BertModel.from_pretrained(args.prot_encoder_path)
|
| 145 |
-
disease_model = BertModel.from_pretrained(args.disease_encoder_path)
|
| 146 |
-
|
| 147 |
-
if args.save_model_path:
|
| 148 |
-
model = GDA_Metric_Learning(prot_model, disease_model, 1280, 768, args)
|
| 149 |
-
|
| 150 |
-
if args.use_adapter:
|
| 151 |
-
prot_model_path = os.path.join(
|
| 152 |
-
args.save_model_path, f"prot_adapter_step_{args.step}"
|
| 153 |
-
)
|
| 154 |
-
disease_model_path = os.path.join(
|
| 155 |
-
args.save_model_path, f"disease_adapter_step_{args.step}"
|
| 156 |
-
)
|
| 157 |
-
model.load_adapters(prot_model_path, disease_model_path)
|
| 158 |
-
else:
|
| 159 |
-
prot_model_path = os.path.join(
|
| 160 |
-
args.save_model_path, f"step_{args.step}_model.bin"
|
| 161 |
-
)# , f"step_{args.step}_model.bin"
|
| 162 |
-
disease_model_path = os.path.join(
|
| 163 |
-
args.save_model_path, f"step_{args.step}_model.bin"
|
| 164 |
-
)
|
| 165 |
-
model.non_adapters(prot_model_path, disease_model_path)
|
| 166 |
-
|
| 167 |
-
model = model.to(args.device)
|
| 168 |
-
prot_model = model.prot_encoder
|
| 169 |
-
disease_model = model.disease_encoder
|
| 170 |
-
print(f"loaded prior model {args.save_model_path}.")
|
| 171 |
-
|
| 172 |
-
def collate_fn_batch_encoding(batch):
|
| 173 |
-
query1, query2, scores = zip(*batch)
|
| 174 |
-
|
| 175 |
-
query_encodings1 = prot_tokenizer.batch_encode_plus(
|
| 176 |
-
list(query1),
|
| 177 |
-
max_length=512,
|
| 178 |
-
padding="max_length",
|
| 179 |
-
truncation=True,
|
| 180 |
-
add_special_tokens=True,
|
| 181 |
-
return_tensors="pt",
|
| 182 |
-
)
|
| 183 |
-
query_encodings2 = disease_tokenizer.batch_encode_plus(
|
| 184 |
-
list(query2),
|
| 185 |
-
max_length=512,
|
| 186 |
-
padding="max_length",
|
| 187 |
-
truncation=True,
|
| 188 |
-
add_special_tokens=True,
|
| 189 |
-
return_tensors="pt",
|
| 190 |
-
)
|
| 191 |
-
scores = torch.tensor(list(scores))
|
| 192 |
-
attention_mask1 = query_encodings1["attention_mask"].bool()
|
| 193 |
-
attention_mask2 = query_encodings2["attention_mask"].bool()
|
| 194 |
-
|
| 195 |
-
return query_encodings1["input_ids"], attention_mask1, query_encodings2["input_ids"], attention_mask2, scores
|
| 196 |
-
|
| 197 |
-
test_examples = disGeNET.get_test_examples(args.test)
|
| 198 |
-
print(f"get test examples: {len(test_examples)}")
|
| 199 |
-
|
| 200 |
-
test_dataloader = DataLoader(
|
| 201 |
-
test_examples,
|
| 202 |
-
batch_size=args.batch_size,
|
| 203 |
-
shuffle=False,
|
| 204 |
-
collate_fn=collate_fn_batch_encoding,
|
| 205 |
-
)
|
| 206 |
-
print( f"dataset loaded: test-{len(test_examples)}")
|
| 207 |
-
|
| 208 |
-
x_test, y_test = get_feature(model, test_dataloader, args)
|
| 209 |
-
|
| 210 |
-
else:
|
| 211 |
-
prot_tokenizer = EsmTokenizer.from_pretrained(args.prot_encoder_path, do_lower_case=False)
|
| 212 |
-
# prot_tokenizer = BertTokenizer.from_pretrained(args.prot_encoder_path, do_lower_case=False)
|
| 213 |
-
print("prot_tokenizer", len(prot_tokenizer))
|
| 214 |
-
disease_tokenizer = BertTokenizer.from_pretrained(args.disease_encoder_path)
|
| 215 |
-
print("disease_tokenizer", len(disease_tokenizer))
|
| 216 |
-
|
| 217 |
-
prot_model = EsmModel.from_pretrained(args.prot_encoder_path)
|
| 218 |
-
# prot_model = BertModel.from_pretrained(args.prot_encoder_path)
|
| 219 |
-
disease_model = BertModel.from_pretrained(args.disease_encoder_path)
|
| 220 |
-
|
| 221 |
-
if args.save_model_path:
|
| 222 |
-
model = GDA_Metric_Learning(prot_model, disease_model, 1280, 768, args)
|
| 223 |
-
|
| 224 |
-
if args.use_adapter:
|
| 225 |
-
prot_model_path = os.path.join(
|
| 226 |
-
args.save_model_path, f"prot_adapter_step_{args.step}"
|
| 227 |
-
)
|
| 228 |
-
disease_model_path = os.path.join(
|
| 229 |
-
args.save_model_path, f"disease_adapter_step_{args.step}"
|
| 230 |
-
)
|
| 231 |
-
model.load_adapters(prot_model_path, disease_model_path)
|
| 232 |
-
else:
|
| 233 |
-
prot_model_path = os.path.join(
|
| 234 |
-
args.save_model_path, f"step_{args.step}_model.bin"
|
| 235 |
-
)# , f"step_{args.step}_model.bin"
|
| 236 |
-
disease_model_path = os.path.join(
|
| 237 |
-
args.save_model_path, f"step_{args.step}_model.bin"
|
| 238 |
-
)
|
| 239 |
-
model.non_adapters(prot_model_path, disease_model_path)
|
| 240 |
-
|
| 241 |
-
model = model.to(args.device)
|
| 242 |
-
prot_model = model.prot_encoder
|
| 243 |
-
disease_model = model.disease_encoder
|
| 244 |
-
print(f"loaded prior model {args.save_model_path}.")
|
| 245 |
-
|
| 246 |
-
def collate_fn_batch_encoding(batch):
|
| 247 |
-
query1, query2, scores = zip(*batch)
|
| 248 |
-
|
| 249 |
-
query_encodings1 = prot_tokenizer.batch_encode_plus(
|
| 250 |
-
list(query1),
|
| 251 |
-
max_length=512,
|
| 252 |
-
padding="max_length",
|
| 253 |
-
truncation=True,
|
| 254 |
-
add_special_tokens=True,
|
| 255 |
-
return_tensors="pt",
|
| 256 |
-
)
|
| 257 |
-
query_encodings2 = disease_tokenizer.batch_encode_plus(
|
| 258 |
-
list(query2),
|
| 259 |
-
max_length=512,
|
| 260 |
-
padding="max_length",
|
| 261 |
-
truncation=True,
|
| 262 |
-
add_special_tokens=True,
|
| 263 |
-
return_tensors="pt",
|
| 264 |
-
)
|
| 265 |
-
scores = torch.tensor(list(scores))
|
| 266 |
-
attention_mask1 = query_encodings1["attention_mask"].bool()
|
| 267 |
-
attention_mask2 = query_encodings2["attention_mask"].bool()
|
| 268 |
-
|
| 269 |
-
return query_encodings1["input_ids"], attention_mask1, query_encodings2["input_ids"], attention_mask2, scores
|
| 270 |
-
|
| 271 |
-
train_examples = disGeNET.get_train_examples(args.test)
|
| 272 |
-
print(f"get training examples: {len(train_examples)}")
|
| 273 |
-
valid_examples = disGeNET.get_val_examples(args.test)
|
| 274 |
-
print(f"get validation examples: {len(valid_examples)}")
|
| 275 |
-
test_examples = disGeNET.get_test_examples(args.test)
|
| 276 |
-
print(f"get test examples: {len(test_examples)}")
|
| 277 |
-
|
| 278 |
-
train_dataloader = DataLoader(
|
| 279 |
-
train_examples,
|
| 280 |
-
batch_size=args.batch_size,
|
| 281 |
-
shuffle=False,
|
| 282 |
-
collate_fn=collate_fn_batch_encoding,
|
| 283 |
-
)
|
| 284 |
-
valid_dataloader = DataLoader(
|
| 285 |
-
valid_examples,
|
| 286 |
-
batch_size=args.batch_size,
|
| 287 |
-
shuffle=False,
|
| 288 |
-
collate_fn=collate_fn_batch_encoding,
|
| 289 |
-
)
|
| 290 |
-
test_dataloader = DataLoader(
|
| 291 |
-
test_examples,
|
| 292 |
-
batch_size=args.batch_size,
|
| 293 |
-
shuffle=False,
|
| 294 |
-
collate_fn=collate_fn_batch_encoding,
|
| 295 |
-
)
|
| 296 |
-
print( f"dataset loaded: train-{len(train_examples)}; valid-{len(valid_examples)}; test-{len(test_examples)}")
|
| 297 |
-
|
| 298 |
-
x_train, y_train = get_feature(model, train_dataloader, args)
|
| 299 |
-
x_valid, y_valid = get_feature(model, valid_dataloader, args)
|
| 300 |
-
x_test, y_test = get_feature(model, test_dataloader, args)
|
| 301 |
-
|
| 302 |
-
# Save input feature to reduce encoding time
|
| 303 |
-
np.savez_compressed(
|
| 304 |
-
input_feat_file,
|
| 305 |
-
x_train=x_train,
|
| 306 |
-
y_train=y_train,
|
| 307 |
-
x_valid=x_valid,
|
| 308 |
-
y_valid=y_valid,
|
| 309 |
-
)
|
| 310 |
-
print(f"save input feature into {input_feat_file}")
|
| 311 |
-
# Save input feature to reduce encoding time
|
| 312 |
-
return x_train, y_train, x_valid, y_valid, x_test, y_test
|
| 313 |
-
|
| 314 |
-
|
| 315 |
-
def train(args):
|
| 316 |
-
# defining parameters
|
| 317 |
-
if args.save_model_path:
|
| 318 |
-
args.model_short = (
|
| 319 |
-
args.save_model_path.split("/")[-1]
|
| 320 |
-
)
|
| 321 |
-
print(f"model name {args.model_short}")
|
| 322 |
-
|
| 323 |
-
else:
|
| 324 |
-
args.model_short = (
|
| 325 |
-
args.disease_encoder_path.split("/")[-1]
|
| 326 |
-
)
|
| 327 |
-
print(f"model name {args.model_short}")
|
| 328 |
-
|
| 329 |
-
# disGeNET = DisGeNETProcessor()
|
| 330 |
-
disGeNET = DisGeNETProcessor(input_csv_path=args.input_csv_path)
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
x_train, y_train, x_valid, y_valid, x_test, y_test = encode_pretrained_feature(args, disGeNET)
|
| 334 |
-
|
| 335 |
-
print("train: ", x_train.shape, y_train.shape)
|
| 336 |
-
print("valid: ", x_valid.shape, y_valid.shape)
|
| 337 |
-
print("test: ", x_test.shape, y_test.shape)
|
| 338 |
-
|
| 339 |
-
params = {
|
| 340 |
-
"task": "train", # "predict" train
|
| 341 |
-
"boosting": "gbdt", # "The options are "gbdt" (traditional Gradient Boosting Decision Tree), "rf" (Random Forest), "dart" (Dropouts meet Multiple Additive Regression Trees), or "goss" (Gradient-based One-Side Sampling). The default is "gbdt"."
|
| 342 |
-
"objective": "binary",
|
| 343 |
-
"num_leaves": args.num_leaves,
|
| 344 |
-
"early_stopping_round": 30,
|
| 345 |
-
"max_depth": args.max_depth,
|
| 346 |
-
"learning_rate": args.lr,
|
| 347 |
-
"metric": "binary_logloss", #"metric": "l2","binary_logloss" "auc"
|
| 348 |
-
"verbose": 1,
|
| 349 |
-
}
|
| 350 |
-
|
| 351 |
-
lgb_train = lgb.Dataset(x_train, y_train)
|
| 352 |
-
lgb_valid = lgb.Dataset(x_valid, y_valid)
|
| 353 |
-
lgb_eval = lgb.Dataset(x_test, y_test, reference=lgb_train)
|
| 354 |
-
|
| 355 |
-
# fitting the model
|
| 356 |
-
model = lgb.train(
|
| 357 |
-
params, train_set=lgb_train, valid_sets=lgb_valid)
|
| 358 |
-
|
| 359 |
-
# prediction
|
| 360 |
-
valid_y_pred = model.predict(x_valid)
|
| 361 |
-
test_y_pred = model.predict(x_test)
|
| 362 |
-
|
| 363 |
-
# predict liver fibrosis
|
| 364 |
-
predictions_df = pd.DataFrame(test_y_pred, columns=["Prediction_score"])
|
| 365 |
-
# data_test = pd.read_csv('/nfs/dpa_pretrain/data/downstream/GDA_Data/test_tdc.csv')
|
| 366 |
-
data_test = pd.read_csv(args.input_csv_path)
|
| 367 |
-
predictions = pd.concat([data_test, predictions_df], axis=1)
|
| 368 |
-
# filtered_dataset = test_dataset_with_predictions[test_dataset_with_predictions['diseaseId'] == 'C0009714']
|
| 369 |
-
predictions.sort_values(by='Prediction_score', ascending=False, inplace=True)
|
| 370 |
-
top_100_predictions = predictions.head(100)
|
| 371 |
-
top_100_predictions.to_csv(args.output_csv_path, index=False)
|
| 372 |
-
|
| 373 |
-
# Accuracy
|
| 374 |
-
y_pred = model.predict(x_test, num_iteration=model.best_iteration)
|
| 375 |
-
y_pred[y_pred >= 0.5] = 1
|
| 376 |
-
y_pred[y_pred < 0.5] = 0
|
| 377 |
-
accuracy = accuracy_score(y_test, y_pred)
|
| 378 |
-
|
| 379 |
-
# AUC
|
| 380 |
-
valid_roc_auc_score = metrics.roc_auc_score(y_valid, valid_y_pred)
|
| 381 |
-
valid_average_precision_score = metrics.average_precision_score(
|
| 382 |
-
y_valid, valid_y_pred
|
| 383 |
-
)
|
| 384 |
-
test_roc_auc_score = metrics.roc_auc_score(y_test, test_y_pred)
|
| 385 |
-
test_average_precision_score = metrics.average_precision_score(y_test, test_y_pred)
|
| 386 |
-
|
| 387 |
-
# AUPR
|
| 388 |
-
valid_aupr = metrics.average_precision_score(y_valid, valid_y_pred)
|
| 389 |
-
test_aupr = metrics.average_precision_score(y_test, test_y_pred)
|
| 390 |
-
|
| 391 |
-
# Fmax
|
| 392 |
-
valid_precision, valid_recall, valid_thresholds = precision_recall_curve(y_valid, valid_y_pred)
|
| 393 |
-
valid_fmax = (2 * valid_precision * valid_recall / (valid_precision + valid_recall)).max()
|
| 394 |
-
test_precision, test_recall, test_thresholds = precision_recall_curve(y_test, test_y_pred)
|
| 395 |
-
test_fmax = (2 * test_precision * test_recall / (test_precision + test_recall)).max()
|
| 396 |
-
|
| 397 |
-
# F1
|
| 398 |
-
valid_f1 = f1_score(y_valid, valid_y_pred >= 0.5)
|
| 399 |
-
test_f1 = f1_score(y_test, test_y_pred >= 0.5)
|
| 400 |
-
|
| 401 |
-
|
| 402 |
-
if __name__ == "__main__":
|
| 403 |
-
args = parse_config()
|
| 404 |
-
if torch.cuda.is_available():
|
| 405 |
-
print("cuda is available.")
|
| 406 |
-
print(f"current device {args}.")
|
| 407 |
-
else:
|
| 408 |
-
args.device = "cpu"
|
| 409 |
-
timestamp_str = datetime.now().strftime("%Y%m%d_%H%M%S")
|
| 410 |
-
random_str = "".join([random.choice(string.ascii_lowercase) for n in range(6)])
|
| 411 |
-
best_model_dir = (
|
| 412 |
-
f"{args.save_path_prefix}{args.save_name}_{timestamp_str}_{random_str}/"
|
| 413 |
-
)
|
| 414 |
-
os.makedirs(best_model_dir)
|
| 415 |
-
args.save_name = best_model_dir
|
| 416 |
-
train(args)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|