andrea.pedrotti
/
gfun_multimodal
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

set test key_prefix in test phase for wandb

This commit is contained in:
Andrea Pedrotti 2023-07-04 10:43:33 +02:00
parent 8354d76513
commit b6b1d33fdb
1 changed files with 12 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
hf_trainer.py
View File

@ -47,8 +47,8 @@ def main(args):
data = load_dataset(
"csv",
data_files = {
"train": expanduser("~/datasets/rai/csv/train-rai-multilingual-2000.csv"),
"test": expanduser("~/datasets/rai/csv/test-rai-multilingual-2000.csv")
"train": expanduser("~/datasets/rai/csv/train-split-rai.csv"),
"test": expanduser("~/datasets/rai/csv/test-split-rai.csv")
}
)
@ -81,7 +81,7 @@ def main(args):
process_sample_rai,
batched=True,
num_proc=4,
load_from_cache_file=False,
load_from_cache_file=True,
remove_columns=RAI_D_COLUMNS,
)
train_val_splits = data["train"].train_test_split(test_size=0.2, seed=42)
@ -103,7 +103,7 @@ def main(args):
recall_metric = evaluate.load("recall")
training_args = TrainingArguments(
output_dir=f"{args.model}-rai-multi-2000",
output_dir=f"hf_models/{args.model}-rai-fewshot",
do_train=True,
evaluation_strategy="steps",
per_device_train_batch_size=args.batch,
@ -123,13 +123,14 @@ def main(args):
fp16=args.fp16,
load_best_model_at_end=False if args.nosave else True,
save_strategy="no" if args.nosave else "steps",
save_total_limit=3,
save_total_limit=2,
eval_steps=args.stepeval,
run_name=f"{args.model}-rai-stratified",
disable_tqdm=False,
log_level="warning",
report_to=["wandb"] if args.wandb else "none",
optim="adamw_torch",
save_steps=args.stepeval
)
@ -163,7 +164,7 @@ def main(args):
if args.wandb:
import wandb
wandb.init(entity="andreapdr", project=f"gfun-rai-hf", name="mbert-sent", config=vars(args))
wandb.init(entity="andreapdr", project=f"gfun-rai-hf", name="mbert-rai", config=vars(args))
trainer = Trainer(
model=model,
@ -176,11 +177,11 @@ def main(args):
callbacks=callbacks,
)
# print("- Training:")
# trainer.train()
print("- Training:")
trainer.train()
print("- Testing:")
test_results = trainer.evaluate(eval_dataset=data["test"])
test_results = trainer.evaluate(eval_dataset=data["test"], metric_key_prefix="test")
print(test_results)
exit()
@ -191,8 +192,8 @@ if __name__ == "__main__":
parser = ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter)
parser.add_argument("--model", type=str, metavar="", default="mbert")
parser.add_argument("--nlabels", type=int, metavar="", default=28)
parser.add_argument("--lr", type=float, metavar="", default=1e-5, help="Set learning rate",)
parser.add_argument("--scheduler", type=str, metavar="", default="linear", help="Accepted: [\"cosine\", \"cosine-reset\", \"cosine-warmup\", \"cosine-warmup-reset\", \"constant\"]")
parser.add_argument("--lr", type=float, metavar="", default=1e-4, help="Set learning rate",)
parser.add_argument("--scheduler", type=str, metavar="", default="cosine", help="Accepted: [\"cosine\", \"cosine-reset\", \"cosine-warmup\", \"cosine-warmup-reset\", \"constant\"]")
parser.add_argument("--batch", type=int, metavar="", default=8, help="Set batch size")
parser.add_argument("--gradacc", type=int, metavar="", default=1, help="Gradient accumulation steps")
parser.add_argument("--epochs", type=int, metavar="", default=100, help="Set epochs")