silk/fine-tuning.py at main · boudinfl/silk · GitHub

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
import sys
import argparse
from tqdm import tqdm
from datasets import load_dataset
from transformers import AutoTokenizer
from transformers import BartForConditionalGeneration, TrainingArguments
from transformers import Trainer
from transformers import DataCollatorForSeq2Seq


parser = argparse.ArgumentParser()

parser.add_argument("-m",
                    "--model",
                    default="taln-ls2n/bart-kp20k")

parser.add_argument("--token",
                    default="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")

parser.add_argument("--batch_size",
                    default=16,
                    type=int)

parser.add_argument("--num_train_epochs",
                    default=3,
                    type=int)

parser.add_argument("--few_shot",
                    default=1000,
                    type=int)

# 200 (100) 400 (1K) 800 (10K)
parser.add_argument("--warmup_steps",
                    default=400,
                    type=int)

parser.add_argument("--ordering",
                    default="top")

parser.add_argument("--seed",
                    default=42,
                    type=int)

parser.add_argument("--gradient_accumulation_steps",
                    default=1,
                    type=int)

parser.add_argument("--learning_rate",
                    default=1e-5,
                    type=float)

parser.add_argument("-i",
                    "--input")

parser.add_argument("-o",
                    "--output")

args = parser.parse_args()

print("Using model: {}".format(args.model))
print("Input file: {}".format(args.input))
print("Output model: {}".format(args.output))
outdir = args.output

dataset = load_dataset("json",
                       data_files=args.input)

args.few_shot = min(len(dataset['train']), args.few_shot)

ids = range(args.few_shot)
if args.ordering == "bottom":
    ids = reversed(range(len(dataset['train'])-args.few_shot, len(dataset['train'])))
elif args.ordering == "random":
    dataset = dataset.shuffle(seed=args.seed)

dataset['train'] = dataset['train'].select(ids)
print("Few-shot: {}".format(args.few_shot))

def join_titles_and_abstracts(dataset, special_token="<s>"):
    dataset["src"] = "{}<s>{}".format(dataset["title"], dataset["abstract"])
    return dataset

dataset = dataset.map(join_titles_and_abstracts)

def join_phrases(dataset, special_token=";"):
    dataset["tgt"] = special_token.join(dataset["keyphrases"])
    return dataset

dataset = dataset.map(join_phrases)

dataset = dataset.remove_columns(["title", "abstract", "keyphrases"])

# create a tokenizer
tokenizer = AutoTokenizer.from_pretrained(args.model,
                                          token=args.token)
def preprocess_function(dataset):

    model_inputs = tokenizer(dataset["src"],
                             max_length=512,
                             truncation=True,
                             padding=True)

    labels = tokenizer(dataset["tgt"],
                       max_length=128,
                       truncation=True,
                       padding=True)

    model_inputs["labels"] = labels["input_ids"]

    return model_inputs

tokenized_datasets = dataset.map(preprocess_function, batched=True)

tokenized_datasets.set_format("torch")

model = BartForConditionalGeneration.from_pretrained(args.model,
                                                     token=args.token)
args = TrainingArguments(
    output_dir=outdir,
    #overwrite_output_dir=True,
    # take parameters from Meng 2023 paper
    learning_rate=args.learning_rate,
    #max_steps=max_steps,
    warmup_steps=args.warmup_steps,
    optim="adamw_torch",
    #adam_beta1=0.9,
    #adam_beta2=0.999,
    #adam_epsilon=1e-8,
    #weight_decay=0.01,
    #max_grad_norm=0.1,
    do_train=True,
    per_device_train_batch_size=args.batch_size,
    #per_device_eval_batch_size=batch_size,
    #save_total_limit=10,
    num_train_epochs=args.num_train_epochs,
    #max_steps=max_steps,
    #logging_steps=logging_steps,
    push_to_hub=False,
    gradient_accumulation_steps=args.gradient_accumulation_steps,
    #evaluation_strategy="epoch", #steps
    #load_best_model_at_end=True,
    #save_strategy="epoch"
)

data_collator = DataCollatorForSeq2Seq(tokenizer, model)

trainer = Trainer(
    model=model,
    args=args,
    data_collator=data_collator,#NLPDataCollator(),
    train_dataset=tokenized_datasets["train"],
    #eval_dataset=tokenized_datasets["test"],
    #callbacks=[EarlyStoppingCallback(early_stopping_patience=5)]
)

result = trainer.train()

trainer.save_model(output_dir=outdir)
tokenizer.save_pretrained(save_directory=outdir)

print(result)