/
h2.py
196 lines (175 loc) · 6.69 KB
/
h2.py
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
196
""" h2 == h0 with aggresive name removing (global naming) """
import sys
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer, DataCollatorForLanguageModeling
from datasets import Dataset, load_dataset
import evaluate
import numpy as np
import os
import ast
import torch
import astunparse
CLSN = "[CLSN]"
INIT = "[INIT]"
NOARG = "[NOARG]"
os.environ["TOKENIZERS_PARALLELISM"] = "true"
class NameRemover(ast.NodeVisitor):
''' TODO; reverse mode or renaming back '''
def __init__(self) -> None:
super().__init__()
self.reset()
def reset(self):
self.mapping = {}
self.class_mapping = {}
self.rev_mapping = {}
self.rev_class_mapping = {}
self.i = 0
self.cls_i = 0
def reset_class(self):
self.class_mapping = {}
self.rev_class_mapping = {}
self.cls_i = 0
def add_to_vocab(self, name, token = None):
if name not in self.mapping:
new_name = self.mapping[name] = token or ("[v" + str(self.i) + "]")
self.rev_mapping[new_name] = name
self.i += 1
return self.mapping[name]
def visit_Name(self, node):
node.id = self.add_to_vocab(node.id)
def generic_visit(self, node):
if isinstance(node, ast.ClassDef):
if node.name not in self.class_mapping:
new_name = self.class_mapping[node.name] = "[CLS" + str(self.cls_i) + "]"
self.rev_class_mapping[new_name] = node.name
self.cls_i += 1
node.name = self.class_mapping[node.name]
elif hasattr(node, 'name'):
node.name = self.add_to_vocab(node.name)
elif hasattr(node, "args") and type(node.args) == list:
for arg in node.args:
if isinstance(arg, ast.arg):
arg.arg = self.add_to_vocab(arg.arg)
return super().generic_visit(node)
name_remover = NameRemover()
name_symbols = set()
def process(line: str):
name_remover.reset_class()
tree = ast.parse(line)
name_remover.visit(tree)
name_symbols.update(name_remover.rev_mapping.keys())
name_symbols.update(name_remover.rev_class_mapping.keys())
res = astunparse.unparse(tree).strip().replace("\n\n", "\n").replace(" ", "\t")
return res.replace(".__init__", INIT).replace("()", NOARG)
import re
symbol_pattern = r"\[(CLS\d+|v\d+)\]"
def unprocess(line: str):
res = re.sub(symbol_pattern, r"\1", line.replace(INIT, ".__init__").replace(NOARG, "()"))
return res #we preserve name of symbols but remove []
ds_name = "dvitel/hearthstone"
out_dir = "out/h2"
result_path = "result/h2"
checkpoint = "distilgpt2"
max_length = 912
batch_size = 4
num_epochs = 200
eval_steps = 1600
learning_rate = 2e-5
seed = 17
np.random.seed(seed)
torch.manual_seed(seed)
def normalize(line:str):
return line.strip().replace("§", "\n").replace(" ", "\t").replace("\\ ", "").replace("\n\n", "\n")
def preprocess0(e):
return {"source":e["source"], "target":[process(normalize(x)) for x in e["target"]]}
ds = load_dataset(ds_name)
ds0 = ds.map(preprocess0, batched = True)
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.add_special_tokens({'additional_special_tokens':[CLSN, INIT, NOARG, *list(name_symbols)]})
def preprocess(e):
alt_bodies = []
for s, t in zip(e["source"], e["target"]):
tgt = t # t.replace("(", "[LPAR]").replace(")", "[RPAR]").replace("[RPAR] [LPAR]", "[RPAR][LPAR]")
alt_bodies.append(s + tokenizer.eos_token + tgt)
# print(alt_bodies)
data = tokenizer(alt_bodies, padding = "max_length", truncation = True, max_length = max_length)
return data
ds1 = ds0.map(preprocess, batched = True, remove_columns = ["source", "target"])
model = AutoModelForCausalLM.from_pretrained(checkpoint, n_ctx = max_length, max_length = max_length)
model.resize_token_embeddings(len(tokenizer))
model.to("cuda")
bleu = evaluate.load("bleu")
codebleu = evaluate.load("dvitel/codebleu")
chrF = evaluate.load("chrf")
exact_match = evaluate.load("exact_match")
def compute_metrics(eval_pred):
shift_labels = eval_pred.label_ids[...,1:]
shift_logits = eval_pred.predictions[..., :-1, :]
prediction_labels = np.argmax(shift_logits, axis=-1)
predictions = []
references = []
first_not_matched = 4
for preds, labels in zip(prediction_labels, shift_labels):
label_map = labels >= 0
labels_view = labels[label_map]
pred_view = preds[label_map]
p_text = unprocess(tokenizer.decode(pred_view))
l_text = unprocess(tokenizer.decode(labels_view))
predictions.append(p_text)
references.append(l_text)
if p_text != l_text and first_not_matched > 0:
print("EV L", l_text)
print("EV P", p_text)
print()
first_not_matched -= 1
accuracy_metric = exact_match.compute(predictions = predictions, references = references)
bleu_metric = bleu.compute(predictions = predictions, references = references)
codebleu_metric = codebleu.compute(predictions = predictions, references = references)
chrf_metric = chrF.compute(predictions = predictions, references = references)
return {"exact_match": accuracy_metric["exact_match"], "bleu": bleu_metric["bleu"], **codebleu_metric, "chrf": chrf_metric['score']}
data_collator = DataCollatorForLanguageModeling(tokenizer, mlm = False)
eos_id = tokenizer.eos_token_id
def custom_data_collator(*args):
''' we do not need to deduce preefix parts - change all labels till first -100 to -100 '''
res = data_collator(*args)
for l in res['labels']:
i = 0
while l[i] != -100:
l[i] = -100
i += 1
return res
args = TrainingArguments(
output_dir=out_dir, overwrite_output_dir = True,
per_device_train_batch_size=batch_size,
per_device_eval_batch_size=batch_size,
evaluation_strategy="steps",
num_train_epochs = num_epochs,
logging_steps=eval_steps,
eval_steps = eval_steps,
eval_accumulation_steps = 4,
gradient_accumulation_steps=1,
weight_decay=0.1,
# warmup_steps=1_000,
lr_scheduler_type="cosine",
learning_rate=learning_rate,
save_steps=eval_steps,
fp16=True,
load_best_model_at_end = True,
metric_for_best_model = "exact_match",
seed = seed, push_to_hub = True,
hub_model_id = "h2"
)
trainer = Trainer(
model=model,
tokenizer=tokenizer,
args=args,
compute_metrics = compute_metrics,
data_collator=custom_data_collator,
train_dataset=ds1["train"],
eval_dataset=ds1["validation"],
)
trainer.train()
output = trainer.predict(ds1["test"])
print(output.metrics) #test set metrics
# trainer.save_model(result_path)
trainer.push_to_hub()