-
Notifications
You must be signed in to change notification settings - Fork 0
/
inference.py
167 lines (135 loc) · 6.64 KB
/
inference.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
import os
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from datasets import load_dataset, Dataset, DatasetDict
from tqdm import tqdm
import argparse
from accelerate.utils import set_seed
import logging
from utils.utils import LANG_TABLE
from utils.build_dataset import get_inter_prompt
logger = logging.getLogger(__name__)
def load_pair_dataset(pair, test_file_path):
test_raw_data = {}
src_lang = pair.split("-")[0]
tgt_lang = pair.split("-")[1]
test_file = os.path.join(test_file_path)
test_raw_data[f"{src_lang}-{tgt_lang}"] = load_dataset(
"json",
data_files={"test": test_file}
)
return test_raw_data
def get_pair_suffix(tgt_lang):
return f"\nFinal {LANG_TABLE[tgt_lang]} Translation: "
def get_plain_suffix(tgt_lang):
return f"\n{LANG_TABLE[tgt_lang]}:"
def clean_outputstring(output, key_word, logger, split_idx):
try:
out = output.split(key_word)[split_idx].split("\n")
if out[0].strip() != "":
return out[0].strip()
elif out[1].strip() != "":
## If there is an EOL directly after the suffix, ignore it
logger.info(f"Detect empty output, we ignore it and move to next EOL: {out[1].strip()}")
return "-------------------"
else:
logger.info(f"Detect empty output AGAIN, we ignore it and move to next EOL: {out[2].strip()}")
return "-------------------"
except:
logger.info(f"Can not recover the translation by moving to the next EOL.. Trying move to the next suffix")
def get_plain_prompt(source_lang, target_lang, ex, shots_eval_dict):
src_fullname = LANG_TABLE[source_lang]
tgt_fullname = LANG_TABLE[target_lang]
prefix = f"Translate this from {src_fullname} to {tgt_fullname}:\n{src_fullname}: "
suffix = f"\n{tgt_fullname}:"
prompt = prefix + ex[source_lang] + suffix
return prompt
# Function to generate model outputs based on the input data
def generate_model_outputs(row, model):
input_ids = torch.tensor(row['input_ids']).unsqueeze(dim=0).to('cuda')
with torch.no_grad():
generate_ids = model.generate(input_ids=input_ids, num_beams=5, max_new_tokens=256, do_sample=True, temperature=0.6, top_p=0.9)
model_output = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
return model_output
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--base-model", type=str, default=None, help="The name of model to use.");
parser.add_argument("--peft-path", type=str, default=None, help="The name of model to use.");
parser.add_argument("--prompt-strategy", type=str, default="intermediate", help="intermediate, plain..");
parser.add_argument("--test-pairs", type=str, default="", help="en-zh,de-en... no space");
parser.add_argument("--test-dir", type=str, default="./pair_corpus");
parser.add_argument("--test-file-path", type=str, default="./pair_corpus");
parser.add_argument("--output-dir", type=str, default=None);
parser.add_argument("--output-file-prefix", type=str, default="test");
parser.add_argument("--seed", type=int, default=0)
args = parser.parse_args()
set_seed(args.seed)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
print(args.output_dir)
# test_pairs = args.test_pairs.split(",")
test_pairs = args.test_pairs
print(test_pairs)
test_raw_data = load_pair_dataset(test_pairs, args.test_file_path)
source_lang = test_pairs.split("-")[0]
target_lang = test_pairs.split("-")[1]
# Load base model and LoRA weights
model = AutoModelForCausalLM.from_pretrained(args.base_model, torch_dtype='auto', device_map = 'auto')
if args.peft_path:
model = PeftModel.from_pretrained(model, args.peft_path, torch_dtype='auto', device_map = 'auto')
model.merge_and_unload()
tokenizer = AutoTokenizer.from_pretrained(args.base_model, padding_side='left')
def tokenize_function_test(examples):
prompts = []
targets = []
for ex in examples['translation']:
if args.prompt_strategy == 'intermediate':
shots_eval_dict = ex['shots']
prompt = get_inter_prompt(source_lang, target_lang, ex, shots_eval_dict)
elif args.prompt_strategy == 'plain':
shots_eval_dict = {}
prompt = get_plain_prompt(source_lang, target_lang, ex, shots_eval_dict)
prompts.append(prompt)
targets.append(prompt + ex[target_lang])
original_padding_side = tokenizer.padding_side
if original_padding_side != "left":
tokenizer.padding_side = "left"
model_inputs = tokenizer(prompts, max_length=1024, padding=True, truncation=True, add_special_tokens=True)
return model_inputs
test_datasets = {}
for lg_pair, sub_raw_data in test_raw_data.items():
test_dataset = sub_raw_data["test"]
test_dataset = test_dataset.map(
tokenize_function_test,
batched=True,
num_proc=8,
remove_columns=["translation"],
desc=f"Running tokenizer {lg_pair} test dataset",
)
test_datasets[lg_pair] = test_dataset
input_ids = test_datasets[lg_pair][0]['input_ids']
decode_check = tokenizer.decode(input_ids, skip_special_tokens=True)
print("-"*50 + "check input" + "-"*50)
print(decode_check)
print("-"*50 + "check input" + "-"*50)
for idx in tqdm(range(len(test_dataset)), desc="Generating Responses"):
row = test_dataset[idx]
output = generate_model_outputs(row, model)
if idx == 0:
print("-" * 50 + "output example" + "-" * 50)
print(output)
print("-" * 50 + "output example" + "-" * 50)
with open(os.path.join(args.output_dir, f"{args.output_file_prefix}-{lg_pair}"), "a", encoding="utf-8") as f:
target_lan = lg_pair.split("-")[1]
if args.prompt_strategy == 'intermediate':
suffix = get_pair_suffix(target_lan)
elif args.prompt_strategy == 'plain':
suffix = get_plain_suffix(target_lan)
suffix_count = output.count(suffix)
split_idx = suffix_count
pred = clean_outputstring(output, suffix, logger, split_idx)
try:
f.writelines([pred, "\n"])
except:
f.writelines(["None", "\n"])