File size: 6,509 Bytes
df96e38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
197
import jsonlines
import argparse
import pprint
import sys
import os
import re
from tqdm import tqdm
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
from human_eval.data import write_jsonl, read_problems, stream_jsonl

if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"

try:
    if torch.backends.mps.is_available():
        device = "mps"
except:
    pass

def read_mbpp(path):
    mbpp_problems = {}
    with jsonlines.open(path, "r") as fin:
        for obj in fin:
            mbpp_problems[obj["task_id"]] = obj
    return mbpp_problems

def extract_text(prompt, remove_lines=True):
    token = '\"\"\"'
    start = token
    end = '>>>'

    start_idx = prompt.find(start) + len(start)
    end_idx = prompt.find(end)

    output = prompt[start_idx: end_idx]
    if remove_lines:
        output = output.replace('\n', ' ')
    output = re.sub(r"\s+", " ", output).strip()

    return output

def generate_prompt(input):
    INSTRUCTION = f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.



### Instruction:

Create a Python script for this problem:

{input}



### Response:"""
    return INSTRUCTION

def get_model(

    load_8bit: bool = False,

    base_model: str = "bigcode/starcoder",

):
    assert base_model, (
        "Please specify a --base_model, e.g. --base_model='bigcode/starcoder'"
    )

    tokenizer = AutoTokenizer.from_pretrained(base_model)
    if device == "cuda":
        model = AutoModelForCausalLM.from_pretrained(
            base_model,
            load_in_8bit=load_8bit,
            torch_dtype=torch.float16,
            device_map="auto",
        )
    elif device == "mps":
        model = AutoModelForCausalLM.from_pretrained(
            base_model,
            device_map={"": device},
            torch_dtype=torch.float16,
        )
    model.config.pad_token_id = tokenizer.pad_token_id

    if not load_8bit:
        model.half()  # seems to fix bugs for some users.

    model.eval()
    if torch.__version__ >= "2" and sys.platform != "win32":
        model = torch.compile(model)
    
    return tokenizer, model


def main():
    parser = argparse.ArgumentParser()

    parser.add_argument('--model', type=str, default='bigcode/starcoder', help="")
    parser.add_argument('--output_path', type=str, help="")
    parser.add_argument('--start_index', type=int, default=0, help="")
    parser.add_argument('--end_index', type=int, default=164, help="")
    parser.add_argument('--temperature', type=float, default=0.8, help="")
    parser.add_argument('--N', type=int, default=200, help="")
    parser.add_argument('--max_len', type=int, default=512, help="")
    parser.add_argument('--decoding_style', type=str, default='sampling', help="")
    parser.add_argument('--num_seqs_per_iter', type=int, default=50, help='')
    parser.add_argument('--overwrite', action='store_true', help='')
    parser.add_argument('--mbpp_path', type=str, help="")

    args = parser.parse_args()

    argsdict = vars(args)
    print(pprint.pformat(argsdict))

    STOP_SEQS = ['\nclass', '\ndef', '\n#', '\nif', '\nprint']

    problems = read_mbpp(args.mbpp_path)

    task_ids = sorted(problems.keys())[args.start_index: args.end_index]
    prompts = []
    for task_id in task_ids:
        prompt = f"\n{problems[task_id]['text']}\nTest examples:"
        if task_id == 493:
            # The test examples are too long. We choose to only include the function name.
            test_example = problems[task_id]['test_list'][0]
            prompt += f"\ncalculate_polygons(startx, starty, endx, endy, radius)"
        else:
            for test_example in problems[task_id]['test_list']:
                prompt += f"\n{test_example}"
        prompts.append(prompt)
    
    num_samples = len(prompts)
    print("Number of samples: {}".format(num_samples))

    tokenizer, model = get_model(base_model=args.model)
    generation_config = GenerationConfig(
        pad_token_id=tokenizer.pad_token_id,
        do_sample=True,
        temperature=args.temperature,
        max_length=args.max_len,
        num_return_sequences=args.num_seqs_per_iter,
        eos_token_id=tokenizer.eos_token_id,
        top_p=0.95
    )

    print(f"Loaded {args.model}.")
    for i in tqdm(range(num_samples), ncols=0, total=num_samples):
        output_file = args.output_path + '/{}.jsonl'.format(args.start_index + i)

        if os.path.exists(output_file) and not args.overwrite:
            print(f'Skip {output_file} as it already exists')
            continue

        prompt = prompts[i].replace('    ', '\t')
        prompt_batch = [generate_prompt(prompt)]

        ids_batch = [task_ids[i]]

        completion_seqs = []

        encoding = tokenizer(prompt_batch, return_tensors="pt", truncation=True, max_length=args.max_len).to(device)

        if args.decoding_style == 'sampling':
            loops = int(args.N / args.num_seqs_per_iter)
        else:
            loops = 1

        for _ in tqdm(range(loops), total=loops, leave=False, ncols=0):

            with torch.no_grad():
                if args.decoding_style == 'sampling':
                    gen_tokens = model.generate(
                        **encoding,
                        generation_config=generation_config
                    )

            if gen_tokens is not None:
                gen_seqs = tokenizer.batch_decode(gen_tokens, skip_special_tokens=True)
            else:
                gen_seqs = None

            if gen_seqs is not None:
                assert len(ids_batch) == 1
                task_id = ids_batch[0]

                for seq_idx, gen_seq in enumerate(gen_seqs):
                    completion_seq = gen_seq.split("### Response:")[-1]
                    completion_seq = completion_seq.replace('\t', '    ')
                    all_code = gen_seq.replace('\t', '    ')

                    completion_seqs.append(
                        {'task_id': task_id,
                         'completion': completion_seq,
                         'all_code': all_code,
                         }
                    )

        print("Saving results to {}".format(output_file))
        write_jsonl(output_file, completion_seqs)


if __name__ == '__main__':
    main()