code-eval/eval.py at main · laetitia-teo/code-eval · 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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
import os
import json
import copy
from tqdm import tqdm
import pathlib

import torch
import numpy as np

from transformers import AutoTokenizer, AutoModelForCausalLM


from utils_test import (
    return_full_prompt,
    prompt_train,
    pass_at_k,
    Prompt_Intstruction,
    judge_parallel,
)


def eval(config, run_id, trainer=None):  # TODO how to get the same run_id than the train if done separately

    print('Evaluating...')

    if config.slurm.gpu == "v100":
        type_use = torch.float16
    else:
        type_use = torch.bfloat16

    model_save_dir = f'outputs/{run_id}/save_models'
    name_json_save_all = f'outputs/{run_id}/save_results/passk.json'
    model_name = config.model_id.split('/')[-1]
    name_json_sol = f'outputs/{run_id}/save_sol/sols.json'
    output_dir = os.path.join(model_save_dir, f'{model_name}_{run_id}')

    sol_folder = str(pathlib.Path(name_json_sol).parent)
    if not os.path.exists(sol_folder):
        os.makedirs(sol_folder)

    if trainer is not None:
        trainer.accelerator.clear()

    import gc
    gc.collect()
    torch.cuda.empty_cache()
    for obj in gc.get_objects():
        if torch.is_tensor(obj):
            obj.cpu()
    gc.collect()
    torch.cuda.empty_cache()

    # testing
    tokenizer = AutoTokenizer.from_pretrained(output_dir)
    model = AutoModelForCausalLM.from_pretrained(
        output_dir,
        torch_dtype=type_use,

        # quantization_config=quantization_config,
        device_map="auto",
        local_files_only=True
    )
    tokenizer.padding_side = 'left'
    tokenizer.pad_token = tokenizer.eos_token
    model.eval()
    model.config.use_cache = True
    if config.compile:
        model = torch.compile(model)

    with open(config.test_path) as f:
        testset = json.load(f)
    curr_idx = 0
    correct_puzz = 0

    num_return_sequences = config.k #n_try
    list_all_passk=[[] for i in range(num_return_sequences)]
    list_passk=[]

    list_puzzle=[]
    list_all_puzzle=[]


    list_testset= [x["program_str"] for x in testset]
    print(f'Length of test set {len(list_testset)}')
    list_puzzle_correct=[]

    with torch.inference_mode():

        for idx in tqdm(range(curr_idx, len(list_testset), config.eval_batch_size)):
            print(f"\n\n============ idx {idx} ==================\n")

            list_prompt = []
            list_prompt_f = []
            subset_test = list_testset[idx:idx+config.eval_batch_size]

            for idx_puz in range(len(subset_test)):
                prompt_f = subset_test[idx_puz].split("def g(")[0]
                list_prompt_f.append(prompt_f)
                prompt = return_full_prompt(model_id=model_name, pb=prompt_f)
                list_prompt.append(prompt)

            inputs = tokenizer(list_prompt, return_tensors="pt", padding=True).to("cuda")

            len_prompt = inputs["input_ids"].shape[1]
            list_puzzle_gen=[[] for _ in range(len(list_prompt))]

            for idx_gen in range(num_return_sequences):
                outputs = model.generate(**inputs,max_new_tokens=512,do_sample=True, temperature=0.7)
                generated_texts = tokenizer.batch_decode(outputs[:,len_prompt:], skip_special_tokens=True)
                for idx_out_gen in range(len(outputs)):
                    list_puzzle_gen[idx_out_gen].append(generated_texts[idx_out_gen])

            list_generated_text = copy.deepcopy(list_puzzle_gen)

            for i in range(len(list_puzzle_gen)): # along the bs
                dic_save = {}
                list_raw_puzzle = []
                list_proc_puzzle = []
                for j in range(len(list_puzzle_gen[i])):
                    prompt_f = list_prompt_f[i]
                    try:
                        list_puzzle_gen[i][j] = list_puzzle_gen[i][j].replace("```python","```")
                        list_puzzle_gen[i][j] = list_puzzle_gen[i][j].replace("```Python","```")

                        if "```" in list_puzzle_gen[i][j]:
                            extract_g=list_puzzle_gen[i][j].split("```")[1].split("assert")[0]
                        else:
                            if "assert" in list_puzzle_gen[i][j]:
                                extract_g=list_puzzle_gen[i][j].split("assert")[0]
                            else:
                                extract_g=list_puzzle_gen[i][j]

                    except:
                        print("error extract g")
                        print(list_puzzle_gen[i][j])

                    extract_g = extract_g + "\nassert f(g()) == True\n"
                    test_fg= prompt_f + extract_g
                    list_puzzle_gen[i][j] = test_fg
                    list_puzzle.append(test_fg)
                    list_proc_puzzle.append(test_fg)
                    list_raw_puzzle.append(prompt_f + list_puzzle_gen[i][j])

                dic_save["raw_puzzle"] = list_raw_puzzle
                dic_save["process_puzzle"] = list_proc_puzzle

                list_valid_puzzles = judge_parallel(list_puzzle_gen[i])
                dic_save["list_valid"] = list_valid_puzzles
                list_all_puzzle.append(dic_save)

                cor_puz= np.sum(list_valid_puzzles)

                n_sample, n_correct = num_return_sequences,cor_puz
                pass_k = pass_at_k(n_sample, n_correct, k=num_return_sequences)
                list_passk.append(pass_k)

                for idx_passk in range(num_return_sequences):
                    pass2add = pass_at_k(n_sample, n_correct, k=idx_passk+1)
                    list_all_passk[idx_passk].append(pass2add)
                    testset[idx + i][f'pass_{idx_passk+1}'] = pass2add

                proba_solved = n_correct / n_sample
                testset[idx + i]['proba_solved'] = float(proba_solved)
                testset[idx + i]['n_sample'] = int(n_sample)
                testset[idx + i]['n_correct'] = int(n_correct)
                testset[idx + i]['generated_text'] = list_generated_text[i]
                testset[idx + i]['parsed_puzzles'] = list_puzzle_gen[i]

            print(f"correct puzzles: {int(np.sum(list_passk))}/{len(list_passk)}")

        for idx_passk in range(num_return_sequences):
            print(f"pass {idx_passk+1}: {np.sum(list_all_passk[idx_passk])}/{len(list_all_passk[idx_passk])}")
        dic_passk={}
        for idx_passk in range(num_return_sequences):
            dic_passk[f"pass_{idx_passk+1}"]=float(np.sum(list_all_passk[idx_passk]))

        with open(name_json_save_all, "r") as outfile:
            json_content = json.load(outfile)
        json_content[run_id] = dic_passk
        with open(name_json_save_all, "w") as outfile:
            json.dump(json_content, outfile, indent=4)

        with open(name_json_sol, "w") as outfile:
            json.dump(testset, outfile, indent=4)