detail_release/cls_llm_layers.py at main · BobbyZhouZijian/detail_release · 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
import numpy as np
import os

from utils import (
    load_model,
    process_raw_data,
    get_context,
    compute_infl,
    remove_data,
    perturb_labels,
    check_answer,
)

import argparse


def get_data(data_input, data_output, bs=200, num_data=10):
    dataset = []
    for i in range(bs):
        cur_input = data_input[i*num_data:(i+1)*num_data]
        cur_output = data_output[i*num_data:(i+1)*num_data]
        cur_eval_input = data_input[i*num_data+num_data:i*num_data+num_data+1]
        cur_eval_output = data_output[i*num_data+num_data:i*num_data+num_data+1]
        cur_demo_data = (cur_input, cur_output)
        cur_eval_data = (cur_eval_input, cur_eval_output)
        dataset.append((cur_demo_data, cur_eval_data))
    return dataset


def run(num_remove, layer_num):
    acc_orig = []
    acc_rem_high = []
    acc_rem_low = []
    acc_rem_random = []

    for seed in range(num_trials):
        data_input, data_output, map_abcd_to_int, map_int_to_abcd = process_raw_data(dataset_name, seed, bs * (icl_dataset_size + 1))
        dataset = get_data(data_input, data_output, bs=bs, num_data=icl_dataset_size)

        num_success_orig = 0
        num_success_rem_high = 0
        num_success_rem_low = 0
        num_success_rem_random = 0


        for demo_data, eval_data in dataset:
            context = get_context(demo_data, eval_data)
            infls = compute_infl(
                model_name,
                model,
                tokenizer,
                map_abcd_to_int,
                demo_data,
                eval_data,
                [layer_num],
                project_dim=project_dim,
                device=device,
                alpha=1.0
            )

            # check the original context
            if check_answer(model, tokenizer, context, eval_data, device=device):
                num_success_orig += 1

            # remove the top {num_remove} data points with the highest influence
            remove_idx = np.argsort(infls)[-num_remove:]
            if corrupt:
                new_demo_data = perturb_labels(demo_data, remove_idx, map_int_to_abcd, map_abcd_to_int)
            else:
                new_demo_data = remove_data(demo_data, remove_idx)
            new_context = get_context(new_demo_data, eval_data)
            if check_answer(model, tokenizer, new_context, eval_data, device=device):
                num_success_rem_high += 1

            # remove the top {num_remove} data points with the lowest influence
            remove_idx = np.argsort(infls)[:num_remove]
            if corrupt:
                new_demo_data = perturb_labels(demo_data, remove_idx, map_int_to_abcd, map_abcd_to_int)
            else:
                new_demo_data = remove_data(demo_data, remove_idx)
            new_context = get_context(new_demo_data, eval_data)
            if check_answer(model, tokenizer, new_context, eval_data, device=device):
                num_success_rem_low += 1

            # remove {num_remove} random data points
            remove_idx = np.random.choice(len(demo_data[0]), num_remove, replace=False)
            if corrupt:
                new_demo_data = perturb_labels(demo_data, remove_idx, map_int_to_abcd, map_abcd_to_int)
            else:
                new_demo_data = remove_data(demo_data, remove_idx)
            new_context = get_context(new_demo_data, eval_data)
            if check_answer(model, tokenizer, new_context, eval_data, device=device):
                num_success_rem_random += 1

        cur_acc_orig = num_success_orig / bs
        cur_acc_rem_high = num_success_rem_high / bs
        cur_acc_rem_low = num_success_rem_low / bs
        cur_acc_rem_random = num_success_rem_random / bs

        acc_orig.append(cur_acc_orig)
        acc_rem_high.append(cur_acc_rem_high)
        acc_rem_low.append(cur_acc_rem_low)
        acc_rem_random.append(cur_acc_rem_random)

        print(f"Acc Original: {cur_acc_orig}")
        print(f"Acc Remove high: {cur_acc_rem_high}")
        print(f"Acc Remove low: {cur_acc_rem_low}")
        print(f"Acc Remove random: {cur_acc_rem_random}")

    return acc_orig, acc_rem_high, acc_rem_low, acc_rem_random


if __name__ == "__main__":

    args = argparse.ArgumentParser()
    args.add_argument("--layer_nums", type=int, nargs="+", default=[15])
    args.add_argument("--model", type=str, default="vicuna-7b")
    args.add_argument("--dataset_name", type=str, default="ag_news")
    args.add_argument("--project_dim", type=int, default=None)
    args.add_argument("--alpha", type=float, default=1.0)
    args.add_argument("--bs", type=int, default=100)
    args.add_argument("--corrupt", action="store_true")
    args = args.parse_args()

    device = "cuda"
    model_name = args.model
    icl_dataset_size = 20
    project_dim = args.project_dim
    bs = args.bs
    alpha = args.alpha
    num_trials = 10
    dataset_name = args.dataset_name
    save_dir = f"results/cls_llm_layers_{dataset_name}/"
    os.makedirs(save_dir, exist_ok=True)
    layer_nums = args.layer_nums
    corrupt = args.corrupt

    if model_name == "vicuna-7b":
        model_path = "lmsys/vicuna-7b-v1.3"
    elif model_name == "Llama-2-7b":
        model_path = "meta-llama/Llama-2-7b-chat-hf"
    elif model_name == "gpt2-xl":
        model_path = "gpt2-xl"
    elif model_name == "mistral-7b":
        model_path = "mistralai/Mistral-7B-v0.1"
    elif model_name == "falcon-7b":
        # model_path = "tiiuae/falcon-7b"
        model_path = "OpenBuddy/openbuddy-falcon-7b-v6-bf16"
    elif model_name == "wizardlm-7b":
        model_path = "WizardLM/WizardMath-7B-V1.1"
    else:
        raise NotImplementedError(f"Model {model_name} not implemented")

    model, tokenizer = load_model(model_path=model_path, device=device)
    # raw_dataset = load_dataset(dataset_name)

    for layer_num in args.layer_nums:
        num_remove = 10
        acc_orig, acc_rem_high, acc_rem_low, acc_rem_random = run(num_remove, layer_num)
        print(f"Finishied running for removing {num_remove} data points")
        # save the results
        np.save(os.path.join(save_dir, f"acc_orig_{model_name}_{num_remove}_{layer_num}_{'remove' if not corrupt else 'corrupt'}.npy"), acc_orig)
        np.save(os.path.join(save_dir, f"acc_rem_high_{model_name}_{num_remove}_{layer_num}_{'remove' if not corrupt else 'corrupt'}.npy"), acc_rem_high)
        np.save(os.path.join(save_dir, f"acc_rem_low_{model_name}_{num_remove}_{layer_num}_{'remove' if not corrupt else 'corrupt'}.npy"), acc_rem_low)
        np.save(os.path.join(save_dir, f"acc_rem_random_{model_name}_{num_remove}_{layer_num}_{'remove' if not corrupt else 'corrupt'}.npy"), acc_rem_random)