Fints/train.py at main · DeepExperience/Fints · 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
import sys

sys.path.append("../../")
sys.path.append("../")
sys.path.append("./")

import torch
from tqdm import tqdm
import os
from ChatModels import (
    LlamaWrapper,
)
import argparse
from dataloader import (
    GenerationDataset,
)
import json

from instruction import SYS_PROMPT_SINGLE

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--layers", nargs="+", type=int, default=list(range(32)))
    parser.add_argument(
        "--data_path",
        type=str,
        default="../pa_back/caa_data",
    )
    parser.add_argument("--data_name", type=str, default="caa_python_LaMP_4_0.15_qwen_others")
    parser.add_argument("--model_name", type=str, default="llama-3.1")
    parser.add_argument("--system_prompt", type=str, default="")
    parser.add_argument("--model_name_or_path", type=str, default="../model_weights/fix/Meta-Llama-3.1-8B-Instruct")
    parser.add_argument("--rerun", action="store_true", default=False, help="Rerun even if files already exist")
    parser.add_argument("--act_location", type=str, choices=['whole', 'attn', 'mlp'], default='whole')

    args = parser.parse_args()
    print(args)
    if "llama" in args.model_name.lower():
        model = (
            LlamaWrapper(args.model_name_or_path)
        )
    elif "qwen" in args.model_name.lower():
        model = (
            QwenWrapper(args.model_name_or_path)
        )
    else:
        raise NotImplementedError(f"Model {args.model_name} not supported")
    tokenizer = model.tokenizer
    pos_activations = dict([(layer, []) for layer in args.layers])
    neg_activations = dict([(layer, []) for layer in args.layers])
    if args.act_location == 'whole':
        get_activations = model.get_last_activations
    elif args.act_location == 'attn':
        model.set_save_internal_decodings(True)
        get_activations = model.get_attn_activations
    elif args.act_location == 'mlp':
        model.set_save_internal_decodings(True)
        get_activations = model.get_mlp_activations

    dataset = GenerationDataset()

    vector_dataset_all = dataset.get_data_for_caa(
        data_path=args.data_path,
        data_name=args.data_name,
        split="train",
    )
    device = model.device
    if args.model_name == "gemma-2-9b" or args.model_name=="llama-3.1":
        output_dir = os.path.join(
            args.data_path, "caa_vector_pt", f"{args.model_name}_{args.data_name}_{args.act_location}"
        )
    else:
        output_dir = os.path.join(
            args.data_path, args.data_name, "caa_vector", f"{args.model_name}_{args.mode}"
        )
    for i, (uid, vector_dataset) in enumerate(vector_dataset_all.items()):
        print(i)
        pos_tokens_list, neg_tokens_list = [], []
        if os.path.exists(os.path.join(output_dir, f"{uid}_{args.layers[0]}.pt")) and not args.rerun:
            continue
        for i in range(len(vector_dataset)):
            ques = vector_dataset[i]["question"]
            chosen = vector_dataset[i]["chosen"]
            rejected = vector_dataset[i]["rejected"]

            if ques and chosen and rejected:
                if 'pwab' in args.data_name:
                    if 'pwab_pos' in args.data_name:
                        chosen = json.dumps(chosen)
                        rejected = json.dumps(rejected)
                    else:
                        ques = ""

                if args.model_name=="llama-3.1":
                    if args.system_prompt != "":
                        if ques is not None:
                            ques = args.system_prompt + " " + ques
                        else:
                            ques = args.system_prompt
                else:
                    raise NotImplementedError


                ques_tokens = tokenizer.encode(ques, return_tensors="pt")
                pos_tokens = tokenizer.encode(ques + chosen, return_tensors="pt")
                neg_tokens = tokenizer.encode(ques + rejected, return_tensors="pt")
                pos_tokens_list.append(
                    {
                        "pos_tokens": pos_tokens.to(device),
                        "ques_tokens_len": ques_tokens.shape[1],
                        "pos_answer_len": pos_tokens.shape[1] - ques_tokens.shape[1],
                    }
                )
                neg_tokens_list.append(
                    {
                        "neg_tokens": neg_tokens.to(device),
                        "ques_tokens_len": ques_tokens.shape[1],
                        "neg_answer_len": neg_tokens.shape[1] - ques_tokens.shape[1],
                    }
                )

        for p_tokens_dict, n_tokens_dict in tqdm(
            zip(pos_tokens_list, neg_tokens_list),
            total=len(pos_tokens_list),
            desc="Processing prompts",
        ):
            p_tokens = p_tokens_dict["pos_tokens"]
            n_tokens = n_tokens_dict["neg_tokens"]
            ques_tokens_len = p_tokens_dict["ques_tokens_len"]

            # Get positive logits
            model.reset_all()
            model.get_logits(p_tokens)

            for layer in args.layers:
                p_activations = get_activations(layer)
                p_activations = p_activations[0, ques_tokens_len:, :].mean(0).detach()
                pos_activations[layer].append(p_activations.cpu())

            # Get negative logits
            model.reset_all()
            model.get_logits(n_tokens)

            for layer in args.layers:
                n_activations = get_activations(layer)
                n_activations = n_activations[0, ques_tokens_len:, :].mean(0).detach()
                neg_activations[layer].append(n_activations.cpu())

        if not os.path.exists(output_dir):
            os.makedirs(output_dir)

        for layer in args.layers:
            if (len(pos_activations[layer]) == 0):
                p_activations = model.get_last_activations(layer)
                vec = torch.zeros(4096)
            else:
                all_pos_layer = torch.stack(pos_activations[layer])
                all_neg_layer = torch.stack(neg_activations[layer])
                vec = (all_pos_layer - all_neg_layer).mean(dim=0).to(model.device)

            torch.save(
                vec.cpu(),
                os.path.join(output_dir, f"{uid}_{layer}.pt"),
            )