lora-attention-interpretability/layer_trends.py at main · mattperls-code/lora-attention-interpretability · 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
import json
import os
import matplotlib.pyplot as plt

def from_heatmap_file(path: str):
    with open(path, "r") as heatmap_file:
        heatmap_data = json.load(heatmap_file)

        return [
            sum(layer) / len(layer)
            for layer in heatmap_data
        ]

def apply_windowing(layer_data: list[float], window_size: int):
    window_results = []

    for layer_index in range(len(layer_data) - window_size + 1):
        window_results.append(0)

        for window_offset in range(window_size):
            window_results[-1] += layer_data[layer_index + window_offset]

        window_results[-1] /= window_size

    return window_results

def plot_layer_data(title: str, ylabel: str, data: list[float], output_path: str, window_size: int = 1):
    plt.clf()

    plt.figure(figsize=(10, 8))
    plt.title(title, pad=20)
    plt.xlabel("Layer Index", labelpad=20)
    plt.ylabel(ylabel, labelpad=20)
    plt.axhline(y=0, color='black', linestyle='--')
    plt.grid(True, axis='both', color='lightgray')

    plt.plot(range(len(data)), data)

    if window_size > 1:
        plt.xticks(
            ticks=range(len(data)),
            labels=[f"{i + 1}-{i + window_size}" for i in range(len(data))],
            rotation=45,
            ha="right"
        )

    else:
        plt.xticks(
            ticks=range(len(data)),
            labels=range(1, len(data) + 1),
            rotation=45,
            ha="right"
        )

    plt.tight_layout()
    plt.savefig(output_path)
    plt.close()

if __name__ == "__main__":
    model_data_paths = [ "base-model", "ft-model" ]

    for model_data_path in model_data_paths:
        for feature_name in os.listdir(f"results/{model_data_path}/transformer-heatmaps"):
            if os.path.isdir(f"results/{model_data_path}/transformer-heatmaps/{feature_name}"):
                os.makedirs(f"results/{model_data_path}/layers/{feature_name}", exist_ok=True)

                feature_data_path = f"results/{model_data_path}/transformer-heatmaps/{feature_name}/normalized.json"

                feature_layer_data = from_heatmap_file(feature_data_path)

                with open(f"results/{model_data_path}/layers/{feature_name}/layer.json", "w") as layer_data_file:
                    json.dump(feature_layer_data, layer_data_file)

                plot_layer_data(
                    f"{feature_name}\n(Average Normalized Feature Attention Per Layer)",
                    "Average Normalized Feature Attention Per Layer",
                    feature_layer_data,
                    f"results/{model_data_path}/layers/{feature_name}/layer.png"
                )

                window_sizes = [ 2, 3, 4, 6 ]

                for window_size in window_sizes:
                    feature_window_data = apply_windowing(feature_layer_data, window_size)

                    with open(f"results/{model_data_path}/layers/{feature_name}/window-size{window_size}.json", "w") as window_data_file:
                        json.dump(feature_window_data, window_data_file)

                    plot_layer_data(
                        f"{feature_name}\n(Average Normalized Feature Attention Per {window_size} Layer Window)",
                        "Average Normalized Feature Attention Per Window",
                        feature_window_data,
                        f"results/{model_data_path}/layers/{feature_name}/window-size{window_size}.png",
                        window_size=window_size
                    )

    data_file_paths = [
        ("layer", "Layer", 1),
        ("window-size2", "Window", 2),
        ("window-size3", "Window", 3),
        ("window-size4", "Window", 4),
        ("window-size6", "Window", 6)
    ]

    for feature_name in os.listdir(f"results/base-model/layers"):
        if os.path.isdir(f"results/ft-model/layers/{feature_name}"):
            os.makedirs(f"results/model-diffs/layers/{feature_name}", exist_ok=True)

            for (data_file_path, grouping_name, window_size) in data_file_paths:
                with open(f"results/base-model/layers/{feature_name}/{data_file_path}.json", "r") as base_model_data_file:
                    with open(f"results/ft-model/layers/{feature_name}/{data_file_path}.json", "r") as ft_model_data_file:
                        base_model_data = json.load(base_model_data_file)
                        ft_model_data = json.load(ft_model_data_file)

                        model_diffs_data = [
                            ft_model_data[layer_index] - base_model_data[layer_index]
                            for layer_index in range(len(base_model_data))
                        ]

                        with open(f"results/model-diffs/layers/{feature_name}/{data_file_path}.json", "w") as diff_data_file:
                            json.dump(model_diffs_data, diff_data_file)

                        plot_layer_data(
                            f"{feature_name}\n(Difference Between Average Normalized Feature Attention in Fine Tuned vs Base Model)",
                            f"Normalized Feature Attention Difference Per {grouping_name}",
                            model_diffs_data,
                            f"results/model-diffs/layers/{feature_name}/{data_file_path}.png",
                            window_size=window_size
                        )