maze_vis_attn.py

import os
import argparse
import pickle
import torch
import math
import numpy as np
import random
from tqdm import tqdm
from torch.nn.utils.rnn import pad_sequence
import matplotlib.pyplot as plt

# Import project modules
from model.transformer import GPTConfig, GPT
from cli_utils import parse_count, format_count


def parse_args():
    parser = argparse.ArgumentParser(description='Probe Attention Weights towards the Target Token.')

    # --- Model & Data Configuration ---
    parser.add_argument('--ckpt_iter', type=int, default=10000)
    parser.add_argument('--config', type=str, default='6_6_384')
    parser.add_argument('--device', type=str, default='cuda:0')
    parser.add_argument('--num_nodes', type=int, default=100)
    parser.add_argument('--num_of_paths', type=int, default=50)
    parser.add_argument('--num_train_dataset', type=parse_count, default="9M")
    parser.add_argument('--tasks', type=str, default='C1', help='Tasks used for filename generation.')
    parser.add_argument('--CL', action=argparse.BooleanOptionalAction, default=False)
    parser.add_argument('--path_type', type=str, default='RWs', choices=['RWc', 'RWa', 'RWs'])
    parser.add_argument('--no_task_tag', action='store_true', default=False)
    parser.add_argument('--local', action='store_true', default=False)
    parser.add_argument('--NLS', action='store_true', default=False,
                        help='Use NLS model checkpoint (adds _NLS suffix to checkpoint/output filenames)')

    # --- Attention Probe Specific ---
    parser.add_argument('--target_task', type=str, default='C', help='Which task to analyze (e.g., A, C, E).')
    parser.add_argument('--num_samples', type=int, default=1000, help='Number of sequences to sample.')
    parser.add_argument('--seq_len', type=int, default=30, help='Exact length of sequences to analyze.')
    parser.add_argument('--batch_size', type=int, default=100)

    return parser.parse_args()


def load_lines(path):
    if not os.path.exists(path): return []
    try:
        with open(path, 'r', encoding='gbk') as f:
            return [line.strip() for line in f if line.strip()]
    except:
        with open(path, 'r', encoding='utf-8') as f:
            return [line.strip() for line in f if line.strip()]


def extract_fixed_length_data(lines, stoi, num_samples, seq_len, no_task_tag, target_task):
    valid_seqs = []
    # 根据是否有 Task Tag 确定目标 Token 的位置
    target_idx = 1 if no_task_tag else 2
    labels_chars = {'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j'}

    for line in lines:
        parts = line.split()
        if len(parts) < seq_len:
            continue

        # 校验序列格式是否合法，确保 target_idx 的下一个 Token 确实是冒号 ":"
        if target_idx + 1 >= len(parts) or parts[target_idx + 1] != ':':
            continue

        colon_idx = target_idx + 1

        # 过滤保留指定的 target_task
        if not no_task_tag:
            if parts[0] != target_task:
                continue
        else:
            # 如果没有 Task Tag，通过冒号后的动作序列特征来启发式识别 Task E 和 Task A
            is_task_e = len(parts) > colon_idx + 2 and parts[colon_idx + 2] in labels_chars
            if target_task == 'E' and not is_task_e:
                continue
            if target_task == 'A' and is_task_e:
                continue

        try:
            tokens = [stoi[p] for p in parts[:seq_len]]
            valid_seqs.append(tokens)
        except KeyError:
            continue

    if len(valid_seqs) < num_samples:
        print(f"Warning: Only found {len(valid_seqs)} valid Task {target_task} sequences. Using all of them.")
        sampled = valid_seqs
    else:
        sampled = random.sample(valid_seqs, num_samples)

    return [torch.tensor(s, dtype=torch.long) for s in sampled], target_idx


# 全局容器，用于存储截获的注意力矩阵
activations = {}


def get_attention_hook(layer_idx, n_head):
    """
    挂载在 c_attn 层上的 Hook：
    c_attn 输出 shape 为 (B, T, 3 * C)，包含 Q, K, V。
    修改：保留所有 Head 的维度，不再取平均。
    """

    def hook(module, input, output):
        B, T, C3 = output.size()
        C = C3 // 3
        hs = C // n_head

        # 拆分 Q, K, V
        q, k, v = output.split(C, dim=2)

        # Reshape & Transpose: (B, nh, T, hs)
        k = k.view(B, T, n_head, hs).transpose(1, 2)
        q = q.view(B, T, n_head, hs).transpose(1, 2)

        # 计算 Attention 分数
        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(hs))

        # 应用因果掩码
        mask = torch.tril(torch.ones(T, T, device=output.device)).view(1, 1, T, T)
        att = att.masked_fill(mask == 0, float('-inf'))

        # Softmax
        att = torch.nn.functional.softmax(att, dim=-1)

        # 修改：保留 Head 维度，shape 变为 (B, n_head, T, T)
        att_full = att.detach().cpu()

        if layer_idx not in activations:
            activations[layer_idx] = []
        activations[layer_idx].append(att_full)

    return hook


def main():
    args = parse_args()
    seed = 42
    torch.manual_seed(seed)
    np.random.seed(seed)
    random.seed(seed)

    tasks_tag = f"{args.tasks}_CL" if args.CL else args.tasks
    tasks_tag = f"{tasks_tag}_{args.path_type}"
    if args.no_task_tag: tasks_tag += "_NT"
    # NLS only affects checkpoint naming (per train_maze.py)
    ckpt_tasks_tag = f"{tasks_tag}_NLS" if args.NLS else tasks_tag
    nls_suffix = '_NLS' if args.NLS else ''

    data_dir = f'data/maze/{args.num_nodes}'
    nt_suffix = '_NT' if args.no_task_tag else ''
    out_dir = f'out/transformer/maze_{args.config}_{args.num_nodes}{nt_suffix}/'
    os.makedirs(out_dir, exist_ok=True)

    def pick_first_existing(candidates):
        for path in candidates:
            if os.path.exists(path): return path
        return candidates[0]

    # 1. Load Meta
    meta = pickle.load(open(pick_first_existing([f'{data_dir}/meta_{tasks_tag}.pkl', f'{data_dir}/meta.pkl']), 'rb'))
    stoi, itos = meta['stoi'], meta['itos']

    # 2. Load Model
    train_label = format_count(args.num_train_dataset)
    ckpt_path = pick_first_existing([os.path.join(out_dir, f'{args.ckpt_iter}_ckpt_maze_{ckpt_tasks_tag}_{train_label}.pt'),
                                     os.path.join(out_dir, f'{args.ckpt_iter}_ckpt_maze_{tasks_tag}_{train_label}.pt'),
                                     os.path.join(out_dir, f'{args.ckpt_iter}_ckpt_maze_{args.num_of_paths}.pt')])
    checkpoint = torch.load(ckpt_path, map_location=args.device)
    conf = GPTConfig(**checkpoint['model_args'])

    # 强制不使用 flash_attention 以便于我们 Hook QKV 时保持计算流清晰
    conf.use_flash = False
    model = GPT(conf)
    model.load_state_dict({k.replace('_orig_mod.', ''): v for k, v in checkpoint['model'].items()})
    model.to(args.device).eval()

    # 3. Register Hooks for all layers' c_attn module
    for i in range(conf.n_layer):
        model.transformer.h[i].attn.c_attn.register_forward_hook(get_attention_hook(i, conf.n_head))

    # 4. Prepare Data
    train_txt = pick_first_existing([os.path.join(data_dir, f"train_{tasks_tag}_{train_label}.txt"),
                                     os.path.join(data_dir, f'train_{args.num_of_paths}.txt')])

    print(f"Loading data from: {train_txt}")
    lines = load_lines(train_txt)
    dataset, target_idx = extract_fixed_length_data(lines, stoi, args.num_samples, args.seq_len, args.no_task_tag, args.target_task)
    print(f"Successfully extracted {len(dataset)} Task {args.target_task} sequences of length {args.seq_len}.")
    print(f"Target token index determined as: {target_idx} (0-indexed)")

    # Print a few sample prefixes
    print(f"\n--- Task {args.target_task} Sequence Samples ---")
    for s in dataset[:3]:
        print(" ".join([itos[x.item()] for x in s]))
    print("------------------------\n")

    # 5. Inference
    for i in tqdm(range(0, len(dataset), args.batch_size), desc="Running Inference"):
        batch = dataset[i: i + args.batch_size]
        x_padded = pad_sequence(batch, batch_first=True, padding_value=0).to(args.device)
        with torch.no_grad():
            model(x_padded)

    # 6. Analyze Attention Weights (Text Output)
    # 注意：这里为了保持文本输出的简洁，我们依然计算该层所有 Head 的平均统计值
    print(f"\n--- Attention Weights towards Target Token (Index: {target_idx}) ---")
    results_text = f"Attention Probe Results - Task {args.target_task} (Seq Len: {args.seq_len}, Samples: {len(dataset)})\n"
    results_text += f"Target Token Index: {target_idx}\n"
    results_text += "Note: Stats below are averaged over all heads for each layer.\n\n"

    for l in range(conf.n_layer):
        # 拼接所有 batch -> (Total_Samples, n_head, T, T)
        layer_att = torch.cat(activations[l], dim=0)

        # 对 head 维度取平均，用于文本统计 -> (Total_Samples, T, T)
        layer_att_mean_heads = layer_att.mean(dim=1)

        layer_header = f"=== Layer {l:02d} ==="
        print(layer_header)
        results_text += layer_header + "\n"

        # Calculate attention weight from every position to the target_idx
        for pos in range(args.seq_len):
            if pos < target_idx:
                line = f"  Pos {pos:02d} : 0.0000 ± 0.0000 (Masked)"
            else:
                # 使用平均后的 attention 计算统计值
                weights = layer_att_mean_heads[:, pos, target_idx].numpy()
                mean_w = np.mean(weights)
                std_w = np.std(weights)
                line = f"  Pos {pos:02d} : {mean_w:.4f} ± {std_w:.4f}"

            print(line)
            results_text += line + "\n"
        print()
        results_text += "\n"

    # 7. Save to txt
    output_txt = os.path.join(out_dir, f"attention_probe_Task{args.target_task}_L{args.seq_len}_iter{args.ckpt_iter}{nls_suffix}.txt")
    with open(output_txt, 'w') as f:
        f.write(results_text)
    print(f"Done! Attention stats saved to: {output_txt}")

    # 8. Visualize Attention Matrices (Per Head)
    print("\n--- Generating Attention Matrix Visualization (Per Head) ---")

    # 布局：
    #   - n_head == 1：行=1，列=层数（层横向排列）
    #   - n_head  > 1：行=层数，列=头数（每行是同一层的多个头）
    if conf.n_head > 1:
        rows = conf.n_layer
        cols = conf.n_head
        layout_by_layer_rows = True
    else:
        rows = 1
        cols = conf.n_layer
        layout_by_layer_rows = False

    fig, axes = plt.subplots(rows, cols, figsize=(3 * cols, 3 * rows), squeeze=False)

    for l in range(conf.n_layer):
        # 拼接所有 batch -> (Total_Samples, n_head, T, T)
        layer_att = torch.cat(activations[l], dim=0)

        for h in range(conf.n_head):
            # 取特定 Head，并在样本维度求平均 -> (T, T)
            mean_att_matrix = layer_att[:, h, :, :].mean(dim=0).numpy()

            if layout_by_layer_rows:
                row_idx, col_idx = l, h
            else:
                row_idx, col_idx = 0, l

            ax = axes[row_idx, col_idx]
            # aspect='equal' 强制像素为正方形，保证矩阵不变形
            im = ax.imshow(mean_att_matrix, cmap='viridis', aspect='equal', vmin=0, vmax=1)

            # 设置标题
            ax.set_title(f'L{l} H{h}', fontsize=10)

            # 仅在最左列显示 Y 轴标签
            if col_idx == 0:
                ax.set_ylabel(f'L{l} Query Pos' if layout_by_layer_rows else 'Query Pos')
            # 仅在最下行显示 X 轴标签
            if row_idx == rows - 1:
                ax.set_xlabel('Key Pos')

            # 标记目标 Token 列
            ax.axvline(x=target_idx, color='red', linestyle='--', alpha=0.5, linewidth=1)

    # 调整布局，留出右侧 colorbar 的空间
    fig.subplots_adjust(right=0.82)
    cbar_ax = fig.add_axes([0.86, 0.15, 0.02, 0.7])
    fig.colorbar(im, cax=cbar_ax, label='Attention Weight')

    plt.suptitle(f'Attention Matrices per Head - Task {args.target_task} (Target Idx: {target_idx})', fontsize=16)

    output_png = os.path.join(out_dir, f"attention_matrix_Task{args.target_task}_L{args.seq_len}_iter{args.ckpt_iter}{nls_suffix}.png")
    plt.savefig(output_png, dpi=300, bbox_inches='tight')
    plt.close()

    print(f"Done! Attention visualization saved to: {output_png}")


if __name__ == "__main__":
    main()