bev-project/debug_transformer_output.py

#!/usr/bin/env python3
"""
调试Transformer解码器输出
验证分割头是否能产生合理的预测
"""

import torch
import torch.nn as nn
import sys
sys.path.append('/workspace/bevfusion')

# 简化的分割头测试
class SimpleTransformerDecoder(nn.Module):
    def __init__(self, nc=6, C=64):
        super().__init__()
        self.nc = nc
        self.C = C

        # 简化的权重初始化
        self.task_weights = nn.Parameter(torch.randn(nc, 1) * 0.1)

        # 简化的refine头
        self.refine = nn.Sequential(
            nn.Conv2d(C, C, kernel_size=3, padding=1),
            nn.BatchNorm2d(C),
            nn.ReLU(inplace=True),
            nn.Conv2d(C, 1, kernel_size=1)
        )

    def forward(self, x_single, imgs):
        # 动态投影
        if x_single.shape[1] != self.C:
            single_proj = nn.Conv2d(x_single.shape[1], self.C, kernel_size=1).to(x_single.device)
            x_proj = single_proj(x_single)
        else:
            x_proj = x_single

        # 类别扩展
        task_features = x_proj.unsqueeze(1).expand(-1, self.nc, -1, -1, -1)

        # 权重应用
        task_weights = torch.sigmoid(self.task_weights).view(1, self.nc, 1, 1, 1)
        task_features = task_features * task_weights

        # 重塑并处理
        B, T, C, H, W = task_features.shape
        task_features = task_features.view(B * T, C, H, W)

        # refine
        seg_logits = self.refine(task_features)

        # 调整尺寸
        if seg_logits.shape[-2:] != (imgs, imgs):
            seg_logits = torch.nn.functional.interpolate(seg_logits, size=(imgs, imgs), mode='nearest')

        # 重塑回 (B, nc, H, W)
        seg_masks = seg_logits.view(B, T, 1, imgs, imgs).squeeze(2)

        return seg_masks, []

def test_transformer_output():
    """测试Transformer解码器输出"""
    print("🔬 测试Transformer解码器输出...")

    # 创建解码器
    decoder = SimpleTransformerDecoder(nc=6, C=64)

    # 测试输入
    batch_size = 2
    C_in = 512
    H, W = 360, 360
    target_size = 360

    x_single = torch.randn(batch_size, C_in, H, W)
    print(f"输入形状: {x_single.shape}")

    # 前向传播
    seg_masks, aux_list = decoder(x_single, target_size)

    print(f"输出形状: {seg_masks.shape}")
    print(f"输出范围: [{seg_masks.min():.3f}, {seg_masks.max():.3f}]")
    print(f"输出均值: {seg_masks.mean():.3f}")
    print(f"输出标准差: {seg_masks.std():.3f}")

    # 检查输出是否合理
    if seg_masks.shape == (batch_size, 6, target_size, target_size):
        print("✅ 输出形状正确")
    else:
        print("❌ 输出形状错误")
        return False

    if not torch.isnan(seg_masks).any():
        print("✅ 无NaN值")
    else:
        print("❌ 包含NaN值")
        return False

    if seg_masks.abs().max() < 100:  # 合理的输出范围
        print("✅ 输出值在合理范围内")
    else:
        print("❌ 输出值过大")
        return False

    # 测试类别差异
    class_means = []
    for i in range(6):
        class_mean = seg_masks[:, i].mean().item()
        class_means.append(class_mean)
        print(f"类别 {i} 均值: {class_mean:.3f}")

    # 检查类别间是否有差异
    if max(class_means) - min(class_means) > 0.01:
        print("✅ 类别间有差异")
    else:
        print("⚠️  类别间差异很小")

    print("🎉 Transformer解码器输出测试通过")
    return True

if __name__ == "__main__":
    success = test_transformer_output()
    if success:
        print("\n✅ Transformer解码器架构正常")
        print("问题可能在于:")
        print("1. 训练数据的加载")
        print("2. loss计算")
        print("3. 优化器设置")
    else:
        print("\n❌ Transformer解码器存在问题")