卷积神经网络可视化

最新推荐文章于 2023-10-23 17:40:22 发布

师太，借个吻

最新推荐文章于 2023-10-23 17:40:22 发布

阅读量528

点赞数

分类专栏：神经网络可视化文章标签：神经网络深度学习机器学习

本文链接：https://blog.csdn.net/qq_42891453/article/details/120141395

版权

神经网络可视化专栏收录该内容

1 篇文章 0 订阅

订阅专栏

conv1_weights = list(model_saved.parameters())[0]
conv1_images = make_grid(conv1_weights, normalize=True).cpu()
plt.figure(figsize=(8, 8))
plt.imshow(conv1_images.permute(1, 2, 0).numpy())

输出：可视化了第一个卷积层的卷积核

可以一步一步的来看：

第一步：

conv1_weights = list(model_saved.parameters())[0]
print(conv1_weights)

tensor([[[[ 1.1864e-01,  9.4069e-02,  9.5435e-02,  ...,  5.5822e-02,
            2.1575e-02,  4.9963e-02],
          [ 7.4882e-02,  3.8940e-02,  5.2979e-02,  ...,  2.5709e-02,
           -1.1299e-02,  4.1590e-03],
          [ 7.5425e-02,  3.8779e-02,  5.4930e-02,  ...,  4.3596e-02,
            1.0225e-02,  1.3251e-02],
          ...,
          [ 9.3155e-02,  1.0374e-01,  6.7547e-02,  ..., -2.0277e-01,
           -1.2839e-01, -1.1220e-01],
          [ 4.3544e-02,  6.4916e-02,  3.6164e-02,  ..., -2.0248e-01,
           -1.1376e-01, -1.0719e-01],
          [ 4.7369e-02,  6.2543e-02,  2.4758e-02,  ..., -1.1844e-01,
           -9.5567e-02, -8.3890e-02]],

         [[-7.2634e-02, -5.7996e-02, -8.0661e-02,  ..., -6.0304e-04,
           -2.5309e-02,  2.5471e-02],
          [-6.9042e-02, -6.7562e-02, -7.6367e-02,  ..., -3.9616e-03,
           -3.0402e-02,  1.0477e-02],
          [-9.9517e-02, -8.5592e-02, -1.0521e-01,  ..., -2.6587e-02,
           -2.2777e-02,  6.6451e-03],
          ...,
          [-1.5121e-01, -8.8735e-02, -9.6737e-02,  ...,  3.0853e-01,
            1.8096e-01,  8.4297e-02],
          [-1.4309e-01, -7.5710e-02, -7.2215e-02,  ...,  2.0417e-01,
            1.6447e-01,  9.5166e-02],
          [-8.5925e-02, -4.0134e-02, -5.1491e-02,  ...,  1.6352e-01,
            1.4822e-01,  1.0196e-01]],

         [[-2.3596e-02, -2.1258e-03, -2.7761e-02,  ...,  3.9940e-02,
           -7.1123e-03,  3.2207e-02],
          [ 2.5705e-04,  2.2468e-02,  8.9070e-03,  ...,  1.8771e-02,
           -1.4155e-02,  1.8275e-02],
          [ 5.4084e-03,  2.9397e-02,  3.3051e-04,  ...,  1.2054e-02,
           -2.5237e-03,  8.3515e-03],
          ...,
          [-6.2826e-02, -1.1655e-02, -6.2080e-02,  ...,  1.0332e-01,
           -9.4987e-03, -7.9570e-02],
          [-4.5691e-02,  3.3726e-03, -3.9632e-02,  ..., -2.6448e-02,
           -3.3500e-02, -7.6398e-02],
          [-1.8700e-02,  1.1365e-02, -3.9671e-02,  ..., -6.8563e-02,
           -4.1289e-02, -5.5473e-02]]],


        [[[-1.9950e-03,  2.9262e-03,  4.8212e-02,  ...,  6.1402e-02,
            2.6121e-02,  1.9558e-02],
          [-1.2579e-02, -4.8879e-03,  1.8490e-02,  ...,  5.3881e-02,
            1.6377e-02,  2.3768e-02],
          [ 3.6561e-03, -7.7510e-04,  2.6360e-02,  ..., -2.5849e-02,
           -6.1798e-02,  2.6103e-02],
          ...,
          [-1.0812e-02, -4.6008e-03,  1.5122e-02,  ...,  2.9561e-02,
            5.3272e-03,  6.8561e-02],
          [ 2.7364e-04, -1.4850e-02,  7.8180e-03,  ...,  2.7172e-02,
           -1.8095e-02,  5.2485e-02],
          [-5.2470e-02, -4.6578e-02, -1.0951e-02,  ...,  4.3038e-03,
           -2.6379e-03,  1.4406e-02]],

         [[ 2.3965e-02,  2.2740e-02,  5.7586e-03,  ...,  7.2087e-03,
           -2.4652e-02,  4.4658e-02],
          [ 2.6914e-02,  4.4892e-02, -1.0872e-03,  ...,  4.4243e-02,
           -2.1168e-02,  6.4538e-02],
          [ 1.2421e-02,  1.0247e-02, -4.1554e-02,  ..., -1.2134e-01,
           -1.6294e-01,  2.6266e-02],
          ...,
          [ 3.5926e-02,  5.3235e-02,  1.1016e-02,  ...,  1.2710e-02,
           -2.9737e-02,  8.5926e-02],
          [ 1.5623e-02,  2.1743e-02, -8.2941e-03,  ..., -3.2744e-03,
           -5.4099e-02,  5.7634e-02],
          [ 7.5254e-02,  8.7784e-02,  5.5804e-02,  ...,  5.2849e-02,
            1.0612e-02,  9.3531e-02]],

         [[-3.6488e-02,  6.6332e-03, -3.9035e-02,  ..., -1.5678e-02,
           -7.9994e-02, -8.8658e-04],
          [-5.1740e-03,  5.7395e-02,  8.9841e-03,  ...,  7.4166e-02,
           -3.1792e-03,  4.2777e-02],
          [-7.9446e-02, -2.2924e-02, -7.3370e-02,  ..., -5.6738e-02,
           -1.2923e-01,  1.8896e-02],
          ...,
          [-3.9394e-02,  3.0981e-02, -2.7901e-02,  ..., -1.6774e-02,
           -1.0236e-01,  4.0128e-02],
          [-6.0751e-02, -2.3034e-02, -7.6838e-02,  ..., -7.9069e-02,
           -1.6195e-01, -1.3746e-02],
          [ 7.9522e-03,  4.6969e-02, -1.2460e-02,  ..., -4.6956e-02,
           -1.0082e-01,  1.9832e-02]]],


        [[[-5.1702e-02,  1.3825e-02,  9.0514e-03,  ..., -9.6401e-02,
           -1.1277e-01, -2.1596e-01],
          [-9.0091e-02, -1.3136e-02, -3.2812e-02,  ..., -7.5263e-02,
           -1.4803e-01, -2.9966e-01],
          [-1.3155e-01, -4.2686e-02, -4.7744e-02,  ...,  2.1429e-01,
            3.2543e-02, -1.7151e-01],
          ...,
          [-1.0621e-01, -9.7966e-02, -2.5551e-01,  ...,  1.2277e-01,
            1.9287e-01,  1.2671e-01],
          [-8.0761e-02, -6.1498e-02, -2.2312e-01,  ...,  3.5376e-02,
            1.0532e-01,  1.0669e-01],
          [ 3.8186e-02,  4.9957e-02, -1.2802e-01,  ..., -3.2927e-02,
            1.8685e-02,  4.7146e-02]],

         [[ 3.9013e-02,  6.4311e-03, -3.1710e-03,  ..., -2.1245e-02,
            4.0516e-02,  1.1092e-01],
          [ 6.5689e-02,  2.2132e-02,  6.6539e-03,  ..., -3.9448e-02,
            2.7749e-02,  1.1404e-01],
          [ 7.7954e-02,  4.0220e-02,  1.4047e-02,  ..., -1.5417e-01,
           -9.2291e-02,  3.4460e-02],
          ...,
          [ 1.2836e-01,  9.4449e-02,  1.4659e-01,  ..., -6.0067e-02,
           -9.0891e-02, -6.1129e-02],
          [ 1.2683e-01,  1.0044e-01,  1.3754e-01,  ..., -2.2507e-02,
           -6.6664e-02, -1.9906e-02],
          [ 8.0509e-02,  7.8203e-02,  9.8934e-02,  ...,  9.2865e-03,
           -3.4635e-02, -1.2395e-02]],

         [[ 1.1535e-02, -2.6993e-02,  1.4820e-02,  ...,  9.4833e-02,
            1.2044e-01,  1.1027e-01],
          [ 9.2629e-03, -2.6680e-02,  1.2218e-02,  ...,  8.7219e-02,
            1.5435e-01,  1.8049e-01],
          [ 6.9946e-02,  1.3250e-02,  4.8007e-02,  ..., -5.6851e-02,
            3.2596e-02,  1.6812e-01],
          ...,
          [-1.2187e-02, -3.3265e-02,  1.1284e-01,  ..., -6.7740e-02,
           -1.0240e-01, -7.6188e-02],
          [-6.0069e-03, -2.8631e-02,  1.1643e-01,  ..., -6.7597e-03,
           -4.3772e-02, -3.1101e-02],
          [-1.3355e-01, -1.4825e-01, -1.0060e-03,  ...,  1.8809e-02,
           -6.4637e-03, -2.7061e-02]]],


        ...,


        [[[ 9.0948e-03,  1.4823e-02,  4.7374e-03,  ...,  1.5540e-02,
           -5.8369e-04, -1.9922e-02],
          [ 2.8962e-04,  2.1229e-02, -1.3210e-02,  ...,  2.4388e-03,
           -5.8485e-03, -2.0373e-02],
          [-1.1050e-02,  1.0094e-02, -2.9625e-02,  ..., -1.4471e-02,
           -1.7187e-02, -3.0534e-02],
          ...,
          [ 1.0013e-01,  9.1407e-02,  1.3077e-01,  ...,  1.5798e-01,
            9.0361e-02,  7.8365e-02],
          [ 1.1610e-01,  8.1846e-02,  8.2892e-02,  ..., -6.0174e-02,
           -6.9412e-02, -5.0151e-02],
          [-1.0564e-01, -1.1848e-01, -1.7681e-01,  ..., -2.0837e-01,
           -1.8036e-01, -1.6691e-01]],

         [[-1.1495e-02,  2.4917e-03, -8.2400e-03,  ..., -7.5865e-03,
           -1.7387e-02, -1.7026e-02],
          [-2.7461e-03, -1.1415e-02, -6.0670e-03,  ..., -2.8248e-02,
           -2.2555e-02, -2.2559e-02],
          [-8.4246e-03, -2.2378e-03, -3.5825e-02,  ..., -1.8462e-02,
           -1.9795e-02, -2.4990e-02],
          ...,
          [ 1.2956e-01,  9.8057e-02,  1.4888e-01,  ...,  1.5655e-01,
            7.9547e-02,  9.6928e-02],
          [ 1.6080e-01,  1.0522e-01,  1.0264e-01,  ..., -6.5226e-02,
           -6.4314e-02, -3.9066e-02],
          [-1.2880e-01, -1.4656e-01, -1.9475e-01,  ..., -2.4177e-01,
           -2.0277e-01, -1.9324e-01]],

         [[-5.4209e-03, -1.7648e-03,  4.3163e-03,  ...,  9.8182e-03,
            5.2347e-03,  6.3336e-03],
          [ 9.3478e-03,  1.5920e-03, -2.1660e-03,  ...,  7.1244e-03,
            9.3521e-04, -5.7647e-03],
          [-1.3668e-03,  3.8899e-03, -7.7412e-03,  ...,  3.0651e-03,
            1.4989e-02, -8.1730e-03],
          ...,
          [ 4.3937e-02,  7.9544e-04,  6.0613e-02,  ...,  7.4787e-02,
            4.3960e-02,  5.6071e-02],
          [ 1.0077e-01,  7.5126e-02,  1.0962e-01,  ...,  4.9886e-03,
            1.0789e-02,  1.3402e-02],
          [-8.8773e-02, -7.2790e-02, -9.2907e-02,  ..., -6.6530e-02,
           -3.8977e-02, -4.8370e-02]]],


        [[[ 4.5712e-03,  4.6908e-02, -1.6075e-02,  ...,  7.7790e-03,
           -1.9798e-02,  6.8000e-03],
          [ 6.2769e-02,  4.5108e-02,  4.7187e-02,  ...,  6.0599e-02,
            2.9275e-02,  5.5794e-02],
          [ 3.6748e-03,  1.2952e-02,  1.8988e-05,  ..., -8.3492e-03,
           -1.9689e-03,  8.0830e-03],
          ...,
          [-4.4365e-02, -5.8858e-02, -2.4772e-02,  ..., -2.8423e-02,
           -3.0897e-02, -5.2936e-02],
          [-9.7572e-03, -4.3227e-02,  9.0068e-03,  ..., -4.2596e-02,
           -1.8114e-02, -2.8028e-02],
          [-2.2007e-02, -3.3594e-02,  1.3479e-02,  ..., -4.1530e-02,
           -1.7819e-02, -5.1977e-02]],

         [[-9.0596e-02, -5.1485e-02, -1.6459e-01,  ..., -1.1970e-01,
           -1.1150e-01, -4.3914e-02],
          [ 1.3835e-02,  2.6007e-02, -1.9440e-02,  ...,  2.3757e-02,
            6.3709e-03,  5.4287e-02],
          [-9.3225e-02, -4.7454e-02, -1.1274e-01,  ..., -8.6459e-02,
           -7.3958e-02, -6.6610e-02],
          ...,
          [ 2.7382e-02,  1.0310e-02,  4.3906e-02,  ...,  2.7094e-02,
            4.4510e-02,  1.5977e-02],
          [ 9.8431e-02,  6.1433e-02,  1.1413e-01,  ...,  9.6398e-02,
            1.0725e-01,  9.5719e-02],
          [-1.5100e-02, -1.1830e-02,  4.8571e-02,  ...,  2.9060e-02,
            5.6323e-02, -2.1631e-03]],

         [[-1.3693e-01, -7.9341e-02, -2.1245e-01,  ..., -1.3633e-01,
           -1.5123e-01, -6.3938e-02],
          [ 1.5745e-02,  5.1443e-02, -1.8209e-02,  ...,  4.9085e-02,
            1.9585e-02,  7.8095e-02],
          [-1.7127e-01, -8.8734e-02, -1.7467e-01,  ..., -1.4431e-01,
           -1.3364e-01, -1.1878e-01],
          ...,
          [ 5.1982e-02,  1.5912e-02,  7.0009e-02,  ...,  4.4396e-02,
            6.5429e-02,  2.6919e-02],
          [ 1.3144e-01,  9.1333e-02,  1.6228e-01,  ...,  1.4230e-01,
            1.5500e-01,  1.3808e-01],
          [ 3.7818e-03, -2.0365e-02,  7.8663e-02,  ...,  8.6037e-02,
            1.2596e-01,  3.8774e-02]]],


        [[[-9.5081e-02,  5.6581e-02,  1.4016e-01,  ..., -1.4826e-02,
            1.0425e-02, -3.7919e-03],
          [ 6.0412e-02,  1.2531e-01, -1.2366e-01,  ...,  7.8628e-02,
           -1.0564e-02, -2.2228e-02],
          [ 1.0304e-01, -1.3654e-01, -1.9601e-01,  ..., -5.1919e-02,
           -8.3287e-02,  3.9678e-02],
          ...,
          [-1.8598e-01,  1.5479e-02,  3.3912e-01,  ...,  2.7509e-01,
            1.0493e-01, -1.6863e-01],
          [ 6.8278e-02,  1.3989e-01,  1.1350e-02,  ..., -7.8728e-02,
           -2.5001e-01, -8.0072e-02],
          [ 5.1736e-03, -6.8372e-02, -9.1940e-02,  ..., -1.0727e-01,
            8.7973e-02,  1.0139e-01]],

         [[-9.8760e-02,  6.2635e-02,  1.1495e-01,  ..., -1.3191e-02,
            1.9574e-02, -1.0864e-03],
          [ 8.3206e-02,  1.0300e-01, -1.5419e-01,  ...,  1.0864e-01,
            1.2592e-02, -2.0278e-02],
          [ 1.1248e-01, -1.6928e-01, -1.8876e-01,  ..., -6.3491e-02,
           -9.3101e-02,  4.9006e-02],
          ...,
          [-1.8867e-01,  6.7374e-02,  4.8128e-01,  ...,  3.1371e-01,
            1.6068e-01, -1.3449e-01],
          [ 1.1207e-01,  2.0871e-01,  4.6815e-02,  ..., -7.0456e-03,
           -2.2978e-01, -6.9528e-02],
          [ 1.7255e-02, -9.1683e-02, -1.5943e-01,  ..., -8.0373e-02,
            8.0035e-02,  1.1819e-01]],

         [[-1.0580e-01,  5.4381e-02,  1.3048e-01,  ..., -3.9546e-02,
            1.1924e-02, -2.7886e-03],
          [ 5.7859e-02,  1.0646e-01, -1.3468e-01,  ...,  1.0645e-01,
            1.8412e-02, -1.3245e-03],
          [ 1.0398e-01, -1.0849e-01, -1.6758e-01,  ..., -4.2554e-02,
           -8.5875e-02,  5.2581e-02],
          ...,
          [-1.8241e-01,  5.4228e-02,  3.9395e-01,  ...,  2.4432e-01,
            1.0216e-01, -1.2876e-01],
          [ 8.5720e-02,  1.8378e-01,  5.0316e-02,  ..., -4.7009e-02,
           -2.1584e-01, -4.2482e-02],
          [ 3.9612e-02, -7.6353e-02, -1.3502e-01,  ..., -4.8630e-02,
            1.0063e-01,  9.0307e-02]]]], device='cuda:0')

第二步：

特征图的每个像素点上的数值范围不是[0,1]，而是可正可负，可大可小，因此需要做一些特殊处理。这里就要用到 torchvision.utils.make_grid( )函数，把输入的特征图做一个归一化，把参数normalize设置为True即可，它能帮我们把数据的输入范围调整至[0, 1]之间

make_grid() 输入的是Tensor，而不是numpy.ndarray
torchvision.utils.make_grid() 将一组图片绘制到一个窗口，其本质是将一组图片拼接成一张图片

conv1_images = make_grid(conv1_weights, normalize=True).cpu()
print(conv1_images)

tensor([[[0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.5244,  ..., 0.4508, 0.0000, 0.0000],
         ...,
         [0.0000, 0.0000, 0.3961,  ..., 0.5143, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000]],

        [[0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.4130,  ..., 0.4338, 0.0000, 0.0000],
         ...,
         [0.0000, 0.0000, 0.4496,  ..., 0.5241, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000]],

        [[0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.4416,  ..., 0.3990, 0.0000, 0.0000],
         ...,
         [0.0000, 0.0000, 0.4963,  ..., 0.5079, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000]]])

最后

imshow（）的输入参数：

真彩色图像输入，指定为 m×n×3 数组。如果指定数据类型为 single 或 double 的真彩色图像，则值应在 [0, 1] 范围内。如果像素值超出此范围，则可以使用rescale函数将像素值缩放到范围 [0, 1] 内。当输入图像为真彩色时，'DisplayRange' 参数无效。

plt.figure(figsize=(8, 8))
plt.imshow(conv1_images.permute(1, 2, 0).numpy())

输出

PyTorch 实现的 AlexNet 第一层卷积核参数的形状是 64×3×11×1164×3×11×11 的四维 Tensor，这样就可以得到上述 6464 个11×1111×11 的图片块了。显然，这些重构出来的图像基本都是关于边缘，条纹以及颜色的信息

以上为卷积核的可视化接下来看看卷积层的可视化：

def visualize(alexnet, input_data, submodule_name, layer_index):
    '''
    alexnet: 模型
    input_data: 输入数据
    submodule_name: 可视化 module 的 name, 专门针对 nn.Sequential
    layer_index: 在 submodule 中的 index
    '''
    x = input_data
    modules = alexnet._modules
    for name in modules.keys():
        if (name == submodule_name):
            module_layers = list(modules[name].children())
            for i in range(layer_index+1):
                if (type(module_layers[i]) == torch.nn.Linear):
                    x = x.reshape(x.size(0), -1)  # 针对线性层
                x = module_layers[i](x)
            return x

        x = modules[name](x)

feature_maps = visualize(model_saved, IMAGE.to(dev), 'features', 0)
feature_images = make_grid(feature_maps.permute(1, 0, 2, 3), normalize=True).cpu()
plt.figure(figsize=(8, 8))
plt.imshow(feature_images.permute(1, 2, 0).numpy())

输出可视化的卷积层：

师太，借个吻

关注

0
点赞
踩
2

收藏

觉得还不错? 一键收藏
0
评论
卷积神经网络可视化

conv1_weights = list(model_saved.parameters())[0]conv1_images = make_grid(conv1_weights, normalize=True).cpu()plt.figure(figsize=(8, 8))plt.imshow(conv1_images.permute(1, 2, 0).numpy())输出：可视化了第一个卷积层的卷积核
复制链接

扫一扫