• 深入浅出PyTorch——PyTorch可视化

    1. 可视化网络结构

            在复杂的网络结构中确定每一层的输入结构，方便我们在短时间内完成debug

    1.1 使用print函数打印模型基础信息

            使用ResNet18的结构进行展示

    1. import torchvision.models as models
    2. model = models.resnet18()
    3. print(model)
    4.  
    5. #打印结果
    6. ResNet(
    7.   (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
    8.   (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    9.   (relu): ReLU(inplace=True)
    10.   (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
    11.   (layer1): Sequential(
    12.     (0): Bottleneck(
    13.       (conv1): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
    14.       (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    15.       (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    16.       (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    17.       (conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
    18.       (bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    19.       (relu): ReLU(inplace=True)
    20.       (downsample): Sequential(
    21.         (0): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
    22.         (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    23.       )
    24.     )
    25.    ... ...
    26.   )
    27.   (avgpool): AdaptiveAvgPool2d(output_size=(1, 1))
    28.   (fc): Linear(in_features=2048, out_features=1000, bias=True)
    29. )

    1.2 使用torchinfo可视化网络结构

    1.2.1 torchinfo的安装

    1. # 安装方法一
    2. pip install torchinfo 
    3. # 安装方法二
    4. conda install -c conda-forge torchinfo

    1.2.2 torchinfo的使用

            （1）方法：torchinfo.summary()

            （2）参数：（这里展示的是函数定义时传入的参数），具体请看参数详解）

    1. def summary(
    2.     model: nn.Module,
    3.     input_size: Optional[INPUT_SIZE_TYPE] = None,
    4.     input_data: Optional[INPUT_DATA_TYPE] = None,
    5.     batch_dim: Optional[int] = None,
    6.     cache_forward_pass: Optional[bool] = None,
    7.     col_names: Optional[Iterable[str]] = None,
    8.     col_width: int = 25,
    9.     depth: int = 3,
    10.     device: Optional[torch.device] = None,
    11.     dtypes: Optional[List[torch.dtype]] = None,
    12.     mode: str | None = None,
    13.     row_settings: Optional[Iterable[str]] = None,
    14.     verbose: int = 1,
    15.     **kwargs: Any,
    16. ) -> ModelStatistics

            （3）实例以ResNet18为例：

    1. import torchvision.models as models
    2. from torchinfo import summary
    3. resnet18 = models.resnet18() # 实例化模型
    4. summary(resnet18, (1, 3, 224, 224)) # 1：batch_size 3:图片的通道数 224: 图片的高宽
    5.  
    6.  
    7. # 结果输出
    8. =========================================================================================
    9. Layer (type:depth-idx)                   Output Shape              Param #
    10. =========================================================================================
    11. ResNet                                   --                        --
    12. ├─Conv2d: 1-1                            [1, 64, 112, 112]         9,408
    13. ├─BatchNorm2d: 1-2                       [1, 64, 112, 112]         128
    14. ├─ReLU: 1-3                              [1, 64, 112, 112]         --
    15. ├─MaxPool2d: 1-4                         [1, 64, 56, 56]           --
    16. ├─Sequential: 1-5                        [1, 64, 56, 56]           --
    17. │    └─BasicBlock: 2-1                   [1, 64, 56, 56]           --
    18. │    │    └─Conv2d: 3-1                  [1, 64, 56, 56]           36,864
    19. │    │    └─BatchNorm2d: 3-2             [1, 64, 56, 56]           128
    20. │    │    └─ReLU: 3-3                    [1, 64, 56, 56]           --
    21. │    │    └─Conv2d: 3-4                  [1, 64, 56, 56]           36,864
    22. │    │    └─BatchNorm2d: 3-5             [1, 64, 56, 56]           128
    23. │    │    └─ReLU: 3-6                    [1, 64, 56, 56]           --
    24. │    └─BasicBlock: 2-2                   [1, 64, 56, 56]           --
    25. │    │    └─Conv2d: 3-7                  [1, 64, 56, 56]           36,864
    26. │    │    └─BatchNorm2d: 3-8             [1, 64, 56, 56]           128
    27. │    │    └─ReLU: 3-9                    [1, 64, 56, 56]           --
    28. │    │    └─Conv2d: 3-10                 [1, 64, 56, 56]           36,864
    29. │    │    └─BatchNorm2d: 3-11            [1, 64, 56, 56]           128
    30. │    │    └─ReLU: 3-12                   [1, 64, 56, 56]           --
    31. ├─Sequential: 1-6                        [1, 128, 28, 28]          --
    32. │    └─BasicBlock: 2-3                   [1, 128, 28, 28]          --
    33. │    │    └─Conv2d: 3-13                 [1, 128, 28, 28]          73,728
    34. │    │    └─BatchNorm2d: 3-14            [1, 128, 28, 28]          256
    35. │    │    └─ReLU: 3-15                   [1, 128, 28, 28]          --
    36. │    │    └─Conv2d: 3-16                 [1, 128, 28, 28]          147,456
    37. │    │    └─BatchNorm2d: 3-17            [1, 128, 28, 28]          256
    38. │    │    └─Sequential: 3-18             [1, 128, 28, 28]          8,448
    39. │    │    └─ReLU: 3-19                   [1, 128, 28, 28]          --
    40. │    └─BasicBlock: 2-4                   [1, 128, 28, 28]          --
    41. │    │    └─Conv2d: 3-20                 [1, 128, 28, 28]          147,456
    42. │    │    └─BatchNorm2d: 3-21            [1, 128, 28, 28]          256
    43. │    │    └─ReLU: 3-22                   [1, 128, 28, 28]          --
    44. │    │    └─Conv2d: 3-23                 [1, 128, 28, 28]          147,456
    45. │    │    └─BatchNorm2d: 3-24            [1, 128, 28, 28]          256
    46. │    │    └─ReLU: 3-25                   [1, 128, 28, 28]          --
    47. ├─Sequential: 1-7                        [1, 256, 14, 14]          --
    48. │    └─BasicBlock: 2-5                   [1, 256, 14, 14]          --
    49. │    │    └─Conv2d: 3-26                 [1, 256, 14, 14]          294,912
    50. │    │    └─BatchNorm2d: 3-27            [1, 256, 14, 14]          512
    51. │    │    └─ReLU: 3-28                   [1, 256, 14, 14]          --
    52. │    │    └─Conv2d: 3-29                 [1, 256, 14, 14]          589,824
    53. │    │    └─BatchNorm2d: 3-30            [1, 256, 14, 14]          512
    54. │    │    └─Sequential: 3-31             [1, 256, 14, 14]          33,280
    55. │    │    └─ReLU: 3-32                   [1, 256, 14, 14]          --
    56. │    └─BasicBlock: 2-6                   [1, 256, 14, 14]          --
    57. │    │    └─Conv2d: 3-33                 [1, 256, 14, 14]          589,824
    58. │    │    └─BatchNorm2d: 3-34            [1, 256, 14, 14]          512
    59. │    │    └─ReLU: 3-35                   [1, 256, 14, 14]          --
    60. │    │    └─Conv2d: 3-36                 [1, 256, 14, 14]          589,824
    61. │    │    └─BatchNorm2d: 3-37            [1, 256, 14, 14]          512
    62. │    │    └─ReLU: 3-38                   [1, 256, 14, 14]          --
    63. ├─Sequential: 1-8                        [1, 512, 7, 7]            --
    64. │    └─BasicBlock: 2-7                   [1, 512, 7, 7]            --
    65. │    │    └─Conv2d: 3-39                 [1, 512, 7, 7]            1,179,648
    66. │    │    └─BatchNorm2d: 3-40            [1, 512, 7, 7]            1,024
    67. │    │    └─ReLU: 3-41                   [1, 512, 7, 7]            --
    68. │    │    └─Conv2d: 3-42                 [1, 512, 7, 7]            2,359,296
    69. │    │    └─BatchNorm2d: 3-43            [1, 512, 7, 7]            1,024
    70. │    │    └─Sequential: 3-44             [1, 512, 7, 7]            132,096
    71. │    │    └─ReLU: 3-45                   [1, 512, 7, 7]            --
    72. │    └─BasicBlock: 2-8                   [1, 512, 7, 7]            --
    73. │    │    └─Conv2d: 3-46                 [1, 512, 7, 7]            2,359,296
    74. │    │    └─BatchNorm2d: 3-47            [1, 512, 7, 7]            1,024
    75. │    │    └─ReLU: 3-48                   [1, 512, 7, 7]            --
    76. │    │    └─Conv2d: 3-49                 [1, 512, 7, 7]            2,359,296
    77. │    │    └─BatchNorm2d: 3-50            [1, 512, 7, 7]            1,024
    78. │    │    └─ReLU: 3-51                   [1, 512, 7, 7]            --
    79. ├─AdaptiveAvgPool2d: 1-9                 [1, 512, 1, 1]            --
    80. ├─Linear: 1-10                           [1, 1000]                 513,000
    81. =========================================================================================
    82. Total params: 11,689,512
    83. Trainable params: 11,689,512
    84. Non-trainable params: 0
    85. Total mult-adds (G): 1.81
    86. =========================================================================================
    87. Input size (MB): 0.60
    88. Forward/backward pass size (MB): 39.75
    89. Params size (MB): 46.76
    90. Estimated Total Size (MB): 87.11
    91. =========================================================================================

    2. CNN可视化

            卷积神经网路——CNN

    2.1 CNN卷积核可视化

            以torchvision自带的VGG11模型为例。

    1. import torch
    2. from torchvision.models import vgg11
    3.  
    4. model = vgg11(pretrained=True)
    5. print(dict(model.features.named_children()))
    6.  
    7.  
    8. # 输出
    9. {'0': Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
    10.  '1': ReLU(inplace=True),
    11.  '2': MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
    12.  '3': Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
    13.  '4': ReLU(inplace=True),
    14.  '5': MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
    15.  '6': Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
    16.  '7': ReLU(inplace=True),
    17.  '8': Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
    18.  '9': ReLU(inplace=True),
    19.  '10': MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
    20.  '11': Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
    21.  '12': ReLU(inplace=True),
    22.  '13': Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
    23.  '14': ReLU(inplace=True),
    24.  '15': MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
    25.  '16': Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
    26.  '17': ReLU(inplace=True),
    27.  '18': Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
    28.  '19': ReLU(inplace=True),
    29.  '20': MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)}

    2.2 CNN特征图可视化方法

            在PyTorch中，提供了一个专用的接口使得网络在前向传播过程中能够获取到特征图，这个接口的名称非常形象，叫做hook.

            首先实现了一个hook类，之后在plot_feature函数中，将该hook类的对象注册到要进行可视化的网络的某层中。model在进行前向传播的时候会调用hook的__call__函数，我们也就是在那里存储了当前层的输入和输出。

    1. class Hook(object):
    2.     def __init__(self):
    3.         self.module_name = []
    4.         self.features_in_hook = []
    5.         self.features_out_hook = []
    6.  
    7.     def __call__(self,module, fea_in, fea_out):
    8.         print("hooker working", self)
    9.         self.module_name.append(module.__class__)
    10.         self.features_in_hook.append(fea_in)
    11.         self.features_out_hook.append(fea_out)
    12.         return None
    13.     
    14.  
    15. def plot_feature(model, idx, inputs):
    16.     hh = Hook()
    17.     model.features[idx].register_forward_hook(hh)
    18.     
    19.     # forward_model(model,False)
    20.     model.eval()
    21.     _ = model(inputs)
    22.     print(hh.module_name)
    23.     print((hh.features_in_hook[0][0].shape))
    24.     print((hh.features_out_hook[0].shape))
    25.     
    26.     out1 = hh.features_out_hook[0]
    27.  
    28.     total_ft  = out1.shape[1]
    29.     first_item = out1[0].cpu().clone()    
    30.  
    31.     plt.figure(figsize=(20, 17))
    32.     
    33.  
    34.     for ftidx in range(total_ft):
    35.         if ftidx > 99:
    36.             break
    37.         ft = first_item[ftidx]
    38.         plt.subplot(10, 10, ftidx+1) 
    39.         
    40.         plt.axis('off')
    41.         #plt.imshow(ft[ :, :].detach(),cmap='gray')
    42.         plt.imshow(ft[ :, :].detach())

    2.3 CNN class activation map可视化方法

    2.3.1 实现方法

            CAM系列操作的实现可以通过开源工具包pytorch-grad-cam来实现。

    2.3.2 安装

    pip install grad-cam

    2.3.3 例子

    1. import torch
    2. from torchvision.models import vgg11,resnet18,resnet101,resnext101_32x8d
    3. import matplotlib.pyplot as plt
    4. from PIL import Image
    5. import numpy as np
    6.  
    7. model = vgg11(pretrained=True)
    8. img_path = './dog.png'
    9. # resize操作是为了和传入神经网络训练图片大小一致
    10. img = Image.open(img_path).resize((224,224))
    11. # 需要将原始图片转为np.float32格式并且在0-1之间 
    12. rgb_img = np.float32(img)/255
    13. plt.imshow(img)
    14.  
    15. ##########################################################################
    16. from pytorch_grad_cam import GradCAM,ScoreCAM,GradCAMPlusPlus,AblationCAM,XGradCAM,EigenCAM,FullGrad
    17. from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
    18. from pytorch_grad_cam.utils.image import show_cam_on_image
    19.  
    20. target_layers = [model.features[-1]]
    21. # 选取合适的类激活图，但是ScoreCAM和AblationCAM需要batch_size
    22. cam = GradCAM(model=model,target_layers=target_layers)
    23. targets = [ClassifierOutputTarget(preds)]   
    24. # 上方preds需要设定，比如ImageNet有1000类，这里可以设为200
    25. grayscale_cam = cam(input_tensor=img_tensor, targets=targets)
    26. grayscale_cam = grayscale_cam[0, :]
    27. cam_img = show_cam_on_image(rgb_img, grayscale_cam, use_rgb=True)
    28. print(type(cam_img))
    29. Image.fromarray(cam_img)

     2.4 使用FlashTorch快速实现CNN可视化

    2.4.1 安装

    pip install flashtorch

    2.4.2 可视化梯度

    1. # Download example images
    2. # !mkdir -p images
    3. # !wget -nv \
    4. #    https://github.com/MisaOgura/flashtorch/raw/master/examples/images/great_grey_owl.jpg \
    5. #    https://github.com/MisaOgura/flashtorch/raw/master/examples/images/peacock.jpg   \
    6. #    https://github.com/MisaOgura/flashtorch/raw/master/examples/images/toucan.jpg    \
    7. #    -P /content/images
    8.  
    9. import matplotlib.pyplot as plt
    10. import torchvision.models as models
    11. from flashtorch.utils import apply_transforms, load_image
    12. from flashtorch.saliency import Backprop
    13.  
    14. model = models.alexnet(pretrained=True)
    15. backprop = Backprop(model)
    16.  
    17. image = load_image('/content/images/great_grey_owl.jpg')
    18. owl = apply_transforms(image)
    19.  
    20. target_class = 24
    21. backprop.visualize(owl, target_class, guided=True, use_gpu=True)

    2.4.3 可视化卷积核

    1. import torchvision.models as models
    2. from flashtorch.activmax import GradientAscent
    3.  
    4. model = models.vgg16(pretrained=True)
    5. g_ascent = GradientAscent(model.features)
    6.  
    7. # specify layer and filter info
    8. conv5_1 = model.features[24]
    9. conv5_1_filters = [45, 271, 363, 489]
    10.  
    11. g_ascent.visualize(conv5_1, conv5_1_filters, title="VGG16: conv5_1")

    3. 使用TensorBoard可视化训练过程

    3.1 TensorBoard安装

    pip install tensorboardX

    3.2 TensorBoard可视化的基本逻辑

           （1）TensorBoard是一个记录员；

            （2）记录我们指定的数据，包括模型每一层的feature map，权重，以及训练loss等等；

            （3）保存在指定的文件夹里；

            （4）程序不断运行TensorBoard会不断记录；

            （5）可以通过网页的形式加以可视化。

    3.3 TensorBoard的配置与启动

    1. # 方法一、
    2. from tensorboardX import SummaryWriter
    3. writer = SummaryWriter('./runs')
    4.  
    5. # 方法二
    6. from torch.utils.tensorboard import SummaryWriter

    3.4 TensorBoard模型结构可视化

    1. import torch.nn as nn
    2.  
    3. class Net(nn.Module):
    4.     def __init__(self):
    5.         super(Net, self).__init__()
    6.         self.conv1 = nn.Conv2d(in_channels=3,out_channels=32,kernel_size = 3)
    7.         self.pool = nn.MaxPool2d(kernel_size = 2,stride = 2)
    8.         self.conv2 = nn.Conv2d(in_channels=32,out_channels=64,kernel_size = 5)
    9.         self.adaptive_pool = nn.AdaptiveMaxPool2d((1,1))
    10.         self.flatten = nn.Flatten()
    11.         self.linear1 = nn.Linear(64,32)
    12.         self.relu = nn.ReLU()
    13.         self.linear2 = nn.Linear(32,1)
    14.         self.sigmoid = nn.Sigmoid()
    15.  
    16.     def forward(self,x):
    17.         x = self.conv1(x)
    18.         x = self.pool(x)
    19.         x = self.conv2(x)
    20.         x = self.pool(x)
    21.         x = self.adaptive_pool(x)
    22.         x = self.flatten(x)
    23.         x = self.linear1(x)
    24.         x = self.relu(x)
    25.         x = self.linear2(x)
    26.         y = self.sigmoid(x)
    27.         return y
    28.  
    29. model = Net()
    30. # 保存模型信息
    31. writer.add_graph(model, input_to_model = torch.rand(1, 3, 224, 224))
    32. writer.close()

    3.5 TensorBoard图像可视化

            （1）对于单张图片的显示使用add_image；

            （2）对于多张图片的显示使用add_images；

            （3）有时需要使用torchvision.utils.make_grid将多张图片拼成一张图片后，用writer.add_image显示。

            使用torchvision的CIFAR10数据集为例：

    1. import torchvision
    2. from torchvision import datasets, transforms
    3. from torch.utils.data import DataLoader
    4.  
    5. transform_train = transforms.Compose(
    6.     [transforms.ToTensor()])
    7. transform_test = transforms.Compose(
    8.     [transforms.ToTensor()])
    9.  
    10. train_data = datasets.CIFAR10(".", train=True, download=True, transform=transform_train)
    11. test_data = datasets.CIFAR10(".", train=False, download=True, transform=transform_test)
    12. train_loader = DataLoader(train_data, batch_size=64, shuffle=True)
    13. test_loader = DataLoader(test_data, batch_size=64)
    14.  
    15. images, labels = next(iter(train_loader))
    16.  
    17. # 仅查看一张图片
    18. writer = SummaryWriter('./pytorch_tb')
    19. writer.add_image('images[0]', images[0])
    20. writer.close()
    21.  
    22. # 将多张图片拼接成一张图片，中间用黑色网格分割
    23. # create grid of images
    24. writer = SummaryWriter('./pytorch_tb')
    25. img_grid = torchvision.utils.make_grid(images)
    26. writer.add_image('image_grid', img_grid)
    27. writer.close()
    28.  
    29. # 将多张图片直接写入
    30. writer = SummaryWriter('./pytorch_tb')
    31. writer.add_images("images",images,global_step = 0)
    32. writer.close()

    3.6 TensorBoard连续变量可视化

            通过add_scalar实现

    1. writer = SummaryWriter('./pytorch_tb')
    2. for i in range(500):
    3.     x = i
    4.     y = x**2
    5.     writer.add_scalar("x", x, i) #日志中记录x在第step i 的值
    6.     writer.add_scalar("y", y, i) #日志中记录y在第step i 的值
    7. writer.close()

    3.7 TensorBoard参数分布可视化

            通过add_histogram实现

    1. import torch
    2. import numpy as np
    3.  
    4. # 创建正态分布的张量模拟参数矩阵
    5. def norm(mean, std):
    6.     t = std * torch.randn((100, 20)) + mean
    7.     return t
    8.  
    9. writer = SummaryWriter('./pytorch_tb/')
    10. for step, mean in enumerate(range(-10, 10, 1)):
    11.     w = norm(mean, 1)
    12.     writer.add_histogram("w", w, step)
    13.     writer.flush()
    14. writer.close()

    参考：PyTorch可视化

  • 相关阅读:
    Pycharm中使用远程JupyterLab以及JupyterHub登录问题
    Google 开源库Guava详解
    unordered_set && unordered_map 的封装
    会计账簿、会计账簿概述、会计账簿的启用与登记要求、会计账簿的格式和登记方法
    java毕业设计成都某4S店销售管理系统Mybatis+系统+数据库+调试部署
    SDP护航车企数智化（一）丨深挖汽车前市场 MOT
    启发式搜索: A*算法
    【技术积累】Linux中的命令行【理论篇】【六】
    pandas实现列转行
    （10）点云数据处理学习——基于手机视频和工具实现三位点云
  • 原文地址：https://blog.csdn.net/qq_51167531/article/details/128043195