Convolution kernel and characteristic graph visualization

Recently read AlexNet This article CNN It's the beginning of the mountain , There is the visualization of convolution kernel of convolution layer , So record that , Other networks can follow suit , Hope to get a valuable praise .

One 、 Convolution kernel Visualization

1、 Prepare a trained model

It is suggested to use a trained model for visualization , The visualized results can help to observe some potential characteristics （ What I'm using here is AlexNet Pre training model ）.

2、 Convolution kernel Visualization

Here is a simple comb ：
First layer convolution kernel ：torch.Size([64, 3, 11, 11]),
Number of output channels ：64, The number of corresponding convolution kernels
Enter the number of channels ：3, Number of channels corresponding to convolution kernel
Convolution kernel width ：11,
Convolution kernel is high ：11
Single channel convolution kernel visualization see the following figure for multi-channel convolution kernel visualization ：

The code is as follows ：

import os
import torch
import torch.nn as nn
from torch.utils.tensorboard import SummaryWriter
import torchvision.utils as vutils
import torchvision.models as models

BASE_DIR = os.path.dirname(os.path.abspath(__file__))
log_dir = os.path.join(BASE_DIR, "results")
writer = SummaryWriter(log_dir=log_dir, filename_suffix="_kernel")

path_state_dict = os.path.join("data", "alexnet-owt-4df8aa71.pth")
alexnet = models.alexnet()
alexnet.load_state_dict(torch.load(path_state_dict))

kernel_num = -1
vis_max = 1

#  Take the first two convolution kernels 
for sub_module in alexnet.modules():
    if not isinstance(sub_module, nn.Conv2d):
        continue
    if kernel_num >= vis_max:
        break
    kernel_num += 1
    kernels = sub_module.weight
    c_out, c_int, k_h, k_w = tuple(kernels.shape)  #  Number of output channels , Enter the number of channels , Convolution kernel width , Convolution kernel is high 
    print(kernels.shape)
    for o_idx in range(c_out):
        kernel_idx = kernels[o_idx, :, :, :].unsqueeze(1)  #  get (3, h, w),  however make_grid need  BCHW, Expand here C Dimension becomes （3, 1, h, w）
        kernel_grid = vutils.make_grid(kernel_idx, normalize=True, scale_each=True, nrow=8)  #  The convolution kernel is visualized in the grid 
        # nrow: Number of images displayed per line 
        writer.add_image('{}_Convlayer_split_in_channel'.format(kernel_num), kernel_grid, global_step=o_idx)
    #  name , picture , Picture number 
    kernel_all = kernels.view(-1, 3, k_h, k_w)  # 3, h, w
    kernel_grid = vutils.make_grid(kernel_all, normalize=True, scale_each=True, nrow=8)  # c, h, w
    writer.add_image('{}_all'.format(kernel_num), kernel_grid, global_step=620)
    print("{}_convlayer shape:{}".format(kernel_num, tuple(kernels.shape)))

Supplementary operation
After running the code, a results file , Next
Execute the following code ：tensorboard --logdir= Stored results File path )
Click the following link to jump to the visual interface

The results are shown in the following figure

Two 、 Visual feature map

Only the first floor is shown here , Input the visualization of the feature layer after the first layer convolution of a picture
The code is as follows ：

BASE_DIR = os.path.dirname(os.path.abspath(__file__))
log_dir = os.path.join(BASE_DIR, "results")

path_state_dict = os.path.join("data", "alexnet-owt-4df8aa71.pth")
alexnet = models.alexnet()
alexnet.load_state_dict(torch.load(path_state_dict))
writer = SummaryWriter(log_dir=log_dir, filename_suffix="_feature map")

#  data 
path_img = os.path.join(BASE_DIR,"data", "tiger cat.jpg")  # your path to image
img_transforms = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.49139968, 0.48215827, 0.44653124], [0.24703233, 0.24348505, 0.26158768])
])

img_pil = Image.open(path_img).convert('RGB')
img_tensor = img_transforms(img_pil)
img_tensor.unsqueeze_(0)  # chw -> bchw


convlayer1 = alexnet.features[0]  #  The first convolution 
fmap_1 = convlayer1(img_tensor)

fmap_1.transpose_(0, 1)  # bchw=(1, 64, 55, 55) --> (64, 1, 55, 55)
fmap_1_grid = vutils.make_grid(fmap_1, normalize=True, scale_each=True, nrow=8)

writer.add_image('feature map in conv1', fmap_1_grid, global_step=620)
writer.close()

The following are the original and effect drawings ：

If there are any deficiencies, please point out , Finally, ask for a little praise again .