计算机视觉可解释性——将特征图的某一通道转化为一张图来可视化

本实验是将某一层的特征图中的某一通道对其进行的可视化，有两种方式一种是使用opencv；另一种是使用tensorboard。

1.使用opencv

import cv2
import torch
import numpy as np
from PIL import Image
from torchvision import models, transforms
import warnings
import argparse

warnings.filterwarnings('ignore')

def parse_args():
    parser = argparse.ArgumentParser(description='Show Feature Map by Pytorch')
    parser.add_mutually_exclusive_group()
    parser.add_argument('--img_root',
                        type=str,
                        default='cat.jpg',
                        help='image path')
    parser.add_argument('--basenet',
                        type=str,
                        default='resnet50',
                        help='train base model')
    parser.add_argument('--cuda',
                        type=str,
                        default=True,
                        help='if we can use cuda')
    return parser.parse_args()

args = parse_args()

# img processing
def get_img_info(img_root):
    img_info = Image.open(img_root).convert('RGB')
    img_transforms = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])
    img_info = img_transforms(img_info)
    # C,H,W -> 1,C,H,W
    img_info = img_info.unsqueeze(0)
    return img_info

# get No. K feature map
# vgg16
def model_get_feature_map_vgg(model, x):
    with torch.no_grad():
        feature_map = model.features(x)
        return feature_map

# resnet50
def model_get_feature_map_resnet(model, x):
    with torch.no_grad():
        x = model.conv1(x)
        x = model.bn1(x)
        x = model.relu(x)
        x = model.maxpool(x)
        feature = x = model.layer1(x)
        x = model.layer2(x)
        x = model.layer3(x)
        x = model.layer4(x)
        return feature, x

def show_k_feature_map(feature_map, k):
    feature_map = feature_map.squeeze(0)
    feature_map = feature_map.cpu().numpy()
    for index, feature_map_i in enumerate(feature_map):
        feature_map_i = np.array(feature_map_i * 255, dtype=np.uint8)
        feature_map_i = cv2.resize(feature_map_i, (224, 224), interpolation=cv2.INTER_NEAREST)
        if k == index + 1:
            feature_map_i = cv2.applyColorMap(feature_map_i, cv2.COLORMAP_JET)
            cv2.imwrite("{}.jpg".format(str(index + 1)), feature_map_i)

if __name__ == '__main__':
    img_info = get_img_info(args.img_root)
    k = 10
    if args.basenet == 'vgg16':
        model = models.vgg16(pretrained=True)
    elif args.basenet == 'resnet50':
        model = models.resnet50(pretrained=True)
    else:
        raise ValueError("the model type is error!")
    if args.cuda == torch.cuda.is_available():
        img_info = img_info.cuda()
        model = model.cuda().eval()
    else:
        img_info = img_info.cpu()
        model = model.cpu().eval()
    if args.basenet == 'vgg16':
        feature_map = model_get_feature_map_vgg(model, img_info)
    elif args.basenet == 'resnet50':
        feature_map, _ = model_get_feature_map_resnet(model, img_info)
    show_k_feature_map(feature_map, k)
    print("finished!")

运行结果

使用的resnet50作为backbone的效果更好，将layer1的特征中第一层通道(256中的第1层)转化为图像

layer1. channels1.

将layer1的特征中第一层通道(256中的第1层)转化为图像，将代码改为：

if __name__ == '__main__':
    ...
k=2

layer1. channels2.

layer1. channels3.

...

layer1. channels10.

...

若要提取layer2的特征，将代码改为：

def model_get_feature_map_resnet(model, x):
    with torch.no_grad():
        x = model.conv1(x)
        x = model.bn1(x)
        x = model.relu(x)
        x = model.maxpool(x)
        x = model.layer1(x)
        feature = x = model.layer2(x)
        x = model.layer3(x)
        x = model.layer4(x)
        return feature, x

2.使用tensorboard

import torch
import numpy as np
from PIL import Image
from torchvision import models, transforms
from torch.utils.tensorboard import SummaryWriter
import warnings
import argparse

warnings.filterwarnings('ignore')

def parse_args():
    parser = argparse.ArgumentParser(description='Show Feature Map by Pytorch')
    parser.add_mutually_exclusive_group()
    parser.add_argument('--img_root',
                        type=str,
                        default='cat.jpg',
                        help='image path')
    parser.add_argument('--basenet',
                        type=str,
                        default='resnet50',
                        help='train base model')
    parser.add_argument('--cuda',
                        type=str,
                        default=True,
                        help='if we can use cuda')
    return parser.parse_args()

args = parse_args()

# img processing
def get_img_info(img_root):
    img_info = Image.open(img_root).convert('RGB')
    img_transforms = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])
    img_info = img_transforms(img_info)
    # C,H,W -> 1,C,H,W
    img_info = img_info.unsqueeze(0)
    return img_info

# get No. K feature map
# vgg16
def model_get_feature_map_vgg(model, x):
    with torch.no_grad():
        feature_map = model.features(x)
        return feature_map

# resnet50
def model_get_feature_map_resnet(model, x):
    with torch.no_grad():
        x = model.conv1(x)
        x = model.bn1(x)
        x = model.relu(x)
        x = model.maxpool(x)
        feature = x = model.layer1(x)
        x = model.layer2(x)
        x = model.layer3(x)
        x = model.layer4(x)
        return feature, x

def show_k_feature_map(feature_map, k):
    writer = SummaryWriter("tensorboard")
    feature_map = feature_map.squeeze(0)
    feature_map = feature_map.cpu().numpy()
    for index, feature_map_i in enumerate(feature_map):
        feature_map_i = np.array(feature_map_i * 255, dtype=np.uint8)
        feature_map_i = np.expand_dims(feature_map_i, axis=0)
        if k == index + 1:
            writer.add_image("feature_map", feature_map_i, k)
        writer.close()

if __name__ == '__main__':
    img_info = get_img_info(args.img_root)
    k = 10
    if args.basenet == 'vgg16':
        model = models.vgg16(pretrained=True)
    elif args.basenet == 'resnet50':
        model = models.resnet50(pretrained=True)
    else:
        raise ValueError("the model type is error!")
    if args.cuda == torch.cuda.is_available():
        img_info = img_info.cuda()
        model = model.cuda().eval()
    else:
        img_info = img_info.cpu()
        model = model.cpu().eval()
    if args.basenet == 'vgg16':
        feature_map = model_get_feature_map_vgg(model, img_info)
    elif args.basenet == 'resnet50':
        feature_map, _ = model_get_feature_map_resnet(model, img_info)
    show_k_feature_map(feature_map, k)
    print("finished!")

运行结果

tensorboard --logdir=tensorboard

将layer1的特征中第10层通道(256中的第1层)转化为图像

layer1. channels10.

successed!