在实时训练网络中显示CAM结果

在实时训练网络中显示CAM结果

paper: Learning Deep Features for Discriminative Localization

主要代码

def ComputeCAM(features, softmax_weight, class_ids):
    class_ids = class_ids.long()
    m = []
    for idx in class_ids:
        wei = softmax_weight[idx, :].unsqueeze(0)
        m.append(wei)
    tensor_m = torch.cat(m, dim=0).unsqueeze(2).unsqueeze(3)
    cam = tensor_m * features
    cam = cam.sum(dim=1).unsqueeze(1)
    cam =  (cam - cam.min()) / (cam.max() - cam.min())
    weight = torch.nn.functional.interpolate(cam, scale_factor=8, mode='bilinear', align_corners=False)  # reshape its size as input
    return weight
    
def GetCam(model, inputs):
    out, features = model(inputs)
    _, idx = torch.max(out, 1)
    params = list(model.parameters())
    softmax_weight = params[-2].squeeze()
    weight = ComputeCAM(features, softmax_weight, idx)
    return weight

if __name__ == "__main__":
    x = torch.rand(10, 3, 32, 32).to(device)
    z = GetCam(net, x)
    print(z.size())

完整代码链接： https://github.com/Jixianrui/CAM-with-ResNet

结果展示

时间原因，该结果为只训练了10个epoch之后的结果