pycharm的模型训练可视化

可视化

Tensorboard的可视化

1.图片可视化
并排可视化几张图片：

    if torch.cuda.is_available():
        edge_predic_X_binary = edge_predic_X_binary.cuda()
        edge_gt_X_binary = edge_gt_X_binary.cuda()
    concatloss = vutils.make_grid(torch.cat((edge_predic_X_binary,edge_gt_X_binary,edge_predic_X_binary-edge_gt_X_binary),0))

    writer.add_image('edge_predic_X_binary and edge_gt_X_binary',
                     concatloss,
                     global_step=1)

前后可视化几张图片

#在‘I1 and I2’窗口中连续显示两张图片，I1_ori_img和I2_ori_img_fig
if glob_iter % 100 == 0:
	I1_ori_img = cv2.normalize(I.cpu().detach().numpy()[0, 0, ...], None, 0, 255, cv2.NORM_MINMAX, cv2.CV_8U)
	I2_ori_img_fig = cv2.normalize(I2_ori_img.cpu().detach().numpy()[0, 0, ...], None, 0, 255, cv2.NORM_MINMAX,cv2.CV_8U)
	writer.add_image('I1 and I2',I1_ori_img,global_step=1, dataformats='HW')
     writer.add_image('I1 and I2',I2_ori_img_fig,  global_step=2,dataformats='HW')

查看某些层数据的分布和梯度的分布

for name, layer in net.named_parameters():
    if layer.requires_grad == True:
        writer.add_histogram(name + '_grad', layer.grad.cpu().data.numpy(), glob_iter)
        writer.add_histogram(name + '_data', layer.cpu().data.numpy(), glob_iter)

查看某些标量

一个窗口绘制多个saclar的变化，或者只绘制一个scalar的变化
writer.add_scalars('Loss_group', {'feature_loss': loss_feature.item()}, glob_iter)
writer.add_scalars('Loss_group', {'L1_loss': l1_loss.item()}, glob_iter)
writer.add_scalar('learning rate', scheduler.get_lr()[0], glob_iter)

TensorWatch的可视化

参考：https://github.com/microsoft/tensorwatch

使用graphviz+torchviz来可视化模型

参考https://blog.csdn.net/qq_27825451/article/details/96856217
官网：https://github.com/szagoruyko/pytorchviz/blob/master/examples.ipynb

权重可视化，特征图可视化，卷积核可视化：参考https://zhuanlan.zhihu.com/p/54947519