Useful Tools
Most useful tools are in tools folder
Visualize Network Structure
You can use visualize_network.py to check your network structure
The method of using the file is as follows:
init network layers in function _init_layers
add parameters in cfg_model if need
run:
cd tools/
python visualize_network.py
tensorboard --logdir runs/XXX # XXX means folder name
Usually you can see the network structure at http://localhost:6006/#
Train Process Monitoring and Training Details Review
You can always know the details while traning (usually at http://localhost:8097/). We currently support visualize loss and image during training process. But at present, we only support manual addition loss and image.
The method of adding loss and image visualization are as follows:
losses['loss_XXX'] = loss_XXX.data.cpu() # adding loss visualization
XXX = normimage(XXX, save_cfg=save_cfg) # adding image visualization
shows.append(XXX.transpose([2, 0, 1]))
Every training process we save a new log file and you can see the option and training details from this log file. Also, you can use tensorboard to visualize the loss curve:
cd XXXX # workdir root
tensorboard --logdir tf_logs/XXX # XXX means file name
Test Time Argument x8
You can ues test_TTAx8.py to use Test Time Argument(TTA)x8. Testing way is same as test.py
Tensorflow to PyTorch
You can use tf2torch.py to convert Tensorflow checkpoint to PyTorch checkpoint, which can be loaded by PyTorch Network.
The method of using the file is as follows:
change label name (checkpoint dir name)
init pytorch based network, it is important that the torch conv layer name should same as tf name!
change save path root.
run.
Output Feature Map
You can use output_featuremap.py to output the feature map of network middle layer. The method of using the file is as follows:
change args.config and args.load_from root
add the layer name in name_list
change the output root and image root
run
python output_featuremap.py
Get WaterNet Data
Copy from Water-Net-Code, it’s a MATLAB code, you can use generate_test_data.m to obtain Histogram Equalization(HE), Gamma Correction(GC) and White Balance(WB) processed image.