search

deepmedic / deepmedic

Efficient Multi-Scale 3D Convolutional Neural Network for Segmentation of 3D Medical Scans
BSD 3-Clause "New" or "Revised" License
1.03k stars 342 forks source link

Plotting test performance #172

Closed samieti02 closed 4 years ago

samieti02 commented 5 years ago

Hello Sir,

Is there a possibility to also integrate "plotTestingProgress" ?

DSC (Full Segment) is empty for training after running the code python plotTrainingProgress.py examples/output/logs/trainSessionConf.txt -d . Did I miss something?

Thanks Sir

Kamnitsask commented 5 years ago

Sorry for the late reply, I am currently focusing on writing up by PhD Thesis and time is ridiculously limited.

Thank you, for the suggestion on plotting test results. It would be nice to implement a collection of plotting/evalution scripts, but currently this is a lower priority than focusing on the main model/components. It's lower priority as evaluation/plotting scripts is something that I believe users could actually make it themselves if needed, given the outputs of the model, while I focus on the model/training part. I will leave this issue open for now, perhaps I ll find time in later months for this.

Regarding the DSC being empty for training, this happens if you have disabled: performFullInferenceOnValidationImagesEveryFewEpochs = False

in the trainConfig, or somehow the training logs are corrupted.

samieti02 commented 5 years ago

I also set the parameter performFullInferenceOnValidationImagesEveryFewEpochs = True, and other parameters to True, but for the training in DSC (Full) nothing appears, but for validation, we can see a graph

Kamnitsask commented 5 years ago

I now understand what you meant.

The DSC on whole volumes is never computed for training cases. It's too expensive and there is no point (I think). That's why it's empty.