Separate data config for different splits

Subject of the issue

When predict with config labeled=true, without providing label files, error raised: WARNING:root:sample_label is not used in predict. It is True if and only if mode == ‘train’. WARNING:root:Using customized configuration.The code might break if the config doesn’t match the saved model. 2021-02-02 19:58:43.570942: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library ‘libcuda.so.1’; dlerror: libcuda.so.1: cannot open shared object file: No such file or directory 2021-02-02 19:58:43.570972: W tensorflow/stream_executor/cuda/cuda_driver.cc:312] failed call to cuInit: UNKNOWN ERROR (303) 2021-02-02 19:58:43.570986: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:156] kernel driver does not appear to be running on this host (precision7750-yhu): /proc/driver/nvidia/version does not exist Traceback (most recent call last): File “demos/paired_ct_lung/demo_predict.py”, line 59, in <module> config_path=config_path, File “/home/yipenghu/git/DeepReg/deepreg/predict.py”, line 238, in predict repeat=False, File “/home/yipenghu/git/DeepReg/deepreg/util.py”, line 42, in build_dataset data_loader = get_data_loader(dataset_config, mode) File “/home/yipenghu/git/DeepReg/deepreg/dataset/load.py”, line 47, in get_data_loader config=data_loader_config, default_args=default_args File “/home/yipenghu/git/DeepReg/deepreg/registry.py”, line 262, in build_data_loader category=DATA_LOADER_CLASS, config=config, default_args=default_args File “/home/yipenghu/git/DeepReg/deepreg/registry.py”, line 141, in build_from_config return cls(**args) File “/home/yipenghu/git/DeepReg/deepreg/dataset/loader/paired_loader.py”, line 71, in init dir_paths=data_dir_paths, name=“moving_labels”, grouped=False File “/home/yipenghu/git/DeepReg/deepreg/dataset/loader/nifti_loader.py”, line 40, in init self.set_data_structure() File “/home/yipenghu/git/DeepReg/deepreg/dataset/loader/nifti_loader.py”, line 65, in set_data_structure ), f"directory {named_dir_path} does not exist" AssertionError: directory demos/paired_ct_lung/dataset/test/moving_labels does not exist

Your environment

• DeepReg version bd3202ace94151d657c3daf9c49ce06563fdc556

• OS: Ubuntu 20.04

• Python Version: 3.7

• TensorFlow

  • TensorFlow Version 2.3

Steps to reproduce

remove fixed_labels and moving_labels in dataset/test folder of the demo, e.g. paired_ct_lung

run python demos/paired_ct_lung/demo_predict.py

Expected behaviour

It should allow prediction without any labels

Actual behaviour

1- It only allows prediction when the config key changed to “false” - this is a confusing behaviour as the labeled key should only control the training.

2- When debugging, the docstring isn’t clear: L163 in network.py: it is not clear when labels are not provided, if the labels should be “None” or else.

3- double check: The availability of the labels should not affect the model output.

@mathpluscode can you have a look.

[mathpluscode]

I will limit this issue as config change, then decouple the train/predict in a separate issue, so that we do not modify 100 files at a time.

[mathpluscode]

I’m now working on this issue.

So the reality is that, we may have different availabilities of labels for

• train/valid
• test

I can see two options

1. For prediction script, we can overwrite labeled flag using flags.
2. Or alternatively, we need to change data config as

dataset:
  train:
    dir: "data/test/nifti/unpaired/train" # required
    labeled: true
  valid:
    dir: "data/test/nifti/unpaired/train" # required
    labeled: true
  test:
    dir: "data/test/nifti/unpaired/train" # required
    labeled: false

We can even allow mixed format like nifit/h5 in the future.

@YipengHu what do u think?

I prefer the “new” config. It is clearer. And I insist to deal with this as a bug fix as the current can be ambiguous. What do you think?

Yes, this new config unlocks more options.

Bug fix or new feature, it’s just different names 😂