Global_net train without regulariser raised error

[YipengHu]

Subject of the issue

Train with global net, without providing regulariser in config file, generates error in config_sanity_check(config)

Yes

Your environment

• DeepReg version (commit hash)

fda6421

• OS (e.g. Ubuntu 20.04, MacOS 10.15, etc.)

  Ubuntu 16.04

• Python Version (3.7, 3.8, etc.)

  3.7

• TensorFlow

  • TensorFlow 2.3
  • CUDA Version 10.1
  • cuDNN Version

Steps to reproduce

python demos/paired_ct_lung/demo_train_global.py

edits: the same for field "'extract_levels'"

[mathpluscode]

global net is essentially a ddf model, so you should have regularize and extract levels?

[YipengHu]

global net is essentially a ddf model, so you should have regularize and extract levels?

It is a bit complicated than it should be.

• global = affine, here anyway
• first, bending energy is invariant to affine transformation
• although gradient norm is not, adding regulariser will penalise the affine, e.g. scaling
• but traditionally, affine registration means optimising affine parameters "freely" - these parameters generate ddf for resampling reasons to warp the moving images, not for regularisation;
• so i think we should do:

1. allow using global without regulariser
2. when with bending !=0, warning or even error
3. when with gradient norms, warning

[mathpluscode]

ok, in #557 I'm now rewriting all these models, and I actually merged affine model into ddf model.

I will solve this issue in the PR or afterwards.

[YipengHu]

ok, in #557 I'm now rewriting all these models, and I actually merged affine model into ddf model.

I will solve this issue in the PR or afterwards.

good - i also would like to add another rigid transform with 6 and 7 degrees of freedom - only the warped_grid function needs an extra option when using the same global_net - we can discuss how to best design the classes/register?

[mathpluscode]

ok, in #557 I'm now rewriting all these models, and I actually merged affine model into ddf model.
I will solve this issue in the PR or afterwards.

good - i also would like to add another rigid transform with 6 and 7 degrees of freedom - only the warped_grid function needs an extra option when using the same global_net - we can discuss how to best design the classes/register?

sure, once #602 is almost ready, will let you know

[mathpluscode]

@YipengHu can you confirm if you stiff have this error?

if yes, plz provide the error log

[YipengHu]

@YipengHu can you confirm if you stiff have this error?

if yes, plz provide the error log

(deepreg) yipenghu@precision7750-yhu:~/git/DeepReg$ python demos/paired_ct_lung/demo_train.py --no-test
2021-01-24 16:30:15.086598: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library 'libcudart.so.10.1'; dlerror: libcudart.so.10.1: cannot open shared object file: No such file or directory
2021-01-24 16:30:15.086633: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.

=======================================================
The training can also be launched using the following command.
deepreg_train --gpu '0' --config_path demos/paired_ct_lung/paired_ct_lung.yaml --log_root demos/paired_ct_lung --log_dir logs_train

Traceback (most recent call last):
File "demos/paired_ct_lung/demo_train.py", line 52, in
log_dir=log_dir,
File "/home/yipenghu/git/DeepReg/deepreg/train.py", line 93, in train
max_epochs=max_epochs,
File "/home/yipenghu/git/DeepReg/deepreg/train.py", line 44, in build_config
config = config_parser.load_configs(config_path)
File "/home/yipenghu/git/DeepReg/deepreg/config/parser.py", line 43, in load_configs
loaded_config = config_sanity_check(config)
File "/home/yipenghu/git/DeepReg/deepreg/config/parser.py", line 112, in config_sanity_check
loss_weight = config["train"]["loss"][name]["weight"]
KeyError: 'regularization'

[YipengHu]

this error can be re-produced by setting config in the demo as follows (name=global and remove regularisation field):
dataset:
dir:
train: "demos/paired_ct_lung/dataset/train"
valid: "demos/paired_ct_lung/dataset/valid"
test: "demos/paired_ct_lung/dataset/test"
format: "nifti"
type: "paired" # paired / unpaired / grouped
labeled: true
moving_image_shape: [96, 96, 104]
fixed_image_shape: [96, 96, 104]
train:
method: "ddf" # the registration method, value should be ddf / dvf / conditional
backbone:
name: "global" # value should be local / global / unet
num_channel_initial: 32 # number of initial channel in local net, controls the size of the network
extract_levels: [0, 1, 2]
loss:
image:
name: "ssd"
weight: 10.0
label:
weight: 1.0
name: "dice"
scales: [0, 1, 2, 4, 8, 16]
preprocess:
data_augmentation:
name: "affine"
batch_size: 4
shuffle_buffer_num_batch: 1
optimizer:
name: "adam"
adam:
learning_rate: 1.0e-5
epochs: 8000
save_period: 1000

[mathpluscode]

ok, will fix it later.