VSCode and jsonargparse error

[nicoloesch]

Hi everyone,

I recently switchted to VSCode from PyCharm and I am facing currently an issue. If I try to debug my main.py file, which uses the LightningCLI (DiffLightningCLi inherits from it and just overwrites add_arguments_to_parser):

import sys
import os

from adiff.utils import train_util
import adiff.model.lightning.base_model


def cli_main():
    # Start Training/Testing based on the args

      data_dir = f'/home/{os.getlogin()}/Data/datasets'
      logging_dir = f'/home/{os.getlogin()}/Data/lightning_logs'
      default_cfg = ["config/debug_cfg.yaml"]
      cmdline_args = [
          "fit", 
              "--model", "DiffGen",
              '--task', 'GEN',
              '--data_dir', f'{data_dir}', 
              '--logging_dir', f'{logging_dir}',
              '--data.dimensions', '2',
              '--num_workers', '0',
              '--data.batch_size=5',
              '--data.dataset=cifar-10',
              '--data.in_channels=3',
              '--model.out_channels=3',
              '--trainer.logger=False',
              '--data.num_train_subj=10',
              '--data.num_val_subj=10',
              '--model.diffusion_mean_type=EPSILON',
              '--model.diffusion_loss_type=SIMPLE',
              '--model.diffusion_var_type=FIXED_LARGE',
              '--model.ema_decay=0.9999',
              '--trainer.max_steps=700000',
              ]
    subcommands = ('fit', 'test', 'validate', 'predict')
    parser_kwargs = {subcommand: {'default_config_files': default_cfg} for subcommand in subcommands}

    cli = train_util.DiffLightningCLI(
            model_class=adiff.model.lightning.base_model.BaseModel,
            subclass_mode_model=True,
            run=run,
            seed_everything_default=123,
            parser_kwargs=parser_kwargs,
            save_config_kwargs={"overwrite": True,
                                "multifile": True},
            args=cmdline_args)

    # Shutdown
    sys.exit()


if __name__ == "__main__":
    import warnings
    from lightning_fabric.utilities.warnings import PossibleUserWarning

    with warnings.catch_warnings():
        warnings.filterwarnings(
            action='ignore',
            category=PossibleUserWarning,
            message='The dataloader,')
        cli_main()

I am aware that passing in the args like this is not recommended and is also only used in debugging and not in the final main.py. Running/debugging the script using the launch.json configuration of VSCode

{
    "version": "0.2.0",
    "configurations": [
        {
            "name": "main.py",
            "type": "python",
            "request": "launch",
            "program": "${workspaceFolder}/ADiff/adiff/main.py",
            "console": "integratedTerminal",
            "justMyCode": false,
            "redirectOutput": false,
        }
    ]
}

I get the following error message:

main.py [options] fit: error: Parser key "data":
  Not a valid subclass of LightningDataModule. Got value: NestedArg(key='dimensions', val='2')
  Subclass types expect one of:
  - a class path (str)
  - a dict with class_path entry
  - a dict without class_path but with init_args entry (class path given previously)

The command under the hood in VSCode uses cd /home/u24417317/Code ; /usr/bin/env /home/u24417317/.conda/envs/phd39/bin/python /home/u24417317/.vscode-server/extensions/ms-python.python-2023.12.0/pythonFiles/lib/python/debugpy/adapter/../../debugpy/launcher 60221 -- /home/u24417317/Code/ADiff/adiff/main.py so the correct file is executed.

However, if I run the exact same file using python directly without VSCode, the code runs as intended (as it was running on PyCharm).

(phd39) u24417317@LabServer:~/Code/ADiff/adiff$ python main.py
Global seed set to 123
GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs

I am debugging remotely using ssh (I am running the same file with and without VSCode), made sure that the exact same conda environment is utilised with the following packages:

jsonargparse: 4.22.1
lightning-utilities: 0.8.0
pytorch-lightning: 2.0.4

It seems to me like a VSCode issue, hence why I am aware that this might not be the best place to ask that question but maybe someone has encountered a similar problem or could share their working launch.json configuration.

I appreciate any input!
Nico

[mauvilsa]

I think this is not related to VSCode. I can reproduce the same when running from command line with:

from lightning.pytorch.cli import LightningCLI
from lightning.pytorch.demos.boring_classes import DemoModel

def cli_main():
    cmdline_args = [
        "fit", 
            "--model", "DemoModel",
            '--data.dimensions', '2',
    ]
    cli = LightningCLI(
        model_class=DemoModel,
        subclass_mode_model=True,
        run=True,
        args=cmdline_args)

if __name__ == "__main__":
    cli_main()

The problem is "a dict without class_path but with init_args entry (class path given previously)". That is, --data.dimensions=2 is given, but it is not known which data class to check if it accepts a dimensions parameter. The error goes away if before that argument, there is --data=SomeDataClass.

The error message could be better. To request to improve that, please do in https://github.com/omni-us/jsonargparse/issues.

[nicoloesch]

@mauvilsa thanks for the information! I forgot to mention that I provide the respective class with init_args in a default_cfg.yaml. As such, the class should be fully instantiated with the default arguments and then with the manually placed args (reflecting command line args) overwritten for debugging.

The main problem I have is the following: if I try to debug main.py through VSCode without any other commands, I get the error message. If I execute main.py just using python with the same conda environment, the code runs as expected (and as it has been doing using PyCharm). As a result, there MUST be a difference running main.py through VSCode and without it.

# This is the configuration file for all models inheriting from BaseModel
# It is specified here as class path despite not requiring most of the args
# as in the instantiation of the CLI, I specified model_class as the BaseModel
# and allow the subclass_mode_model
model:
  class_path: adiff.model.lightning.base_model.BaseModel
  init_args:
    activation: 'ReLU'
    act_args:
      - false
    attention: true
    attention_heads: 1
    attention_res:
      - 16
    channel_factor:
      - 1
      - 2
      - 2
      - 2
    classifier_ckpt: null
    concat_sc: true
    conv_zero_init: false
    diffusion_gradient_scale: 1.0
    diffusion_log_var: true
    diffusion_loss_type: SIMPLE
    diffusion_mean_type: EPSILON
    diffusion_noising: 'linear'
    diffusion_steps: 1000
    diffusion_var_type: FIXED_LARGE
    diffusion_verbose: true
    dropout: 10
    hidden_channels: 128
    kernel_size: 3
    learning_rate: 1e-4
    norm: 'group'
    num_fid_samples: 50000
    num_res_blocks: 2
    out_channels: 4
    padding: true
    padding_mode: 'zeros'
    preactivation: true
    spatial_factor: 2
    upd_conv: true
    ema_decay: 0.0
data:
  class_path: adiff.input.data_loader.DataModuleBraTS
  init_args:
    augment: false
    batch_size: 2
    brats_2023_subtask: null
    conditional_sampling: false
    #dataset: 'cifar-10'
    dataset: 'BraTS_2020'
    dimensions: 3
    fg_prob: 3
    in_channels: 4
    multiclass_pred: true
    num_train_subj: 2  # null enables the utilisation of 60%, otherwise explicit
    num_val_subj: 2  # null enables the utilisation of 20%, otherwise explicit
    num_test_subj: 2  # null enables the utilisation of 20%, otherwise explicit
    patches_per_subj: 1
    patch_overlap: 8
    patch_size: 32
    resample: false
    to_ras: false
trainer:
  accelerator: auto
  check_val_every_n_epoch: 1
  detect_anomaly: true
  devices: auto
  deterministic: true
  benchmark: false
  enable_checkpointing: true
  enable_progress_bar: true
  enable_model_summary: true
  fast_dev_run: false
  gradient_clip_val: null
  gradient_clip_algorithm: 'norm'
  inference_mode: true
  limit_train_batches: null
  limit_val_batches: null
  limit_test_batches: null
  limit_predict_batches: null
  log_every_n_steps: 1
  logger:
    class_path: adiff.utils.callbacks.Logger
  max_epochs: 5
  max_steps: -1
  min_epochs: 1
  num_nodes: 1
  num_sanity_val_steps: 0
  precision: 32-true
  strategy: auto
  use_distributed_sampler: false
progress_bar:
  refresh_rate: 1
  process_position: 0
#optimizer:
#  class_path: torch.optim.Adam
#  init_args:
#    lr: 1e-5
#    weight_decay: 0.0
#lr_scheduler:
#  class_path: adiff.utils.callbacks.WarmUpScheduler
#  init_args:
#    warmup: 20

[nicoloesch]

@mauvilsa I figured it out. It turned out that my default configuration file was not found due to an error in the launch.json file of VSCode. I provided the configuration as a relative path to the main.py file. However, the cwd was set to the ${workspaceFolder}, which is two folders above it (as I want to have all my different projects in the same workspace). This lead to not finding the respective default configuration file in _get_default_config_files of jsonargparse/core.py.
Your suggestion of having a more distinct error message could be helpful in the future but is included in the current error messages if one has more experience with the jsonargparse module.

Thank you for your help!
Nico

---

Solution: Setting
"cwd": "${workspaceFolder}/<path_to_main_file>" in launch.json