Problem logging an experiment to wandb

Describe the bug Logging a training run works fine, however when logging an experiment run (the evaluation part) with the default parameters, there is a OSError 22 relating to “probabilities_<UNK>.csv”

OSError: [Errno 22] Invalid argument: '/results/classification_Distilbert_2/label_probabilities_<UNK>.csv

To Reproduce Steps to reproduce the behavior:

1. Run an experiment from the CLI with default parameters.
2. Configure the model for text classification.
3. Add the wandb flag

Model config

preprocessing:
    text:
        word_tokenizer: hf_tokenizer

input_features:
    -
        name: feature1
        type: text
        level: word
        encoder: distilbert
        pretrained_model_name_or_path: distilbert-base-uncased
    -
        name: feature2
        type: text
        level: word
        encoder: distilbert
        pretrained_model_name_or_path: distilbert-base-uncased
        tied_weights: feature1
    -
        name: feature3
        type: category

output_features:
    -
        name: label
        top_k: 5
        type: category

training:
    epochs: 3
    batch_size: 16
    learning_rate: 0.000005

wandb metadata

{
    "os": "Windows-10-10.0.19041-SP0",
    "python": "3.9.0",
    "heartbeatAt": "2022-03-10T14:21:23.463980",
    "startedAt": "2022-03-10T14:21:17.632459",
    "docker": null,
    "gpu": "NVIDIA GeForce RTX 2070",
    "gpu_count": 1,
    "cpu_count": 6,
    "cuda": null,
    "args": [
        "experiment",
        "--output_directory",
        "results",
        "--experiment_name",
        "classification",
        "--model_name",
        "Distilbert",
        "--training_set",
        "TRAIN_en.arrow",
        "--validation_set",
        "VAL_en.arrow",
        "--test_set",
        "EVAL_en.arrow",
        "--data_format",
        "feather",
        "--config_file",
        "model1.yaml",
        "--wandb"
    ],
    "state": "running",
    "program": "-m __main__",
    "executable": "I:\\Anaconda\\envs\\ludwig\\python.exe"
}

Expected behavior The csv file for the UNK token should be created.

Environment

• OS: Windows-10-10.0.19041-SP0
• Python version: 3.9
• Ludwig version: 0.4.1

Additional context The problem can be avoided with the ludwig API and passing the parameter skip_save_predictions=True (not possible with the CLI).

[brightsparc]

Thanks for this detailed error log @danielduckworth. We have identified the issue as a result of converting from a url which has the Posix file system format, to a local windows path. I have PR #1830 which should resolve this issue.

Please note whilst best effort is made to work on windows, it is not an officially supported platform at the moment.

[brightsparc]

No worries, you can run the filesystem tests on windows using the filesystem marker. Let me know if they catch anything as I’m unable to repo locally.

pytest -m filesystem