Trying to finetune on a dataset

[marcopeix]

• Working in Google Colab
• Using the latest version and cloned the repo inside Colab

I managed to prepare the data. I manually split my dataset into a train/val set and saved the dataset using HF in the data directory` at the root of the repo like this:

url = "https://raw.githubusercontent.com/marcopeix/TimeSeriesForecastingUsingFoundationModels/refs/heads/main/data/walmart_sales_small.csv"
df = pd.read_csv(url, index_col='Date', parse_dates=True)
df = df[df['Store'] == 1]
finetune_df = df[:-32]
train_df, val_df = finetune_df[:-16], finetune_df[-16:]

def multivar_gen_func(df) -> Generator[dict[str, Any], None, None]:
    features_df = df.drop(['Store', 'Weekly_Sales'], axis=1)
    yield {
        "target": df['Weekly_Sales'].to_numpy(),
        "features": features_df.to_numpy().T,
        "start": df.index[0],
        "freq": pd.infer_freq(df.index),
        "item_id": "1",
    }
features = Features(
    dict(
        target=Sequence(Value("float32")),  
        features=Sequence(
            Sequence(Value("float32")), 
            length=len(df.columns) - 2  
        ),
        start=Value("timestamp[s]"),
        freq=Value("string"),
        item_id=Value("string"),
    )
)
hf_train_ds = datasets.Dataset.from_generator(
    lambda: multivar_gen_func(train_df), 
    features=features
)
hf_val_ds = datasets.Dataset.from_generator(
    lambda: multivar_gen_func(train_df),  # Pass df to the generator function
    features=features
)
hf_train_ds.save_to_disk("/content/uni2ts/data/store_ds")
hf_val_ds.save_to_disk("/content/uni2ts//data/store_ds_eval")

Then, I created a YAML configuration file for both training and validation. They are respectively in /content/uni2ts/cli/conf/finetune/data/store_finetune.yaml and /content/uni2ts/cli/conf/finetune/val_data/store_finetune.yaml.

The training YAML file is:

_target_: uni2ts.data.builder.simple.SimpleDatasetBuilder
dataset: store_ds
weight: 1000
storage_path: /content/uni2ts/data

and the validation YAML file is:

_target_: uni2ts.data.builder.ConcatDatasetBuilder
_args_:
  _target_: uni2ts.data.builder.simple.generate_eval_builders
  dataset: store_ds_eval
  offset: 95
  eval_length: 8
  prediction_lengths: [8]
  context_lengths: [16]
  patch_sizes: [16]

Then, I run the training command as specified:

python -m cli.train \
  -cp conf/finetune \
  run_name=store_sales_finetune \
  model=moirai_1.0_R_small \
  data=store_finetune \
  val_data=store_finetune

And it throws this error:

Error executing job with overrides: ['run_name=store_sales_finetune', 'model=moirai_1.0_R_small', 'data=store_finetune', 'val_data=store_finetune']
Error in call to target 'uni2ts.data.builder.simple.SimpleDatasetBuilder':
TypeError('expected str, bytes or os.PathLike object, not NoneType')
full_key: data

My questions:

1. What am I missing to fix this error?
2. How can I specify the maximum number of training steps?
3. What is the offset parameter in the validation YAML file?

Thanks in advance for your help!

[chenghaoliu89]

Hi @marcopeix

1. Have you add your datapath into .env file? You can check the readme file for fine tuning guideline echo "CUSTOM_DATA_PATH=PATH_TO_SAVE" >> .env
2. you can add the argument in your cmd or directly modify it from conf/finetune/default.yaml file by setting trainer.max_epochs
3. date_offset (datetime string) or offset (integer) options determines the last time step of the fine-tuning train set. The validation set will be saved as DATASET_NAME_eval.

[marcopeix]

Yes, I specified the CUSTOM_DATA_PATH in an .env file at the root of the repository. When I do !cat .env, I get CUSTOM_DATA_PATH="/content/uni2ts/data", but I still get the error:

/usr/local/lib/python3.11/dist-packages/uni2ts/common/env.py:43: UserWarning: Failed to load .env file.
  warnings.warn("Failed to load .env file.")
Error executing job with overrides: ['run_name=store_sales_finetune', 'model=moirai_1.0_R_small', 'data=store_finetune', 'val_data=store_finetune']
Error in call to target 'uni2ts.data.builder.simple.generate_eval_builders':
TypeError('expected str, bytes or os.PathLike object, not NoneType')
full_key: val_data._args_

[chenghaoliu89]

Yes, I specified the CUSTOM_DATA_PATH in an .env file at the root of the repository. When I do !cat .env, I get CUSTOM_DATA_PATH="/content/uni2ts/data", but I still get the error:

/usr/local/lib/python3.11/dist-packages/uni2ts/common/env.py:43: UserWarning: Failed to load .env file.
  warnings.warn("Failed to load .env file.")
Error executing job with overrides: ['run_name=store_sales_finetune', 'model=moirai_1.0_R_small', 'data=store_finetune', 'val_data=store_finetune']
Error in call to target 'uni2ts.data.builder.simple.generate_eval_builders':
TypeError('expected str, bytes or os.PathLike object, not NoneType')
full_key: val_data._args_

Could you print self.storage_path and self.dataset in SimpleDatasetBuilder.load_dataset when you run fine-tuning script