Visualize Training Command Guide¶

The sequifier visualize-training command reads the structured metric files generated during training and hyperparameter search to create interactive Plotly HTML visualizations of the training and validation losses. It supports viewing a single model’s progress or comparing multiple models side-by-side.

Usage¶

# Visualize a single model
sequifier visualize-training my-model-name

# Visualize multiple models side-by-side
sequifier visualize-training model-A,model-B,model-C

# Visualize every run from a hyperparameter search
sequifier visualize-training my-hyperparameter-search

# Visualize models listed in a text file
sequifier visualize-training path/to/models.txt --log-scale

Arguments¶

Unlike other commands that rely on a YAML config, visualize-training is configured directly via command-line arguments.

Argument	Type	Default	Description
`models`	`str`	Required	A model name, hyperparameter-search name, comma-separated list of model names, or path to a `.txt` file containing model names (one per line). A search name includes all models named `[SEARCH]-run-[NUMBER]`.
`--log-scale`	`flag`	`False`	Use a logarithmic scale on the y-axis for the loss curves.
`--bucket-training-batches`	`int`	`null`	Smooths the training loss curve by averaging the loss over a specified number of batches. Must be a multiple of the logged batch interval used during training.
`--project-root`	`str`	`.`	The root directory of your Sequifier project.

For a single-dataset model, the command reads logs/[MODEL_NAME]/[MODEL_NAME]-training-full.csv and logs/[MODEL_NAME]/[MODEL_NAME]-validation-full.csv. The corresponding files without -full contain only condensed global-loss records.

Outputs¶

The interactive HTML reports are saved in the outputs/visualization/ directory.

Single Model: outputs/visualization/[MODEL_NAME]-training-visualization.html (Includes global losses and normalized variable validation losses if applicable).
Multiple Models: outputs/visualization/multi-model-training-visualization.html (Side-by-side comparison of validation and training losses across all specified models).
Hyperparameter Search: outputs/visualization/[SEARCH_NAME].html (Includes all valid runs and lists skipped invalid runs and their reasons).

If every run in a hyperparameter search is invalid, Sequifier still creates the report with an empty plot and the invalid-run list.

When comparing multiple models, their initial baseline validation loss must match unless SKIP_BASELINE_CHECK or SEQUIFIER_SKIP_BASELINE_CHECK is set.