| --- |
| license: mit |
| tags: |
| - fashion-mnist |
| - image-classification |
| --- |
| |
| # CO3133 course project checkpoints, group DeepDive |
|
|
| Selected runs for the three course assignments. Weights are never committed to git; this |
| repository is the only place they live. |
|
|
| ``` |
| <assignment>/runs/<run_id>/ |
| checkpoint.pt weights, selected by the checkpoint rule in config.yaml |
| config.yaml the resolved configuration that produced them |
| environment.json commit hash, library versions, hardware, thread count |
| history.json one record per epoch |
| summary.json parameters, best epoch, training time, selected metric |
| checksums.json SHA-256 of every file above |
| ``` |
|
|
| ## Reconstructing a checkpoint |
|
|
| Clone [the repository](https://github.com/longlephamtien/DL261-Assignment) at the commit |
| recorded in `environment.json`, then run the model configuration named in `config.yaml` with |
| the seed it records: |
|
|
| ```sh |
| cd assignments |
| python -m venv .venv && source .venv/bin/activate |
| pip install -e . |
| |
| cd assignment-1 |
| python -m src.data --config configs/base.yaml |
| python -m src.train --config configs/<model>.yaml --set seed=<seed> |
| ``` |
|
|
| The dataset split is fixed and committed, so the run reproduces the reported accuracy and |
| macro-F1 on any machine. Wall-clock times depend on the hardware recorded in |
| `environment.json`. |
|
|
