text_classification: registry + configs + prompts + assemble loader for 70 re-collected cells (token embedding frozen)
11d8066 | # Text classification: reassembling pool checkpoints | |
| Each text-classification cell `<arch>__<dataset>` has | |
| | file | content | | |
| |---|---| | |
| | `zoo/<dataset>/<arch>/<model_id>.pth` | base checkpoint (full `model_state_dict`, training config, val metric; word-level models also store the vocabulary `stoi`) | | |
| | `ckpt/text_classification/<arch>__<dataset>/<model_id>/collect_000.pth ... collect_199.pth` | 200 pool checkpoints | | |
| | `metadata/configs/text_classification/<cell>/<model_id>/config.json` | base training config | | |
| | `metadata/configs/text_classification/<cell>/<model_id>/collect_config.json` | collection protocol | | |
| | `metadata/registry/text_classification.json` | one record per cell: base training info, collect config, per-checkpoint logs, pool statistics | | |
| ## The token embedding is not in the pool files | |
| The pool is collected by continuing training from the converged base model with a fresh AdamW (small learning rate, | |
| `collect_lr` in `collect_config.json`) and saving one checkpoint every `save_every` steps. During collection the | |
| vocabulary **token-embedding matrix is frozen**, so it is identical in all 200 pool members and in the base model. | |
| It is therefore stored only once, in the base checkpoint, and left out of every pool file. | |
| A pool file is a dict with | |
| - `model_state_dict`: every tensor of the model except the frozen keys (float32); | |
| - `frozen_keys`: the omitted keys, e.g. `["bert.embeddings.word_embeddings.weight"]`; for `t5small` the three tied copies | |
| `transformer.shared.weight`, `transformer.encoder.embed_tokens.weight`, `transformer.decoder.embed_tokens.weight`; | |
| - `frozen_source`: the local path of the base checkpoint at collection time (on the Hub it is the `ckpt_path` of the | |
| registry record, `zoo/<dataset>/<arch>/<model_id>.pth`); | |
| - `collect_step`, `train_step`, `val_metric`, `val_loss`. | |
| The parameters a generator has to produce are `generated_params` in the registry record (total minus `frozen_params`). | |
| ## Reassembling a full model | |
| ```python | |
| from assemble import assemble | |
| sd = assemble("zoo/agnews/bertbase/bertbase_010.pth", | |
| "ckpt/text_classification/bertbase__agnews/bertbase_010/collect_000.pth") | |
| model.load_state_dict(sd, strict=True) | |
| ``` | |
| `assemble(base_ckpt_path, pool_ckpt_path)` copies each key in `frozen_keys` from the base's `model_state_dict` into the | |
| pool `model_state_dict` and returns the full state_dict. The same function applies to generated weights: fill the | |
| generated tensors, then add the frozen keys from the base checkpoint. For every cell the collection script checked that | |
| the reassembled `collect_000.pth` reproduces its logged validation metric (`training_info.assemble_check` in the registry). | |