Pretrained Prophet checkpoint fails to load due to embedding size mismatch

[vsarsam]

When initializing a Prophet model with one of the pretrained checkpoints from HuggingFace and the published iv/cell-line embedding files from the Mendeley dataset (as referenced in your README), loading the model fails with a RuntimeError due to a mismatch in embedding matrix size.

I am following the official example from your tutorial like that:

pretrained_checkpoint_path = '../epoch=29-step=45360.ckpt'
model = Prophet(
    iv_emb_path='../prophet/embeddings/global_iv_scaledv3.csv',
    cl_emb_path='../prophet/embeddings/cell_line_embedding_full_ccle_300_scaled.csv',
    ph_emb_path=None,
    model_pth=pretrained_checkpoint_path,
)

Error message

RuntimeError: Error(s) in loading state_dict for TransformerPredictor:
    size mismatch for learnable_embedding.weight:
        copying a param with shape torch.Size([1000, 512]) from checkpoint,
        the shape in current model is torch.Size([2000, 512]).

The model construction fails due to a size mismatch in learnable_embedding.weight, suggesting that:

• the checkpoint expects 1000 embeddings,
• while the provided embedding files define 2000 embeddings.

This makes it unclear whether:

1. the published embedding CSVs https://data.mendeley.com/datasets/g7z3pw3bfw/1 match the pretrained models on HuggingFace, or

2. the pretrained checkpoints were trained with different internal embedding sizes.