search

facebookresearch / spiritlm

Inference code for the paper "Spirit-LM Interleaved Spoken and Written Language Model".
Other
828 stars 53 forks source link

colab - loading the expressive model. AttributeError: 'Spiritlm' object has no attribute 'is_expressive_model' #20

Open rowntreerob opened 1 week ago

rowntreerob commented 1 week ago

running in colab, loading model from drive with:

spirit_lm = Spiritlm("/content/drive/MyDrive/data/checkpoints/spiritlm_model/spirit-lm-expressive-7b")

then, running generation step from standard example , it has error below:

image

and the logger has following : INFO:root:test INFO:spiritlm.model.spiritlm_model:Loading SPIRIT-LM model from the path /content/drive/MyDrive/data/checkpoints/spiritlm_model/spirit-lm-expressive-7b... INFO:spiritlm.model.spiritlm_model:SPIRIT-LM model is loaded. INFO:spiritlm.model.spiritlm_model:Loading SPIRIT-LM speech tokenizers ... INFO:spiritlm.model.spiritlm_model:SPIRIT-LM speech tokenizers are loaded. INFO:DET:testRR

rowntreerob commented 1 week ago

ban_expressivity_tokens=True

without the rest of the "if then" @233 works fine.

hitchhicker commented 1 week ago

Hey @rowntreerob ! Thanks for raising the issue, it is indeed an issue, the variable is_expressive_model is not initialised correctly in here when the model name is a path (in your example "/content/drive/MyDrive/data/checkpoints/spiritlm_model/spirit-lm-expressive-7b"), we can reuse the function _ensure_model_name from here to get correct model name.

You are welcome to create a PR to fix it.

kingfener commented 3 days ago

fix file in spiritlm/model/spiritlm_model.py :change 184 and 187: from : if name == to : if os.path.basename(name) ==