Open ivanhe123 opened 2 weeks ago
Hi @ivanhe123, thanks for opening this issue!
From the error, it looks like the keys in the state dict new_model
do not match those in the model SeamlessM4TModel
. You can check the expected keys in the model by doing model_seam.state_dict().keys()
.
Note, it's not necessary for you to download and load in a pretrained checkpoint and then load in new weights. You can initialize a new model with the same architecture and empty weights by just downloading the config:
import torch
from accelerate import init_empty_weights
from transformers import AutoConfig, SeamlessM4TModel
config = AutoConfig.from_pretrained("facebook/hf-seamless-m4t-medium")
with init_empty_weights():
model = SeamlessM4TModel(config)
new_model = torch.load("./expt4_m4tM.pt")
model.load_state_dict(new_model)
System Info
Python version 3.11
transformers
version: 4.42.3Who can help?
No response
Information
Tasks
examples
folder (such as GLUE/SQuAD, ...)Reproduction
Expected behavior
Expected to load the new fintuned model and then save it to a new model file.