How to change input signature
bytosaur opened this issue · comments
hey there!
thanks for making this repository! This may be a huge help for me.
When I download the model from https://tfhub.dev/vasudevgupta7/wav2vec2/1 the saved_model_cli says that the input signature for the model is actually (None, 50000) and not (None, 246000)... however when using tfhub to load the model into a keras layer (as done in this cloab ) it is (None, 246000)
i am confused... please help :)
thanks a lot!
Hello @bytosaur,
It's kinda strange that the warning is asking for shape (None, 50_000)
, the model was exported with (None, 246_000)
only. I am not sure why that's happening. But here is the link to code dependent checkpoint (https://huggingface.co/vasudevgupta/gsoc-wav2vec2), in case you would like to export yourself.
I hope this would help in your project!
thanks a lot! I ll try it out right away :)
cool! I was able to convert the model to a sequence length of 80000 using these commands (you can change conda for virtualenv or whatever):
git clone https://github.com/vasudevgupta7/gsoc-wav2vec2
cd gsoc-wav2vec2
conda create -n wav2vec2 python=3.7
conda activate wav2vec2
pip install -r requirements.txt
pip install torch tensorflow transformers
python src/convert_torch_to_tf.py --hf_model_id "facebook/wav2vec2-base"
python src/export2hub.py --model_id tf-wav2vec2-base --saved_model_dir gsoc-wav2vec2/saved-model --seqlen 80000
# show the signatures
saved_model_cli show --dir gsoc-wav2vec2/saved-model/ --tag_set serve --signature_def serving_default
Note: as of now, it is hf_model_id
in the first script but model_id
in the second one
as a side note: tensorflow==2.6 worked for me :)
Glad to know that!