Skip to content

Latest commit

 

History

History
52 lines (40 loc) · 2.19 KB

DeepSpeech2.md

File metadata and controls

52 lines (40 loc) · 2.19 KB
Raw

PyTorch DeepSpeech

Setup AI Model Efficiency Toolkit (AIMET)

Please install and setup AIMET before proceeding further. This model was tested with the torch_gpu variant of AIMET 1.22.2.

Setup DeepSpeech2

  • Clone the SeanNaren DeepSpeech2 Repo
    git clone https://github.com/SeanNaren/deepspeech.pytorch.git

  • checkout this commit id:
    cd deepspeech.pytorch
    git checkout 78f7fb791f42c44c8a46f10e79adad796399892b

  • Append the repo locations to your PYTHONPATH with the following:

export PYTHONPATH=$PYTHONPATH:<path to parent>/deepspeech.pytorch
export PYTHONPATH=$PYTHONPATH:<path to parent>/aimet-model-zoo
  • Install requirements :
    pip install -r aimet_zoo_torch/deepspeech2/evaluators/requirements.txt

Obtain the Test Dataset

  • The evaluation script will automatically download the model checkpoint from here.

  • Run the command below to download the dataset and format the csv needed for the test-manifest flag.

python3 deepspeech.pytorch/data/librispeech.py --files-to-use test-clean.tar.gz

Details are available on the Datasets Section in the SeanNaren Repo.

Usage

  • To run evaluation with QuantSim in AIMET, use the following
python deepspeech2_quanteval.py \
  --test-manifest=<path to test manifest csv>

Quantization Configuration

In the evaluation script included, we have manually configured the quantizer ops with the following assumptions:

  • Weight quantization: 8 bits, per tensor asymmetric quantization
  • Bias parameters are not quantized
  • Model inputs are quantized
  • Activation quantization: 8 bits, asymmetric quantization
    • Inputs to Conv layers are quantized
    • Input and recurrent activations for LSTM layers are quantized
  • Quantization scheme is tf enhanced
  • Operations which shuffle data such as reshape or transpose do not require additional quantizers