edna_mode

Installation / Setup

1. create venv: python3 -m venv env and activate it with source env/bin/activate.
2. download requirements: pip install requirements.txt and then install tf_slim by first cd slim and then pip install -e ..
   • you might need to pip install -U pip and pip install wheel

Training

1. Navigate to the speech commands directory e.g. tensorflow/tensorflow/tensorflow/examples/speech_commands
2. Run training script with the following command train.py --model_architecture mobilenet_embedding --embedding_size 100
   • Arden: currently using python ./tensorflow/tensorflow/tensorflow/examples/speech_commands/train.py --model_architecture mobilenet_embedding --data_dir /home/arden/Classes/EE292D/dataset/vox1_dev_wav_flat/ --optimizer adam --background_frequency 0 --embedding_size 100 --batch_size 10 --how_many_training_steps "100000,5000" --learning_rate "0.001,0.0001"
   • Phil: currently using python ./tensorflow/tensorflow/tensorflow/examples/speech_commands/train.py --model_architecture mobilenet_embedding --data_dir /Users/philipmateopfeffer/Desktop/stanford/Y5Q1/cs329e/edna_mode/dataset --optimizer adam --background_frequency 0 --embedding_size 50 --batch_size 5 --preprocess micro --how_many_training_steps=100,100
   • Arden: use --verbosity debug for more debug output
   • Arden: use tensorboard --logdir /tmp/retrain_logs/train
   • If we want to quantize, use --quantize True flag.
   • With new model architecture, Phil: python ./tensorflow/tensorflow/tensorflow/examples/speech_commands/train.py --model_architecture "tiny_embedding_conv" --data_dir /Users/philipmateopfeffer/Desktop/stanford/Y5Q1/cs329e/edna_mode/dataset --optimizer adam --background_frequency 0 --embedding_size 20 --batch_size 10 --num_samples 3 --how_many_training_steps 100,100 --loss triplet --use_cpu

Inference

1. (Optional) Run inference on a new example:
   • Call python demo/inference.py --input_path=PATH --embedding_size=N

Conversion

1. Freeze model: please save into demo/frozen_models/ directory!

   • Check that demo/CONSTANTS.py is correct
   • Call: python demo/freeze_model.py --save_path=PATH --embedding_size=N
   • e.g. python demo/freeze_model.py --save_path=/Users/philipmateopfeffer/Desktop/stanford/Y5Q1/cs329e/edna_mode/demo/frozen_models/tiny_embedding_conv.ckpt-200 --embedding_size=20

2. Convert to tflite

   • Call: python demo/convert_to_tflite.py --data_dir=PATH --saved_model_dir=PATH --embedding_size=N
   • e.g. python demo/convert_to_tflite.py --data_dir=/Users/philipmateopfeffer/Desktop/stanford/Y5Q1/cs329e/edna_mode/dataset --saved_model_dir=/Users/philipmateopfeffer/Desktop/stanford/Y5Q1/cs329e/edna_mode/demo/frozen_models/tiny_embedding_conv.ckpt-200 --embedding_size=20

3. Calculate new mean embeddings:

   • Modify CONSTANTS.MODEL_CHECKPOINT_PATH if model retrained, then
   • Call with and without --run_quantized flag.
   • Call: python demo/create_mean_embeddings.py --embedding_size=N

4. Compare float and quantized model accuracies:

   • Call: python demo/test_tflite.py --data_dir=PATH --saved_model_dir=PATH --embedding_size=N
   • Use --run_quantized to run the quantized model, or omit it to run the float model.
   • e.g. python demo/test_tflite.py --data_dir=/Users/philipmateopfeffer/Desktop/stanford/Y5Q1/cs329e/edna_mode/dataset --saved_model_dir=/Users/philipmateopfeffer/Desktop/stanford/Y5Q1/cs329e/edna_mode/demo/frozen_models/tiny_embedding_conv.ckpt-200 --embedding_size=20

5. Convert to tfmicro:

   • Run the following in terminal (if you haven't before): apt-get update && apt-get -qq install xxd
   • Call: python demo/convert_to_tfmicro.py --saved_model_dir=PATH
   • e.g. python demo/convert_to_tfmicro.py --saved_model_dir=/Users/philipmateopfeffer/Desktop/stanford/Y5Q1/cs329e/edna_mode/demo/frozen_models/tiny_embedding_conv.ckpt-200