mcomunita / google_research_fad

repo used as a submodule in ccwavegan

This repo is derived from:

https://github.com/google-research/google-research/tree/master/frechet_audio_distance

and is used as a submodule for:

https://github.com/mcomunita/ccwavegan

---

Instruction on how to compute the Frèchet Audio Distance are available from the source repo here, but to make it work and to reduce the repo to the bare minimum to compute FAD, I did the following:

clone repo

$ git clone https://github.com/google-research/google-research.git
$ cd google-research

remove unnecessarty stuff

• remove .git folder
• remove all folders but frechet_audio_distance

+ rm .travis.yml
+ rm compile_protos.sh

init new repo

$ git init
$ git add .
$ git commit -m “removed all unnecessary files for fad”

prepare venv

$ python3 -m venv .venv_fad
$ source .venv_fad/bin/activate
$ python -m pip install --upgrade pip
$ pip install apache-beam numpy scipy tensorflow

get vgg-ish architecture

$ mkdir tensorflow_models
$ touch tensorflow_models/__init__.py
$ svn export https://github.com/tensorflow/models/trunk/research/audioset tensorflow_models/audioset/
$ rm -r tensorflow_models/audioset/yamnet
$ touch tensorflow_models/audioset/__init__.py
$ touch tensorflow_models/audioset/vggish/__init__.py

get pre-trained vgg-ish

$ mkdir -p data
$ curl -o data/vggish_model.ckpt https://storage.googleapis.com/audioset/vggish_model.ckpt

modify frechet_audio_distance/audioset_model.py:

• lines 25:27

from tensorflow_models.audioset.vggish import mel_features
from tensorflow_models.audioset.vggish import vggish_params
from tensorflow_models.audioset.vggish import vggish_slim

• line 90

for i in range(0, samples - vggish_params.SAMPLE_RATE + 1,

modify tensorflow_models/audioset/vggish/vggish_slim.py:

• line 36

import tensorflow_models.audioset.vggish.vggish_params as params

modify frechet_audio_distance/fad_utils.py:

• line 37:

tf_record = tf.python_io.tf_record_iterator(filename).__next__()

also

pip install tf_slim

Test FAD computation

$ python -m frechet_audio_distance.gen_test_files --test_files "test_audio"

$ ls test_audio/background/*  > test_audio/test_files_background.csv
$ ls test_audio/test1/*  > test_audio/test_files_test1.csv
$ ls test_audio/test2/*  > test_audio/test_files_test2.csv

$ mkdir -p stats
$ python -m frechet_audio_distance.create_embeddings_main --input_files test_audio/test_files_background.csv --stats stats/background_stats
$ python -m frechet_audio_distance.create_embeddings_main --input_files test_audio/test_files_test1.csv --stats stats/test1_stats
$ python -m frechet_audio_distance.create_embeddings_main --input_files test_audio/test_files_test2.csv --stats stats/test2_stats

$ python -m frechet_audio_distance.compute_fad --background_stats stats/background_stats --test_stats stats/test1_stats
$ python -m frechet_audio_distance.compute_fad --background_stats stats/background_stats --test_stats stats/test2_stats

Good Luck!