GitLab

Update to TF 2 for deep_speech

PiperOrigin-RevId: 326665731

* Merged commit includes the following changes:
326369548  by Andre Araujo:

    Fix import issues.

--
326159826  by Andre Araujo:

    Changed the implementation of the cosine weights from Keras layer to tf.Variable to manually control for L2 normalization.

--
326139082  by Andre Araujo:

    Support local feature matching using ratio test.

    To allow for easily choosing which matching type to use, we rename a flag/argument and modify all related files to avoid breakages.

    Also include a small change when computing nearest neighbors for geometric matching, to parallelize computation, which saves a little bit of time during execution (argument "n_jobs=-1").

--
326119848  by Andre Araujo:

    Option to measure DELG latency taking binarization into account.

--
324316608  by Andre Araujo:

    DELG global features training.

--
323693131  by Andre Araujo:

    PY3 conversion for delf public lib.

--
321046157  by Andre Araujo:

    Purely Google refactor

--

PiperOrigin-RevId: 326369548

* Added export of delg_model module.
Co-authored-by: Andre Araujo <andrearaujo@google.com>

PiperOrigin-RevId: 326541072

PiperOrigin-RevId: 326528933

PiperOrigin-RevId: 326483694

* Input/Output tweaks for YAMNet and VGGish.

- Waveform input for YAMNet is now padded so that we get at least
  one patch of log mel spectrogram. The VGGish TF-Hub exporter
  uses YAMNet's feature computation so the VGGish export will
  also pad waveform input similarly.
- Added a 1024-D embedding output to YAMNet so we now produce
  predicted scores, log mel spectrogram features, and embeddings,
  to satisfy a variety of uses: class prediction, acoustic
  feature visualization, semantic feature extraction.
- Simplified usage of YAMNet in inference mode. Instead of trying
  to work around implicit batch size issues in the Model.predict()
  API, we simply __call__() the Model.
- Switched inference.py to TF 2 and Eager execution.
- Updated the visualization notebook: now uses TF2/Eager and
  can be loaded and run in Google Colab.

* Responded to DAn's comments in https://github.com/tensorflow/models/pull/9092

- Merged spectrogram computation and framing into a single function
  that returns both spectrogram and framed features.
- Extended waveform padding to pad up to an integral number of hops
  in addition to the final STFT analysis window.

* Added a TF-Hub-compatible SavedModel exporter for VGGish.

* Responded to DAn's comments, switched linspace to arange everywhere.

PiperOrigin-RevId: 325797820

Fixed a long-standing bug where the released VGGish model used
post-activation embedding output while the released embeddings
were pre-activation. There are still discrepancies due to
other reasons: differences in choice of YouTube transcode,
repeated resamplings with different resamplers, slight differences
in feature computation, etc.

* Made VGGish and YAMNet work in TF2 without disabling TF2 behavior.

Allowed TF2 behavior and allowed passing in a features tensor into the
VGGish model definition. Both of these changes are needed for making
TF-Hub exports of these models. Lifted constraints on TF versions since
tf_slim has been updated to work with TF 2.

* Responded to DAn's comments in https://github.com/tensorflow/models/pull/9077

* Fixed typo in comment.

PiperOrigin-RevId: 325621239

PiperOrigin-RevId: 325536056

PiperOrigin-RevId: 325470519

This is the likely reason for jobs that take forever.

PiperOrigin-RevId: 325275789