Updated Automatic Speech Recognition using CTC example for Keras v3 #1768
Conversation
examples/audio/ctc_asr.py
Outdated
| @@ -244,16 +249,74 @@ def encode_single_sample(wav_file, label): | |||
| """ | |||
|
|
|||
|
|
|||
| # Reference: https://github.com/keras-team/keras/blob/ec67b760ba25e1ccc392d288f7d8c6e9e153eea2/keras/legacy/backend.py#L674-L711 | |||
| def ctc_label_dense_to_sparse(labels, label_lengths): | |||
There was a problem hiding this comment.
Rather than rewriting this code, you can just use the built-in Keras 3 loss function keras.losses.CTC. I expect it will also enable the code example to run with all backends.
There was a problem hiding this comment.
Thanks for the feedback 👍
After removing the legacy code we still have some references to tf in the example and I'm not sure this can be made backend-agnostic.
Please let me know if I should substitute the remaining tf references.
examples/audio/ctc_asr.py
Outdated
| @@ -320,7 +307,7 @@ def build_model(input_dim, output_dim, rnn_layers=5, rnn_units=128): | |||
| # Optimizer | |||
| opt = keras.optimizers.Adam(learning_rate=1e-4) | |||
| # Compile the model and return | |||
| model.compile(optimizer=opt, loss=CTCLoss) | |||
| model.compile(optimizer=opt, loss=keras.losses.ctc) | |||
There was a problem hiding this comment.
Prefer using CTC() (ends up running the same thing but it's more idiomatic)
| input_length = tf.cast(input_length, tf.int32) | ||
|
|
||
| if greedy: | ||
| (decoded, log_prob) = tf.nn.ctc_greedy_decoder( |
There was a problem hiding this comment.
So, we're going to have to use TF for this and ctc_beam_search_decoder I guess, unless we implement them as new backend ops.
There was a problem hiding this comment.
Again, thanks for the feedback 👍
I created an issue to address this.
Please let me know if I should change the description or add/remove details.
Thanks!
Updates the "Automatic Speech Recognition using CTC" example to support Keras v3.