Where to find 0.6.0a8 model

Deep Speech
#1

Where can I find models for the 0.6.0a8 binaries? (Specifically for the pi 4)

#2

We have not yet released a v0.6 compatible model. So you would have to train and re-export yourself. This patch could help you re-use v0.5.1 checkpoints for that: https://gist.github.com/reuben/b68b9085f7b293580f8431156a33daa9

#3

Do you have any guess on when 0.6 model might be available?

#4

We do not yet, sorry.

#5

@lissyx is there a way to continue training from 0.5.1 checkpoint using the 0.6.0a14 implementation? I am mainly interested in performing/adding the noise augmentation over a previously trained 0.5.1 checkpoint.

#6

I shared that above already

#7

Instead of applying patch in 0.5.1 i made following changes in 0.6a11

@@ -486,9 +487,19 @@ def train():
     }

     # Checkpointing
+    def fixup(name):
+        if name.startswith('cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/'):
+            return name.replace('cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/', 'lstm_fused_cell/')
+        return name
+    mapping = {fixup(v.op.name): v for v in tfv1.global_variables()}
+
+    checkpoint_saver = tfv1.train.Saver(max_to_keep=FLAGS.max_to_keep,var_list = mapping)
+
     checkpoint_saver = tfv1.train.Saver(max_to_keep=FLAGS.max_to_keep)
     checkpoint_path = os.path.join(FLAGS.checkpoint_dir, 'train')

+    best_dev_saver = tfv1.train.Saver(max_to_keep=1,var_list = mapping)

It starts the training using ds051 checkpoint and later saved few checkpoints
so i thought once it has saved , i can remove patch and use new checkpoints to resume
the training

but if i remove the patch it do not load the newely generated checkpoint , and throws below error

Traceback (most recent call last):
  File "DeepSpeech.py", line 949, in <module>
    absl.app.run(main)
  File "/usr/local/lib/python3.6/dist-packages/absl/app.py", line 299, in run
    _run_main(main, args)
  File "/usr/local/lib/python3.6/dist-packages/absl/app.py", line 250, in _run_main
    sys.exit(main(argv))
  File "DeepSpeech.py", line 933, in main
    train()
  File "DeepSpeech.py", line 566, in train
    loaded = try_loading(session, checkpoint_saver, 'checkpoint', 'most recent')
  File "DeepSpeech.py", line 403, in try_loading
    saver.restore(session, checkpoint_path)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/training/saver.py", line 1302, in restore
    err, "a Variable name or other graph key that is missing")
tensorflow.python.framework.errors_impl.NotFoundError: Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Key cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/bias not found in checkpoint
         [[node save/RestoreV2 (defined at DeepSpeech.py:498) ]]

anything i am missing here, ?

#8

do we need this patch

#9

Note that the patch just makes it load the weights, but the resulting model will perform really badly. You should not rely on models created using that flag.