Error while testing over pretrained model

[piyush101]

I am using pytorch v1.0, python 3.6 and using the same commit as the release of the librispeech pretrained model.
When I try to run on librispeech pretrained model as-
python3 test.py --model-path librispeech.pth --test-manifest an4_val_manifest.csv

I get this error-
RuntimeError: Error(s) in loading state_dict for DeepSpeech: Missing key(s) in state_dict: "conv.seq_module.0.weight", "conv.seq_module.0.bias", "conv.seq_module.1.weight", "conv.seq_module.1.bias", "conv.seq_module.1.running_mean", "conv.seq_module.1.running_var", "conv.seq_module.3.weight", "conv.seq_module.3.bias", "conv.seq_module.4.weight", "conv.seq_module.4.bias", "conv.seq_module.4.running_mean", "conv.seq_module.4.running_var", "rnns.0.rnn.bias_ih_l0", "rnns.0.rnn.bias_hh_l0", "rnns.0.rnn.bias_ih_l0_reverse", "rnns.0.rnn.bias_hh_l0_reverse", "rnns.1.rnn.bias_ih_l0", "rnns.1.rnn.bias_hh_l0", "rnns.1.rnn.bias_ih_l0_reverse", "rnns.1.rnn.bias_hh_l0_reverse" Unexpected key(s) in state_dict: "conv.0.weight", "conv.0.bias", "conv.1.weight", "conv.1.bias", "conv.1.running_mean", "conv.1.running_var", "conv.3.weight", "conv.3.bias", "conv.4.weight", "conv.4.bias", "conv.4.running_mean", "conv.4.running_var". size mismatch for rnns.0.rnn.weight_ih_l0: copying a param with shape torch.Size([2400, 672]) from checkpoint, the shape in current model is torch.Size([2400, 1312]). size mismatch for rnns.0.rnn.weight_ih_l0_reverse: copying a param with shape torch.Size([2400, 672]) from checkpoint, the shape in current model is torch.Size([2400, 1312]).

Can anyone please help, How to fix this error?

[mertyildiran]

@piyush101 I'm getting the same error. This is the first time I try this library.

[boyuangong]

You should checkout the corresponding commitment hash.

[piyush101]

You should checkout the corresponding commitment hash.

@boyuangong I am using the same commit hash as the release for the libripeech model i.e. e2c2d83

[piyush101]

@SeanNaren @ryanleary How can I overcome this issue?

[lyc1993]

hi, have you fix this issue?

[piyush101]

@lyc1993 No. Do you experiencing the same?

[jybaek]

@SeanNaren Looks like the hash(e2c2d83) in the pretrained model is wrong. People with errors should use at least the following version of the hash to try out the pre-trained model. The parameter is changed in conv_layer.

80a060f#diff-93ff136fae812392eb0f68d1ce89b7feL133

[lucgeo]

Hi,

Using the 80a060f I got two other different errors:

size mismatch for rnns.0.rnn.weight_ih_l0: copying a param of torch.Size([2400, 1312]) from checkpoint, where the shape is torch.Size([2400, 672]) in current model.
size mismatch for rnns.0.rnn.weight_ih_l0_reverse: copying a param of torch.Size([2400, 1312]) from checkpoint, where the shape is torch.Size([2400, 672]) in current model.

Lucian

[Kabur]

@piyush101 @mertyildiran @lucgeo

Have you guys managed to overcome this issue?

@jybaek
Getting the same error on this hash, as well as the master. Same case as for lucgeo it seems like. Any recommendations by any chance?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.