Tacotron-2: CBHG_postnet_highwaynet_1/H/bias not found in checkpoint

[zuzana]

Hi,

I was trying to use your pre-trained model from #30 but I'm getting following error:

Tacotron_model/inference/CBHG_postnet/CBHG_postnet_highwaynet_1/H/bias not found in checkpoint

I'm running following command:

python synthesize.py  --model='Tacotron-2'  --mode=eval 

Could you please let me know if I'm making some obvious mistake anywhere?

Thanks.

[cliuxinxin]

I think the pre-trained model is too old. They dont have CBHG part.

[zuzana]

Thanks!

[mrgloom]

Where is to find newer model?

[anasvaf]

Hello. I have tried to train the model from the current master branch for 10 iterations. However when I am trying to restore the checkpoint as follows:

model_folder = 'logs-Tacotron-2/taco_pretrained'
checkpoint = tf.train.get_checkpoint_state(model_folder)
input_checkpoint = checkpoint.model_checkpoint_path
print(input_checkpoint)
output_graph = 'tacotron2_frozen.pb'

tf.compat.v1.reset_default_graph()

with tf.compat.v1.variable_scope('datafeeder') as scope:
    inputs = tf.placeholder(tf.int32, [1, None], 'inputs')
    input_lengths = tf.placeholder(tf.int32, [1], 'input_lengths')
    split_infos = tf.placeholder(tf.int32, shape=(hparams.tacotron_num_gpus, None), name='split_infos')
with tf.compat.v1.variable_scope('model') as scope:
    model = create_model('Tacotron', hparams)
    model.initialize(inputs, input_lengths, split_infos=split_infos)

graph = tf.compat.v1.get_default_graph()
saver = tf.compat.v1.train.Saver()

with tf.Session(graph=graph, config=tf.ConfigProto(allow_soft_placement=True)) as sess:
    sess.run(tf.global_variables_initializer())
    saver.restore(sess, input_checkpoint)

I get the same error regarding missing CBHG_postnet_highwaynet_1/H/bias from the checkpoint. Does anybody have an idea?