-
Notifications
You must be signed in to change notification settings - Fork 79
Inference is giving ValueError: When input_signature is provided, all inputs to the Python function must be convertible to tensors: #32
Comments
Have you fixed the issue anyway? |
Hi, No didn't fix this issue. Not sure how to fix this |
@tumusudheer What did the word-error-rate look like during your training? Mine does not look very promising: |
Finally, the issue is fixed following the 2nd solution in the issue of tensorflow repo. iterator = iter(train_dataset)
@tf.function(input_signature=[iterator.element_spec])
def train_step(dataset_inputs):
def step_fn(inputs):
# ...
for batch, inputs in enumerate(train_dataset):
loss, metrics_results = train_step(next(iterator)) |
Hi, but it didn't work the bug is here. Traceback (most recent call last): During handling of the above exception, another exception occurred: Traceback (most recent call last): |
Hi,
I started training the model using the entire Common Voice dataset given in the github page.
I'm using tensorflow 2.2.0 with python 3.6. The training command used
python run_rnnt.py --mode train --data_dir data_trail/preprocessed --batch_size 8 --eval_size 100
using 1080Ti single GPU. I got OOM error after about 18k steps (still in Epoch 0) and my loss was about 116.7. The Accuracy graph in tensorboard is showing about 0.42.
Since a checkpoint is getting saved for every 1000 steps, I tried to run evaluation:
python transcribe_file.py --checkpoint model/checkpoint_15000_109.9516.hdf5 --i data_trail/clips/common_voice_en_19945797.wav
But I'm getting the following error:
Is this because here (hparams is not a tensor but a json) ?
The text was updated successfully, but these errors were encountered: