[Tensorflow] ValueError: Cannot feed value of shape (1,) for Tensor 'Inputs/InputFrame:0', which has shape '(1, 257, None, 1)'

[peiwenhuang27]

Environment: Google Colab, Tensorflow 2.2.0, Intel Tensorflow 2.2.0

I used a custom dataset to store the data, and used common.dataloader to load the dataset.
quantizer.calib_dataloader = common.DataLoader(dataset, batch_size=1)
Later, I input a pre-trained pb model for quantization,
and the model has an input node of shape (1, 257, None, 1).

Upon running the quantization, the following error occured:

Traceback (most recent call last):
File "main.py", line 134, in
evaluate_opt_graph.run()
File "main.py", line 116, in run
q_model = quantizer()
File "/usr/local/lib/python3.7/dist-packages/lpot/experimental/quantization.py", line 170, in call
self.strategy.traverse()
File "/usr/local/lib/python3.7/dist-packages/lpot/strategy/strategy.py", line 287, in traverse
self.baseline = self._evaluate(self.model)
File "/usr/local/lib/python3.7/dist-packages/lpot/strategy/strategy.py", line 425, in _evaluate
val = self.objective.evaluate(eval_func, model)
File "/usr/local/lib/python3.7/dist-packages/lpot/objective.py", line 213, in evaluate
acc = eval_func(model)
File "/usr/local/lib/python3.7/dist-packages/lpot/utils/create_obj_from_config.py", line 127, in eval_func
tensorboard, fp32_baseline)
File "/usr/local/lib/python3.7/dist-packages/lpot/adaptor/tensorflow.py", line 211, in evaluate
predictions = model.sess.run(output_tensor, feed_dict)
File "/usr/local/lib/python3.7/dist-packages/tensorflow/python/client/session.py", line 958, in run
run_metadata_ptr)
File "/usr/local/lib/python3.7/dist-packages/tensorflow/python/client/session.py", line 1157, in _run
(np_val.shape, subfeed_t.name, str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (1,) for Tensor 'Inputs/InputFrame:0', which has shape '(1, 257, None, 1)'

I'm not sure why this error occurred, though I suspect it's due to the None dimension.
Is there any workaround to it without having to set the None dimension to be fixed? Thanks.

[ftian1]

May we know the domain the model works on? is NLP? It's rarely seen the 3rd dim is None on CV domain and need user to specify.

I think when you construct the calibration dataloader, you need specify collate_fn. It's used to concat each batch into a big mini-batch along 3rd dim. if you are not familiar with what collate_fn means, pls refer to https://pytorch.org/docs/stable/data.html

quantizer.calib_dataloader = common.DataLoader(dataset, batch_size=1, collate_fn=xxx) # implement a collate_fn as the 3rd input of common.DataLoader().

[peiwenhuang27]

I am working with audio processing, and the 3rd dim represents the length of time. Because of varying time series length, I want to have this dimension be flexible.

Thank you, I did not know collate_fn can be used for this circumstance, I will look into it.