dataset patch error

[superhg2012]

Hi, an error thrown when run train, could you help to see this problem?

Traceback (most recent call last):
File "main.py", line 414, in
tf.app.run(main=train_or_predict)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 48, in run
_sys.exit(main(_sys.argv[:1] + flags_passthrough))
File "main.py", line 405, in train_or_predict
hparams=params,
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/learn_runner.py", line 218, in run
return _execute_schedule(experiment, schedule)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/learn_runner.py", line 46, in _execute_schedule
return task()
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/experiment.py", line 625, in train_and_evaluate
self.train(delay_secs=0)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/experiment.py", line 367, in train
hooks=self._train_monitors + extra_hooks)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/experiment.py", line 807, in _call_train
hooks=hooks)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/estimator/estimator.py", line 302, in train
loss = self._train_model(input_fn, hooks, saving_listeners)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/estimator/estimator.py", line 708, in _train_model
input_fn, model_fn_lib.ModeKeys.TRAIN)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/estimator/estimator.py", line 577, in _get_features_and_labels_from_input_fn
result = self._call_input_fn(input_fn, mode)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/estimator/estimator.py", line 663, in _call_input_fn
return input_fn(**kwargs)
File "main.py", line 218, in input_fn
padded_shapes=([None, None, None],[])
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/data/ops/dataset_ops.py", line 695, in padded_batch
return PaddedBatchDataset(self, batch_size, padded_shapes, padding_values)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/data/ops/dataset_ops.py", line 1292, in init
input_dataset.output_shapes, _partial_shape_to_tensor, padded_shapes)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/data/util/nest.py", line 512, in map_structure_up_to
assert_shallow_structure(shallow_tree, input_tree)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/data/util/nest.py", line 372, in assert_shallow_structure
check_types=check_types)
File "/data0/anaconda3/lib/python3.6/site-packages/tensorflow/python/data/util/nest.py", line 356, in assert_shallow_structure
"Input has type: %s." % type(input_tree))
TypeError: If shallow structure is a sequence, input must also be a sequence. Input has type: <class 'list'>.

[superhg2012]

@naktinis can you help to see this issue? thanks in advance!

[naktinis]

@superhg2012 the respository is not actively maintained. If you paste here the exact commands you executed when you got this error, I may try to replicate it when I have time.

[superhg2012]

@naktinis
I use the voxforge corpus only. and I use the code from the main branch.
I run this train command:

python main.py --model-dir /data0/cnn_models --model cnn --image-dir /data0/vox_pngs/ --train-set /data0/vox_pngs/train-set.csv --eval-set /data0/vox_pngs/eval-set.csv

the problem seems caused by this tuple ( {'sgam' : features}, labels): there are two code snaps in the project.

1 dataset.py

@staticmethod
  def instance_as_tensor(image_name, label=None):
      """Convert a single training/prediction instance into a tensor."""
      print(image_name)
      img_file = tf.read_file(image_name)
      image = tf.image.decode_png(img_file, channels=0)
      image = tf.cast(image, tf.float32)
      image_data = tf.transpose(image[:128, :] / 256.)

      if label is not None:
          label = tf.cast(label, tf.int32)

      return {'sgram': image_data}, label

2 main.py

    def input_fn():
        dataset = tf.data.Dataset.from_tensor_slices(usable_data)
        dataset = dataset.shuffle(params.shuffle_buffer_size)
        dataset = dataset.map(data.instance_as_tensor)
        dataset = dataset.repeat(epochs)
        dataset = dataset.padded_batch(params.batch_size, padded_shapes=([None, None, None], []), padding_values=0.0)

        iterator = dataset.make_initializable_iterator()
        init_hook.iterator_init = iterator.initializer
        next_example, next_label = iterator.get_next()
        return {'sgram': next_example}, next_label

when I modify {'sgram': next_example}, next_label to next_example, next_label and

{'sgram': image_data}, label to image_data, label

the train process can run without problem.

I do not know why.

Another question:

the spectrogram_width is set to 858, pngs are generated from audios which is ? X 128 in size. ? is different for audios with different duration. So some pngs is less than 858 in width, so when train cnn model, these pngs will fail since cnn require fixed filter size. I see that the tensorflow dataset Api is used to padding batch to same size, Is this is used to process the pngs whose spectrogram_width is less than 858? if not, how can I deal with those small pngs?
dataset = dataset.padded_batch(params.batch_size, padded_shapes=([None, None, None], []))

[superhg2012]

@naktinis

When run with predict, another exception throws:

tensorflow.python.framework.errors_impl.InvalidArgumentError: Index out of range using input dim 3; input has only 3 dims for 'strided_slice_2' (op: 'StridedSlice') with input shapes: [?,?,?], [4], [4], [4] and with computed input tensors: input[3] = <1 1 1 1>.