run.py fails with CNN model

[sechkova]

Describe the bug

Running run.py with the same dataset works with LSTM model but crashes when the chosen model is CNN.

Reproduction steps

python run.py test_data_annotated.csv test_data_raw.csv CNN Both

Traceback (most recent call last):
  File "~/ml-conversational-analytic-tool/ml-conversational-analytic-tool/run.py", line 92, in <module>
    run_res_constructive = run(args.annotated_filename, args.dataset_filename, 'Constructive', encodingType,
  File "~/ml-conversational-analytic-tool/ml-conversational-analytic-tool/run.py", line 47, in run
    model.makeModel(obs[0].shape)
  File "~/ml-conversational-analytic-tool/ml-conversational-analytic-tool/baseCNN.py", line 27, in makeModel
    self.model.add(keras.layers.MaxPooling1D(2))
  File "~/ml-conversational-analytic-tool/.env/lib/python3.10/site-packages/tensorflow/python/training/tracking/base.py", line 629, in _method_wrapper
    result = method(self, *args, **kwargs)
  File "~/ml-conversational-analytic-tool/.env/lib/python3.10/site-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "~/ml-conversational-analytic-tool/.env/lib/python3.10/site-packages/keras/backend.py", line 6065, in pool2d
    x = tf.compat.v1.nn.max_pool(
ValueError: Exception encountered when calling layer "max_pooling1d" (type MaxPooling1D).

Negative dimension size caused by subtracting 2 from 1 for '{{node max_pooling1d/MaxPool}} = MaxPool[T=DT_FLOAT, data_format="NHWC", explicit_paddings=[], ksize=[1, 2, 1, 1], padding="VALID", strides=[1, 2, 1, 1]](max_pooling1d/ExpandDims)' with input shapes: [?,1,1,32].

Call arguments received:
  • inputs=tf.Tensor(shape=(None, 1, 32), dtype=float32)

[sechkova]

Update: After looking at the code and the arguments of the run.py command, it is possible to run successfully the above by adding -pad. Using both -roleRelevant -pad however still raises errors caused by some dimensionality mismatch.

It seems like either the data preprocessing steps need to be improved or the documentation has to describe strictly the expected format of the input data since it is not clear.