KNIME DL Python on Mac Silicon

[mck_sbs]

Hi,
creating a DL Environment with KNIME on Mac Silicon is not possible. Now I wanted to solve DL Problems with DL Python Network Creator Node in KNIME instead of using Keras nodes. But the execution gives me the error:

from pandas.util.testing import network

ModuleNotFoundError: No module named ‘pandas.util.testing’

My Code inside DL Python Network Creator:

import tensorflow as tf

inputs = tf.keras.Input(shape=(4,))
x = tf.keras.layers.Dense(10)(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)

output_network = tf.keras.Model(inputs, outputs)

[steffen_KNIME]

Hi @mck_sbs,

it looks like some of the modules (which? please send the knime log - maybe that gives a hint) is using pandas.

In the changelog of pandas version 2.0.0, we see:

Removed pandas.util.testing in favor of pandas.testing (GH30745)

So I guess the module which uses pandas somewhere uses pandas.util.testing. Maybe you can update your module(s) hoping that a more current version uses pandas accordingly. But easier could be in you case do downgrade pandas in your Python environment to a version <2.0.0.

Does that help?

Best regards
Steffen

[mck_sbs]

Hi and thank you.

now I got this error:

ERROR DL Python Network Creator 4:1        Execute failed: Tensors are unhashable (this tensor: KerasTensor(type_spec=TensorSpec(shape=(None, 4), dtype=tf.float32, name='input_1'), name='input_1', description="created by layer 'input_1'")). Instead, use tensor.ref() as the key.
Traceback (most recent call last):
  File "<string>", line 5, in <module>
  File "/Applications/KNIME 5.1.2.app/Contents/Eclipse/plugins/org.knime.dl.python_5.1.0.v202303061416/py/DLPythonNetworkSpecToDataFrameConverter.py", line 92, in get_layer_data_specs_as_data_frames
    extractor = DLPythonNetworkSpecToDataFrameConverter(network.spec)
  File "/Applications/KNIME 5.1.2.app/Contents/Eclipse/plugins/org.knime.dl.keras_5.1.0.v202303061416/py/DLKerasNetwork.py", line 150, in spec
    self._spec = self._extract_model_spec()
  File "/Applications/KNIME 5.1.2.app/Contents/Eclipse/plugins/org.knime.dl.keras_5.1.0.v202303061416/py/DLKerasTensorFlowNetwork.py", line 87, in _extract_model_spec
    return DLKerasTensorFlowNetworkSpecExtractor(self._model).extract_spec()
  File "/Applications/KNIME 5.1.2.app/Contents/Eclipse/plugins/org.knime.dl.keras_5.1.0.v202303061416/py/DLKerasNetworkSpecExtractor.py", line 95, in extract_spec
    self._extract_node_input_tensor_specs(layer, node_idx)
  File "/Applications/KNIME 5.1.2.app/Contents/Eclipse/plugins/org.knime.dl.keras_5.1.0.v202303061416/py/DLKerasNetworkSpecExtractor.py", line 149, in _extract_node_input_tensor_specs
    if input_tensor in self._model_inputs and input_tensor not in self._visited_inputs:
  File "/Users/metinkaratas/anaconda3/envs/py38/lib/python3.8/site-packages/keras/engine/keras_tensor.py", line 268, in __hash__
    raise TypeError(
TypeError: Tensors are unhashable (this tensor: KerasTensor(type_spec=TensorSpec(shape=(None, 4), dtype=tf.float32, name='input_1'), name='input_1', description="created by layer 'input_1'")). Instead, use tensor.ref() as the key.

[mck_sbs]

It is a workflow with one node

[steffen_KNIME]

Maybe these hints in the web provide help, can you try them out?
Several suggestions:

TypeError: Tensors are unhashable. (KerasTensor(type_spec=TensorSpec(shape=(None, 224, 224, 3), dtype=tf.float32, name='input_2'), name='input_2', description="created by layer 'input_2'"))Instead, use tensor.ref() as the key.

Suggests downgrading tf:

Tensorflow Tutorial Error. "Tensor is unhashable. Instead, use tensor.ref() as the key"

Best regards
Steffen

PS: what did you change exactly to eliminate the first error?

[mlauber71]

@mck_sbs from my perspective there currently is no solution to run a tensorflow or Keras task with knime nodes.

Deep Learning Python Environment Mac OS (M2)

@trj I have not yet found a suitable way to get KNIME DL integration and Python Deep Learning packages (Keras, Tensorflow) up and running on an Apple Silicon machine. The current KNIME DL nodes seem to expect older versions of the Python packages which are not there for M1 and M2 systems. You can read about my general efforts here: I also tried to replace several nodes that would want old TF packages with ‘pure’ Python nodes - which kind of is not what a low-code would look like, but did not…

[mck_sbs]

Oh, thank you. I will try with Python nodes.