LSTM: ONNX weights seems not to match KERAS weights?

[ed-mr]

Firstly, I want to thank all the ones doing the good work on the project. Anyway, I am a newbee, I might not understand the ONNX LSTM concept rightly.

Solved:

That was the issue: I did not read the documents: LSTM Operator
There it is clearly spelled out, the order is Riofc and not Rifco, as with Keras.

I assumed, that the weights of the keras LSTM building block and the weights of the ONNX are equal, but that seem to be wrong. What do I wrong, or what is the wrong assumption?

Please have a look at the following code:

import os
import tensorflow as tf
import tf2onnx
import numpy as np
import numpy.testing as npt
import array

##################### Keras Model #####################

model = tf.keras.Sequential()
model.add(tf.keras.layers.LSTM(20, input_shape=(1,10), batch_size=1))
model.compile(loss='mean_absolute_error', optimizer='adam',
                metrics=['mean_squared_error'])
nweights = []
np.random.seed(0)
for W in model.get_weights():
    nweights.append((np.random.randint(20, size=W.shape)-10)/10)
model.set_weights(nweights)

input_signature = [tf.TensorSpec(model.input_shape, tf.float32, name='x')]
onnx_model, _ = tf2onnx.convert.from_keras(model, input_signature, opset=16)

weights_keras = model.layers[0].get_weights()
initializers = onnx_model.graph.initializer
weights_onnx = {}
for val in initializers:
    if val.data_type == 1:
        print(val.name, val.dims)
        weights_onnx[val.name] = np.array(array.array('f', val.raw_data)).reshape(val.dims)

layer_0_reccurent_keras = weights_keras[1]
layer_0_recurrent_onnx = weights_onnx["R0__28"][0,:,:].T
layer_0_reccurent_keras.shape == layer_0_recurrent_onnx.shape

try:
    npt.assert_array_equal(layer_0_reccurent_keras, layer_0_recurrent_onnx)
except AssertionError as ae:
    print("AssertionError:\n", ae)

The output shows the results:

const_fold_opt__46 [1, 1, 20]
W0__27 [1, 80, 10]
R0__28 [1, 80, 20]
B0__29 [1, 160]
AssertionError:

Arrays are not equal

Mismatched elements: 1142 / 1600 (71.4%)
Max absolute difference: 1.9
Max relative difference: 11.
 x: array([[-0.5,  0.5, -0.4, ..., -0.7,  0.4, -0.8],
       [-0.5, -0.3,  0.5, ...,  0.6, -0.4, -0.7],
       [ 0.6, -1. , -0.9, ..., -0.3,  0.6, -0.8],...
 y: array([[-0.5,  0.5, -0.4, ..., -0.5, -0.6, -0.6],
       [-0.5, -0.3,  0.5, ...,  0.2, -0.4, -0.2],
       [ 0.6, -1. , -0.9, ..., -0.4,  0.1, -0.8],...

Thanks for reading the stuff.