[bugfix][relay] Fix alpha attribute with None in ELU

[jikechao]

This patch fixes a bug in ELU. when the alpha=None, It will lead to a wrong compilation and give a confusing inference result.
For example, the following script will produce inconsistent inference results between TVM and Keras. However, we hope tvm can reject it immediately in the model loading stage. Because the model is invalid (see Line 210-212 in keras source code file).

This PR is a similar fix to the pr-14707. pr-14707 fixed a similar bug in LeakyReLU.

Reproducible script

import tvm
import tvm.relay as relay
import numpy as np
from tensorflow import keras
from tensorflow.keras import layers, models


input_shape = (1, 2, 3, 4)
input_data = np.random.random(input_shape)
x = layers.Input(shape=input_shape[1:], dtype='float32')

layer = keras.layers.ELU(alpha=None)
layer.set_weights(layer.get_weights())

y = layer(x)
model = models.Model(x, y)

res_keras = model.predict(input_data)

shape_dict = {'input_1': input_shape}
mod, params = relay.frontend.from_keras(model, shape_dict)

with tvm.transform.PassContext(opt_level=3):
    model = relay.build_module.create_executor("graph", mod, tvm.cpu(0), 'llvm', params).evaluate()


test_x_tvm = input_data
res_tvm = model(tvm.nd.array(test_x_tvm.astype('float32'))).numpy()

np.testing.assert_allclose(res_keras, res_tvm, atol=1e-3, rtol=1e-3)

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @shingjan _{See #10317 for details}

_{Generated by tvm-bot}

[jikechao]

Hi @echuraev @AndrewZhaoLuo, could you help me review it? Thank you!

[echuraev]

LGTM. Thank you for your PR.

[jikechao]

@tvm-bot rerun

[jikechao]

@tvm-bot rerun