AssertionError: batch_normalization_1/keras_learning_phase:0 is disconnected

[monteksingh]

I trained an Efficientnet B2 model using [https://pypi.org/project/efficientnet/]

I am using tensorflow==1.14.0 numpy==1.17.0 onnx==1.7.0 Keras==2.2.4 efficientnet==0.0.4

During the conversion, I get multiple warnings

WARN: No corresponding ONNX op matches the tf.op node swish_69/swish_f32 of type swish_f32
The generated ONNX model needs run with the custom op supports.

WARNING: Logging before flag parsing goes to stderr.

W0731 16:03:10.044219 139835338614528 parser.py:303] WARN: No corresponding ONNX op matches the tf.op node swish_69/swish_f32 of type swish_f32
The generated ONNX model needs run with the custom op supports.
WARN: No corresponding ONNX op matches the tf.op node drop_connect_16/cond/Merge of type Merge
The generated ONNX model needs run with the custom op supports.

W0731 16:03:10.045672 139835338614528 parser.py:303] WARN: No corresponding ONNX op matches the tf.op node drop_connect_16/cond/Merge of type Merge                 
The generated ONNX model needs run with the custom op supports.
WARN: No corresponding ONNX op matches the tf.op node drop_connect_16/cond/Switch_1 of type Switch
The generated ONNX model needs run with the custom op supports.

W0731 16:03:10.045809 139835338614528 parser.py:303] WARN: No corresponding ONNX op matches the tf.op node drop_connect_16/cond/Switch_1 of type Switch
The generated ONNX model needs run with the custom op supports.
WARN: No corresponding ONNX op matches the tf.op node drop_connect_16/cond/Shape/Switch of type Switch

The generated ONNX model needs run with the custom op supports.
W0731 16:03:10.045921 139835338614528 parser.py:303] WARN: No corresponding ONNX op matches the tf.op node drop_connect_16/cond/Shape/Switch of type Switch
The generated ONNX model needs run with the custom op supports.

And the conversion fails with the following error. AssertionError: batch_normalization_1/keras_learning_phase:0 is disconnected, check the parsing log for more details.

[jiafatom]

The example for efficientNet conversion can be found in our nightly build here. It needs tensorflow >= 2.1.0 and set keras.backend.set_learning_phase(0).

[monteksingh]

I tried to convert the Model using the following dependencies

tensorflow_gpu = 2.3.0 keras=2.4.3 efficientnet= 0.0.4

Code

import keras2onnx
from keras.models import Model, load_model
import efficientnet
model = load_model(model_path, compile=False)
# converting to ONNX
onnx_model = keras2onnx.convert_keras(model, name=model.name)
keras2onnx.save_model(onnx_model, onnx_model_path)

After doing so I get

WARN: No corresponding ONNX op matches the tf.op node model_2/swish_69/IdentityN of type IdentityN The generated ONNX model needs run with the custom op supports. The ONNX operator number change on the optimization: 982 -> 601

Pls help me understand how can I resolve this issue.

[jiafatom]

tensorflow 2.3.0 support is ongoing, please use tf 2.2 at this moment. This IdentityN is recently supported, so please pull keras2onnx from master source code rather than pypi.

[jiafatom]

See my previous answer, you need pull keras2onnx from master source code: pip install -U git+https://github.com/onnx/keras-onnx

[monteksingh]

Thanks for your help.

I followed all the steps and the conversion was successful.

But the issue now is, I wanted to import this model to Snap Lens Studio. But when I import the ONNX model, I get this error 'Transpose' layer type is not supported I know this issue is not related to this repo. But if anyone is facing the same issue, as I am. Pls help....

[ch-hristov]

Should we still use TF 2.2 for this?

Also here: https://github.com/onnx/keras-onnx/blob/master/applications/nightly_build/test_efn.py#L27

is this the standard efficientnet from tf.keras?

my error is the following: AssertionError: functional_3/efficientnetb6/block7c_se_reduce/BiasAdd:0 is disconnected, check the parsing log for more details.

I'm using concat of BERT model & ENet.