KeyError: key "conv2d_9_2/kernel:0" not found

[WenJieWan]

I train a simple cnn model by tensorflow(1.15), and use keras-onnx tool to convert model to onnx.
Next, i load onnx and occur KeyError: key "conv2d_9_2/kernel:0" not found
But, i print the msgs when keras-onnx converts model, and the key "conv2d_9_2/kernel:0" has existed.
The message segments are shown as follows:
ir_version: 6
producer_name: "keras2onnx"
producer_version: "1.6.1"
domain: "onnx"
model_version: 0
doc_string: ""
graph {
node {
input: "conv2d_9_input"
output: "adjusted_input3"
name: "Transpose8"
op_type: "Transpose"
attribute {
name: "perm"
ints: 0
ints: 3
ints: 1
ints: 2
type: INTS
}
doc_string: ""
domain: ""
}
node {
input: "adjusted_input3"
input: "conv2d_9_2/kernel:0"
input: "conv2d_9_2/bias:0"
output: "convolution_output3"
name: "conv2d_9"
op_type: "Conv"
attribute {
name: "auto_pad"
s: "NOTSET"
type: STRING
}
..................................................
..................................................
The model network structure:
mpadding = 'same'
model = Sequential()
model.add(Convolution2D(nb_filters, kernel_size, strides=strides, padding=mpadding, activation='relu', input_shape=input_shape))
model.add(Convolution2D(nb_filters, kernel_size, strides=strides, padding=mpadding, activation='relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Convolution2D(nb_filters2, kernel_size, strides=strides, padding=mpadding, activation='relu'))
model.add(Convolution2D(nb_filters2, kernel_size, strides=strides, padding=mpadding, activation='relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Flatten())
model.add(Dense(100, activation='relu'))
model.add(Dropout(0.25))
model.add(Dense(nb_classes, activation='softmax'))
model.compile(optimizer=Adam(lr=0.0001),loss='categorical_crossentropy',metrics=['accuracy'])
model = LSUVinit(model,x_train[:batch_size,:,:,:])
model.fit(x_train, y_train, batch_size=batch_size, epochs=nb_epoch, validation_data=(x_test,y_test), shuffle=True)
model.save('test.h5')

keras2onnx code:
import keras2onnx,onnx
model = load_model('./test.h5')
onnx_model = keras2onnx.convert_keras(model,model.name)
onnx.save_model(onnx_model,"test.onnx")

[RKowalsky]

Is was your Keras model created in TensorFlow 1.x or 2.x? We are having a similar issue at our company with reading ONNX files that were written using the keras2onnx package. From the keras2onnx webpage: "keras2onnx has been tested on Python 3.5, 3.6, and 3.7, with TensorFlow 1.x (CI build). It does not support Python 2.x.". Perhaps this matters?

[andevellicus]

Similar issue. I have TF 2.1, ran keras2onnx on my model, and then tried to load into Julia. Got conv3d_0/kernal:0 not found.