investigating convert error by padding

Makefile

WEBDNN = /Users/yuji/OpenSources/webdnn/bin/convert_keras.py
PLUGIN = padding_channel_plugin.py
PLUGIN_OPTION = --plugin padding_channel_plugin.py
OUTDIR = output

$(OUTDIR)/kernels_webassembly.js: model.h5 $(PLUGIN)
	DEBUG=1 CPLUS_INCLUDE_PATH=/usr/local/include/eigen3 python3 $(WEBDNN) $< $(PLUGIN_OPTION) --input_shape '(1,4,1)' --out $(OUTDIR)

model.h5: model.py padding_channel.py
	python3 model.py

model.py

import sys
import tensorflow as tf
from keras import backend as K
from keras.models import Model
from keras.layers import Input, Reshape, Lambda
from padding_channel import PaddingChannel

# インプットを直接繋げば変換に成功する。
def model_can_be_converted1():
    x = Input(shape=(4, 1), name="x")
    y = PaddingChannel(padding=1)(x)
    return Model(inputs=[x], outputs=[y])

# 前段でReshapeすると変換に失敗する。
def model_cannot_be_converted1():
    x = Input(shape=(4, 1), name="x")
    tmp = Reshape((2, 2, 1))(x)
    y = PaddingChannel(padding=1)(tmp)
    return Model(inputs=[x], outputs=[y])

# Reshapeのみで変換できることを確認した。
def model_can_be_converted2():
    x = Input(shape=(4, 1), name="x")
    y = Reshape((2, 2, 1))(x)
    return Model(inputs=[x], outputs=[y])

# 2次元で変換できることを確認した。
def model_can_be_converted2():
    x = Input(shape=(2, 2, 1), name="x")
    y = PaddingChannel(padding=1)(x)
    return Model(inputs=[x], outputs=[y])

if __name__ == '__main__':
    #model = model_can_be_converted2()
    model = model_cannot_be_converted1()
    model.save('model.h5')

padding_channel.py

import tensorflow as tf
from keras.engine import Layer

class PaddingChannel(Layer):

    def __init__(self, padding=0, **kwargs):
        super(PaddingChannel, self).__init__(**kwargs)
        self.padding = padding

    def call(self, x):
        paddings = [[0, 0] for _ in range(len(x.shape))]
        paddings[-1][1] = self.padding
        return tf.pad(x, paddings)

    def compute_output_shape(self, input_shape):
        assert input_shape and len(input_shape) >= 2
        assert input_shape[-1]
        output_shape = list(input_shape)
        output_shape[-1] = input_shape[-1] + self.padding
        return tuple(output_shape)

padding_channel_plugin.py

import sys
sys.path.append('.')

from padding_channel import PaddingChannel