卷积算子实现的bug

[WangFengtu1996]

• keras model

def make_model(input_shape):
    inputs = keras.Input(shape=input_shape)
    x = layers.Conv2D(16, (1, 5), strides=(1, 3), activation='relu')(inputs)
    x = layers.Conv2D(16, (1, 5), strides=(1, 3), activation='relu')(x)
    x = layers.Conv2D(16, (1, 5), strides=(1, 3), activation='relu')(x)
    x = layers.Conv2D(1, (2, 1), strides=(1, 1))(x)
    outputs = x
    return Model(inputs, outputs)

• 与tflite 结果不一致，对齐不了。

[Zepan]

就是不完全一致的，加快运算速度的折中，想要完全一致就会变慢。代码很少，可以进去查看