UFF model fails when there is more than 1 convolution2d layer

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

hi,

this is my keras model:

inputs = Input(shape=(1,256,1024))
    x = Conv2D(1, (15,15), padding='same', strides=(1,1), name = 'conv1', data_format="channels_first")(inputs)
    x = Activation('relu', name='actv1')(x)
    x = MaxPooling2D((2, 2), strides=(2, 2), name = 'maxpool1')(x)
    x = Conv2D(32, (3,3), padding='same', strides=(1,1), name = 'conv2', data_format="channels_first")(x)
    x = Activation('relu', name='actv2')(x)
    x = MaxPooling2D((2, 2), strides=(2, 2), name = 'maxpool2')(x)
    x = Conv2D(64, (3,3), padding='same', strides=(1,1), name = 'conv3', data_format="channels_first")(x)
    x = MaxPooling2D((2, 2), strides=(2, 2), name = 'maxpool3')(x)
    x = Conv2D(128, (3,3), padding='same', strides=(1,1), name = 'conv4', data_format="channels_first")(x)
    x = MaxPooling2D((2, 2), strides=(2, 2), name = 'maxpool4')(x)
    x = Flatten()(x)
    x = Dense(1000, activation='relu', name = 'dense1')(x)
    x = Dense(500, activation='relu',name = 'dense2')(x)
    x = Dense(2, activation='softmax',name = 'dense3')(x)

My generated UFF model gave a result that was very off from the correct tensorflow results. However, when i removed the blocks of convolution layers:

inputs = Input(shape=(1,256,1024))
    x = Conv2D(1, (15,15), padding='same', strides=(1,1), name = 'conv1', data_format="channels_first")(inputs)
    x = Activation('relu', name='actv1')(x)
    x = MaxPooling2D((2, 2), strides=(2, 2), name = 'maxpool1')(x)
    x = Flatten()(x)
    x = Dense(1000, activation='relu', name = 'dense1')(x)
    x = Dense(500, activation='relu',name = 'dense2')(x)
    x = Dense(2, activation='softmax',name = 'dense3')(x)

The generated UFF model was able to produce consistent results with the tensorflow model.

Is there a bug in TensorRT that is causing some form of wrong optimisation for the convolution block layers? Could someone advise? thanks so much.

2

Hi,
I’ve fixed the problem by changing the keras flatten layer to tf.reshape. However, I’m not sure why this solved the problem, can someone advise as to what is happening? Thanks!

3

Hi,

I have a similar probelm as jay. See [url]https://devtalk.nvidia.com/default/topic/1038304/tensorrt/tensorrt-gives-different-output-to-original-tensorflow-model/?offset=3#5276125[/url]

I can just say it isn’t connected to the number of convolution layers, it depends on the filter parameters in the convolution layer. With “filters”=1 (as in your first layer) it works fine. But when filters > 1 it gives a different output.

Can you please provide the code to replace the flatten layer thanks!

Best Johannes