Explicit padding for Transpose Convolution fails about armnn

[TeresaARM]

Hello,
TF accepts both outputs shapes as valid: [1, 66, 226, 16] and [1, 65, 226, 16]

input  = tf.constant(1, shape=[1, 32, 224, 16]) # NHWCin
kernel = tf.constant(2, shape=[3, 3, 16, 16]) # HWCoutCin

x = tf.nn.conv2d_transpose(input,
                           kernel,
                           output_shape=[1, 66, 226, 16], #NHWCout
                           strides=(2, 1),
                           padding="VALID",
                           data_format='NHWC',
                           dilations=None,
                           name=None)

x.shape

input  = tf.constant(1, shape=[1, 32, 224, 16]) # NHWCin
kernel = tf.constant(2, shape=[3, 3, 16, 16]) # HWCoutCin

x = tf.nn.conv2d_transpose(input,
                           kernel,
                           output_shape=[1, 65, 226, 16], #NHWCout
                           strides=(2, 1),
                           padding="VALID",
                           data_format='NHWC',
                           dilations=None,
                           name=None)

x.shape

In ArmNN we calculate the output this way:
Out = ( In - 1 ) * stride + filter - padding.
65 = ( 32 - 1 ) * 2 + 3 - 0

I am trying to see how we can support both output shapes, as TF does.

[TeresaARM]

Hi @riestmo-nxp ,
if you want to attach the model that would help us to debug the issue.

[riestmo-nxp]

Hi @TeresaARM,
sure, that's the model: model.zip

[xiaotongnii]

It seems that inferShape erro happened when Convert this model to armnn ir. And TfliteParser maybe not consider that VALD Padding mode.
This mode datalayout is NHWC or NCHW?
about TransposeConv padding mode :https://www.tensorflow.org/api_docs/python/tf/nn/conv2d_transpose

[riestmo-nxp]

Hi @xiaotongnii,
the data format is NHWC (as usual in TFLite).

[xiaotongnii]

Hi，@riestmo-nxp Convert model.tflite to model.armnn and everythins is OK!
./ArmnnConverter -f tflite-binary -m model.tflite -i serving_default_input:0 -s 1,512,256,32 -o PartitionedCall:0 -p model.armnn
check your armnn version pls！

[xiaotongnii]

[tracyn-arm]

Tflite padding is more complex than what we have in Arm NN. Take a look at:
https://github.com/tensorflow/tensorflow/blob/be8160ab0a925311041f7a53b7c7c148a3e51f35/tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc#L1206
In the VALID case (one my test was hitting the error with your model), the calculations are:

        *padding_top = *padding_bottom = *padding_left = *padding_right = 0;
        *adjustment_height = (output_height - kernel_height) % stride_height;
        *adjustment_width = (output_width - kernel_width) % stride_width;

The CalcPadding() code inserted above reflects the first line. The difference occurs with the adjustment calculations. From the tflite code:

/// @param adjustment_height - additional elements in the bottom of the 2D output data.
/// @param adjustment_width - additional elements to the right of the 2D output data.

In the case that is failing with the provided model, the adjustment_height = 1 # (64 - 3) % 2 , adjustment_width = 0

We use the TransposeConv2d from the Compute Library and I could not find support for the adjustment parameters unfortunately and this limits the type of padding we can have in Arm NN.

It's a complex problem so please let me know if you think I might have missed something. Thanks.