Padding error when creating TensorRT Plan

Robotics & Edge Computing Jetson & Embedded Systems Jetson TX2
1

I get the following error when I run my trained networks’ UFF file to make a plan file:

Step 3/5: Parsing UFF Model…
UFFParser: parsing input/IteratorGetNext
UFFParser: parsing reshape/Reshape/shape
UFFParser: parsing reshape/Reshape
UFFParser: parsing reshape/transpose
UFFParser: parsing layer3/kernel/Variable
UFFParser: parsing layer3/Conv2D
UFFParser: parsing layer3/bias/Variable
UFFParser: parsing layer3/BiasAdd
UFFParser: parsing layer3/Relu
UFFParser: parsing layer3/MaxPool
UFFParser: parsing layer3/Reshape/shape
UFFParser: parsing layer3/Reshape

make_trt_plan: Network.h:103: virtual nvinfer1::DimsHW nvinfer1::NetworkDefaultConvolutionFormula::compute(nvinfer1::DimsHW, nvinfer1::DimsHW, nvinfer1::DimsHW, nvinfer1::DimsHW, nvinfer1::DimsHW, const char*): Assertion `(input.h() + padding.h() * 2) >= dkh && “Image height with padding must always be at least the height of the dilated filter.”’ failed.

Here is the layer in question:

with tf.name_scope('layer3'):
    w = tf.Variable(tf.truncated_normal([l1, l2, l3, l4], stddev=0.1, name='weights'))
    b = tf.Variable(tf.constant(0.0, shape=[l4], name='biases'))
    conv1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x_reshape, w, strides=[1, 1, 1, 1], padding='VALID'), b))
    pool1 = tf.nn.max_pool(conv1, ksize=[1, l5, l6, 1], strides=[1, l5, l6, 1], padding='VALID')
    pool1_flat =tf.reshape(pool1, [-1, np.prod(pool1.shape.as_list()[-3:])])

I tested it on my x86 machine and the plan file is able to be created and works like a charm. When I create it on the Jetson, I get that error.

Any help is appreciated!

2

Hi,

PLAN has some dependencies on GPU architecture.

You need to make the PLAN on TX2 directly.
PLAN created on other GPU architecture will crash due to different GPU layout.

Thanks.

3

Hi @AastaLL,

Sorry if I was unclear.

I get the above error when I create the PLAN on the TX2 directly.

I do not get the error when I create the PLAN on my x86.

4

Hi,

This issue looks related to the tensor dimension.

Could you share the value of these parameters: l1, l2, l3, l4, l5, l6
And the dimension of x_reshape with us?

Thanks.