Can tf.split layer be converted to tensorRt ?

Hello. When I was trying to convert my TF model to uff format with uff.from_tensorflow_frozen_model method., I got an error : KeyError: u’convolutional_alexnet/conv5/split:1’ (tensorRT 3.0 + tensorflow 1.4.0) , I wanna kown that Can I convert tf.split layer into tensorRt ?

Error Info :

Using output node convolutional_alexnet/conv5/concat
Converting to UFF graph
Traceback (most recent call last):
File “ckpt2pb.py”, line 51, in
uff_model = uff.from_tensorflow(frozen_graph,output)
File “/home/xucheng/local/anaconda2/lib/python2.7/site-packages/uff/converters/tensorflow/conversion_helpers.py”, line 75, in from_tensorflow
name=“main”)
File “/home/xucheng/local/anaconda2/lib/python2.7/site-packages/uff/converters/tensorflow/converter.py”, line 64, in convert_tf2uff_graph
uff_graph, input_replacements)
File “/home/xucheng/local/anaconda2/lib/python2.7/site-packages/uff/converters/tensorflow/converter.py”, line 51, in convert_tf2uff_node
op, name, tf_node, inputs, uff_graph, tf_nodes=tf_nodes)
File “/home/xucheng/local/anaconda2/lib/python2.7/site-packages/uff/converters/tensorflow/converter.py”, line 32, in convert_layer
return cls.registry_[op](name, tf_node, inputs, uff_graph, **kwargs)
File “/home/xucheng/local/anaconda2/lib/python2.7/site-packages/uff/converters/tensorflow/converter_functions.py”, line 244, in convert_conv2d
return _conv2d_helper(name, tf_node, inputs, uff_graph, func=“conv2d”, **kwargs)
File “/home/xucheng/local/anaconda2/lib/python2.7/site-packages/uff/converters/tensorflow/converter_functions.py”, line 260, in _conv2d_helper
tf_node, inputs, kwargs[“tf_nodes”])
File “/home/xucheng/local/anaconda2/lib/python2.7/site-packages/uff/converters/tensorflow/converter.py”, line 114, in apply_fused_padding
tf_lhs_node = tf_nodes[inputs[0]]
KeyError: u’convolutional_alexnet/conv5/split:1’

And here is my tensorflow code :

with tf.variable_scope(scope, ‘convolutional_alexnet’, [inputs], reuse=reuse) as sc:
end_points_collection = sc.name + ‘_end_points’
with slim.arg_scope([slim.conv2d, slim.max_pool2d],
outputs_collections=end_points_collection):
net = inputs
net = slim.conv2d(net, 96, [11, 11], 2, scope=‘conv1’)
net = slim.max_pool2d(net, [3, 3], 2, scope=‘pool1’)
with tf.variable_scope(‘conv2’):
b1, b2 = tf.split(net, 2, 3)
b1 = slim.conv2d(b1, 128, [5, 5], scope=‘b1’)

The original implementation has bias terms for all convolution, but

it actually isn’t necessary if the convolution layer is followed by a batch

normalization layer since batch norm will subtract the mean.

b2 = slim.conv2d(b2, 128, [5, 5], scope=‘b2’)
net = tf.concat([b1, b2], 3)
net = slim.max_pool2d(net, [3, 3], 2, scope=‘pool2’)
net = slim.conv2d(net, 384, [3, 3], 1, scope=‘conv3’)
with tf.variable_scope(‘conv4’):
b1, b2 = tf.split(net, 2, 3)
b1 = slim.conv2d(b1, 192, [3, 3], 1, scope=‘b1’)
b2 = slim.conv2d(b2, 192, [3, 3], 1, scope=‘b2’)
net = tf.concat([b1, b2], 3)

Conv 5 with only convolution, has bias

with tf.variable_scope(‘conv5’):
with slim.arg_scope([slim.conv2d],
activation_fn=None, normalizer_fn=None):
b1, b2 = tf.split(net, 2, 3 , name = ‘split’)
b1 = slim.conv2d(b1, 128, [3, 3], 1, scope=‘b1’)
b2 = slim.conv2d(b2, 128, [3, 3], 1, scope=‘b2’)
net = tf.concat([b1, b2], 3)

Convert end_points_collection into a dictionary of end_points.

end_points = slim.utils.convert_collection_to_dict(end_points_collection)
return net, end_points

Can you help me fix this problem ? Thanks

Hi,

Split layer is not supported by TensorRT and UFF parser.
You can find the detail supported layer here:
[url]Developer Guide :: NVIDIA Deep Learning TensorRT Documentation

For a non-supported layer, you can implement it with plugin API from TensorRT4.0.

Thanks.