kernel weights has count 5514240 but 96499200 was expected

Provide details on the platforms you are using:
Linux distro and version:ubuntu 16.04
GPU type :gtx 1080
nvidia driver version :390.25
CUDA version :9.0
CUDNN version :7.2.1
Python version [if using python] :3.5
Tensorflow version :1.11
TensorRT version :
If Jetson, OS, hw versions

Describe the problem

here is my code: = (1,4,70,192)
cnn_out =
cnn_output_shape = tf.shape(cnn_out)
batch_size = cnn_output_shape[0]
cnn_output_h = cnn_output_shape[1]
cnn_output_w = cnn_output_shape[2]
cnn_output_channel = cnn_output_shape[3]

cnn_out_transposed = tf.transpose(cnn_out, [0, 2, 1, 3], name='f_t')
#cnn_out_transposed = (1,70,4,192)
cnn_out_reshaped = tf.reshape(cnn_out_transposed,
                                      [batch_size, cnn_output_w, 1, cnn_output_h * cnn_output_channel],

cnn_shape = cnn_out.get_shape().as_list()
cnn_out_reshaped.set_shape([cnn_shape[0], cnn_shape[2], 1, cnn_shape[1] * cnn_shape[3]])

# cnn_out_reshaped = (1,70,1,768)
# weights = (1,1,768,7180)=5514240

logits = slim.conv2d(cnn_out_reshaped, 7180, [1, 1], activation_fn=None)
# logits=(1,70,1,7180)
logits = tf.squeeze(logits, [2])
# logits=(1,70,7180)
probs = tf.nn.softmax(logits, name='probs')

I use convert-to-uff to convert .pb to .uff successfuly,but when I execute this code

model_file = '../output/crnn.pb'
uff_model = uff.from_tensorflow_frozen_model(model_file, ["probs"], list_nodes=False, quiet=False,

It is throw errors below:

[TensorRT] ERROR(code line 21): Conv/Conv2D: kernel weights has count 5514240 but 96499200 was expected
[TensorRT] ERROR: UFFParser: Parser error: Conv/BiasAdd: The input to the Scale Layer is required to have a minimum of 3 dimensions.
[TensorRT] ERROR: Failed to parse UFF model stream

the 96499200 = 701927180,I think TensorRT reshape op could compute wrong result!


to help us debug, can you share the .pb and uff file? (use dropbox or google drive if size is an issue).

Thank you for response me!
I have uploaded two file here


to help us debug, can you share the complete code? Asking because the pb does not seem to match the code provided above in #1.


The ONNX model now works with latest TRT. Engineering suggests this as an alternative to UFF.