Minimal input size of Faster RCNN in TensorRT 4

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

Hi,

When I changed a Faster RCNN model with ZF feature extractor (trained with Caffe) in TensorRT 4, I got the following error:

Weights for layer rpn_conv/3x3 doesn't exist
ERROR: CaffeParser: ERROR: Attempting to access NULL weights
Weights for layer rpn_conv/3x3 doesn't exist
ERROR: CaffeParser: ERROR: Attempting to access NULL weights
ERROR: Parameter check failed at: ../builder/Network.cpp::addConvolution::40, condition: kernelWeights.values != NULL
error parsing layer type Convolution index 14
End parsing model...

The input image size of the ZF model I used is 224x 224, after the final convolution, the size of feature map is 15x 15. I’m wondering if 15x 15 is too small for layer rpn_con/3x3, since the default input size to VGG16 feature extractor is 375x 500, whose feature map is 24x 32.

The following link contains the model and .prototxt I used:
https://drive.google.com/drive/folders/1v6oIfUPsRZRcNePh3cqx16WdRSVMetfq?usp=sharing

2

hello,

Per engineering, this looks like a custom version of the fasterRCNN model. The error seems to occur due to missing weights. It could be an issue with the model since it is failing to parse a convolution layer.

3

Dear NVES,

Thank you for the reply, I’ve retrained the ZF mode, and the error was solved.