How to use tensorrt6 to parse a 3D convolution layer and 3D Pooling layer in a caffe deploy file?

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

In my caffe deploy file the 3D Pooling is set as following:

   layer {

name: “globalpool”
bottom: “res50”
top: “globalpool”
type: “Pooling”
pooling_param {
pool: AVE
kernel_size: [3, 3, 3]
stride: [1, 1, 1]
}
}
The following error happened:
Error parsing text-format ditcaffe.NetParameter: 2131:18: Expected integer, got: [
CaffeParser: Could not parse deploy file

Could you give me some suggention?

2

It seems the “kernel_size” and “stride” values are incorrect, causing the parse error. Values should be integer instead of array in case of Caffe.

I was able to just find below example on web for 3D caffe implementation which might help:
https://github.com/ltnghia/3DFCN/blob/master/3DFCN/examples/c3d_ucf101_finetuning/train_resnet18_r2.prototxt

3

When I revise the 3dpooling layer in my deploy file as follow:
layer {
name: “globalpool”
bottom: “res50”
top: “globalpool”
type: “Pooling”
pooling_param {
pool: AVE
kernel_size: 3
kernel_depth: 3
stride: 1
temporal_stride: 1
}
}
The following error happened:
Message type “ditcaffe.PoolingParameter” has no field named “kernel_depth”.

Could you supply an example to show how to use the function of 3d convolution and 3d pooling

4

We currently do not support Caffe models for 3D layers.
3D layers are supported in ONNX/uff model or TensorRT python/C++ API.