Problem with image_infer

Hello.

I don't know, maybe it is somehow mxnet related or cuda related, but when i'm trying to launch image_infer (little bit different since i've placed antelope face detector) i'm getting following error
```
Could not load library libcudnn_cnn_train.so.8. Error: /home/daddywesker/anaconda3/envs/insightface/bin/../lib/libcudnn_ops_train.so.8: undefined symbol: _Z22cudnnGenericOpTensorNdILi3EE13cudnnStatus_tP12cudnnContext16cudnnGenericOp_t21cudnnNanPropagation_tPKdPKvPK17cudnnTensorStructS8_S8_SB_S8_S8_SB_Pv, version libcudnn_ops_infer.so.8
Please make sure libcudnn_cnn_train.so.8 is in your library path!
```

on the line
`self.model.forward(db, is_train=False)`

I've tried to install cuda 11.0, 11.2, 11.3 via Nvidia's runfile (local) installation.
I've tried to apply cudnn 8.1 from nvidia site using tar installation.
I've tried to install cudnn using anaconda installation `conda install -c conda-forge cudnn`
Moreover, if i'm not installing cudnn and cudatoolkit inside anaconda using conda install it won't find my /usr/local/cuda installation. Probably problem with anaconda environment or so.
I can't try to use cuda 10.2 since i'm using laptop with 3070 and 10.2 doesn't support 3070. I'm actually out of ideas right now. Only if use dlib's landmarks instead, which is kinda slower. Any suggestions?

Start by reproducing image_infer with the reported antelope face detector and inspect the Anaconda environment's CUDA and cuDNN versions and library paths. Focus on the failure at self.model.forward(db, is_train=False) and compare the installed libraries with the versions required by the project. Done means image_infer loads the model and completes the forward pass without the cuDNN undefined-symbol error.

Written by the indexing model from the issue text.