Hello,

I am using a Raspberry Pi with DepthAI to detect passing planks. I have tried to make a YoloV7 model in the past, but realized that it could be too intensive for this simple task and the limited amount of computer resource I have (RPi).

I have tried running the YoloV4-tiny with custom data tutorial. However, the tutorial uses Tensorflow 1.x, which is unusable with Google Colab. Then, I have tried running it with Tensorflow 2.x. Everything ran smooth, until the training, where it gave me a cuDNN error :

cuDNN status Error in: file: ./src/convolutional_kernels.cu function: forward_convolutional_layer_gpu() line: 543
cuDNN Error: CUDNN_STATUS_BAD_PARAM
Darknet error location: ./src/convolutional_kernels.cu, forward_convolutional_layer_gpu(), line #543
cuDNN Error: CUDNN_STATUS_BAD_PARAM: Success

I could make the YoloV7 work, but I feel like it is unnecessary for the work I need the detection to do. But, if it is the only way to make it work due to the new tensorflow requirements, it'll have to do.

Therefore, here are my three questions :

  1. Is it possible to still use the yolov4-tiny tutorial even with the new TensorFlow requirements?
  2. Will there be an updated version, or a yoloV5-6-7-tiny tutorial?
  3. Unrelated, but can I use 300x300 images for training? The tutorial says the resolutions have to be divisible by 32. I have used 320x320 instead to make it work, but I am unsure if it's going to work.

Thanks a lot and have a nice day!

    Hey FrancisGuindon ,

    1/ YoloV4-tiny tutorial requires the TF1 if I remember correctly. So you can make it work but it would have to be trained on your computer since Colab does not support it anymore.

    2/ Yes, we do have all the tutorials here. I'd suggest you to use the YoloV6-n model for which the tutorial is located here (https://colab.research.google.com/github/luxonis/depthai-ml-training/blob/master/colab-notebooks/YoloV6_training.ipynb).

    3/ I am not sure if possible - most of the Yolo models are coded in a way that requires you to use image size that is divisible by 32. Some repositories/versions might support training with different resolutions and will resize the images automatically. In such cases it should be fine -- so if the training works then feel free to use it. When you export the model you will have to choose a shape divisible by 32 though.