Florian Vahl comments

Results 184 comments of


                                            Florian Vahl

train on coco, 100 epochs and 20% mAP

Lowering the learning rate should counterintuitively improve your convergence speed. I did some tests a while ago. Lower it by a factor of 10 and try again. I also suspect...

Why is the output of the model has no grads?

Are you in test or training mode? The model is not differentiable in test mode as it directly outputs a set of boxes.

convert this project's .pth to darknet's .weights

We should add this file as a script to this git. With some command-line arguments.

Using the model as an argument to a class method

I am not able to reproduce this issue. Is it still relevant / occurring and are you sure that there is no other PyTorch code in your class that might...

RuntimeError: CUDA error: an illegal memory access was encountered

Is this issue still relevant/occurring?

Does this implementation can create a model from YOLOv3-spp.cfg

I don't know, did you try an spp cfg file?

Training is slow on Google Cloud Nvidia Tesla P100

Is this issue still relevant/occurring?

mAP and loss value is fluctuated

You could also increase the gradient accumulations

low mAP on coco val2014

What confidence threshold did you use? The yolo paper uses afaik relatively low thresholds such as `0.001` combined with another test dataset called `test-dev`.

low mAP on coco val2014

Have a look at this list https://cocodataset.org/#download