I tried all the training parameters suggested by the authors, e.g. samples_per_gpu=4,workers_per_gpu=8, changing the learning rate, adding gradient cropping, changing the dataset, etc. But a learning rate of 4e-4 will always result in IOU=0, and a learning rate set to 2e-4 or 1e-4 trains a model that is far less effective than the author's model.
So has anyone trained the same model results as the authors? And how are the parameters set? Very much looking forward to the guidance!
I tried all the training parameters suggested by the authors, e.g. samples_per_gpu=4,workers_per_gpu=8, changing the learning rate, adding gradient cropping, changing the dataset, etc. But a learning rate of 4e-4 will always result in IOU=0, and a learning rate set to 2e-4 or 1e-4 trains a model that is far less effective than the author's model. So has anyone trained the same model results as the authors? And how are the parameters set? Very much looking forward to the guidance!