StephenRUK

When a (TensorFlow) checkpoint is stored, the current weights are saved. If they aren't saved in a checkpoint, there's no way to get them back at a later time. You...

There doesn't seem to be an option, see what `lumi eval --help` tells you. I would also like to see a better direct interaction with the TensorFlow checkpoints here. Even...

Edit: I previously also mentioned the fact that `lumi eval` runs against already evaluated checkpoints. I forgot that the argument `--from-global-step` exists where you can specify the checkpoint/step number to...

@MathGaron did you ever follow up on this? I am getting different results (lower score) than with the official COCO API (assuming it's the same as VOC). Also the precision...

I will brush up my PR-curve theory and create an example to check the results against. It is possible that the differences are due to minor differences in the algorithms...

While we did not find an error in this implementation, we discovered a reason for the difference to our existing results. The library we were using, Luminoth, takes _all_ predictions...

Cool! Also if you have a better idea for the method name, go ahead 😁

Hi, you should look at the author's article on Medium and the notebook for training. You can replace the base network with your own easily, see `nn_base` in the training...

You can find the notebook in this repository. https://github.com/RockyXu66/Faster_RCNN_for_Open_Images_Dataset_Keras/blob/master/frcnn_train_vgg.ipynb

You can use pretrained weights for common models e.g. VGG16, ResNet or such. See the Keras docs for available models with ImageNet weights, if you wish to use them https://keras.io/applications/