Potter Hsu

To add a new backbone, you should do the following steps: 1. Create an inherited class from `backbone.base.Base`, then override the method `features` ```python class DenseNet161(backbone.base.Base): def __init__(self, pretrained: bool):...

No, currently I assume that the program is run with GPU.

Here are more trained models: [Resnet-18](https://drive.google.com/open?id=1lEzMw9X1vVq_t-RODoX4NtltZIAhIpwh) trained on VOC2007, got mAP = 0.6791 [Resnet-50](https://drive.google.com/open?id=1QaxWHbjM5oX_b1nb8ZKuDxWJ4qX3zgpc) trained on VOC2007, got mAP = 0.7438 [Resnet-50](https://drive.google.com/open?id=16W1k3WeOjsWJ-4GD8l0VTQOBIEvLsQg2) trained on COCO2017, got AP = 0.3252 [Resnet-50](https://drive.google.com/open?id=1MOoviXCaq7BRU_RnJ38ydp2O6OixqK_Y)...

1. First of all, the function [`nms` has handled boxes sorting](https://github.com/potterhsu/easy-faster-rcnn.pytorch/blob/2c38b785eb488cb2dc366cbbdbe269bc5bf00ef2/support/src/cuda/nms.cu#L70-L75) 2. So that in [`generate_detections` boxes doesn't require to sort before `nms` is called](https://github.com/potterhsu/easy-faster-rcnn.pytorch/blob/2c38b785eb488cb2dc366cbbdbe269bc5bf00ef2/model.py#L204) 3. You might be confused...

Right, and this threshold is determined by reference to other popular repos, such as [maskrcnn-benchmark](https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/maskrcnn_benchmark/config/defaults.py#L200)

No, but you can easily extend for your needs.

These sorts of errors usually occurred when your configs between training and inference are inconsistent, for example: * You were training with anchor_sizes=[64,128,256,512] * But are inferring with anchor_sizes=[128,256,512]

I've not tried yet, my environment as follows, just for your reference: OS: Ubuntu 14.04 RAM: 32GB GPU: GTX 780 Ti 3GB Memory It takes about 14 hours for training.

If you’re using Python 2, try to replace with `float(len(self._loader.dataset))` at `evaluator.py` #47, or you could check the value of `num_corrects`

Could you please provide a minimal reproduction of your issue? Also have you followed our sample guide: `inference_sample.ipynb` and `inference_outside_sample.ipynb`?