This project implements Faster RCNN proposed in Ren et al. (2015). The code is completely PyTorch. Every single line of the code is written from scratch (except for the PyTorch's built-in ResNet). The programming style is different from the authors' implementation and other existing implementations, in the sense of improved efficiency and readability.
Clone the repository
git clone https://github.com/symphonylyh/Faster_RCNN.git
Pre-trained ResNet101 and PASCAL VOC 2007 Dataset will be automatically downloaded.
Train the network or resume training from last saved model.
python train.py or python train.py --resume
Model checkpoint and training statistics will be saved in /logs. Tensorboard can be used to visualize the training process
tensorboard --logdir=logs
This is my first experience implementing a non-trivial model architecture, so here are some notes/practice for PyTorch programming
- Within layers, we may need to declare new variables. I was using like
newTensor = torch.zeros(N,M). It works for CPU execution, however, when running on GPU, it will complain thatnewTensoris on CPU device thus is not compatible with other GPU variables. To fix this, we can usenewTensor = torch.zeros(N,M).to(oldTensor.device)given thatoldTensoris a variable on the device we want to move to. This will allow simple inheritance of device info instead of keeping a global variable specifying the correct device.