libtorch-SSD

SSD net Training on pytorch and Implementation on libtorch

Download ssd.pytorch to train and get the model.

Requirements

1. LibTorch v1.0.0
2. Cuda
3. OpenCV

Train SSD Pytorch model

I use ssd.pytorch to train ssd and get a model, c++ code for detection_layer is same to the pytorch code.
ssd_epoch_11_.pth got in this part.

Get traced model for libtorch

to get the traced model , I modify the ssd.py in ssd.pytorch,

1. ssd.py modified around Line 194

if phase != "test" and phase != "train" and phase != "trace":
        print("ERROR: Phase: " + phase + " not recognized")
        return

2. ssd.py add around Line 103

elif self.phase == "trace":
        output = torch.cat((loc.view(loc.size(0), -1, 4), self.softmax(conf.view(conf.size(0), -1, self.num_classes))), 2)

3. python ssd_trace.py

Get traced models ssd_voc.pt by ssd_epoch_11_.pth download models, u46v

Running the libtorch-ssd

configuration path in ssd-app.cpp and make

./ ssd_app

to do

1. only support ssd300, to expand to ssd500
2. not support batch test, the test data of is 1x3x300x300