Bilinear-CNN-for-fine-grained-recognition

细分类双线性池化模型（ICCV2015）
1.Dataset :
I trained it on bird dataset
2.Requirement
python3.5
torchvision0.4
3.train
1)first fine-tune fc layers:
lr=1.0,weight decay=1e-8,
I achieved 74.543%
first you need delete this all rows:this rows can load your pretrained parameters:
state_dict="/home/lyh2017/code/bili/bilinear_09_02/model_09_02/100.pkl"
pretrained_dict = torch.load(state_dict)
model_dict = net.state_dict()
pretrained_dict = {k[7: ]: v for k, v in pretrained_dict.items() if k[7:] in model_dict}
model_dict.update(pretrained_dict)
net.load_state_dict(model_dict)
2)fune-tune all layers:
you need change net=BCNN(200,True)
add delete rows :
state_dict="/home/lyh2017/code/bili/bilinear_09_02/model_09_02/100.pkl"
pretrained_dict = torch.load(state_dict)
model_dict = net.state_dict()
pretrained_dict = {k[7: ]: v for k, v in pretrained_dict.items() if k[7:] in model_dict}
model_dict.update(pretrained_dict)
net.load_state_dict(model_dict)
lr=0.01,weight decay=1e-5