ouceduxzk/fine_grained_classification

Fined grained classification On Car dataset

Fine_Grained_Classification

Task : Classification of 196 classes of cars with less than 9k images for training. It's intend for replication of the work :

Monza: Image Classification of Vehicle Make and Model Using Convolutional Neural Networks and Transfer Learning

1. use pre-trained googlenet
2. train with normal rates
3. Achieved 87% top1 accuracy and 97% top5 acc with 10000 iterations

Lessons learned :

1.  Training from scratch is very hard with less data and more class.
2.  Bounding box of the car really helps
3.  fine tuning need to select a good lr and make the bottom layers learn slow, but the last few layers learn fast.

fine_grained_classification's Issues

can you share the models?

How to pre-train APNs ?

I trained this model on my dataset but i am not getting much improvement in accuracy than directly using fine-tuned googLenet. Maybe because i didn't use pre-trained APNs. So how do i train APN?

I have following questions regarding pretraining APN :
1. In RA-CNN paper what does it mean to select the square from original image with the highest response value in the last conv5_4 layer? (conv5_4 blobs.data.shape is 32, 512, 14, 14)
2. Which dataset you are using to train these APNs?

make error，how can this problem be solved?

../lib/libcaffe.so.1.0.0：对‘caffe::RankLoss2Layer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
../lib/libcaffe.so.1.0.0：对‘caffe::RankLoss2Layer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用

how does the backward of attention_crop_layer work?

hi~ I'm trying to reproduce the RA-CNN paper, I read your code of attention_crop_layer, I am puzzled the idea of your Backward part. I don't know why you compute the max top diff, then use a coefficient of 0.0000001?

The Rank_Loss in train_val_fixcls.prototxt?

Hi ,guy
I have a question about the rank_loss in train_val_fixcls.prototxt.In the RA-CNN/alternate_training/train_val_fixcls.prototxt, you concat prob1,prob2,prob3 compute rank_loss.I think there are two rank_loss.The first rank_loss between prob1,prob2 and the second rank_loss between prob2,prob3. @ouceduxzk

where is the 'rank_loss2_layer.cu'

Training Process of RA-CNN

Hi, I'm confused about the your training process when reproducing RA-CNN.

I wonder whether it is appropriate to pretrain APN for first 50k iterations, then fix parameters in APN and train CNN for next 50k iterations, and further fix parameters in CNN and train APN for next iterations, and finally train them all for the final 50k iterations.

Thanks a lot !

How the grad flow back while training APN?

Hi, im trying to reimplement RACNN, and your src code helps me a lot.

But here are some questions confuse me.

1. How the grad flow back to APN while training APN? I mean that APN have 2 inputs: last image and coordinates[tx, ty, tl] from FC layer, and the output is cropped finer image. But while backprop, how to compute grad of [tx, ty, tl]?

2. Rank loss indeed takes as inputs two probabilities in paper, how to optimize 3 APN networks from 2 loss?

Thanks a lot.