@@ -17,87 +17,62 @@ Running sample code in this directory requires PaddelPaddle Fluid v0.14.0 and la
...
@@ -17,87 +17,62 @@ Running sample code in this directory requires PaddelPaddle Fluid v0.14.0 and la
## Data preparation
## Data preparation
Caltech-UCSD Birds 200 (CUB-200) is an image dataset including 200 bird species. We use it to conduct the metric learning experiments. More details of this dataset can be found from its [official website](http://www.vision.caltech.edu/visipedia/CUB-200.html). First of all, preparation of CUB-200 data can be done as:
Stanford Online Product(SOP) dataset contains 120,053 images of 22,634 products downloaded from eBay.com. We use it to conduct the metric learning experiments. For training, 59,5511 out of 11,318 classes are used, and 11,316 classes(60,502 images) are held out for testing. First of all, preparation of SOP data can be done as:
```
```
cd data/
cd data/
sh download_cub200.sh
sh download_sop.sh
```
The script ```data/split.py``` is used to split train/valid set. In our settings, we use images from first 100 classes(001-100) as training data while the other 100 classes are validation data. After the splitting, there are two label files which contain train and validation image labels respectively:
**CUB200_train.txt*: label file of CUB-200 training set, with each line seperated by ```SPACE```, like:
To train a metric learning model, one need to set the neural network as backbone and the metric loss function to optimize. One example of training triplet loss using ResNet-50 is shown below:
To train a metric learning model, one need to set the neural network as backbone and the metric loss function to optimize. We train meiric learning model using softmax or [arcmargin](https://arxiv.org/abs/1801.07698) loss firstly, and then fine-turned the model using other metric learning loss, such as triplet, [quadruplet](https://arxiv.org/abs/1710.00478) and [eml](https://arxiv.org/abs/1212.6094) loss. One example of training using arcmargin loss is shown below:
***total_iter_num**: total number of training iterations. Default: 30000.
***use_gpu**: whether to use GPU or not. Default: True.
***use_gpu**: whether to use GPU or not. Default: True.
***model_save_dir**: the directory to save trained model. Default: "output".
***lr**: initialized learning rate. Default: 0.1.
***pretrained_model**: model path for pretraining. Default: None.
***pretrained_model**: model path for pretraining. Default: None.
***model_save_dir**: the directory to save trained model. Default: "output".
**training log:** the log from training ResNet-50 based triplet loss is like:
***loss_name**: loss fortraining model. Default: "softmax".
```
***arc_scale**: parameter of arcmargin loss. Default: 80.0.
Pass 0, trainbatch 0, lr 9.99999974738e-05, loss_metric 0.0700866878033, loss_cls 5.23635625839, acc1 0.0, acc5 0.100000008941, time 0.16 sec
***arc_margin**: parameter of arcmargin loss. Default: 0.15.
Pass 0, trainbatch 10, lr 9.99999974738e-05, loss_metric 0.0752244070172, loss_cls 5.30303478241, acc1 0.0, acc5 0.100000008941, time 0.14 sec
***arc_easy_margin**: parameter of arcmargin loss. Default: False.
Pass 0, trainbatch 20, lr 9.99999974738e-05, loss_metric 0.0840565115213, loss_cls 5.41880941391, acc1 0.0, acc5 0.0333333350718, time 0.14 sec
Pass 0, trainbatch 30, lr 9.99999974738e-05, loss_metric 0.0698839947581, loss_cls 5.35385560989, acc1 0.0, acc5 0.0333333350718, time 0.14 sec
Pass 0, trainbatch 40, lr 9.99999974738e-05, loss_metric 0.0596057735384, loss_cls 5.34744024277, acc1 0.0, acc5 0.0, time 0.14 sec
Pass 0, trainbatch 50, lr 9.99999974738e-05, loss_metric 0.067836754024, loss_cls 5.37124729156, acc1 0.0, acc5 0.0333333350718, time 0.14 sec
Pass 0, trainbatch 60, lr 9.99999974738e-05, loss_metric 0.0637686774135, loss_cls 5.47412204742, acc1 0.0, acc5 0.0333333350718, time 0.14 sec
Pass 0, trainbatch 70, lr 9.99999974738e-05, loss_metric 0.0772982165217, loss_cls 5.38295936584, acc1 0.0, acc5 0.0, time 0.14 sec
Pass 0, trainbatch 80, lr 9.99999974738e-05, loss_metric 0.0861896127462, loss_cls 5.41250753403, acc1 0.0, acc5 0.0, time 0.14 sec
Pass 0, trainbatch 90, lr 9.99999974738e-05, loss_metric 0.0653102770448, loss_cls 5.53133153915, acc1 0.0, acc5 0.0, time 0.14 sec
...
```
## Finetuning
## Finetuning
Finetuning is to finetune model weights in a specific task by loading pretrained weights. After initializing ```path_to_pretrain_model```, one can finetune a model as:
Finetuning is to finetune model weights in a specific task by loading pretrained weights. After training model using softmax or arcmargin loss, one can finetune the model using triplet, quadruplet or eml loss. One example of fine-turned using eml loss is shown below:
For comparation, many metric learning models with different neural networks and loss functions are trained using corresponding experiential parameters. Recall@Rank-1 is used as evaluation metric and the performance is listed in the table. Pretrained models can be downloaded by clicking related model names.
For comparation, many metric learning models with different neural networks and loss functions are trained using corresponding experiential parameters. Recall@Rank-1 is used as evaluation metric and the performance is listed in the table. Pretrained models can be downloaded by clicking related model names.
