add loss

Author: Qian Liu

README.md

@@ -42,7 +42,7 @@ Now we have supported:
 - Running the code on Google Colab with Free GPU. Check [Here](https://github.com/layumi/Person_reID_baseline_pytorch/tree/master/colab) (Thanks to @ronghao233)
 - [DG-Market](https://github.com/NVlabs/DG-Net#dg-market) (10x Large Synethic Dataset from Market **CVPR 2019 Oral**)
 - [Swin Transformer](https://github.com/microsoft/Swin-Transformer) / [EfficientNet](https://github.com/lukemelas/EfficientNet-PyTorch) / [HRNet](https://github.com/HRNet)
-- Circle Loss (**CVPR 2020 Oral**), Triplet Loss, Contrastive Loss, Sphere Loss, Lifted Loss and Instance Loss
+- Circle Loss, Triplet Loss, Contrastive Loss, Sphere Loss, Lifted Loss, Arcface, Cosface  and Instance Loss
 - Float16 to save GPU memory based on [apex](https://github.com/NVIDIA/apex)
 - Part-based Convolutional Baseline(PCB)
 - Multiple Query Evaluation
@@ -203,6 +203,7 @@ I do not optimize the hyper-parameters. You are free to tune them for better per
 | CE + Sphere [[Paper]](https://openaccess.thecvf.com/content_cvpr_2017/papers/Liu_SphereFace_Deep_Hypersphere_CVPR_2017_paper.pdf) | 92.01% | 79.39% | `python train.py --warm_epoch 5 --stride 1 --erasing_p 0.5 --batchsize 32 --lr 0.08 --name warm5_s1_b32_lr8_p0.5_sphere100 --sphere --total 100; python test.py --name warm5_s1_b32_lr8_p0.5_sphere100` |
 | CE + Triplet [[Paper]](https://arxiv.org/pdf/1703.07737) | 92.40%	| 79.71% | `python train.py --warm_epoch 5 --stride 1 --erasing_p 0.5 --batchsize 32 --lr 0.08 --name warm5_s1_b32_lr8_p0.5_triplet100 --triplet --total 100; python test.py  --name warm5_s1_b32_lr8_p0.5_triplet100` |
 | CE + Lifted [[Paper]](https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Song_Deep_Metric_Learning_CVPR_2016_paper.pdf)|  91.78% | 79.77% | `python train.py --warm_epoch 5 --stride 1 --erasing_p 0.5 --batchsize 32 --lr 0.08 --name warm5_s1_b32_lr8_p0.5_lifted100 --lifted --total 100; python test.py --name warm5_s1_b32_lr8_p0.5_lifted100` |
+| CE + Instance [[Paper]](https://zdzheng.xyz/files/TOMM20.pdf) | 92.49% | 80.51% | `python train.py --warm_epoch 5 --stride 1 --erasing_p 0.5 --batchsize 32 --lr 0.08 --name warm5_s1_b32_lr8_p0.5_instance100_gamma32 --instance --ins_gamma 32 --total 100 ; python test.py  --name  warm5_s1_b32_lr8_p0.5_instance100_gamma32`|
 | CE + Contrast [[Paper]](https://zdzheng.xyz/files/TOMM18.pdf) | 92.28% | 81.42% | `python train.py --warm_epoch 5 --stride 1 --erasing_p 0.5 --batchsize 32 --lr 0.08 --name warm5_s1_b32_lr8_p0.5_contrast100 --contrast  --total 100; python test.py  --name warm5_s1_b32_lr8_p0.5_contrast100`|
 | CE + Circle [[Paper]](https://arxiv.org/abs/2002.10857) | 92.46% | 81.70% | `python train.py --warm_epoch 5 --stride 1 --erasing_p 0.5 --batchsize 32 --lr 0.08 --name warm5_s1_b32_lr8_p0.5_circle100 --circle --total 100 ; python test.py  --name  warm5_s1_b32_lr8_p0.5_circle100` |
 | CE + Contrast + Sphere | 92.79% | 82.02% | `python train.py --warm_epoch 5 --stride 1 --erasing_p 0.5 --batchsize 32 --lr 0.08 --name warm5_s1_b32_lr8_p0.5_cs100 --contrast --sphere --total 100; python test.py --name warm5_s1_b32_lr8_p0.5_cs100`|

train.py

@@ -63,6 +63,7 @@
 parser.add_argument('--cosface', action='store_true', help='use CosFace loss' )
 parser.add_argument('--contrast', action='store_true', help='use contrast loss' )
 parser.add_argument('--instance', action='store_true', help='use instance loss' )
+parser.add_argument('--ins_gamma', default=32, type=int, help='gamma for instance loss')
 parser.add_argument('--triplet', action='store_true', help='use triplet loss' )
 parser.add_argument('--lifted', action='store_true', help='use lifted loss' )
 parser.add_argument('--sphere', action='store_true', help='use sphere loss' )
@@ -216,7 +217,7 @@ def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
     if opt.contrast: 
         criterion_contrast = losses.ContrastiveLoss(pos_margin=0, neg_margin=1)
     if opt.instance:
-        criterion_instance = InstanceLoss(gamma=8)
+        criterion_instance = InstanceLoss(gamma = opt.ins_gamma)
     if opt.sphere:
         criterion_sphere = losses.SphereFaceLoss(num_classes=opt.nclasses, embedding_size=512, margin=4)
     for epoch in range(num_epochs):
@@ -283,7 +284,7 @@ def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
                     if opt.contrast:
                         loss +=  criterion_contrast(ff, labels) #/now_batch_size
                     if opt.instance:
-                        loss += criterion_instance(ff, labels)
+                        loss += criterion_instance(ff, labels) /now_batch_size
                     if opt.sphere:
                         loss +=  criterion_sphere(ff, labels)/now_batch_size
                 elif opt.PCB:  #  PCB