From 3a22aa52f3b99664121b243d1d3615301e251250 Mon Sep 17 00:00:00 2001
From: liaoxingyu <xyliao1993@qq.com>
Date: Thu, 26 Jul 2018 08:18:25 +0800
Subject: [PATCH] update

---
 README.md    | 23 +++++++----------------
 main_reid.py | 19 +++++++++++--------
 2 files changed, 18 insertions(+), 24 deletions(-)

diff --git a/README.md b/README.md
index f34591f..3b9e4b0 100644
--- a/README.md
+++ b/README.md
@@ -1,25 +1,16 @@
-# reid_baseline
-reid baseline model for exploring softmax and triplet hard loss's influence.
+# ReID_baseline
+Baseline model (with bottleneck) for person ReID (using softmax and triplet loss).
 
-### Classification
-<img src='https://ws3.sinaimg.cn/large/006tKfTcly1fs76ysvu3zj30kg0d60t5.jpg' height='200'>  <img src='https://ws2.sinaimg.cn/large/006tKfTcly1fs76zbtfxcj30js0d674m.jpg' height='200'>
-
-
-### Triplet Hard
-<img src='https://ws2.sinaimg.cn/large/006tNc79ly1fs3sxc54xjj30ka0d6dgd.jpg' height='200'> <img src='https://ws2.sinaimg.cn/large/006tNc79ly1fs3tpat6emj30k00d2t93.jpg' height=200>
-
-
-### Classification + Triplet Hard
-
-<img src='https://ws2.sinaimg.cn/large/006tKfTcly1fs79doog34j30ja0cudg5.jpg' height='200'> <img src='https://ws2.sinaimg.cn/large/006tNc79ly1fs3tpat6emj30k00d2t93.jpg' height=200>
+## Learning rate
+This is the warmpup strategy learning rate
 
+<img src='https://ws3.sinaimg.cn/large/006tNc79ly1fthmcjwoaaj31kw0natad.jpg' height='200'>
 
 ## Results
 
 | loss | rank1 | map |
 | --- | --| ---|
-| triplet hard | 89.9% | 76.8% | 
 | softmax | 87.9% | 70.1% |
-|triplet + softmax | 92% | 78.1% |
-
+| triplet | 88.8% | 74.8% | 
+|triplet + softmax | 92.0% | 78.1% |
 
diff --git a/main_reid.py b/main_reid.py
index 6de54d8..f299563 100644
--- a/main_reid.py
+++ b/main_reid.py
@@ -27,7 +27,7 @@ from datasets.data_loader import ImageData
 from datasets.samplers import RandomIdentitySampler
 from models import get_baseline_model
 from trainers import clsTrainer, cls_tripletTrainer, tripletTrainer, ResNetEvaluator
-from utils.loss import TripletLoss
+from utils.loss import TripletLoss, CrossEntropyLabelSmooth
 from utils.serialization import Logger
 from utils.serialization import save_checkpoint
 from utils.transforms import TrainTransform, TestTransform
@@ -94,7 +94,8 @@ def train(**kwargs):
     print('model size: {:.5f}M'.format(sum(p.numel()
                                            for p in model.parameters()) / 1e6))
 
-    xent_criterion = nn.CrossEntropyLoss()
+    # xent_criterion = nn.CrossEntropyLoss()
+    xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids)
     tri_criterion = TripletLoss(opt.margin)
 
     def cls_criterion(cls_scores, targets):
@@ -113,7 +114,7 @@ def train(**kwargs):
 
     # get optimizer
     optimizer = torch.optim.Adam(
-        optim_policy, lr=opt.lr, weight_decay=opt.weight_decay,
+        optim_policy, lr=opt.lr, weight_decay=opt.weight_decay
     )
 
     def adjust_lr(optimizer, ep):
@@ -121,13 +122,15 @@ def train(**kwargs):
             lr = 1e-4 * (ep + 1) / 2
         elif ep < 80:
             lr = 1e-3 * opt.num_gpu
-        elif 80 <= ep <= 180:
+        elif ep < 180:
             lr = 1e-4 * opt.num_gpu
-        elif 180 <= ep <= 300:
+        elif ep < 300:
             lr = 1e-5 * opt.num_gpu
-        elif 300 <= ep <= 320:
-            lr = 1e-4 * (ep - 300 + 1) / 2 * opt.num_gpu
-        elif 380 <= ep <= 480:
+        elif ep < 320:
+            lr = 1e-5 * 0.1 ** ((ep - 320) / 80) * opt.num_gpu
+        elif ep < 400:
+            lr = 1e-6
+        elif ep < 480:
             lr = 1e-4 * opt.num_gpu
         else:
             lr = 1e-5 * opt.num_gpu