Merge pull request #4 from beesk135/cosface

Cosface

.vscode/settings.json

@@ -0,0 +1,4 @@
+{
+    "python.linting.pylintEnabled": true,
+    "python.linting.enabled": true
+}

Experiment-AGW-market.sh

@@ -8,5 +8,5 @@
 # weight regularized triplet loss
 # generalized mean pooling
 # non local blocks
-python3 tools/main.py --config_file='configs/AGW_baseline.yml' MODEL.DEVICE_ID "('3')" \
-DATASETS.NAMES "('market1501')" OUTPUT_DIR "('./log/market1501/Experiment-AGW-baseline')"
+python3 tools/main.py --config_file='configs/AGW_baseline.yml' MODEL.DEVICE_ID "('0')" \
+DATASETS.NAMES "('market1501')" OUTPUT_DIR "('./log/market1501/cosface-1')"

config/defaults.py

@@ -22,7 +22,7 @@
 # ID number of GPU
 _C.MODEL.DEVICE_ID = '0'
 # Name of backbone
-_C.MODEL.NAME = 'resnet50'
+_C.MODEL.BACKBONE = 'resnet50'
 # Last stride of backbone
 _C.MODEL.LAST_STRIDE = 1
 # Path to pretrained model of backbone
@@ -36,9 +36,14 @@
 _C.MODEL.CENTER_FEAT_DIM = 2048
 # If train with weighted regularized triplet loss, options: 'on', 'off'
 _C.MODEL.WEIGHT_REGULARIZED_TRIPLET = 'off'
-# If train with generalized mean pooling, options: 'on', 'off'
-_C.MODEL.GENERALIZED_MEAN_POOL = 'off'
-
+# custom config
+_C.MODEL.POOL_TYPE = 'avg'
+_C.MODEL.COSINE_LOSS_TYPE = ''
+_C.MODEL.SCALING_FACTOR = 60.0
+_C.MODEL.MARGIN = 0.35
+_C.MODEL.USE_BNBIAS = False 
+_C.MODEL.USE_DROPOUT = True
+_C.MODEL.USE_SESTN = False 
 
 # -----------------------------------------------------------------------------
 # INPUT

configs/AGW_baseline.yml

@@ -1,13 +1,19 @@
 MODEL:
-  PRETRAIN_CHOICE: 'self'
-  PRETRAIN_PATH: "/content/drive/Shared drives/REID/ReId-Survey-Paper/ReID-Survey/log/market1501/Experiment-AGW-baseline/resnet50_nl_model_160.pth"
-  # PRETRAIN_PATH: 'C:/Users/60310249/.torch/models/resnet50-19c8e357.pth'
+  PRETRAIN_CHOICE: 'imagenet'
+  # PRETRAIN_PATH: "/content/drive/Shared drives/REID/ReId-Survey-Paper/ReID-Survey/log/market1501/local-AGW-baseline/resnet50_nl_model_120.pth"
+  PRETRAIN_PATH: '/home/thanit456/torch/weights/resnet50-19c8e357.pth'
   TRANSFER_MODE : 'off'
   CENTER_LOSS: 'on'
   CENTER_FEAT_DIM: 2048
-  NAME: 'resnet50_nl'
+  BACKBONE: 'resnet50_nl'
   WEIGHT_REGULARIZED_TRIPLET: 'on'
-  GENERALIZED_MEAN_POOL: 'on'
+  POOL_TYPE: 'avg'
+  COSINE_LOSS_TYPE: 'CosFace'
+  SCALING_FACTOR: 60.0
+  MARGIN: 0.35
+  USE_BNBIAS: False 
+  USE_DROPOUT: False
+  USE_SESTN: False 
 
 INPUT:
   IMG_SIZE: [256, 128]
@@ -32,6 +38,8 @@ SOLVER:
 
   WEIGHT_DECAY: 0.0005
   IMS_PER_BATCH: 64
+  # IMS_PER_BATCH: 32
+
 
   STEPS: [40, 70]
   GAMMA: 0.1

evaluate/__init__.py

@@ -0,0 +1,12 @@
+import torch 
+
+from .eval_reid import eval_func
+
+def euclidean_dist(x, y):
+    m, n = x.size(0), y.size(0)
+    xx = torch.pow(x, 2).sum(1, keepdim=True).expand(m, n)
+    yy = torch.pow(y, 2).sum(1, keepdim=True).expand(n, m).t()
+    dist = xx + yy
+    dist.addmm_(1, -2, x, y.t())
+    dist = dist.clamp(min=1e-12).sqrt()
+    return dist

evaluate/eval_reid.py

@@ -0,0 +1,65 @@
+import numpy as np
+
+def eval_func(distmat, q_pids, g_pids, q_camids, g_camids, max_rank=200):
+    """Evaluation with market1501 metric
+        Key: for each query identity, its gallery images from the same camera view are discarded.
+        """
+    num_q, num_g = distmat.shape
+    if num_g < max_rank:
+        max_rank = num_g
+        print("Note: number of gallery samples is quite small, got {}".format(num_g))
+    indices = np.argsort(distmat, axis=1)
+    matches = (g_pids[indices] == q_pids[:, np.newaxis]).astype(np.int32)
+
+    # compute cmc curve for each query
+    all_cmc = []
+    all_AP = []
+    num_valid_q = 0.  # number of valid query
+    for q_idx in range(num_q):
+        # get query pid and camid
+        q_pid = q_pids[q_idx]
+        q_camid = q_camids[q_idx]
+
+        # remove gallery samples that have the same pid and camid with query
+        order = indices[q_idx]
+        remove = (g_pids[order] == q_pid) & (g_camids[order] == q_camid)
+        keep = np.invert(remove)
+
+        # compute cmc curve
+        # binary vector, positions with value 1 are correct matches
+        orig_cmc = matches[q_idx][keep]
+        if not np.any(orig_cmc):
+            # this condition is true when query identity does not appear in gallery
+            # [update:20191029] divide by query
+            all_AP.append(0)
+            continue
+
+        cmc = orig_cmc.cumsum()
+        cmc[cmc > 1] = 1
+
+        all_cmc.append(cmc[:max_rank])
+        num_valid_q += 1.
+
+        # compute average precision
+        # reference: https://en.wikipedia.org/wiki/Evaluation_measures_(information_retrieval)#Average_precision
+        # [update:20191029] support for map@max_rank
+        orig_cmc = orig_cmc[:max_rank]
+        if not np.any(orig_cmc):
+            all_AP.append(0)
+            continue
+        num_rel = orig_cmc.sum()
+        tmp_cmc = orig_cmc.cumsum()
+        tmp_cmc = [x / (i + 1.) for i, x in enumerate(tmp_cmc)]
+        tmp_cmc = np.asarray(tmp_cmc) * orig_cmc
+        AP = tmp_cmc.sum() / num_rel
+        all_AP.append(AP)
+
+    # assert num_valid_q > 0, "Error: all query identities do not appear in gallery"
+
+    all_cmc = np.asarray(all_cmc).astype(np.float32)
+    # [update:20191029] divide by query
+    all_cmc = all_cmc.sum(0) / num_q
+
+    mAP = np.mean(all_AP)
+
+    return all_cmc, mAP,all_AP

modeling/__init__.py

@@ -3,6 +3,16 @@
 from .baseline import Baseline
 
 def build_model(cfg, num_classes):
-    model = Baseline(num_classes, cfg.MODEL.LAST_STRIDE, cfg.MODEL.PRETRAIN_PATH, cfg.MODEL.NAME,
-                     cfg.MODEL.GENERALIZED_MEAN_POOL, cfg.MODEL.PRETRAIN_CHOICE)
+    model = Baseline(num_classes=num_classes, 
+                    last_stride=cfg.MODEL.LAST_STRIDE, 
+                    model_path=cfg.MODEL.PRETRAIN_PATH,
+                    backbone=cfg.MODEL.BACKBONE,
+                    pool_type=cfg.MODEL.POOL_TYPE,
+                    use_dropout=cfg.MODEL.USE_DROPOUT,
+                    cosine_loss_type=cfg.MODEL.COSINE_LOSS_TYPE,
+                    s=cfg.MODEL.SCALING_FACTOR,
+                    m=cfg.MODEL.MARGIN,
+                    use_bnbias=cfg.MODEL.USE_BNBIAS,
+                    use_sestn=cfg.MODEL.USE_SESTN,
+                    pretrain_choice=cfg.MODEL.PRETRAIN_CHOICE)
     return model

modeling/backbones/resnet_ibn_a.py

@@ -3,6 +3,7 @@
 import math
 import torch.utils.model_zoo as model_zoo
 
+from ..layer.attention import SESTNLayer
 
 __all__ = ['ResNet_IBN', 'resnet50_ibn_a', 'resnet101_ibn_a',
            'resnet152_ibn_a']
@@ -76,7 +77,7 @@ def forward(self, x):
 
 class ResNet_IBN(nn.Module):
 
-    def __init__(self, last_stride, block, layers, num_classes=1000):
+    def __init__(self, last_stride, block, layers, num_classes=1000, use_sestn=False):
         scale = 64
         self.inplanes = scale
         super(ResNet_IBN, self).__init__()
@@ -92,6 +93,11 @@ def __init__(self, last_stride, block, layers, num_classes=1000):
         self.avgpool = nn.AvgPool2d(7)
         self.fc = nn.Linear(scale * 8 * block.expansion, num_classes)
 
+        self.use_sestn = use_sestn
+        if use_sestn:
+            self.sestn1 = SESTNLayer(256, 16)
+            self.sestn2 = SESTNLayer(512, 32)
+
         for m in self.modules():
             if isinstance(m, nn.Conv2d):
                 n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
@@ -130,7 +136,15 @@ def forward(self, x):
         x = self.maxpool(x)
 
         x = self.layer1(x)
+
+        if self.use_sestn:
+            x = self.sestn1(x)
+
         x = self.layer2(x)
+
+        if self.use_sestn:
+            x = self.sestn2(x)
+
         x = self.layer3(x)
         x = self.layer4(x)