gradient clip

demo.py

@@ -7,13 +7,14 @@
 
 def test():
     import os
-    im_file = 'demo/004545.jpg'
+    # im_file = 'demo/004545.jpg'
     im_file = 'data/VOCdevkit2007/VOC2007/JPEGImages/009936.jpg'
-    im_file = '/media/longc/Data/data/2DMOT2015/test/ETH-Crossing/img1/000100.jpg'
+    # im_file = '/media/longc/Data/data/2DMOT2015/test/ETH-Crossing/img1/000100.jpg'
     image = cv2.imread(im_file)
 
     # model_file = '/media/longc/Data/models/VGGnet_fast_rcnn_iter_70000.h5'
     model_file = '/media/longc/Data/models/faster_rcnn_pytorch/faster_rcnn_30000.h5'
+    # model_file = '/media/longc/Data/models/faster_rcnn_pytorch2/faster_rcnn_2000.h5'
     detector = FasterRCNN()
     network.load_net(model_file, detector)
     detector.cuda()

experiments/cfgs/faster_rcnn_end2end.yml

@@ -12,7 +12,7 @@ TRAIN:
   DISPLAY: 10
   SNAPSHOT_ITERS: 5000
   HAS_RPN: True
-  LEARNING_RATE: 0.0001
+  LEARNING_RATE: 0.001
   MOMENTUM: 0.9
   GAMMA: 0.1
   STEPSIZE: 60000

faster_rcnn/faster_rcnn.py

@@ -226,7 +226,7 @@ def loss(self):
         # print self.loss_box
         # print self.rpn.cross_entropy
         # print self.rpn.loss_box
-        return self.cross_entropy + self.loss_box * 10 + self.rpn.loss
+        return self.cross_entropy + self.loss_box * 10
 
     def forward(self, im_data, im_info, gt_boxes=None, dontcare_areas=None):
         features, rois = self.rpn(im_data, im_info, gt_boxes, dontcare_areas)
@@ -258,18 +258,18 @@ def build_loss(self, cls_score, bbox_pred, roi_data):
         fg_cnt = torch.sum(label.data.ne(0))
         bg_cnt = label.data.numel() - fg_cnt
 
-        # ce_weights = torch.ones(cls_score.size()[1])
-        # ce_weights[0] = float(fg_cnt) / bg_cnt / 3.0
-        # # ce_weights[0] = 1./50
-        # ce_weights = ce_weights.cuda()
+        ce_weights = torch.ones(cls_score.size()[1])
+        ce_weights[0] = float(fg_cnt) / bg_cnt
+        # ce_weights[0] = 1./50
+        ce_weights = ce_weights.cuda()
 
         maxv, predict = cls_score.data.max(1)
         self.tp = torch.sum(predict[:fg_cnt].eq(label.data[:fg_cnt]))
         self.tf = torch.sum(predict[fg_cnt:].eq(label.data[fg_cnt:]))
         self.fg_cnt = fg_cnt
         self.bg_cnt = bg_cnt
         # print predict
-        cross_entropy = F.cross_entropy(cls_score, label)
+        cross_entropy = F.cross_entropy(cls_score, label, weight=ce_weights)
         # print cross_entropy
 
         # bounding box regression L1 loss

faster_rcnn/network.py

@@ -62,8 +62,27 @@ def set_trainable(model, requires_grad):
 
 
 def weights_normal_init(model, dev=0.01):
-    for m in model.modules():
-        if isinstance(m, nn.Conv2d):
-            m.weight.data.normal_(0.0, dev)
-        elif isinstance(m, nn.Linear):
-            m.weight.data.normal_(0.0, dev)
+    if isinstance(model, list):
+        for m in model:
+            weights_normal_init(m, dev)
+    else:
+        for m in model.modules():
+            if isinstance(m, nn.Conv2d):
+                m.weight.data.normal_(0.0, dev)
+            elif isinstance(m, nn.Linear):
+                m.weight.data.normal_(0.0, dev)
+
+
+def clip_gradient(model, clip_norm):
+    """Computes a gradient clipping coefficient based on gradient norm."""
+    totalnorm = 0
+    for p in model.parameters():
+        if p.requires_grad:
+            modulenorm = p.grad.data.norm()
+            totalnorm += modulenorm ** 2
+    totalnorm = np.sqrt(totalnorm)
+
+    norm = clip_norm / max(totalnorm, clip_norm)
+    for p in model.parameters():
+        if p.requires_grad:
+            p.grad.mul_(norm)

train.py

@@ -32,11 +32,12 @@
 data_layer = RoIDataLayer(roidb, imdb.num_classes)
 
 net = FasterRCNN(classes=imdb.classes)
-network.weights_normal_init(net, dev=0.01)
-net.rpn.features.load_from_npy_file(pretrained_model)
-# model_file = '/media/longc/Data/models/VGGnet_fast_rcnn_iter_70000.h5'
+# network.weights_normal_init(net, dev=0.01)
+# net.rpn.features.load_from_npy_file(pretrained_model)
+model_file = '/media/longc/Data/models/VGGnet_fast_rcnn_iter_70000.h5'
 # model_file = '/media/longc/Data/models/faster_rcnn_pytorch/faster_rcnn_10000.h5'
-# network.load_net(model_file, net)
+network.load_net(model_file, net)
+network.weights_normal_init([net.bbox_fc, net.score_fc, net.fc6, net.fc7], dev=0.01)
 
 # net = net.rpn
 
@@ -46,10 +47,12 @@
 params = list(net.parameters())
 for p in params:
     print p.size()
-# optimizer = torch.optim.Adam(params[8:], lr=lr)
-optimizer = torch.optim.SGD(params[8:-8], lr=lr, momentum=0.9, weight_decay=0.0005)
-train_all = False
-target_net = net.rpn
+# optimizer = torch.optim.Adam(params[-8:], lr=lr)
+optimizer = torch.optim.SGD(params[-8:], lr=lr, momentum=0.9, weight_decay=0.0005)
+train_all = True
+# target_net = net.rpn
+target_net = net
+network.set_trainable(net.rpn, False)
 
 if not os.path.exists(output_dir):
     os.mkdir(output_dir)
@@ -85,6 +88,7 @@
     # backward
     optimizer.zero_grad()
     loss.backward()
+    network.clip_gradient(target_net, 10.)
     optimizer.step()
 
     if step % log_interval == 0:
@@ -103,12 +107,15 @@
         tp, tf, fg, bg = 0., 0., 0, 0
         t.tic()
 
-    if step % 20000 == 0 and step > 0:
+    if step % 10000 == 0 and step > 0:
         save_name = os.path.join(output_dir, 'faster_rcnn_{}.h5'.format(step))
         network.save_net(save_name, net)
         print('save model: {}'.format(save_name))
-        lr /= 3
-        optimizer = torch.optim.SGD(params[8:], lr=lr, momentum=0.9, weight_decay=0.0005)
-        # optimizer = torch.optim.Adam(params[8:], lr=lr)
-        train_all = True
-        target_net = net
+        # lr /= 10
+        # optimizer = torch.optim.SGD(params[-8:], lr=lr, momentum=0.9, weight_decay=0.0005)
+        # if step >= 20000:
+        #     lr /= 3
+        #     optimizer = torch.optim.SGD(params[8:], lr=lr, momentum=0.9, weight_decay=0.0005)
+        #     # optimizer = torch.optim.Adam(params[8:], lr=lr)
+        #     train_all = True
+        #     target_net = net