add support for auto lr-decay

.gitignore

@@ -12,3 +12,5 @@ checkpoints
 model/utils/build/
 
 imgs/
+*.png
+*.jpg

model/faster_rcnn.py

@@ -338,7 +338,7 @@ class FasterRCNN(nn.Module):
             cls_bbox[:, 0::2] = (cls_bbox[:, 0::2]).clamp(min =  0, max = size[0])
             cls_bbox[:, 1::2] = (cls_bbox[:, 1::2]).clamp(min=0,max=size[1])
 
-            prob = at.tonumpy(F.softmax(at.tovariable(roi_score)))
+            prob = at.tonumpy(F.softmax(at.tovariable(roi_score),dim=1))
 
             raw_cls_bbox = at.tonumpy(cls_bbox)
             raw_prob = at.tonumpy(prob)
@@ -353,7 +353,7 @@ class FasterRCNN(nn.Module):
         return bboxes, labels, scores
 
 
-    def get_optimizer_3(self):
+    def get_optimizer_adam(self):
         self.lr1,self.lr2,self.lr3 = opt.lr1,opt.lr2,opt.lr3
         param_groups = [
             {'params':[param for param in self.extractor.parameters() if param.requires_grad], 'lr':opt.lr1},

model/faster_rcnn_vgg16.py

@@ -37,8 +37,8 @@ def decom_vgg16_chainer(pretrained=True):
     classifier = nn.Sequential(*classifier)
 
     # chainer ceil mode = True for maxpooling
-    for idx in [4,9,16,23]:
-        features[idx].ceil_mode=True
+    # for idx in [4,9,16,23]:
+    #     features[idx].ceil_mode=True
     # 
     # del classifier._modules['6']
 
@@ -47,8 +47,8 @@ def decom_vgg16_chainer(pretrained=True):
     #     for p in layer.parameters():
     #         p.requires_grad=False
 
-        
     return nn.Sequential(*features),classifier
+
 def decom_vgg16bn(pretrained=True):
     # the 30th layer of features is relu of conv5_3
     model = vgg16_bn(pretrained)
@@ -130,7 +130,7 @@ class FasterRCNNVGG16(FasterRCNN):
                  min_size=600, max_size=1000,
                  ratios=[0.5, 1, 2], anchor_scales=[8, 16, 32]
                  ):
-        extractor,classifier = decom_vgg16(not opt.load_path)
+        extractor,classifier = decom_vgg16_chainer(not opt.load_path)
 
         rpn = RegionProposalNetwork(
             512, 512,

train.py

@@ -52,7 +52,7 @@ def train(**kwargs):
     test_dataloader = data_.DataLoader(testset,
                                 batch_size=1,
                                 num_workers=2,
-                                shuffle=True,\
+                                shuffle=False,\
                                 # pin_memory=True
                                 )
 
@@ -64,7 +64,7 @@ def train(**kwargs):
         print('load pretrained model from %s' %opt.load_path)
     
     trainer.vis.text(dataset.db.label_names,win='labels')
-
+    best_map = 0
     for epoch in range(opt.epoch):
         trainer.reset_meters()
         for ii,(img, bbox_, label_, scale, ori_img) in tqdm(enumerate(dataloader)):
@@ -107,9 +107,16 @@ def train(**kwargs):
             trainer.faster_rcnn.update_optimizer(opt.lr_decay)
 
         eval_result  = eval(test_dataloader,faster_rcnn)
+
+        if eval_result['map']>best_map:
+            best_path = trainer.save()
+            best_map = eval_result['map']
+        else:
+            trainer.load(best_path)
+            trainer.faster_rcnn.update_optimizer(opt.lr_decay)
+
         trainer.vis.plot('test_map', eval_result['map'])
         trainer.vis.log('map:{},loss:{},roi_cm:{}'.format(str(eval_result),str(trainer.get_meter_data()),str(trainer.rpn_cm.conf.tolist())))
-        trainer.save()
         # t.save(trainer.state_dict(),'checkpoints/fasterrcnn_%s.pth' %epoch)
         # t.vis.save([opt.env])