Make changes for mobilenet. (Training is still not working)

Author: Ruotian Luo

README.md

@@ -88,7 +88,7 @@ Additional features not mentioned in the [report](https://arxiv.org/pdf/1702.021
 ### Setup data
 Please follow the instructions of py-faster-rcnn [here](https://github.com/rbgirshick/py-faster-rcnn#beyond-the-demo-installation-for-training-and-testing-models) to setup VOC and COCO datasets (Part of COCO is done). The steps involve downloading data and optionally creating soft links in the ``data`` folder. Since faster RCNN does not rely on pre-computed proposals, it is safe to ignore the steps that setup proposals.
 
-If you find it useful, the ``data/cache`` folder created on Xinlei's side is also shared [here](http://ladoga.graphics.cs.cmu.edu/xinleic/tf-faster-rcnn/cache.tgz).
+If you find it useful, the ``data/cache`` folder created on Xinlei's side is also shared [here](https://drive.google.com/drive/folders/0B1_fAEgxdnvJSmF3YUlZcHFqWTQ).
 
 ### Demo and Test with pre-trained models
 1. Download pre-trained model (only google drive works)
@@ -173,6 +173,15 @@ This script will create a `.pth` file with the same name in the same folder as t
    cd ../..
    ```
 
+   For Mobilenet V1, you can set up like:
+   ```Shell
+   mkdir -p data/imagenet_weights
+   cd data/imagenet_weights
+   # download from my gdrive (https://drive.google.com/open?id=0B7fNdx_jAqhtZGJvZlpVeDhUN1k)
+   mv mobilenet_v1_1.0_224.pth.pth mobile.pth
+   cd ../..
+   ```
+
 2. Train (and test, evaluation)
   ```Shell
   ./experiments/scripts/train_faster_rcnn.sh [GPU_ID] [DATASET] [NET]

lib/model/train_val.py

@@ -128,7 +128,7 @@ def construct_graph(self):
         if 'bias' in key:
           params += [{'params':[value],'lr':lr*(cfg.TRAIN.DOUBLE_BIAS + 1), 'weight_decay': cfg.TRAIN.BIAS_DECAY and cfg.TRAIN.WEIGHT_DECAY or 0}]
         else:
-          params += [{'params':[value],'lr':lr, 'weight_decay': cfg.TRAIN.WEIGHT_DECAY}]
+          params += [{'params':[value],'lr':lr, 'weight_decay': getattr(value, 'weight_decay', cfg.TRAIN.WEIGHT_DECAY)}]
     self.optimizer = torch.optim.SGD(params, momentum=cfg.TRAIN.MOMENTUM)
     # Write the train and validation information to tensorboard
     self.writer = tb.writer.FileWriter(self.tbdir)

lib/nets/mobilenet_v1.py

@@ -229,22 +229,18 @@ def set_bn_fix(m):
     self.mobilenet.apply(set_bn_fix)
 
     # Add weight decay
-    def l2_regularizer(m, wd):
+    def l2_regularizer(m, wd, regu_depth):
       if m.__class__.__name__.find('Conv') != -1:
-        m.weight.weight_decay = cfg.MOBILENET.WEIGHT_DECAY
-    if cfg.MOBILENET.REGU_DEPTH:
-      self.mobilenet.apply(lambda x: l2_regularizer(x, cfg.MOBILENET.WEIGHT_DECAY))
-    else:
-      self.mobilenet.apply(lambda x: l2_regularizer(x, 0))
-      # always set the first conv layer
-      list(self.mobilenet.children())[0].apply(lambda x: l2_regularizer(x, cfg.MOBILENET.WEIGHT_DECAY))
+        if regu_depth or m.groups == 1:
+            m.weight.weight_decay = wd
+        else:
+            m.weight.weight_decay = 0
+    self.mobilenet.apply(lambda x: l2_regularizer(x, cfg.MOBILENET.WEIGHT_DECAY, cfg.MOBILENET.REGU_DEPTH))
 
     # Build mobilenet.
     self._layers['head'] = nn.Sequential(*list(self.mobilenet.children())[:12])
     self._layers['tail'] = nn.Sequential(*list(self.mobilenet.children())[12:])
 
   def load_pretrained_cnn(self, state_dict):
-    # TODO
     print('Warning: No available pretrained model yet')
-    return
-    self.mobilenet.load_state_dict({k: state_dict[k] for k in list(self.resnet.state_dict())})
+    self.mobilenet.load_state_dict({k: state_dict['features.'+k] for k in list(self.mobilenet.state_dict())})