fix resnet training

examples/ResNet/imagenet-resnet.py

@@ -93,7 +93,7 @@ class Model(ModelDesc):
             50: ([3,4,6,3], bottleneck),
             101: ([3,4,23,3], bottleneck)
         }
-        defs, block_func = cfg[50]
+        defs, block_func = cfg[34]
 
         with argscope(Conv2D, nl=tf.identity, use_bias=False,
                 W_init=variance_scaling_initializer(mode='FAN_OUT')):
@@ -119,8 +119,7 @@ class Model(ModelDesc):
         add_moving_summary(tf.reduce_mean(wrong, name='train-error-top5'))
 
         # weight decay on all W of fc layers
-        wd_w = tf.train.exponential_decay(1e-4, get_global_step_var(),
-                                          200000, 0.7, True)
+        wd_w = 1e-4
         wd_cost = tf.mul(wd_w, regularize_cost('.*/W', tf.nn.l2_loss), name='l2_regularize_loss')
         add_moving_summary(loss, wd_cost)
         self.cost = tf.add_n([loss, wd_cost], name='cost')
@@ -135,7 +134,8 @@ def get_data(train_or_test):
     image_std = np.array([0.229, 0.224, 0.225], dtype='float32')
 
     if isTrain:
-        def resize_func(img):
+        class Resize(imgaug.ImageAugmentor):
+            def _augment(self, img, _):
                 # crop 8%~100% of the original image
                 # See `Going Deeper with Convolutions` by Google.
                 h, w = img.shape[:2]
@@ -157,7 +157,7 @@ def get_data(train_or_test):
                 return out
 
         augmentors = [
-            imgaug.MapImage(resize_func),
+            Resize(),
             imgaug.RandomOrderAug(
                 [imgaug.Brightness(30, clip=False),
                  imgaug.Contrast((0.8, 1.2), clip=False),
@@ -228,8 +228,6 @@ def eval_on_ILSVRC12(model_file, data_dir):
         acc5.feed(o[1].sum(), batch_size)
     print("Top1 Error: {}".format(acc1.ratio))
     print("Top5 Error: {}".format(acc5.ratio))
-    print("Top1 Error: {}".format(acc1.ratio))
-    print("Top5 Error: {}".format(acc5.ratio))
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()