[FasterRCNN] update LR schedule

examples/FasterRCNN/README.md

@@ -39,8 +39,8 @@ To predict on an image (and show output in a window):
 
 ## Results
 
-+ trainval35k/minival, FASTRCNN_BATCH=256: 32.9
-+ trainval35k/minival, FASTRCNN_BATCH=64: 31.6. Takes less than one day on 8 Maxwell TitanX.
++ trainval35k/minival, FASTRCNN_BATCH=256: 33.4. Takes 49h on 8 TitanX.
++ trainval35k/minival, FASTRCNN_BATCH=64: 32.2. Takes 31h on 8 TitanX.
 
 The hyperparameters are not carefully tuned. You can probably get better performance by e.g.  training longer.
 

examples/FasterRCNN/data.py

@@ -191,9 +191,9 @@ def get_rpn_anchor_input(im, boxes, klass, is_crowd):
 def read_and_augment_images(ds):
     def mapf(dp):
         fname = dp[0]
-        im = cv2.imread(fname, cv2.IMREAD_COLOR).astype('float32')
-        assert im is not None, dp[0]
-        dp[0] = im
+        im = cv2.imread(fname, cv2.IMREAD_COLOR)
+        assert im is not None, fname
+        dp[0] = im.astype('float32')
 
         # assume floatbox as input
         assert dp[1].dtype == np.float32

examples/FasterRCNN/train.py

@@ -286,19 +286,19 @@ if __name__ == '__main__':
                 # linear warmup
                 ScheduledHyperParamSetter(
                     'learning_rate',
-                    [(0, 0.003), (warmup_epoch * factor, 0.01)], interp='linear'),
+                    [(0, 3e-3), (warmup_epoch * factor, 1e-2)], interp='linear'),
                 # step decay
                 ScheduledHyperParamSetter(
                     'learning_rate',
-                    [(warmup_epoch * factor, 0.01),
-                     (120000 * factor // stepnum, 1e-3),
-                     (180000 * factor // stepnum, 1e-4)]),
+                    [(warmup_epoch * factor, 1e-2),
+                     (150000 * factor // stepnum, 1e-3),
+                     (210000 * factor // stepnum, 1e-4)]),
                 HumanHyperParamSetter('learning_rate'),
                 EvalCallback(),
                 GPUUtilizationTracker(),
             ],
             steps_per_epoch=stepnum,
-            max_epoch=205000 * factor // stepnum,
+            max_epoch=230000 * factor // stepnum,
             session_init=get_model_loader(args.load) if args.load else None,
             nr_tower=get_nr_gpu()
         )