misc small changes and fix #688

examples/FasterRCNN/README.md

@@ -64,7 +64,7 @@ The two 360k models have identical configurations with
 `R50-C4-2x` configuration in
 [Detectron Model Zoo](https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md#end-to-end-faster--mask-r-cnn-baselines).
 They get the __same performance__ with the official models, and are about 14% slower than the official implementation,
-probably due to the lack of specialized ops in TensorFlow.
+probably due to the lack of specialized ops (e.g. AffineChannel, ROIAlign) in TensorFlow.
 
 ## Notes
 

examples/FasterRCNN/config.py

@@ -21,7 +21,7 @@ FREEZE_AFFINE = False   # do not train affine parameters inside BN
 
 # schedule -----------------------
 BASE_LR = 1e-2
-WARMUP = 500    # in steps
+WARMUP = 1000    # in steps
 STEPS_PER_EPOCH = 500
 LR_SCHEDULE = [150000, 230000, 280000]
 # LR_SCHEDULE = [120000, 160000, 180000]    # "1x" schedule in detectron

scripts/checkpoint-prof.py

@@ -21,7 +21,7 @@ if __name__ == '__main__':
     parser.add_argument('--print-timing', action='store_true')
     args = parser.parse_args()
 
-    tf.train.import_meta_graph(args.meta)
+    tf.train.import_meta_graph(args.meta, clear_devices=True)
     G = tf.get_default_graph()
     with tf.Session(config=get_default_sess_config()) as sess:
         init = get_model_loader(args.model)

scripts/dump-model-params.py

@@ -17,7 +17,7 @@ if __name__ == '__main__':
     parser.add_argument(dest='output', help='output model file, can be npz or TF checkpoint')
     args = parser.parse_args()
 
-    tf.train.import_meta_graph(args.meta)
+    tf.train.import_meta_graph(args.meta, clear_devices=True)
 
     # loading...
     init = get_model_loader(args.input)

tensorpack/models/regularize.py

@@ -125,8 +125,9 @@ def Dropout(x, *args, **kwargs):
     if 'is_training' in kwargs:
         kwargs['training'] = kwargs.pop('is_training')
     if len(args) > 0:
+        if args[0] != 0.5:
             logger.warn(
-            "The first positional argument to tensorpack.Dropout is the probability to keep rather than to drop. "
+                "The first positional argument to tensorpack.Dropout is the probability to keep, rather than to drop. "
                 "This is different from the rate argument in tf.layers.Dropout due to historical reasons. "
                 "To mimic tf.layers.Dropout, explicitly use keyword argument 'rate' instead")
         rate = 1 - args[0]