fix test_time_augmentation (#324)

Summary:
also found that the AP (without TTA) for the largest model is computed on a different checkpoint.
Pull Request resolved: fairinternal/detectron2#324

Differential Revision: D18418965

Pulled By: ppwwyyxx

fbshipit-source-id: 8ede8d9f11cc52c771aefd8c5a7f73e5643aaee4

MODEL_ZOO.md

@@ -858,8 +858,8 @@ A few very large models trained for a long time, for demo purposes:
  <tr><td align="left"><a href="configs/Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv.yaml">Mask R-CNN X152</a></td>
 <td align="center">0.281</td>
 <td align="center">15.1</td>
-<td align="center">49.3</td>
-<td align="center">43.2</td>
+<td align="center">50.2</td>
+<td align="center">44.0</td>
 <td align="center"></td>
 <td align="center">18131413</td>
 <td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv/18131413/model_0039999_e76410.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv/18131413/metrics.json">metrics</a></td>
@@ -868,8 +868,8 @@ A few very large models trained for a long time, for demo purposes:
  <tr><td align="left">above + test-time aug.</td>
 <td align="center"></td>
 <td align="center"></td>
-<td align="center">51.4</td>
-<td align="center">45.5</td>
+<td align="center">51.9</td>
+<td align="center">45.9</td>
 <td align="center"></td>
 <td align="center"></td>
 <td align="center"></td>

detectron2/engine/hooks.py

@@ -233,6 +233,19 @@ class AutogradProfiler(HookBase):
     """
     A hook which runs `torch.autograd.profiler.profile`.
 
+    Examples:
+
+    .. code-block:: python
+
+        hooks.AutogradProfiler(
+             lambda trainer: trainer.iter > 10 and trainer.iter < 20, self.cfg.OUTPUT_DIR
+        )
+
+    The above example will run the profiler for iteration 10~20 and dump
+    results to ``OUTPUT_DIR``. We did not profile the first few iterations
+    because they are typically slower than the rest.
+    The result files can be loaded in the ``chrome://tracing`` page in chrome browser.
+
     Note:
         When used together with NCCL on older version of GPUs,
         autograd profiler may cause deadlock because it unnecessarily allocates

detectron2/modeling/test_time_augmentation.py

@@ -194,7 +194,8 @@ def _inference_one_image(self, input):
 
         # 1.3: select from the union of all results
         num_classes = self.cfg.MODEL.ROI_HEADS.NUM_CLASSES
-        all_scores_2d = torch.zeros(num_boxes, num_classes, device=all_boxes.device)
+        # +1 because fast_rcnn_inference expects background scores as well
+        all_scores_2d = torch.zeros(num_boxes, num_classes + 1, device=all_boxes.device)
         for idx, cls, score in zip(count(), all_classes, all_scores):
             all_scores_2d[idx, cls] = score
 

docs/tutorials/models.md

@@ -23,7 +23,7 @@ corresponds to information about one image.
 
 The dict may contain the following keys:
 
-* "image": `Tensor` in (C, H, W) format.
+* "image": `Tensor` in (C, H, W) format. The meaning of channels are defined by `cfg.INPUT.FORMAT`.
 * "instances": an `Instances` object, with the following fields:
 	+ "gt_boxes": `Boxes` object storing N boxes, one for each instance.
 	+ "gt_classes": `Tensor` of long type, a vector of N labels, in range [0, num_categories).