fix checkpoint & runner bugs

mmcv/image/transforms/geometry.py

@@ -3,7 +3,7 @@ from __future__ import division
 import cv2
 import numpy as np
 
-__all__ = ['imflip', 'imrotate', 'imcrop', 'impad', 'impad_to_multiple']
+__all__ = ['imflip', 'imrotate', 'imcrop', 'impad', 'impad_to_multiple', 'bbox_flip']
 
 
 def imflip(img, direction='horizontal'):
@@ -111,6 +111,20 @@ def bbox_scaling(bboxes, scale, clip_shape=None):
         return scaled_bboxes
 
 
+def bbox_flip(bboxes, img_shape):
+    """Flip bboxes horizontally
+    Args:
+        bboxes(ndarray): shape (..., 4*k)
+        img_shape(tuple): (height, width)
+    """
+    assert bboxes.shape[-1] % 4 == 0
+    w = img_shape[1]
+    flipped = bboxes.copy()
+    flipped[..., 0::4] = w - bboxes[..., 2::4] - 1
+    flipped[..., 2::4] = w - bboxes[..., 0::4] - 1
+    return flipped
+
+
 def imcrop(img, bboxes, scale_ratio=1.0, pad_fill=None):
     """Crop image patches.
 

mmcv/torchpack/runner/runner.py

@@ -201,7 +201,7 @@ class Runner(object):
         else:
             meta.update(epoch=self.epoch + 1, iter=self.iter)
 
-        filename = osp.join(out_dir, filename_tmpl.format(self.epoch))
+        filename = osp.join(out_dir, filename_tmpl.format(self.epoch + 1))
         linkname = osp.join(out_dir, 'latest.pth')
         optimizer = self.optimizer if save_optimizer else None
         save_checkpoint(self.model, filename, optimizer=optimizer, meta=meta)
@@ -213,7 +213,6 @@ class Runner(object):
         self.data_loader = data_loader
         self._max_iters = self._max_epochs * len(data_loader)
         self.call_hook('before_train_epoch')
-
         for i, data_batch in enumerate(data_loader):
             self._inner_iter = i
             self.call_hook('before_train_iter')
@@ -347,7 +346,16 @@ class Runner(object):
         if checkpoint_config is None:
             checkpoint_config = {}
         self.register_lr_hooks(lr_config)
+ 
+        if isinstance(grad_clip_config, Hook):
+            self.register_hook(grad_clip_config)
+        elif isinstance(grad_clip_config, dict):
             self.register_hook(OptimizerStepperHook(**grad_clip_config))
+        else:
+            raise TypeError(
+                "OptimizerStepperHook should be a Hook object or dict, not {}".
+                format(type(grad_clip_config)))
+
         self.register_hook(CheckpointSaverHook(**checkpoint_config))
         self.register_hook(IterTimerHook())
         if log_config is not None: