Merge pull request open-mmlab#257 from open-mmlab/pytorch-1.0

Support Pytorch 1.0

Author: hellock

INSTALL.md

@@ -4,13 +4,13 @@
 
 - Linux (tested on Ubuntu 16.04 and CentOS 7.2)
 - Python 3.4+
-- PyTorch 0.4.1
+- PyTorch 1.0
 - Cython
-- [mmcv](https://github.com/open-mmlab/mmcv)
+- [mmcv](https://github.com/open-mmlab/mmcv) >= 0.2.2
 
 ### Install mmdetection
 
-a. Install PyTorch 0.4.1 and torchvision following the [official instructions](https://pytorch.org/).
+a. Install PyTorch 1.0 and torchvision following the [official instructions](https://pytorch.org/).
 
 b. Clone the mmdetection repository.
 

MODEL_ZOO.md

@@ -10,11 +10,17 @@
 ### Software environment
 
 - Python 3.6 / 3.7
-- PyTorch 0.4.1
+- PyTorch 1.0
 - CUDA 9.0.176
 - CUDNN 7.0.4
 - NCCL 2.1.15
 
+Note: The train time was measured with PyTorch 0.4.1. We will update it later, which should be about 0.02s ~ 0.05s faster.
+
+## Mirror sites
+
+We use AWS as the main site to host our model zoo, and maintain a mirror on aliyun.
+You can replace `https://s3.ap-northeast-2.amazonaws.com` with `https://open-mmlab.oss-cn-beijing.aliyuncs.com` in model urls.
 
 ## Common settings
 

README.md

@@ -3,6 +3,9 @@
 
 ## Introduction
 
+The master branch works with **PyTorch 1.0**. If you would like to use PyTorch 0.4.1,
+please checkout to the [pytorch-0.4.1](https://github.com/open-mmlab/mmdetection/tree/pytorch-0.4.1) branch.
+
 mmdetection is an open source object detection toolbox based on PyTorch. It is
 a part of the open-mmlab project developed by [Multimedia Laboratory, CUHK](http://mmlab.ie.cuhk.edu.hk/).
 
@@ -36,6 +39,9 @@ This project is released under the [Apache 2.0 license](LICENSE).
 
 ## Updates
 
+v0.6rc0(06/02/2019)
+- Migrate to PyTorch 1.0.
+
 v0.5.7 (06/02/2019)
 - Add support for Deformable ConvNet v2. (Many thanks to the authors and [@chengdazhi](https://github.com/chengdazhi))
 - This is the last release based on PyTorch 0.4.1.

mmdet/core/loss/losses.py

@@ -34,13 +34,21 @@ def sigmoid_focal_loss(pred,
                        weight,
                        gamma=2.0,
                        alpha=0.25,
-                       reduction='elementwise_mean'):
+                       reduction='mean'):
     pred_sigmoid = pred.sigmoid()
     pt = (1 - pred_sigmoid) * target + pred_sigmoid * (1 - target)
     weight = (alpha * target + (1 - alpha) * (1 - target)) * weight
     weight = weight * pt.pow(gamma)
-    return F.binary_cross_entropy_with_logits(
-        pred, target, weight, reduction=reduction)
+    loss = F.binary_cross_entropy_with_logits(
+        pred, target, reduction='none') * weight
+    reduction_enum = F._Reduction.get_enum(reduction)
+    # none: 0, mean:1, sum: 2
+    if reduction_enum == 0:
+        return loss
+    elif reduction_enum == 1:
+        return loss.mean()
+    elif reduction_enum == 2:
+        return loss.sum()
 
 
 def weighted_sigmoid_focal_loss(pred,
@@ -62,22 +70,22 @@ def mask_cross_entropy(pred, target, label):
     inds = torch.arange(0, num_rois, dtype=torch.long, device=pred.device)
     pred_slice = pred[inds, label].squeeze(1)
     return F.binary_cross_entropy_with_logits(
-        pred_slice, target, reduction='elementwise_mean')[None]
+        pred_slice, target, reduction='mean')[None]
 
 
-def smooth_l1_loss(pred, target, beta=1.0, reduction='elementwise_mean'):
+def smooth_l1_loss(pred, target, beta=1.0, reduction='mean'):
     assert beta > 0
     assert pred.size() == target.size() and target.numel() > 0
     diff = torch.abs(pred - target)
     loss = torch.where(diff < beta, 0.5 * diff * diff / beta,
                        diff - 0.5 * beta)
-    reduction = F._Reduction.get_enum(reduction)
-    # none: 0, elementwise_mean:1, sum: 2
-    if reduction == 0:
+    reduction_enum = F._Reduction.get_enum(reduction)
+    # none: 0, mean:1, sum: 2
+    if reduction_enum == 0:
         return loss
-    elif reduction == 1:
+    elif reduction_enum == 1:
         return loss.sum() / pred.numel()
-    elif reduction == 2:
+    elif reduction_enum == 2:
         return loss.sum()