updated instance discrimination model with Softmax-CE and 65536 negatives

zhirongw · zhirongw · commit f7cfe298357c · 2020-08-24T19:12:14.000+08:00
diff --git a/README.md b/README.md
@@ -2,30 +2,28 @@
 
 This repo constains the pytorch implementation for the CVPR2018 unsupervised learning paper [(arxiv)](https://arxiv.org/pdf/1805.01978.pdf).
 
-```
-@inproceedings{wu2018unsupervised,
-  title={Unsupervised Feature Learning via Non-Parametric Instance Discrimination},
-  author={Wu, Zhirong and Xiong, Yuanjun and Stella, X Yu and Lin, Dahua},
-  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
-  year={2018}
-}
-```
+## Updated Pretrained Model
 
-## Highlight
+An updated instance discrimination model with memory bank implementation and with nce-k=65536 negatives is provided.
+The updated model is trained with Softmax-CE loss as in CPC/MoCo instead of the original NCE loss.
 
-- We formulate unsupervised learning from a completely different non-parametric perspective.
-- Feature encodings can be as compact as 128 dimension for each image.
-- Enjoys the benefit of advanced architectures and techniques from supervised learning.
-- Runs seamlessly with nearest neighbor classifiers.
+- [ResNet 50](https://frontiers.blob.core.windows.net/pretraining/checkpoints/pil_pretrained_models/lemniscate/lemniscate_resnet50_update.pth) (Linear ImageNet Acc 58.5%)
 
-## Pretrained Model
 
-Currently, we provide pretrained models of ResNet 18 and ResNet 50. 
+**Oldies**: original releases of ResNet18 and ResNet50 trained with 4096 negatives and the NCE loss.
 Each tar ball contains the feature representation of all ImageNet training images (600 mb) and model weights (100-200mb).
 You can also get these representations by forwarding the network for the entire ImageNet images.
 
-- [ResNet 18](http://zhirongw.westus2.cloudapp.azure.com/models/lemniscate_resnet18.pth.tar) (top 1 accuracy 41.0%)
-- [ResNet 50](http://zhirongw.westus2.cloudapp.azure.com/models/lemniscate_resnet50.pth.tar) (top 1 accuracy 46.8%)
+- [ResNet 18](https://frontiers.blob.core.windows.net/pretraining/checkpoints/pil_pretrained_models/lemniscate/lemniscate_resnet18.pth) (top 1 nearest neighbor accuracy 41.0%)
+- [ResNet 50](https://frontiers.blob.core.windows.net/pretraining/checkpoints/pil_pretrained_models/lemniscate/lemniscate_resnet50.pth) (top 1 nearest neighbor accuracy 46.8%)
+
+
+## Highlight
+
+- We formulate unsupervised learning from a completely different non-parametric perspective.
+- Feature encodings can be as compact as 128 dimension for each image.
+- Enjoys the benefit of advanced architectures and techniques from supervised learning.
+- Runs seamlessly with nearest neighbor classifiers.
 
 ## Nearest Neighbor
 
@@ -65,6 +63,17 @@ Please refer to the official repo for details of data preparation and hardware c
   `python cifar.py --nce-k 0 --nce-t 0.1 --lr 0.03`
 
 
+## Citation
+
+```
+@inproceedings{wu2018unsupervised,
+  title={Unsupervised Feature Learning via Non-Parametric Instance Discrimination},
+  author={Wu, Zhirong and Xiong, Yuanjun and Stella, X Yu and Lin, Dahua},
+  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
+  year={2018}
+}
+```
+
 ## Contact
 
 For any questions, please feel free to reach 
