Update README.md

Author: anxiangsir

recognition/partial_fc/README.md

@@ -8,7 +8,7 @@ Partial FC is a distributed deep learning training framework for face recognitio
 [Partial FC](https://arxiv.org/abs/2010.05222)
 - [Largest Face Recognition Dataset: **Glint360k**](#Glint360K)
 - [Docker](#Docker)
-- [Performance On Million Identities](#Benchmark)  
+- [Performance On Million Identities](#Benchmark)
 - [FAQ](#FAQ)
 - [Citation](#Citation)
 
@@ -67,7 +67,7 @@ Use [unpack_glint360k.py](./unpack_glint360k.py) to unpack.
 - [x] [**Baidu Drive**](https://pan.baidu.com/s/1sd9ZRsV2c_dWHW84kz1P1Q) (code:befi)
 - [x] [**Google Drive**](https://drive.google.com/drive/folders/1WLjDzEs1wC1K1jxDHNJ7dhEmQ3rOOILl?usp=sharing)
 
-| Framework       |  backbone                      | sample_rate  | IJBC@e4 | IFRT@e6 |
+| Framework       |  backbone                      | negative class centers sample_rate  | IJBC@e4 | IFRT@e6 |
 |  :---           | :---                           | :---         |  :---   |  :---   | 
 |  mxnet   | [R100](https://drive.google.com/drive/folders/1YPqIkOZWrmbli4GWfMJO2b0yiiZ7UCsP?usp=sharing) |1.0|97.3|-|
 |  mxnet   | [R100](https://drive.google.com/drive/folders/1-gF5sDwNoRcjwmpPSTNLpaZJi5N91BvL?usp=sharing) |0.1|97.3|-|

recognition/partial_fc/mxnet/README.md

@@ -1,6 +1,4 @@
-## [中文版本请点击这里](./README_CN.md)
-
-## Train
+## Training
 ### 1.Requirements
 python==3.6  
 cuda==10.1    
@@ -90,5 +88,32 @@ sudo mkdir /train_tmp
 mount -t tmpfs -o size=140G  tmpfs /train_tmp
 ```
 
+## Our Method
+![Image text](https://github.com/nttstar/insightface-resources/blob/master/images/partial_fc.png)
+
+### 1. The classification layer model is parallel
+Class centers are evenly distributed across different GPUs. It only takes three communications to complete 
+loss-free Softmax calculations.
+
+#### 1. Synchronization of features
+Make sure each GPU has all the GPU features on it, as is shown in `AllGather(x_i)`.
+
+#### 2. Synchronization of denominator of the softmax function
+We can first calculate the local sum of each GPU, and then compute the global sum through communication, as is shown
+in `Allreduce(sum(exp(logits_i)))`
+
+#### 3. Synchronization the gradients of feature
+The gradient of logits can be calculated independently, so is the gradient of the feature. finally, we collect all the 
+gradients on GPU and send them back to backbone, as is shown in `Allreduce(deta(X))`
+
+### 2. Softmax approximate
+
+Just a subset of class centers can approximate softmax's computation(positive class centers must in these class centers),
+this can be done with the following code:
+```python
+centers_p = func_positive(label)                 # select the positive class centers by the label of the sample
+centers_n = func_negative(centers_p)             # negative class centers are randomly sampled after excluding positive classes
+centers_final = concat(centers_n, centers_p)     # class centers that participate in softmax calculations
+```
 
 

recognition/partial_fc/pytorch/README.md

@@ -1,15 +1,5 @@
 # Parital FC
 
-## TODO
-
-- [x] **No BUG** Sampling  
-- [ ] Pytorch Experiments (Glint360k, 1.0/0.1)   
-- [ ] Mixed precision training  
-- [ ] Pipeline Parallel  
-- [ ] Checkpoint  
-- [ ] Docker  
-- [ ] A Wonderful Documents  
-
 ## Results
 We employ ResNet100 as the backbone.