Update PP-LCNet docs

docs/en/ImageNet_models_en.md

@@ -59,10 +59,10 @@ Accuracy and inference time of the prtrained models based on SSLD distillation a
 * Note: `Reference Top-1 Acc` means accuracy of pretrained models which are trained on ImageNet1k dataset.
 
 
-<a name="PPLCNet_series"></a>
-### PPLCNet_series
+<a name="PP-LCNet_series"></a>
+### PP-LCNet_series
 
-Accuracy and inference time metrics of PPLCNet series models are shown as follows. More detailed information can be refered to [PPLCNet series tutorial](../en/models/PPLCNet_en.md).
+Accuracy and inference time metrics of PPLCNet series models are shown as follows. More detailed information can be refered to [PPLCNet series tutorial](../en/models/PP-LCNet_en.md).
 
 | Model           | Top-1 Acc | Top-5 Acc | Intel-Xeon-Gold-6148 time(ms)<br>bs=1 | FLOPs(M) | Params(M) | Download Address |
 |:--:|:--:|:--:|:--:|:--:|:--:|:--:|

docs/en/models/PP-LCNet_en.md

+# PPLCNet series
+
+## Overview
+
+The PPLCNet series is a network that has excellent performance on Intel-CPU proposed by the Baidu PaddleCV team. The author summarizes some methods that can improve the accuracy of the model on Intel-CPU but hardly increase the inference time. The author combines these methods into a new network, namely PPLCNet. Compared with other lightweight networks, PPLCNet can achieve higher accuracy with the same inference time. PPLCNet has shown strong competitiveness in image classification, object detection, and semantic segmentation.
+
+
+
+## Accuracy, FLOPS and Parameters
+
+| Models           | Top1 | Top5 | FLOPs<br>(M) | Parameters<br>(M) |
+|:--:|:--:|:--:|:--:|:--:|
+| PPLCNet_x0_25        |0.5186           | 0.7565           | 18    | 1.5  |
+| PPLCNet_x0_35        |0.5809           | 0.8083           | 29    | 1.6  |
+| PPLCNet_x0_5         |0.6314           | 0.8466           | 47    | 1.9  |
+| PPLCNet_x0_75        |0.6818           | 0.8830           | 99    | 2.4  |
+| PPLCNet_x1_0         |0.7132           | 0.9003           | 161   | 3.0  |
+| PPLCNet_x1_5         |0.7371           | 0.9153           | 342   | 4.5  |
+| PPLCNet_x2_0         |0.7518           | 0.9227           | 590   | 6.5  |
+| PPLCNet_x2_5         |0.7660           | 0.9300           | 906   | 9.0  |
+| PPLCNet_x0_5_ssld    |0.6610           | 0.8646           | 47    | 1.9  |
+| PPLCNet_x1_0_ssld    |0.7439           | 0.9209           | 161   | 3.0  |
+| PPLCNet_x2_5_ssld    |0.8082           | 0.9533           | 906   | 9.0  |
+
+
+
+## Inference speed based on Intel(R)-Xeon(R)-Gold-6148-CPU
+
+| Models                 | Crop Size | Resize Short Size | FP32<br>Batch Size=1<br>(ms) |
+|------------------|-----------|-------------------|--------------------------|
+| PPLCNet_x0_25        | 224       | 256               | 1.74                    |
+| PPLCNet_x0_35        | 224       | 256               | 1.92                    |
+| PPLCNet_x0_5         | 224       | 256               | 2.05                    |
+| PPLCNet_x0_75        | 224       | 256               | 2.29                    |
+| PPLCNet_x1_0         | 224       | 256               | 2.46                    |
+| PPLCNet_x1_5         | 224       | 256               | 3.19                    |
+| PPLCNet_x2_0         | 224       | 256               | 4.27                    |
+| PPLCNet_x2_5         | 224       | 256               | 5.39                    |
+| PPLCNet_x0_5_ssld    | 224       | 256               | 2.05                    |
+| PPLCNet_x1_0_ssld    | 224       | 256               | 2.46                    |
+| PPLCNet_x2_5_ssld    | 224       | 256               | 5.39                    |

docs/zh_CN/models/PP-LCNet.md

@@ -7,12 +7,12 @@
 ## 介绍
 
 近年来，有很多轻量级的骨干网络问世，尤其最近两年，各种NAS搜索出的网络层出不穷，这些网络要么主打FLOPs或者Params上的优势，要么主打ARM设备上的推理速度的优势，很少有网络专门针对Intel CPU做特定的优化，导致这些网络在Intel CPU端的推理速度并不是很完美。基于此，我们针对Intel CPU设备以及其加速库MKLDNN设计了特定的骨干网络PP-LCNet，比起其他的轻量级的SOTA模型，该骨干网络可以在不增加推理时间的情况下，进一步提升模型的性能，最终大幅度超越现有的SOTA模型。与其他模型的对比图如下。
-![](../../images/PP-LCNet/PP-LCNet-Acc.png)
+<img src="../../images/PP-LCNet/PP-LCNet-Acc.png" style="zoom:30%;" />
 
 ## 方法
 
 网络结构整体如下图所示。
-![](../../images/PP-LCNet/PP-LCNet.png)
+<img src="../../images/PP-LCNet/PP-LCNet.png" style="zoom:30%;" />
 
 我们经过大量的实验发现，在基于Intel CPU设备上，尤其当启用MKLDNN加速库后，很多看似不太耗时的操作反而会增加延时，比如elementwise-add操作、split-concat结构等。所以最终我们选用了结构尽可能精简、速度尽可能快的block组成我们的BaseNet（类似MobileNetV1）。基于BaseNet，我们通过实验，总结了四条几乎不增加延时但是可以提升模型精度的方法，融合这四条策略，我们组合成了PP-LCNet。下面对这四条策略一一介绍：