update

docs/zh_CN/models/ImageNet1k/MobileViTv3.md

+# MobileviTv3
+-----
+
+## 目录
+
+- [1. 模型介绍](#1)
+    - [1.1 模型简介](#1.1)
+    - [1.2 模型指标](#1.2)
+- [2. 模型快速体验](#2)
+- [3. 模型训练、评估和预测](#3)
+- [4. 模型推理部署](#4)
+  - [4.1 推理模型准备](#4.1)
+  - [4.2 基于 Python 预测引擎推理](#4.2)
+  - [4.3 基于 C++ 预测引擎推理](#4.3)
+  - [4.4 服务化部署](#4.4)
+  - [4.5 端侧部署](#4.5)
+  - [4.6 Paddle2ONNX 模型转换与预测](#4.6)
+
+<a name='1'></a>
+
+## 1. 模型介绍
+
+<a name='1.1'></a>
+
+### 1.1 模型简介
+
+MobileViTv3 是一个结合 CNN 和 ViT 的轻量级模型，用于移动视觉任务。通过 MobileViTv3-block 解决了 MobileViTv1 的扩展问题并简化了学习任务，从而得倒了 MobileViTv3-XXS、XS 和 S 模型，在 ImageNet-1k、ADE20K、COCO 和 PascalVOC2012 数据集上表现优于 MobileViTv1。
+通过将提出的融合块添加到 MobileViTv2 中，创建 MobileViTv3-0.5、0.75 和 1.0 模型，在ImageNet-1k、ADE20K、COCO和PascalVOC2012数据集上给出了比 MobileViTv2 更好的准确性数据。[论文地址](https://arxiv.org/abs/2209.15159)。
+
+<a name='1.2'></a>
+
+### 1.2 模型指标
+
+| Models           | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) |
+|:--:|:--:|:--:|:--:|:--:|:--:|:--:|
+| MobileViTv3_XXS    | 0.7087 | 0.8976 | 0.7098 | - |  289.02 | 1.25 |
+| MobileViTv3_XS     | 0.7663 | 0.9332 | 0.7671 | - |  926.98 | 2.49 |
+| MobileViTv3_S      | 0.7928 | 0.9454 | 0.7930 | - | 1841.39 | 5.76 |
+| MobileViTv3_XXS_L2 | 0.7028 | 0.8942 | 0.7023 | - |  256.97 | 1.15 |
+| MobileViTv3_XS_L2  | 0.7607 | 0.9300 | 0.7610 | - |  852.82 | 2.26 |
+| MobileViTv3_S_L2   | 0.7907 | 0.9440 | 0.7906 | - | 1651.96 | 5.17 |
+| MobileViTv3_x0_5   | 0.7200 | 0.9083 | 0.7233 | - |  481.33 | 1.43 |
+| MobileViTv3_x0_75  | 0.7626 | 0.9308 | 0.7655 | - | 1064.48 | 3.00 |
+| MobileViTv3_x1_0   | 0.7838 | 0.9421 | 0.7864 | - | 1875.96 | 5.14 |
+
+**备注：** PaddleClas 所提供的该系列模型的预训练模型权重，均是基于其官方提供的权重转得。
+
+<a name="2"></a>  
+
+## 2. 模型快速体验
+
+安装 paddlepaddle 和 paddleclas 即可快速对图片进行预测，体验方法可以参考[ResNet50 模型快速体验](./ResNet.md#2)。
+
+<a name="3"></a>
+
+## 3. 模型训练、评估和预测
+
+此部分内容包括训练环境配置、ImageNet数据的准备、该模型在 ImageNet 上的训练、评估、预测等内容。在 `ppcls/configs/ImageNet/MobileViTv3/` 中提供了该模型的训练配置，启动训练方法可以参考：[ResNet50 模型训练、评估和预测](./ResNet.md#3-模型训练评估和预测)。
+
+**备注：** 由于 MobileViT 系列模型默认使用的 GPU 数量为 8 个，所以在训练时，需要指定8个GPU，如`python3 -m paddle.distributed.launch --gpus="0,1,2,3,4,5,6,7" tools/train.py -c xxx.yaml`, 如果使用 4 个 GPU 训练，默认学习率需要减小一半，精度可能有损。
+
+<a name="4"></a>
+
+## 4. 模型推理部署
+
+<a name="4.1"></a>
+
+### 4.1 推理模型准备
+
+Paddle Inference 是飞桨的原生推理库， 作用于服务器端和云端，提供高性能的推理能力。相比于直接基于预训练模型进行预测，Paddle Inference可使用 MKLDNN、CUDNN、TensorRT 进行预测加速，从而实现更优的推理性能。更多关于Paddle Inference推理引擎的介绍，可以参考[Paddle Inference官网教程](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/infer/inference/inference_cn.html)。
+
+Inference 的获取可以参考 [ResNet50 推理模型准备](./ResNet.md#4.1) 。
+
+<a name="4.2"></a>
+
+### 4.2 基于 Python 预测引擎推理
+
+PaddleClas 提供了基于 python 预测引擎推理的示例。您可以参考[ResNet50 基于 Python 预测引擎推理](./ResNet.md#4.2) 完成模型的推理预测。
+
+<a name="4.3"></a>
+
+### 4.3 基于 C++ 预测引擎推理
+
+PaddleClas 提供了基于 C++ 预测引擎推理的示例，您可以参考[服务器端 C++ 预测](../../deployment/image_classification/cpp/linux.md)来完成相应的推理部署。如果您使用的是 Windows 平台，可以参考[基于 Visual Studio 2019 Community CMake 编译指南](../../deployment/image_classification/cpp/windows.md)完成相应的预测库编译和模型预测工作。
+
+<a name="4.4"></a>
+
+### 4.4 服务化部署
+
+Paddle Serving 提供高性能、灵活易用的工业级在线推理服务。Paddle Serving 支持 RESTful、gRPC、bRPC 等多种协议，提供多种异构硬件和多种操作系统环境下推理解决方案。更多关于Paddle Serving 的介绍，可以参考[Paddle Serving 代码仓库](https://github.com/PaddlePaddle/Serving)。
+
+PaddleClas 提供了基于 Paddle Serving 来完成模型服务化部署的示例，您可以参考[模型服务化部署](../../deployment/image_classification/paddle_serving.md)来完成相应的部署工作。
+
+<a name="4.5"></a>
+
+### 4.5 端侧部署
+
+Paddle Lite 是一个高性能、轻量级、灵活性强且易于扩展的深度学习推理框架，定位于支持包括移动端、嵌入式以及服务器端在内的多硬件平台。更多关于 Paddle Lite 的介绍，可以参考[Paddle Lite 代码仓库](https://github.com/PaddlePaddle/Paddle-Lite)。
+
+PaddleClas 提供了基于 Paddle Lite 来完成模型端侧部署的示例，您可以参考[端侧部署](../../deployment/image_classification/paddle_lite.md)来完成相应的部署工作。
+
+<a name="4.6"></a>
+
+### 4.6 Paddle2ONNX 模型转换与预测
+
+Paddle2ONNX 支持将 PaddlePaddle 模型格式转化到 ONNX 模型格式。通过 ONNX 可以完成将 Paddle 模型到多种推理引擎的部署，包括TensorRT/OpenVINO/MNN/TNN/NCNN，以及其它对 ONNX 开源格式进行支持的推理引擎或硬件。更多关于 Paddle2ONNX 的介绍，可以参考[Paddle2ONNX 代码仓库](https://github.com/PaddlePaddle/Paddle2ONNX)。
+
+PaddleClas 提供了基于 Paddle2ONNX 来完成 inference 模型转换 ONNX 模型并作推理预测的示例，您可以参考[Paddle2ONNX 模型转换与预测](../../deployment/image_classification/paddle2onnx.md)来完成相应的部署工作。

docs/zh_CN/models/ImageNet1k/README.md

@@ -796,15 +796,24 @@ DeiT（Data-efficient Image Transformers）系列模型的精度、速度指标
 
 <a name="MobileViT"></a>
 
-## MobileViT 系列 <sup>[[42](#ref42)]</sup>
+## MobileViT 系列 <sup>[[42](#ref42)][[51](#ref51)]</sup>
 
-关于 MobileViT 系列模型的精度、速度指标如下表所示，更多介绍可以参考：[MobileViT 系列模型文档](MobileViT.md)。
+关于 MobileViT 系列模型的精度、速度指标如下表所示，更多介绍可以参考：[MobileViT 系列模型文档](MobileViT.md), [MobileViTv3 系列模型文档](MobileViTv3.md)。
 
 | 模型       | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(M) | Params(M) | 预训练模型下载地址                                               | inference模型下载地址                                      |
 | ---------- | --------- | --------- | ---------------- | ---------------- | -------- | --------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ |
 |  MobileViT_XXS     | 0.6867 | 0.8878 | - | - | - |  337.24 | 1.28 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/MobileViT_XXS_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/MobileViT_XXS_infer.tar) |
 |  MobileViT_XS      | 0.7454 | 0.9227 | - | - | - |  930.75 | 2.33 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/MobileViT_XS_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/MobileViT_XS_infer.tar) |
 |  MobileViT_S       | 0.7814 | 0.9413 | - | - | - | 1849.35 | 5.59 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/MobileViT_S_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/MobileViT_S_infer.tar) |
+| MobileViTv3_XXS    | 0.7087 | 0.8976 | - | - | - |  289.02 | 1.25 | [下载链接]() | [下载链接]() |
+| MobileViTv3_XS     | 0.7663 | 0.9332 | - | - | - |  926.98 | 2.49 | [下载链接]() | [下载链接]() |
+| MobileViTv3_S      | 0.7928 | 0.9454 | - | - | - | 1841.39 | 5.76 | [下载链接]() | [下载链接]() |
+| MobileViTv3_XXS_L2 | 0.7028 | 0.8942 | - | - | - |  256.97 | 1.15 | [下载链接]() | [下载链接]() |
+| MobileViTv3_XS_L2  | 0.7607 | 0.9300 | - | - | - |  852.82 | 2.26 | [下载链接]() | [下载链接]() |
+| MobileViTv3_S_L2   | 0.7907 | 0.9440 | - | - | - | 1651.96 | 5.17 | [下载链接]() | [下载链接]() |
+| MobileViTv3_x0_5   | 0.7200 | 0.9083 | - | - | - |  481.33 | 1.43 | [下载链接]() | [下载链接]() |
+| MobileViTv3_x0_75  | 0.7626 | 0.9308 | - | - | - | 1064.48 | 3.00 | [下载链接]() | [下载链接]() |
+| MobileViTv3_x1_0   | 0.7838 | 0.9421 | - | - | - | 1875.96 | 5.14 | [下载链接]() | [下载链接]() |
 
 <a name='reference'></a>
 
@@ -910,3 +919,5 @@ TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE.
 <a name="ref49">[49]</a>Mingyuan Mao, Renrui Zhang, Honghui Zheng, Peng Gao, Teli Ma, Yan Peng, Errui Ding, Baochang Zhang, Shumin Han. Dual-stream Network for Visual Recognition.
 
 <a name="ref50">[50]</a>Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, Baining Guo. Swin Transformer V2: Scaling Up Capacity and Resolution
+
+<a name="ref50">[51]</a>Wadekar, Shakti N. and Chaurasia, Abhishek. MobileViTv3: Mobile-Friendly Vision Transformer with Simple and Effective Fusion of Local, Global and Input Features

ppcls/arch/backbone/__init__.py

@@ -78,7 +78,7 @@ from .model_zoo.cae import cae_base_patch16_224, cae_large_patch16_224
 from .model_zoo.cvt import CvT_13_224, CvT_13_384, CvT_21_224, CvT_21_384, CvT_W24_384
 from .model_zoo.micronet import MicroNet_M0, MicroNet_M1, MicroNet_M2, MicroNet_M3
 from .model_zoo.mobilenext import MobileNeXt_x0_35, MobileNeXt_x0_5, MobileNeXt_x0_75, MobileNeXt_x1_0, MobileNeXt_x1_4
-from .model_zoo.mobilevit_v3 import MobileViTv3_XXS, MobileViTv3_XS, MobileViTv3_S, MobileViTv3_x0_5, MobileViTv3_x0_75, MobileViTv3_x1_0
+from .model_zoo.mobilevit_v3 import MobileViTv3_XXS, MobileViTv3_XS, MobileViTv3_S, MobileViTv3_XXS_L2, MobileViTv3_XS_L2, MobileViTv3_S_L2, MobileViTv3_x0_5, MobileViTv3_x0_75, MobileViTv3_x1_0
 
 from .variant_models.resnet_variant import ResNet50_last_stage_stride1
 from .variant_models.resnet_variant import ResNet50_adaptive_max_pool2d

ppcls/arch/backbone/model_zoo/mobilevit_v3.py

@@ -29,6 +29,9 @@ MODEL_URLS = {
     "MobileViTv3_XXS": "",
     "MobileViTv3_XS": "",
     "MobileViTv3_S": "",
+    "MobileViTv3_XXS_L2": "",
+    "MobileViTv3_XS_L2": "",
+    "MobileViTv3_S_L2": "",
     "MobileViTv3_x0_5": "",
     "MobileViTv3_x0_75": "",
     "MobileViTv3_x1_0": "",
@@ -319,7 +322,7 @@ class MobileViTv3Block(nn.Layer):
         transposed_fm = reshaped_fm.transpose([0, 3, 2, 1])
         # [B, P, N, C] --> [BP, N, C]
         patches = transposed_fm.reshape(
-            [batch_size * patch_area, num_patches, -1])
+            [batch_size * patch_area, num_patches, in_channels])
 
         info_dict = {
             "orig_size": (orig_h, orig_w),
@@ -339,7 +342,7 @@ class MobileViTv3Block(nn.Layer):
         # [BP, N, C] --> [B, P, N, C]
         patches = patches.reshape([
             info_dict["batch_size"], self.patch_area,
-            info_dict["total_patches"], -1
+            info_dict["total_patches"], patches.shape[2]
         ])
 
         batch_size, pixels, num_patches, channels = patches.shape
@@ -399,6 +402,7 @@ class MobileViTv3Block(nn.Layer):
 class LinearSelfAttention(nn.Layer):
     def __init__(self, embed_dim, attn_dropout=0.0, bias=True):
         super().__init__()
+        self.embed_dim = embed_dim
         self.qkv_proj = nn.Conv2D(
             embed_dim, 1 + (2 * embed_dim), 1, bias_attr=bias)
         self.attn_dropout = nn.Dropout(p=attn_dropout)
@@ -428,7 +432,7 @@ class LinearSelfAttention(nn.Layer):
 
         # combine context vector with values
         # [B, d, P, N] * [B, d, P, 1] --> [B, d, P, N]
-        out = F.relu(value) * context_vector.expand_as(value)
+        out = F.relu(value) * context_vector
         out = self.out_proj(out)
         return out
 
@@ -552,10 +556,11 @@ class MobileViTv3Block_v2(nn.Layer):
         # [B, C, H, W] --> [B, C, P, N]
         patches = F.unfold(
             feature_map,
-            kernel_sizes=(self.patch_h, self.patch_w),
-            stride=(self.patch_h, self.patch_w))
+            kernel_sizes=[self.patch_h, self.patch_w],
+            strides=[self.patch_h, self.patch_w])
+        n_patches = img_h * img_w // (self.patch_h * self.patch_w)
         patches = patches.reshape(
-            [batch_size, in_channels, self.patch_h * self.patch_w, -1])
+            [batch_size, in_channels, self.patch_h * self.patch_w, n_patches])
 
         return patches, (img_h, img_w)
 
@@ -567,9 +572,9 @@ class MobileViTv3Block_v2(nn.Layer):
 
         feature_map = F.fold(
             patches,
-            output_sizes=output_size,
-            kernel_sizes=(self.patch_h, self.patch_w),
-            stride=(self.patch_h, self.patch_w))
+            output_size,
+            kernel_sizes=[self.patch_h, self.patch_w],
+            strides=[self.patch_h, self.patch_w])
 
         return feature_map
 
@@ -656,16 +661,16 @@ class MobileViTv3(nn.Layer):
             cfg=mobilevit_config["layer5"],
             dilate=dilate_l5)
 
-        in_channels = out_channels
-        exp_channels = min(mobilevit_config["last_layer_exp_factor"] *
-                           in_channels, 960)
         if self.mobilevit_v2_based:
             self.conv_1x1_exp = nn.Identity()
         else:
+            in_channels = out_channels
+            out_channels = min(mobilevit_config["last_layer_exp_factor"] *
+                               in_channels, 960)
             self.conv_1x1_exp = nn.Sequential(
                 ('conv', nn.Conv2D(
-                    in_channels, exp_channels, 1, bias_attr=False)),
-                ('norm', nn.BatchNorm2D(exp_channels)), ('act', nn.Silu()))
+                    in_channels, out_channels, 1, bias_attr=False)),
+                ('norm', nn.BatchNorm2D(out_channels)), ('act', nn.Silu()))
 
         self.classifier = nn.Sequential()
         self.classifier.add_sublayer(
@@ -675,7 +680,7 @@ class MobileViTv3(nn.Layer):
             self.classifier.add_sublayer(
                 name="dropout", sublayer=nn.Dropout(p=classifier_dropout))
         self.classifier.add_sublayer(
-            name="fc", sublayer=nn.Linear(exp_channels, class_num))
+            name="fc", sublayer=nn.Linear(out_channels, class_num))
 
         # weight initialization
         self.apply(self._init_weights)
@@ -1022,6 +1027,216 @@ def MobileViTv3_XXS(pretrained=False, use_ssld=False, **kwargs):
     return model
 
 
+def MobileViTv3_S_L2(pretrained=False, use_ssld=False, **kwargs):
+    mv2_exp_mult = 4
+    mobilevit_config = {
+        "layer0": {
+            "img_channels": 3,
+            "out_channels": 16,
+        },
+        "layer1": {
+            "out_channels": 32,
+            "expand_ratio": mv2_exp_mult,
+            "num_blocks": 1,
+            "stride": 1,
+            "block_type": "mv2"
+        },
+        "layer2": {
+            "out_channels": 64,
+            "expand_ratio": mv2_exp_mult,
+            "num_blocks": 3,
+            "stride": 2,
+            "block_type": "mv2"
+        },
+        "layer3": {  # 28x28
+            "out_channels": 128,
+            "transformer_channels": 144,
+            "ffn_dim": 288,
+            "transformer_blocks": 2,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "layer4": {  # 14x14
+            "out_channels": 256,
+            "transformer_channels": 192,
+            "ffn_dim": 384,
+            "transformer_blocks": 2,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "layer5": {  # 7x7
+            "out_channels": 320,
+            "transformer_channels": 240,
+            "ffn_dim": 480,
+            "transformer_blocks": 3,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "last_layer_exp_factor": 4
+    }
+
+    model = MobileViTv3(mobilevit_config, **kwargs)
+
+    _load_pretrained(
+        pretrained, model, MODEL_URLS["MobileViTv3_S_L2"], use_ssld=use_ssld)
+    return model
+
+
+def MobileViTv3_XS_L2(pretrained=False, use_ssld=False, **kwargs):
+    mv2_exp_mult = 4
+    mobilevit_config = {
+        "layer0": {
+            "img_channels": 3,
+            "out_channels": 16,
+        },
+        "layer1": {
+            "out_channels": 32,
+            "expand_ratio": mv2_exp_mult,
+            "num_blocks": 1,
+            "stride": 1,
+            "block_type": "mv2"
+        },
+        "layer2": {
+            "out_channels": 48,
+            "expand_ratio": mv2_exp_mult,
+            "num_blocks": 3,
+            "stride": 2,
+            "block_type": "mv2"
+        },
+        "layer3": {  # 28x28
+            "out_channels": 96,
+            "transformer_channels": 96,
+            "ffn_dim": 192,
+            "transformer_blocks": 2,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "layer4": {  # 14x14
+            "out_channels": 160,
+            "transformer_channels": 120,
+            "ffn_dim": 240,
+            "transformer_blocks": 2,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "layer5": {  # 7x7
+            "out_channels": 160,
+            "transformer_channels": 144,
+            "ffn_dim": 288,
+            "transformer_blocks": 3,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "last_layer_exp_factor": 4
+    }
+
+    model = MobileViTv3(mobilevit_config, **kwargs)
+
+    _load_pretrained(
+        pretrained, model, MODEL_URLS["MobileViTv3_XS_L2"], use_ssld=use_ssld)
+    return model
+
+
+def MobileViTv3_XXS_L2(pretrained=False, use_ssld=False, **kwargs):
+    mv2_exp_mult = 2
+    mobilevit_config = {
+        "layer0": {
+            "img_channels": 3,
+            "out_channels": 16,
+        },
+        "layer1": {
+            "out_channels": 16,
+            "expand_ratio": mv2_exp_mult,
+            "num_blocks": 1,
+            "stride": 1,
+            "block_type": "mv2"
+        },
+        "layer2": {
+            "out_channels": 24,
+            "expand_ratio": mv2_exp_mult,
+            "num_blocks": 3,
+            "stride": 2,
+            "block_type": "mv2"
+        },
+        "layer3": {  # 28x28
+            "out_channels": 64,
+            "transformer_channels": 64,
+            "ffn_dim": 128,
+            "transformer_blocks": 2,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "layer4": {  # 14x14
+            "out_channels": 80,
+            "transformer_channels": 80,
+            "ffn_dim": 160,
+            "transformer_blocks": 2,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "layer5": {  # 7x7
+            "out_channels": 128,
+            "transformer_channels": 96,
+            "ffn_dim": 192,
+            "transformer_blocks": 3,
+            "patch_h": 2,
+            "patch_w": 2,
+            "stride": 2,
+            "mv_expand_ratio": mv2_exp_mult,
+            "head_dim": None,
+            "num_heads": 4,
+            "block_type": "mobilevit"
+        },
+        "last_layer_exp_factor": 4
+    }
+
+    model = MobileViTv3(mobilevit_config, **kwargs)
+
+    _load_pretrained(
+        pretrained, model, MODEL_URLS["MobileViTv3_XXS_L2"], use_ssld=use_ssld)
+    return model
+
+
 def MobileViTv3_x1_0(pretrained=False, use_ssld=False, **kwargs):
     mobilevit_config = {
         "layer0": {

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_S.yaml

@@ -99,7 +99,7 @@ DataLoader:
       cls_label_path: ./dataset/ILSVRC2012/val_list.txt
       transform_ops:
         - DecodeImage:
-            to_rgb: False
+            to_rgb: True
             channel_first: False
         - ResizeImage:
             resize_short: 288

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_S_L2.yaml

+# global configs
+Global:
+  checkpoints: null
+  pretrained_model: null
+  output_dir: ./output/
+  device: gpu
+  save_interval: 1
+  eval_during_train: True
+  eval_interval: 1
+  epochs: 300
+  print_batch_step: 10
+  use_visualdl: False
+  # used for static mode and model export
+  image_shape: [3, 256, 256]
+  save_inference_dir: ./inference
+  use_dali: False
+
+# mixed precision training
+AMP:
+  scale_loss: 65536
+  use_dynamic_loss_scaling: True
+  # O1: mixed fp16
+  level: O1
+
+# model ema
+EMA:
+  decay: 0.9995
+
+# model architecture
+Arch:
+  name: MobileViTv3_S_L2
+  class_num: 1000
+  dropout: 0.1
+
+# loss function config for traing/eval process
+Loss:
+  Train:
+    - CELoss:
+        weight: 1.0
+        epsilon: 0.1
+  Eval:
+    - CELoss:
+        weight: 1.0
+
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.999
+  epsilon: 1e-8
+  weight_decay: 0.01
+  lr:
+    # for 8 cards
+    name: Cosine
+    learning_rate: 0.002
+    eta_min: 0.0002
+    warmup_epoch: 1  # 3000 iterations
+    warmup_start_lr: 0.0002
+    # by_epoch: True
+
+# data loader for train and eval
+DataLoader:
+  Train:
+    dataset:
+      name: MultiScaleDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/train_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+        - RandCropImage:
+            size: 256
+            interpolation: bilinear
+            use_log_aspect: True
+        - RandFlipImage:
+            flip_code: 1
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+    # support to specify width and height respectively:
+    # scales: [(256,256) (160,160), (192,192), (224,224) (288,288) (320,320)]
+    sampler:
+      name: MultiScaleSampler
+      scales: [256, 160, 192, 224, 288, 320]
+      # first_bs: batch size for the first image resolution in the scales list
+      # divide_factor: to ensure the width and height dimensions can be devided by downsampling multiple
+      first_bs: 48
+      divided_factor: 32
+      is_training: True
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+  Eval:
+    dataset: 
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/val_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+        - ResizeImage:
+            resize_short: 288
+            interpolation: bilinear
+        - CropImage:
+            size: 256
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 48
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+Infer:
+  infer_imgs: docs/images/inference_deployment/whl_demo.jpg
+  batch_size: 10
+  transforms:
+    - DecodeImage:
+        to_rgb: True
+        channel_first: False
+    - ResizeImage:
+        resize_short: 288
+        interpolation: bilinear
+    - CropImage:
+        size: 256
+    - NormalizeImage:
+        scale: 1.0/255.0
+        mean: [0.0, 0.0, 0.0]
+        std: [1.0, 1.0, 1.0]
+        order: ''
+    - ToCHWImage:
+  PostProcess:
+    name: Topk
+    topk: 5
+    class_id_map_file: ppcls/utils/imagenet1k_label_list.txt
+
+Metric:
+  Train:
+    - TopkAcc:
+        topk: [1, 5]
+  Eval:
+    - TopkAcc:
+        topk: [1, 5]

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_XS.yaml

@@ -99,7 +99,7 @@ DataLoader:
       cls_label_path: ./dataset/ILSVRC2012/val_list.txt
       transform_ops:
         - DecodeImage:
-            to_rgb: False
+            to_rgb: True
             channel_first: False
         - ResizeImage:
             resize_short: 288

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_XS_L2.yaml

+# global configs
+Global:
+  checkpoints: null
+  pretrained_model: null
+  output_dir: ./output/
+  device: gpu
+  save_interval: 1
+  eval_during_train: True
+  eval_interval: 1
+  epochs: 300
+  print_batch_step: 10
+  use_visualdl: False
+  # used for static mode and model export
+  image_shape: [3, 256, 256]
+  save_inference_dir: ./inference
+  use_dali: False
+
+# mixed precision training
+AMP:
+  scale_loss: 65536
+  use_dynamic_loss_scaling: True
+  # O1: mixed fp16
+  level: O1
+
+# model ema
+EMA:
+  decay: 0.9995
+
+# model architecture
+Arch:
+  name: MobileViTv3_XS_L2
+  class_num: 1000
+  dropout: 0.1
+
+# loss function config for traing/eval process
+Loss:
+  Train:
+    - CELoss:
+        weight: 1.0
+        epsilon: 0.1
+  Eval:
+    - CELoss:
+        weight: 1.0
+
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.999
+  epsilon: 1e-8
+  weight_decay: 0.01
+  lr:
+    # for 8 cards
+    name: Cosine
+    learning_rate: 0.002
+    eta_min: 0.0002
+    warmup_epoch: 1  # 3000 iterations
+    warmup_start_lr: 0.0002
+    # by_epoch: True
+
+# data loader for train and eval
+DataLoader:
+  Train:
+    dataset:
+      name: MultiScaleDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/train_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+        - RandCropImage:
+            size: 256
+            interpolation: bilinear
+            use_log_aspect: True
+        - RandFlipImage:
+            flip_code: 1
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+    # support to specify width and height respectively:
+    # scales: [(256,256) (160,160), (192,192), (224,224) (288,288) (320,320)]
+    sampler:
+      name: MultiScaleSampler
+      scales: [256, 160, 192, 224, 288, 320]
+      # first_bs: batch size for the first image resolution in the scales list
+      # divide_factor: to ensure the width and height dimensions can be devided by downsampling multiple
+      first_bs: 48
+      divided_factor: 32
+      is_training: True
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+  Eval:
+    dataset: 
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/val_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+        - ResizeImage:
+            resize_short: 288
+            interpolation: bilinear
+        - CropImage:
+            size: 256
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 48
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+Infer:
+  infer_imgs: docs/images/inference_deployment/whl_demo.jpg
+  batch_size: 10
+  transforms:
+    - DecodeImage:
+        to_rgb: True
+        channel_first: False
+    - ResizeImage:
+        resize_short: 288
+        interpolation: bilinear
+    - CropImage:
+        size: 256
+    - NormalizeImage:
+        scale: 1.0/255.0
+        mean: [0.0, 0.0, 0.0]
+        std: [1.0, 1.0, 1.0]
+        order: ''
+    - ToCHWImage:
+  PostProcess:
+    name: Topk
+    topk: 5
+    class_id_map_file: ppcls/utils/imagenet1k_label_list.txt
+
+Metric:
+  Train:
+    - TopkAcc:
+        topk: [1, 5]
+  Eval:
+    - TopkAcc:
+        topk: [1, 5]

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_XXS.yaml

@@ -99,7 +99,7 @@ DataLoader:
       cls_label_path: ./dataset/ILSVRC2012/val_list.txt
       transform_ops:
         - DecodeImage:
-            to_rgb: False
+            to_rgb: True
             channel_first: False
         - ResizeImage:
             resize_short: 288

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_XXS_L2.yaml

+# global configs
+Global:
+  checkpoints: null
+  pretrained_model: null
+  output_dir: ./output/
+  device: gpu
+  save_interval: 1
+  eval_during_train: True
+  eval_interval: 1
+  epochs: 300
+  print_batch_step: 10
+  use_visualdl: False
+  # used for static mode and model export
+  image_shape: [3, 256, 256]
+  save_inference_dir: ./inference
+  use_dali: False
+
+# mixed precision training
+AMP:
+  scale_loss: 65536
+  use_dynamic_loss_scaling: True
+  # O1: mixed fp16
+  level: O1
+
+# model ema
+EMA:
+  decay: 0.9995
+
+# model architecture
+Arch:
+  name: MobileViTv3_XXS_L2
+  class_num: 1000
+  dropout: 0.1
+
+# loss function config for traing/eval process
+Loss:
+  Train:
+    - CELoss:
+        weight: 1.0
+        epsilon: 0.1
+  Eval:
+    - CELoss:
+        weight: 1.0
+
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.999
+  epsilon: 1e-8
+  weight_decay: 0.01
+  lr:
+    # for 8 cards
+    name: Cosine
+    learning_rate: 0.002
+    eta_min: 0.0002
+    warmup_epoch: 1  # 3000 iterations
+    warmup_start_lr: 0.0002
+    # by_epoch: True
+
+# data loader for train and eval
+DataLoader:
+  Train:
+    dataset:
+      name: MultiScaleDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/train_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+        - RandCropImage:
+            size: 256
+            interpolation: bilinear
+            use_log_aspect: True
+        - RandFlipImage:
+            flip_code: 1
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+    # support to specify width and height respectively:
+    # scales: [(256,256) (160,160), (192,192), (224,224) (288,288) (320,320)]
+    sampler:
+      name: MultiScaleSampler
+      scales: [256, 160, 192, 224, 288, 320]
+      # first_bs: batch size for the first image resolution in the scales list
+      # divide_factor: to ensure the width and height dimensions can be devided by downsampling multiple
+      first_bs: 48
+      divided_factor: 32
+      is_training: True
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+  Eval:
+    dataset: 
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/val_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+        - ResizeImage:
+            resize_short: 288
+            interpolation: bilinear
+        - CropImage:
+            size: 256
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 48
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+Infer:
+  infer_imgs: docs/images/inference_deployment/whl_demo.jpg
+  batch_size: 10
+  transforms:
+    - DecodeImage:
+        to_rgb: True
+        channel_first: False
+    - ResizeImage:
+        resize_short: 288
+        interpolation: bilinear
+    - CropImage:
+        size: 256
+    - NormalizeImage:
+        scale: 1.0/255.0
+        mean: [0.0, 0.0, 0.0]
+        std: [1.0, 1.0, 1.0]
+        order: ''
+    - ToCHWImage:
+  PostProcess:
+    name: Topk
+    topk: 5
+    class_id_map_file: ppcls/utils/imagenet1k_label_list.txt
+
+Metric:
+  Train:
+    - TopkAcc:
+        topk: [1, 5]
+  Eval:
+    - TopkAcc:
+        topk: [1, 5]

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_x0_5.yaml

@@ -47,48 +47,68 @@ Optimizer:
   beta1: 0.9
   beta2: 0.999
   epsilon: 1e-8
-  weight_decay: 0.01
+  weight_decay: 0.05
+  no_weight_decay_name: .bias norm
+  one_dim_param_no_weight_decay: True
   lr:
     # for 8 cards
     name: Cosine
     learning_rate: 0.002
     eta_min: 0.0002
-    warmup_epoch: 1  # 3000 iterations
-    warmup_start_lr: 0.0002
+    warmup_epoch: 20  # 20000 iterations
+    warmup_start_lr: 1e-6
     # by_epoch: True
+  clip_norm: 10
 
 # data loader for train and eval
 DataLoader:
   Train:
     dataset:
-      name: MultiScaleDataset
+      name: ImageNetDataset
       image_root: ./dataset/ILSVRC2012/
       cls_label_path: ./dataset/ILSVRC2012/train_list.txt
       transform_ops:
         - DecodeImage:
             to_rgb: True
             channel_first: False
+            backend: pil
         - RandCropImage:
             size: 256
-            interpolation: bilinear
+            interpolation: bicubic
+            backend: pil
             use_log_aspect: True
         - RandFlipImage:
             flip_code: 1
+        - TimmAutoAugment:
+            config_str: rand-m9-mstd0.5-inc1
+            interpolation: bicubic
+            img_size: 256
         - NormalizeImage:
             scale: 1.0/255.0
             mean: [0.0, 0.0, 0.0]
             std: [1.0, 1.0, 1.0]
             order: ''
-    # support to specify width and height respectively:
-    # scales: [(256,256) (160,160), (192,192), (224,224) (288,288) (320,320)]
+        - RandomErasing:
+            EPSILON: 0.25
+            sl: 0.02
+            sh: 1.0/3.0
+            r1: 0.3
+            attempt: 10
+            use_log_aspect: True
+            mode: pixel
+      batch_transform_ops:
+        - OpSampler:
+            MixupOperator:
+              alpha: 0.2
+              prob: 0.5
+            CutmixOperator:
+              alpha: 1.0
+              prob: 0.5
     sampler:
-      name: MultiScaleSampler
-      scales: [256, 160, 192, 224, 288, 320]
-      # first_bs: batch size for the first image resolution in the scales list
-      # divide_factor: to ensure the width and height dimensions can be devided by downsampling multiple
-      first_bs: 48
-      divided_factor: 32
-      is_training: True
+      name: DistributedBatchSampler
+      batch_size: 128
+      drop_last: False
+      shuffle: False
     loader:
       num_workers: 4
       use_shared_memory: True
@@ -99,11 +119,13 @@ DataLoader:
       cls_label_path: ./dataset/ILSVRC2012/val_list.txt
       transform_ops:
         - DecodeImage:
-            to_rgb: False
+            to_np: False
             channel_first: False
+            backend: pil
         - ResizeImage:
             resize_short: 288
-            interpolation: bilinear
+            interpolation: bicubic
+            backend: pil
         - CropImage:
             size: 256
         - NormalizeImage:
@@ -113,7 +135,7 @@ DataLoader:
             order: ''
     sampler:
       name: DistributedBatchSampler
-      batch_size: 48
+      batch_size: 128
       drop_last: False
       shuffle: False
     loader:
@@ -125,11 +147,13 @@ Infer:
   batch_size: 10
   transforms:
     - DecodeImage:
-        to_rgb: True
+        to_np: False
         channel_first: False
+        backend: pil
     - ResizeImage:
         resize_short: 288
-        interpolation: bilinear
+        interpolation: bicubic
+        backend: pil
     - CropImage:
         size: 256
     - NormalizeImage:

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_x0_75.yaml

+# global configs
+Global:
+  checkpoints: null
+  pretrained_model: null
+  output_dir: ./output/
+  device: gpu
+  save_interval: 1
+  eval_during_train: True
+  eval_interval: 1
+  epochs: 300
+  print_batch_step: 10
+  use_visualdl: False
+  # used for static mode and model export
+  image_shape: [3, 256, 256]
+  save_inference_dir: ./inference
+  use_dali: False
+
+# mixed precision training
+AMP:
+  scale_loss: 65536
+  use_dynamic_loss_scaling: True
+  # O1: mixed fp16
+  level: O1
+
+# model ema
+EMA:
+  decay: 0.9995
+
+# model architecture
+Arch:
+  name: MobileViTv3_x0_75
+  class_num: 1000
+  classifier_dropout: 0.
+
+# loss function config for traing/eval process
+Loss:
+  Train:
+    - CELoss:
+        weight: 1.0
+        epsilon: 0.1
+  Eval:
+    - CELoss:
+        weight: 1.0
+
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.999
+  epsilon: 1e-8
+  weight_decay: 0.05
+  no_weight_decay_name: .bias norm
+  one_dim_param_no_weight_decay: True
+  lr:
+    # for 8 cards
+    name: Cosine
+    learning_rate: 0.002
+    eta_min: 0.0002
+    warmup_epoch: 20  # 20000 iterations
+    warmup_start_lr: 1e-6
+    # by_epoch: True
+  clip_norm: 10
+
+# data loader for train and eval
+DataLoader:
+  Train:
+    dataset:
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/train_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+            backend: pil
+        - RandCropImage:
+            size: 256
+            interpolation: bicubic
+            backend: pil
+            use_log_aspect: True
+        - RandFlipImage:
+            flip_code: 1
+        - TimmAutoAugment:
+            config_str: rand-m9-mstd0.5-inc1
+            interpolation: bicubic
+            img_size: 256
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+        - RandomErasing:
+            EPSILON: 0.25
+            sl: 0.02
+            sh: 1.0/3.0
+            r1: 0.3
+            attempt: 10
+            use_log_aspect: True
+            mode: pixel
+      batch_transform_ops:
+        - OpSampler:
+            MixupOperator:
+              alpha: 0.2
+              prob: 0.5
+            CutmixOperator:
+              alpha: 1.0
+              prob: 0.5
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 128
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+  Eval:
+    dataset: 
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/val_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_np: False
+            channel_first: False
+            backend: pil
+        - ResizeImage:
+            resize_short: 288
+            interpolation: bicubic
+            backend: pil
+        - CropImage:
+            size: 256
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 128
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+Infer:
+  infer_imgs: docs/images/inference_deployment/whl_demo.jpg
+  batch_size: 10
+  transforms:
+    - DecodeImage:
+        to_np: False
+        channel_first: False
+        backend: pil
+    - ResizeImage:
+        resize_short: 288
+        interpolation: bicubic
+        backend: pil
+    - CropImage:
+        size: 256
+    - NormalizeImage:
+        scale: 1.0/255.0
+        mean: [0.0, 0.0, 0.0]
+        std: [1.0, 1.0, 1.0]
+        order: ''
+    - ToCHWImage:
+  PostProcess:
+    name: Topk
+    topk: 5
+    class_id_map_file: ppcls/utils/imagenet1k_label_list.txt
+
+Metric:
+  Train:
+    - TopkAcc:
+        topk: [1, 5]
+  Eval:
+    - TopkAcc:
+        topk: [1, 5]

ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_x1_0.yaml

+# global configs
+Global:
+  checkpoints: null
+  pretrained_model: null
+  output_dir: ./output/
+  device: gpu
+  save_interval: 1
+  eval_during_train: True
+  eval_interval: 1
+  epochs: 300
+  print_batch_step: 10
+  use_visualdl: False
+  # used for static mode and model export
+  image_shape: [3, 256, 256]
+  save_inference_dir: ./inference
+  use_dali: False
+
+# mixed precision training
+AMP:
+  scale_loss: 65536
+  use_dynamic_loss_scaling: True
+  # O1: mixed fp16
+  level: O1
+
+# model ema
+EMA:
+  decay: 0.9995
+
+# model architecture
+Arch:
+  name: MobileViTv3_x1_0
+  class_num: 1000
+  classifier_dropout: 0.
+
+# loss function config for traing/eval process
+Loss:
+  Train:
+    - CELoss:
+        weight: 1.0
+        epsilon: 0.1
+  Eval:
+    - CELoss:
+        weight: 1.0
+
+Optimizer:
+  name: AdamW
+  beta1: 0.9
+  beta2: 0.999
+  epsilon: 1e-8
+  weight_decay: 0.05
+  no_weight_decay_name: .bias norm
+  one_dim_param_no_weight_decay: True
+  lr:
+    # for 8 cards
+    name: Cosine
+    learning_rate: 0.002
+    eta_min: 0.0002
+    warmup_epoch: 20  # 20000 iterations
+    warmup_start_lr: 1e-6
+    # by_epoch: True
+  clip_norm: 10
+
+# data loader for train and eval
+DataLoader:
+  Train:
+    dataset:
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/train_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_rgb: True
+            channel_first: False
+            backend: pil
+        - RandCropImage:
+            size: 256
+            interpolation: bicubic
+            backend: pil
+            use_log_aspect: True
+        - RandFlipImage:
+            flip_code: 1
+        - TimmAutoAugment:
+            config_str: rand-m9-mstd0.5-inc1
+            interpolation: bicubic
+            img_size: 256
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+        - RandomErasing:
+            EPSILON: 0.25
+            sl: 0.02
+            sh: 1.0/3.0
+            r1: 0.3
+            attempt: 10
+            use_log_aspect: True
+            mode: pixel
+      batch_transform_ops:
+        - OpSampler:
+            MixupOperator:
+              alpha: 0.2
+              prob: 0.5
+            CutmixOperator:
+              alpha: 1.0
+              prob: 0.5
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 128
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+  Eval:
+    dataset: 
+      name: ImageNetDataset
+      image_root: ./dataset/ILSVRC2012/
+      cls_label_path: ./dataset/ILSVRC2012/val_list.txt
+      transform_ops:
+        - DecodeImage:
+            to_np: False
+            channel_first: False
+            backend: pil
+        - ResizeImage:
+            resize_short: 288
+            interpolation: bicubic
+            backend: pil
+        - CropImage:
+            size: 256
+        - NormalizeImage:
+            scale: 1.0/255.0
+            mean: [0.0, 0.0, 0.0]
+            std: [1.0, 1.0, 1.0]
+            order: ''
+    sampler:
+      name: DistributedBatchSampler
+      batch_size: 128
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+Infer:
+  infer_imgs: docs/images/inference_deployment/whl_demo.jpg
+  batch_size: 10
+  transforms:
+    - DecodeImage:
+        to_np: False
+        channel_first: False
+        backend: pil
+    - ResizeImage:
+        resize_short: 288
+        interpolation: bicubic
+        backend: pil
+    - CropImage:
+        size: 256
+    - NormalizeImage:
+        scale: 1.0/255.0
+        mean: [0.0, 0.0, 0.0]
+        std: [1.0, 1.0, 1.0]
+        order: ''
+    - ToCHWImage:
+  PostProcess:
+    name: Topk
+    topk: 5
+    class_id_map_file: ppcls/utils/imagenet1k_label_list.txt
+
+Metric:
+  Train:
+    - TopkAcc:
+        topk: [1, 5]
+  Eval:
+    - TopkAcc:
+        topk: [1, 5]

test_tipc/configs/MobileViTv3/MobileViTv3_S_train_infer_python.txt

+===========================train_params===========================
+model_name:MobileViTv3_S
+python:python3.7
+gpu_list:0|0,1
+-o Global.device:gpu
+-o Global.auto_cast:null
+-o Global.epochs:lite_train_lite_infer=2|whole_train_whole_infer=120
+-o Global.output_dir:./output/
+-o DataLoader.Train.sampler.first_bs:8
+-o Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./dataset/ILSVRC2012/val
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py -c ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_S.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False -o Global.print_batch_step=1 -o Global.eval_during_train=False -o Global.save_interval=2
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:tools/eval.py -c ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_S.yaml
+null:null
+##
+===========================infer_params==========================
+-o Global.save_inference_dir:./inference
+-o Global.pretrained_model:
+norm_export:tools/export_model.py -c ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_S.yaml
+quant_export:null
+fpgm_export:null
+distill_export:null
+kl_quant:null
+export2:null
+inference_dir:null
+infer_model:../inference/
+infer_export:True
+infer_quant:Fasle
+inference:python/predict_cls.py -c configs/inference_cls.yaml -o PreProcess.transform_ops.0.ResizeImage.resize_short=288 -o PreProcess.transform_ops.1.CropImage.size=256 -o PreProcess.transform_ops.2.NormalizeImage.mean=[0.,0.,0.] -o PreProcess.transform_ops.2.NormalizeImage.std=[1.,1.,1.]
+-o Global.use_gpu:True|False
+-o Global.enable_mkldnn:False
+-o Global.cpu_num_threads:1
+-o Global.batch_size:1
+-o Global.use_tensorrt:False
+-o Global.use_fp16:False
+-o Global.inference_model_dir:../inference
+-o Global.infer_imgs:../dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
+-o Global.save_log_path:null
+-o Global.benchmark:False
+null:null
+null:null
+===========================train_benchmark_params==========================
+batch_size:128
+fp_items:fp32
+epoch:1
+model_type:norm_train
+--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
+flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,256,256]}]

test_tipc/configs/MobileViTv3/MobileViTv3_x1_0_train_infer_python.txt

+===========================train_params===========================
+model_name:MobileViTv3_x1_0
+python:python3.7
+gpu_list:0|0,1
+-o Global.device:gpu
+-o Global.auto_cast:null
+-o Global.epochs:lite_train_lite_infer=2|whole_train_whole_infer=120
+-o Global.output_dir:./output/
+-o DataLoader.Train.sampler.batch_size:8
+-o Global.pretrained_model:null
+train_model_name:latest
+train_infer_img_dir:./dataset/ILSVRC2012/val
+null:null
+##
+trainer:norm_train
+norm_train:tools/train.py -c ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_x1_0.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False -o Global.print_batch_step=1 -o Global.eval_during_train=False -o Global.save_interval=2
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##
+===========================eval_params===========================
+eval:tools/eval.py -c ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_x1_0.yaml
+null:null
+##
+===========================infer_params==========================
+-o Global.save_inference_dir:./inference
+-o Global.pretrained_model:
+norm_export:tools/export_model.py -c ppcls/configs/ImageNet/MobileViTv3/MobileViTv3_x1_0.yaml
+quant_export:null
+fpgm_export:null
+distill_export:null
+kl_quant:null
+export2:null
+inference_dir:null
+infer_model:../inference/
+infer_export:True
+infer_quant:Fasle
+inference:python/predict_cls.py -c configs/inference_cls.yaml -o PreProcess.transform_ops.0.ResizeImage.resize_short=288 -o PreProcess.transform_ops.1.CropImage.size=256 -o PreProcess.transform_ops.2.NormalizeImage.mean=[0.,0.,0.] -o PreProcess.transform_ops.2.NormalizeImage.std=[1.,1.,1.]
+-o Global.use_gpu:True|False
+-o Global.enable_mkldnn:False
+-o Global.cpu_num_threads:1
+-o Global.batch_size:1
+-o Global.use_tensorrt:False
+-o Global.use_fp16:False
+-o Global.inference_model_dir:../inference
+-o Global.infer_imgs:../dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
+-o Global.save_log_path:null
+-o Global.benchmark:False
+null:null
+null:null
+===========================train_benchmark_params==========================
+batch_size:128
+fp_items:fp32
+epoch:1
+model_type:norm_train
+--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
+flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,256,256]}]