Skip to content

P
PaddleClas

PaddlePaddle / PaddleClas
大约 1 年 前同步成功

通知 115
Star 4999
Fork 1114
P
PaddleClas
提交 fdaf24ee 编写于 作者: weixin_46524038's avatar weixin_46524038 提交者： cuicheng01
浏览文件
操作

add models

上级 885e1bc0
显示空白变更内容
内联 并排
Showing with 454 addition and 73 deletion
docs/zh_CN/models/ImageNet1k/ConvNeXt.md
浏览文件 @ fdaf24ee
...@@ -35,6 +35,11 @@ ConvNeXt（Cross Stage Partial Network）系列模型是 Meta 在 2022 年提出 ...@@ -35,6 +35,11 @@ ConvNeXt（Cross Stage Partial Network）系列模型是 Meta 在 2022 年提出
| Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) | | Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) |
|:--:|:--:|:--:|:--:|:--:|:--:|:--:| |:--:|:--:|:--:|:--:|:--:|:--:|:--:|
| ConvNeXt_tiny | 0.8203 | 0.9590 | 0.821 | - | 4.458 | 28.583 | | ConvNeXt_tiny | 0.8203 | 0.9590 | 0.821 | - | 4.458 | 28.583 |
| ConvNeXt_small | 0.8313 | 0.9643 | 0.831 | - | 8.688 | 50.210 |
| ConvNeXt_base_224 | 0.8384 | 0.9676 | 0.838 | - | 15.360 | 88.573 |
| ConvNeXt_base_384 | 0.8490 | 0.9727 | 0.851 | - | 45.138 | 88.573 |
| ConvNeXt_large_224 | 0.8426 | 0.9690 | 0.843 | - | 34.340 | 197.740 |
| ConvNeXt_large_384 | 0.8527 | 0.9749 | 0.855 | - | 101.001 | 197.740 |
### 1.3 Benchmark ### 1.3 Benchmark
......
docs/zh_CN/models/ImageNet1k/README.md
浏览文件 @ fdaf24ee
...@@ -329,8 +329,11 @@ ResNeSt 系列模型的精度、速度指标如下表所示，更多关于该系 ...@@ -329,8 +329,11 @@ ResNeSt 系列模型的精度、速度指标如下表所示，更多关于该系
| 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 | | 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 |
|------------------------|-----------|-----------|------------------|------------------|----------|-----------|------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------| |------------------------|-----------|-----------|------------------|------------------|----------|-----------|------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------|
| ResNeSt50_<br>fast_1s1x64d | 0.8035 | 0.9528 | 2.73 | 5.33 | 8.24 | 4.36 | 26.27 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt50_fast_1s1x64d_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNeSt50_fast_1s1x64d_infer.tar) | | ResNeSt50_<br>fast_1s1x64d | 0.8061 | 0.9527 | 2.73 | 5.33 | 8.24 | 4.36 | 26.27 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt50_fast_1s1x64d_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNeSt50_fast_1s1x64d_infer.tar) |
| ResNeSt50 | 0.8083 | 0.9542 | 7.36 | 10.23 | 13.84 | 5.40 | 27.54 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt50_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNeSt50_infer.tar) | | ResNeSt50 | 0.8102 | 0.9546 | 7.36 | 10.23 | 13.84 | 5.40 | 27.54 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt50_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNeSt50_infer.tar) |
| ResNeSt101 | 0.8279 | 0.9642 | | | | 10.25 | 48.40 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt101_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNeSt101_infer.tar) |
| ResNeSt200 | 0.8418 | 0.9698 | | | | 17.50 | 70.41 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt200_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNeSt200_infer.tar) |
| ResNeSt269 | 0.8444 |0.9698 | | | | 22.54 | 111.23 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt269_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ResNeSt269_infer.tar) |
<a name="RegNet"></a> <a name="RegNet"></a>
...@@ -340,7 +343,19 @@ RegNet 系列模型的精度、速度指标如下表所示，更多关于该系 ...@@ -340,7 +343,19 @@ RegNet 系列模型的精度、速度指标如下表所示，更多关于该系
| 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 | | 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 |
|------------------------|-----------|-----------|------------------|------------------|----------|-----------|------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------| |------------------------|-----------|-----------|------------------|------------------|----------|-----------|------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------|
| RegNetX_4GF | 0.785 | 0.9416 | 6.46 | 8.48 | 11.45 | 4.00 | 22.23 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_4GF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_4GF_infer.tar) | | RegNetX_200MF | 0.680 | 0.8842 | | | | 0.20 | 2.74 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_200MF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_200MF_infer.tar) |
| RegNetX_400MF | 0.723 | 0.9078 | | | | 0.40 | 5.19 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_400MF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_400MF_infer.tar) |
| RegNetX_600MF | 0.737 | 0.9198 | | | | 0.61 | 6.23 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_600MF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_600MF_infer.tar) |
| RegNetX_800MF | 0.751 | 0.9250 | | | | 0.81 | 7.30 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_800MF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_800MF_infer.tar) |
| RegNetX_1600MF | 0.767 | 0.9329 | | | | 1.62 | 9.23 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_1600MF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_1600MF_infer.tar) |
| RegNetX_3200MF | 0.781 | 0.9413 | | | | 3.20 | 15.36 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_3200MF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_3200MF_infer.tar) |
| RegNetX_4GF | 0.785 | 0.9416 | 6.46 | 8.48 | 11.45 | 3.99 | 22.16 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_4GF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_4GF_infer.tar) |
| RegNetX_6400MF | 0.790 | 0.9461 | | | | 6.49 | 26.28 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_6400MF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_6400MF_infer.tar) |
| RegNetX_8GF | 0.793 | 0.9464 | | | | 8.02 | 39.66 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_8GF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_8GF_infer.tar) |
| RegNetX_12GF | 0.797 | 0.9501 | | | | 12.13 | 46.20 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_12GF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_12GF_infer.tar) |
| RegNetX_16GF | 0.801 | 0.9505 | | | | 15.99 | 54.39 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_16GF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_16GF_infer.tar) |
| RegNetX_32GF | 0.803 | 0.9526 | | | | 32.33 | 130.67 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_32GF_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RegNetX_32GF_infer.tar) |
<a name="RepVGG"></a> <a name="RepVGG"></a>
...@@ -359,7 +374,9 @@ RegNet 系列模型的精度、速度指标如下表所示，更多关于该系 ...@@ -359,7 +374,9 @@ RegNet 系列模型的精度、速度指标如下表所示，更多关于该系
| RepVGG_B1g2 | 0.7732 | 0.9359 | | | | 8.82 | 41.36 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B1g2_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B1g2_infer.tar) | | RepVGG_B1g2 | 0.7732 | 0.9359 | | | | 8.82 | 41.36 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B1g2_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B1g2_infer.tar) |
| RepVGG_B1g4 | 0.7675 | 0.9335 | | | | 7.31 | 36.13 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B1g4_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B1g4_infer.tar) | | RepVGG_B1g4 | 0.7675 | 0.9335 | | | | 7.31 | 36.13 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B1g4_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B1g4_infer.tar) |
| RepVGG_B2g4 | 0.7881 | 0.9448 | | | | 11.34 | 55.78 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B2g4_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B2g4_infer.tar) | | RepVGG_B2g4 | 0.7881 | 0.9448 | | | | 11.34 | 55.78 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B2g4_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B2g4_infer.tar) |
| RepVGG_B3g4 | 0.7965 | 0.9485 | | | | 16.07 | 75.63 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B3g4_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B3g4_infer.tar) | | RepVGG_B3 | 0.8031 | 0.9517 | | | | 29.16 | 123.19 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B3_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B3_infer.tar) |
| RepVGG_B3g4 | 0.8005 | 0.9502 | | | | 17.89 | 83.93 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B3g4_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_B3g4_infer.tar) |
| RepVGG_D2se | 0.8339 | 0.9665 | | | | 36.54 | 133.47 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_D2se_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/RepVGG_D2se_infer.tar) |
<a name="MixNet"></a> <a name="MixNet"></a>
...@@ -441,6 +458,11 @@ RegNet 系列模型的精度、速度指标如下表所示，更多关于该系 ...@@ -441,6 +458,11 @@ RegNet 系列模型的精度、速度指标如下表所示，更多关于该系
| 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 | | 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 |
| ---------- | --------- | --------- | ---------------- | ---------------- | -------- | --------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | | ---------- | --------- | --------- | ---------------- | ---------------- | -------- | --------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ |
| ConvNeXt_tiny | 0.8203 | 0.9590 | - | - | - | 4.458 | 28.583 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_tiny_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ConvNeXt_tiny_infer.tar) | | ConvNeXt_tiny | 0.8203 | 0.9590 | - | - | - | 4.458 | 28.583 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_tiny_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ConvNeXt_tiny_infer.tar) |
| ConvNeXt_small | 0.8313 | 0.9643 | - | - | - | 8.688 | 50.210 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_small_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ConvNeXt_small_infer.tar) |
| ConvNeXt_base_224 | 0.8384 | 0.9676 | - | - | - | 15.360 | 88.573 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_base_224_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ConvNeXt_base_224_infer.tar) |
| ConvNeXt_base_384 | 0.8490 | 0.9727 | - | - | - | 45.138 | 88.573 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_base_384_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ConvNeXt_base_384_infer.tar) |
| ConvNeXt_large_224 | 0.8426 | 0.9690 | - | - | - | 34.340 | 197.740 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_large_224_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ConvNeXt_large_224_infer.tar) |
| ConvNeXt_large_384 | 0.8527 | 0.9749 | - | - | - | 101.001 | 197.740 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_large_384_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/ConvNeXt_large_384_infer.tar) |
<a name="VAN"></a> <a name="VAN"></a>
...@@ -451,6 +473,9 @@ RegNet 系列模型的精度、速度指标如下表所示，更多关于该系 ...@@ -451,6 +473,9 @@ RegNet 系列模型的精度、速度指标如下表所示，更多关于该系
| 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 | | 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | time(ms)<br/>bs=8 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 |
| ---------- | --------- | --------- | ---------------- | ---------------- | -------- | --------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | | ---------- | --------- | --------- | ---------------- | ---------------- | -------- | --------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ |
| VAN_B0 | 0.7535 | 0.9299 | - | - | - | 0.880 | 4.110 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B0_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/VAN_B0_infer.tar) | | VAN_B0 | 0.7535 | 0.9299 | - | - | - | 0.880 | 4.110 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B0_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/VAN_B0_infer.tar) |
| VAN_B1 | 0.8102 | 0.9562 | - | - | - | 2.518 | 13.869 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B1_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/VAN_B1_infer.tar) |
| VAN_B2 | 0.8280 | 0.9620 | - | - | - | 5.032 | 26.592 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B2_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/VAN_B2_infer.tar) |
| VAN_B3 | 0.8389 | 0.9668 | - | - | - | 8.987 | 44.790 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B3_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/VAN_B3_infer.tar) |
<a name="PeleeNet"></a> <a name="PeleeNet"></a>
...@@ -699,7 +724,8 @@ DeiT（Data-efficient Image Transformers）系列模型的精度、速度指标 ...@@ -699,7 +724,8 @@ DeiT（Data-efficient Image Transformers）系列模型的精度、速度指标
| 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 | | 模型 | Top-1 Acc | Top-5 Acc | time(ms)<br>bs=1 | time(ms)<br>bs=4 | FLOPs(G) | Params(M) | 预训练模型下载地址 | inference模型下载地址 |
| ---------- | --------- | --------- | ---------------- | ---------------- | -------- | --------- | ------------------------------------------------------------ | ------------------------------------------------------------ | | ---------- | --------- | --------- | ---------------- | ---------------- | -------- | --------- | ------------------------------------------------------------ | ------------------------------------------------------------ |
| TNT_small | 0.8121 |0.9563 | | | 4.83 | 23.68 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/TNT_small_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/TNT_small_infer.tar) | | TNT_small | 0.8148 |0.9580 | | | 4.83 | 23.69 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/TNT_small_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/TNT_small_infer.tar) |
| TNT_base | 0.8276 |0.9617 | | | 13.40 | 65.30 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/TNT_base_pretrained.pdparams) | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/TNT_base_infer.tar) |
**注**：TNT 模型的数据预处理部分 `NormalizeImage` 中的 `mean` 与 `std` 均为 0.5。 **注**：TNT 模型的数据预处理部分 `NormalizeImage` 中的 `mean` 与 `std` 均为 0.5。
......
docs/zh_CN/models/ImageNet1k/RegNet.md
浏览文件 @ fdaf24ee
...@@ -35,7 +35,18 @@ RegNet 是由 facebook 于 2020 年提出，旨在深化设计空间理念的概 ...@@ -35,7 +35,18 @@ RegNet 是由 facebook 于 2020 年提出，旨在深化设计空间理念的概
| Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) | | Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) |
|:--:|:--:|:--:|:--:|:--:|:--:|:--:| |:--:|:--:|:--:|:--:|:--:|:--:|:--:|
| RegNetX_4GF | 0.7850 | 0.9416| 0.7860 | -| 8.0 | 22.1 | | RegNetX_200MF | 0.6804 | 0.8842| 0.6821 | -| 0.2 | 2.7 |
| RegNetX_400MF | 0.7225 | 0.9078| 0.7228 | -| 0.4 | 5.2 |
| RegNetX_600MF | 0.7366 | 0.9198| 0.7286 | -| 0.6 | 6.2 |
| RegNetX_800MF | 0.7512 | 0.9250| 0.7494 | -| 0.8 | 7.3 |
| RegNetX_1600MF | 0.7673 | 0.9329| 0.7671 | -| 1.6 | 9.2 |
| RegNetX_3200MF | 0.7809 | 0.9413| 0.7819 | -| 3.2 | 15.3 |
| RegNetX_4GF | 0.7850 | 0.9416| 0.7860 | -| 4.0 | 22.2 |
| RegNetX_6400MF | 0.7897 | 0.9461| 0.7915 | -| 6.5 | 26.2 |
| RegNetX_8GF | 0.7928 | 0.9464| 0.7938 | -| 8.0 | 39.7 |
| RegNetX_12GF | 0.7972 | 0.9501| 0.8000 | -| 12.1 | 46.2 |
| RegNetX_16GF | 0.8013 | 0.9505| 0.8012 | -| 16.0 | 54.4 |
| RegNetX_32GF | 0.8032 | 0.9526| 0.8052 | -| 32.33 | 130.67 |
### 1.3 Benchmark ### 1.3 Benchmark
......
docs/zh_CN/models/ImageNet1k/RepVGG.md
浏览文件 @ fdaf24ee
...@@ -41,7 +41,9 @@ RepVGG(Making VGG-style ConvNets Great Again)系列模型是由清华大学(丁 ...@@ -41,7 +41,9 @@ RepVGG(Making VGG-style ConvNets Great Again)系列模型是由清华大学(丁
| RepVGG_B1g2 | 0.7732 | 0.9359 | 0.7778 | - | - | - | | RepVGG_B1g2 | 0.7732 | 0.9359 | 0.7778 | - | - | - |
| RepVGG_B1g4 | 0.7675 | 0.9335 | 0.7758 | - | - | - | | RepVGG_B1g4 | 0.7675 | 0.9335 | 0.7758 | - | - | - |
| RepVGG_B2g4 | 0.7881 | 0.9448 | 0.7938 | - | - | - | | RepVGG_B2g4 | 0.7881 | 0.9448 | 0.7938 | - | - | - |
| RepVGG_B3g4 | 0.7965 | 0.9485 | 0.8021 | - | - | - | | RepVGG_B3 | 0.8031 | 0.9517 | 0.8052 | - | - | - |
| RepVGG_B3g4 | 0.8005 | 0.9502 | 0.8021 | - | - | - |
| RepVGG_D2se | 0.8339 | 0.9665 | 0.8355 | - | - | - |
关于 Params、FLOPs、Inference speed 等信息，敬请期待。 关于 Params、FLOPs、Inference speed 等信息，敬请期待。
......
docs/zh_CN/models/ImageNet1k/ResNeSt.md
浏览文件 @ fdaf24ee
...@@ -35,8 +35,11 @@ ResNeSt 系列模型是在 2020 年提出的，在原有的 resnet 网络结构 ...@@ -35,8 +35,11 @@ ResNeSt 系列模型是在 2020 年提出的，在原有的 resnet 网络结构
| Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) | | Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) |
|:--:|:--:|:--:|:--:|:--:|:--:|:--:| |:--:|:--:|:--:|:--:|:--:|:--:|:--:|
| ResNeSt50_fast_1s1x64d | 0.8035 | 0.9528| 0.8035 | -| 8.68 | 26.3 | | ResNeSt50_fast_1s1x64d | 0.8061 | 0.9527| 0.8035 | -| 5.40 | 26.3 |
| ResNeSt50 | 0.8083 | 0.9542| 0.8113 | -| 10.78 | 27.5 | | ResNeSt50 | 0.8102 | 0.9546| 0.8103 | -| 5.40 | 27.5 |
| ResNeSt101 | 0.8279 | 0.9642| 0.8283 | -| 10.25 | 48.4 |
| ResNeSt200 | 0.8418 | 0.9698| 0.8384 | -| 17.50 | 70.4 |
| ResNeSt269 | 0.8444 | 0.9698| 0.8454 | -| 22.54 | 111.2 |
### 1.3 Benchmark ### 1.3 Benchmark
......
docs/zh_CN/models/ImageNet1k/TNT.md
浏览文件 @ fdaf24ee
...@@ -34,7 +34,8 @@ PaddleClas 所提供的该系列模型的预训练模型权重，均是基于其 ...@@ -34,7 +34,8 @@ PaddleClas 所提供的该系列模型的预训练模型权重，均是基于其
| Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) | | Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) |
|:--:|:--:|:--:|:--:|:--:|:--:|:--:| |:--:|:--:|:--:|:--:|:--:|:--:|:--:|
| TNT_small | 0.8121 | 0.9563 | - | - | 5.2 | 23.8 | | TNT_small | 0.8148 | 0.9580 | 0.815 | - | 4.8 | 23.7 |
| TNT_base | 0.8276 | 0.9617 | 0.829 | - | 13.4 | 65.3 |
**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。 **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。
......
docs/zh_CN/models/ImageNet1k/VAN.md
浏览文件 @ fdaf24ee
...@@ -35,6 +35,9 @@ VAN（Visual Attention Network）系列模型是在 2022 年提出的 CNN 架构 ...@@ -35,6 +35,9 @@ VAN（Visual Attention Network）系列模型是在 2022 年提出的 CNN 架构
| Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) | | Models | Top1 | Top5 | Reference<br>top1 | Reference<br>top5 | FLOPs<br>(G) | Params<br>(M) |
|:--:|:--:|:--:|:--:|:--:|:--:|:--:| |:--:|:--:|:--:|:--:|:--:|:--:|:--:|
| VAN-B0 | 0.7535 | 0.9299 | 0.754 | - | 0.880 | 4.110 | | VAN-B0 | 0.7535 | 0.9299 | 0.754 | - | 0.880 | 4.110 |
| VAN-B1 | 0.8102 | 0.9562 | 0.811 | - | 2.518 | 13.869 |
| VAN-B2 | 0.8280 | 0.9620 | 0.828 | - | 5.032 | 26.592 |
| VAN-B3 | 0.8389 | 0.9668 | 0.839 | - | 8.987 | 44.790 |
### 1.3 Benchmark ### 1.3 Benchmark
......
ppcls/arch/backbone/model_zoo/convnext.py
浏览文件 @ fdaf24ee
...@@ -23,6 +23,16 @@ from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_fro ...@@ -23,6 +23,16 @@ from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_fro
MODEL_URLS = { MODEL_URLS = {
"ConvNeXt_tiny": "ConvNeXt_tiny":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_tiny_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_tiny_pretrained.pdparams",
"ConvNeXt_small":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_small_pretrained.pdparams",
"ConvNeXt_base_224":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_base_224_pretrained.pdparams",
"ConvNeXt_base_384":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_base_384_pretrained.pdparams",
"ConvNeXt_large_224":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_large_224_pretrained.pdparams",
"ConvNeXt_large_384":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ConvNeXt_large_384_pretrained.pdparams"
} }
__all__ = list(MODEL_URLS.keys()) __all__ = list(MODEL_URLS.keys())
...@@ -231,3 +241,42 @@ def ConvNeXt_tiny(pretrained=False, use_ssld=False, **kwargs): ...@@ -231,3 +241,42 @@ def ConvNeXt_tiny(pretrained=False, use_ssld=False, **kwargs):
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["ConvNeXt_tiny"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["ConvNeXt_tiny"], use_ssld=use_ssld)
return model return model
def ConvNeXt_small(pretrained=False, use_ssld=False, **kwargs):
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[96, 192, 384, 768], **kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["ConvNeXt_small"], use_ssld=use_ssld)
return model
def ConvNeXt_base_224(pretrained=False, use_ssld=False, **kwargs):
model = ConvNeXt(
depths=[3, 3, 27, 3], dims=[128, 256, 512, 1024], **kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["ConvNeXt_base_224"], use_ssld=use_ssld)
return model
def ConvNeXt_base_384(pretrained=False, use_ssld=False, **kwargs):
model = ConvNeXt(
depths=[3, 3, 27, 3], dims=[128, 256, 512, 1024], **kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["ConvNeXt_base_384"], use_ssld=use_ssld)
return model
def ConvNeXt_large_224(pretrained=False, use_ssld=False, **kwargs):
model = ConvNeXt(
depths=[3, 3, 27, 3], dims=[192, 384, 768, 1536], **kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["ConvNeXt_large_224"], use_ssld=use_ssld)
return model
def ConvNeXt_large_384(pretrained=False, use_ssld=False, **kwargs):
model = ConvNeXt(
depths=[3, 3, 27, 3], dims=[192, 384, 768, 1536], **kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["ConvNeXt_large_384"], use_ssld=use_ssld)
return model
\ No newline at end of file
ppcls/arch/backbone/model_zoo/regnet.py
浏览文件 @ fdaf24ee
...@@ -34,8 +34,26 @@ from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_fro ...@@ -34,8 +34,26 @@ from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_fro
MODEL_URLS = { MODEL_URLS = {
"RegNetX_200MF": "RegNetX_200MF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_200MF_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_200MF_pretrained.pdparams",
"RegNetX_400MF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_400MF_pretrained.pdparams",
"RegNetX_600MF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_600MF_pretrained.pdparams",
"RegNetX_800MF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_800MF_pretrained.pdparams",
"RegNetX_1600MF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_1600MF_pretrained.pdparams",
"RegNetX_3200MF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_3200MF_pretrained.pdparams",
"RegNetX_4GF": "RegNetX_4GF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_4GF_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_4GF_pretrained.pdparams",
"RegNetX_6400MF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_6400MF_pretrained.pdparams",
"RegNetX_8GF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_8GF_pretrained.pdparams",
"RegNetX_12GF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_12GF_pretrained.pdparams",
"RegNetX_16GF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_16GF_pretrained.pdparams",
"RegNetX_32GF": "RegNetX_32GF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_32GF_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetX_32GF_pretrained.pdparams",
"RegNetY_200MF": "RegNetY_200MF":
...@@ -43,7 +61,7 @@ MODEL_URLS = { ...@@ -43,7 +61,7 @@ MODEL_URLS = {
"RegNetY_4GF": "RegNetY_4GF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetY_4GF_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetY_4GF_pretrained.pdparams",
"RegNetY_32GF": "RegNetY_32GF":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetY_32GF_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RegNetY_32GF_pretrained.pdparams"
} }
__all__ = list(MODEL_URLS.keys()) __all__ = list(MODEL_URLS.keys())
...@@ -106,7 +124,7 @@ class ConvBNLayer(nn.Layer): ...@@ -106,7 +124,7 @@ class ConvBNLayer(nn.Layer):
padding=padding, padding=padding,
groups=groups, groups=groups,
weight_attr=ParamAttr(name=name + ".conv2d.output.1.w_0"), weight_attr=ParamAttr(name=name + ".conv2d.output.1.w_0"),
bias_attr=ParamAttr(name=name + ".conv2d.output.1.b_0")) bias_attr=False)
bn_name = name + "_bn" bn_name = name + "_bn"
self._batch_norm = BatchNorm( self._batch_norm = BatchNorm(
num_filters, num_filters,
...@@ -354,6 +372,81 @@ def RegNetX_200MF(pretrained=False, use_ssld=False, **kwargs): ...@@ -354,6 +372,81 @@ def RegNetX_200MF(pretrained=False, use_ssld=False, **kwargs):
return model return model
def RegNetX_400MF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet(
w_a=24.48,
w_0=24,
w_m=2.54,
d=22,
group_w=16,
bot_mul=1.0,
q=8,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_400MF"], use_ssld=use_ssld)
return model
def RegNetX_600MF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet(
w_a=36.97,
w_0=48,
w_m=2.24,
d=16,
group_w=24,
bot_mul=1.0,
q=8,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_600MF"], use_ssld=use_ssld)
return model
def RegNetX_800MF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet(
w_a=35.73,
w_0=56,
w_m=2.28,
d=16,
group_w=16,
bot_mul=1.0,
q=8,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_800MF"], use_ssld=use_ssld)
return model
def RegNetX_1600MF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet(
w_a=34.01,
w_0=80,
w_m=2.25,
d=18,
group_w=24,
bot_mul=1.0,
q=8,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_1600MF"], use_ssld=use_ssld)
return model
def RegNetX_3200MF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet(
w_a=26.31,
w_0=88,
w_m=2.25,
d=25,
group_w=48,
bot_mul=1.0,
q=8,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_3200MF"], use_ssld=use_ssld)
return model
def RegNetX_4GF(pretrained=False, use_ssld=False, **kwargs): def RegNetX_4GF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet( model = RegNet(
w_a=38.65, w_a=38.65,
...@@ -369,63 +462,75 @@ def RegNetX_4GF(pretrained=False, use_ssld=False, **kwargs): ...@@ -369,63 +462,75 @@ def RegNetX_4GF(pretrained=False, use_ssld=False, **kwargs):
return model return model
def RegNetX_32GF(pretrained=False, use_ssld=False, **kwargs): def RegNetX_6400MF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet( model = RegNet(
w_a=69.86, w_a=60.83,
w_0=320, w_0=184,
w_m=2.0, w_m=2.07,
d=17,
group_w=56,
bot_mul=1.0,
q=8,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_6400MF"], use_ssld=use_ssld)
return model
def RegNetX_8GF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet(
w_a=49.56,
w_0=80,
w_m=2.88,
d=23, d=23,
group_w=168, group_w=120,
bot_mul=1.0, bot_mul=1.0,
q=8, q=8,
**kwargs) **kwargs)
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_32GF"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["RegNetX_8GF"], use_ssld=use_ssld)
return model return model
def RegNetY_200MF(pretrained=False, use_ssld=False, **kwargs): def RegNetX_12GF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet( model = RegNet(
w_a=36.44, w_a=73.36,
w_0=24, w_0=168,
w_m=2.49, w_m=2.37,
d=13, d=19,
group_w=8, group_w=112,
bot_mul=1.0, bot_mul=1.0,
q=8, q=8,
se_on=True,
**kwargs) **kwargs)
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_32GF"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["RegNetX_12GF"], use_ssld=use_ssld)
return model return model
def RegNetY_4GF(pretrained=False, use_ssld=False, **kwargs): def RegNetX_16GF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet( model = RegNet(
w_a=31.41, w_a=55.59,
w_0=96, w_0=216,
w_m=2.24, w_m=2.1,
d=22, d=22,
group_w=64, group_w=128,
bot_mul=1.0, bot_mul=1.0,
q=8, q=8,
se_on=True,
**kwargs) **kwargs)
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_32GF"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["RegNetX_16GF"], use_ssld=use_ssld)
return model return model
def RegNetY_32GF(pretrained=False, use_ssld=False, **kwargs): def RegNetX_32GF(pretrained=False, use_ssld=False, **kwargs):
model = RegNet( model = RegNet(
w_a=115.89, w_a=69.86,
w_0=232, w_0=320,
w_m=2.53, w_m=2.0,
d=20, d=23,
group_w=232, group_w=168,
bot_mul=1.0, bot_mul=1.0,
q=8, q=8,
se_on=True,
**kwargs) **kwargs)
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["RegNetX_32GF"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["RegNetX_32GF"], use_ssld=use_ssld)
......
ppcls/arch/backbone/model_zoo/repvgg.py
浏览文件 @ fdaf24ee
...@@ -17,6 +17,7 @@ ...@@ -17,6 +17,7 @@
import paddle.nn as nn import paddle.nn as nn
import paddle import paddle
import paddle.nn.functional as F
import numpy as np import numpy as np
from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_from_url from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_from_url
...@@ -40,8 +41,12 @@ MODEL_URLS = { ...@@ -40,8 +41,12 @@ MODEL_URLS = {
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B1g4_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B1g4_pretrained.pdparams",
"RepVGG_B2g4": "RepVGG_B2g4":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B2g4_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B2g4_pretrained.pdparams",
"RepVGG_B3":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B3_pretrained.pdparams",
"RepVGG_B3g4": "RepVGG_B3g4":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B3g4_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_B3g4_pretrained.pdparams",
"RepVGG_D2se":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/RepVGG_D2se_pretrained.pdparams"
} }
__all__ = list(MODEL_URLS.keys()) __all__ = list(MODEL_URLS.keys())
...@@ -76,6 +81,33 @@ class ConvBN(nn.Layer): ...@@ -76,6 +81,33 @@ class ConvBN(nn.Layer):
return y return y
class SEBlock(nn.Layer):
def __init__(self, input_channels, internal_neurons):
super(SEBlock, self).__init__()
self.down = nn.Conv2D(
in_channels=input_channels,
out_channels=internal_neurons,
kernel_size=1,
stride=1,
bias_attr=True)
self.up = nn.Conv2D(
in_channels=internal_neurons,
out_channels=input_channels,
kernel_size=1,
stride=1,
bias_attr=True)
self.input_channels = input_channels
def forward(self, inputs):
x = F.avg_pool2d(inputs, kernel_size=inputs.shape[3])
x = self.down(x)
x = F.relu(x)
x = self.up(x)
x = F.sigmoid(x)
x = x.reshape([-1, self.input_channels, 1, 1])
return inputs * x
class RepVGGBlock(nn.Layer): class RepVGGBlock(nn.Layer):
def __init__(self, def __init__(self,
in_channels, in_channels,
...@@ -85,7 +117,8 @@ class RepVGGBlock(nn.Layer): ...@@ -85,7 +117,8 @@ class RepVGGBlock(nn.Layer):
padding=0, padding=0,
dilation=1, dilation=1,
groups=1, groups=1,
padding_mode='zeros'): padding_mode='zeros',
use_se=False):
super(RepVGGBlock, self).__init__() super(RepVGGBlock, self).__init__()
self.is_repped = False self.is_repped = False
...@@ -105,6 +138,11 @@ class RepVGGBlock(nn.Layer): ...@@ -105,6 +138,11 @@ class RepVGGBlock(nn.Layer):
self.nonlinearity = nn.ReLU() self.nonlinearity = nn.ReLU()
if use_se:
self.se = SEBlock(
out_channels, internal_neurons=out_channels // 16)
else:
self.se = nn.Identity()
self.rbr_identity = nn.BatchNorm2D( self.rbr_identity = nn.BatchNorm2D(
num_features=in_channels num_features=in_channels
) if out_channels == in_channels and stride == 1 else None ) if out_channels == in_channels and stride == 1 else None
...@@ -132,7 +170,7 @@ class RepVGGBlock(nn.Layer): ...@@ -132,7 +170,7 @@ class RepVGGBlock(nn.Layer):
else: else:
id_out = self.rbr_identity(inputs) id_out = self.rbr_identity(inputs)
return self.nonlinearity( return self.nonlinearity(
self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out) self.se(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out))
def rep(self): def rep(self):
if not hasattr(self, 'rbr_reparam'): if not hasattr(self, 'rbr_reparam'):
...@@ -198,14 +236,12 @@ class RepVGG(nn.Layer): ...@@ -198,14 +236,12 @@ class RepVGG(nn.Layer):
num_blocks, num_blocks,
width_multiplier=None, width_multiplier=None,
override_groups_map=None, override_groups_map=None,
class_num=1000): class_num=1000,
use_se=False):
super(RepVGG, self).__init__() super(RepVGG, self).__init__()
assert len(width_multiplier) == 4 assert len(width_multiplier) == 4
self.override_groups_map = override_groups_map or dict() self.override_groups_map = override_groups_map or dict()
assert 0 not in self.override_groups_map assert 0 not in self.override_groups_map
self.in_planes = min(64, int(64 * width_multiplier[0])) self.in_planes = min(64, int(64 * width_multiplier[0]))
self.stage0 = RepVGGBlock( self.stage0 = RepVGGBlock(
...@@ -213,20 +249,33 @@ class RepVGG(nn.Layer): ...@@ -213,20 +249,33 @@ class RepVGG(nn.Layer):
out_channels=self.in_planes, out_channels=self.in_planes,
kernel_size=3, kernel_size=3,
stride=2, stride=2,
padding=1) padding=1,
use_se=use_se)
self.cur_layer_idx = 1 self.cur_layer_idx = 1
self.stage1 = self._make_stage( self.stage1 = self._make_stage(
int(64 * width_multiplier[0]), num_blocks[0], stride=2) int(64 * width_multiplier[0]),
num_blocks[0],
stride=2,
use_se=use_se)
self.stage2 = self._make_stage( self.stage2 = self._make_stage(
int(128 * width_multiplier[1]), num_blocks[1], stride=2) int(128 * width_multiplier[1]),
num_blocks[1],
stride=2,
use_se=use_se)
self.stage3 = self._make_stage( self.stage3 = self._make_stage(
int(256 * width_multiplier[2]), num_blocks[2], stride=2) int(256 * width_multiplier[2]),
num_blocks[2],
stride=2,
use_se=use_se)
self.stage4 = self._make_stage( self.stage4 = self._make_stage(
int(512 * width_multiplier[3]), num_blocks[3], stride=2) int(512 * width_multiplier[3]),
num_blocks[3],
stride=2,
use_se=use_se)
self.gap = nn.AdaptiveAvgPool2D(output_size=1) self.gap = nn.AdaptiveAvgPool2D(output_size=1)
self.linear = nn.Linear(int(512 * width_multiplier[3]), class_num) self.linear = nn.Linear(int(512 * width_multiplier[3]), class_num)
def _make_stage(self, planes, num_blocks, stride): def _make_stage(self, planes, num_blocks, stride, use_se=False):
strides = [stride] + [1] * (num_blocks - 1) strides = [stride] + [1] * (num_blocks - 1)
blocks = [] blocks = []
for stride in strides: for stride in strides:
...@@ -238,7 +287,8 @@ class RepVGG(nn.Layer): ...@@ -238,7 +287,8 @@ class RepVGG(nn.Layer):
kernel_size=3, kernel_size=3,
stride=stride, stride=stride,
padding=1, padding=1,
groups=cur_groups)) groups=cur_groups,
use_se=use_se))
self.in_planes = planes self.in_planes = planes
self.cur_layer_idx += 1 self.cur_layer_idx += 1
return nn.Sequential(*blocks) return nn.Sequential(*blocks)
...@@ -367,6 +417,17 @@ def RepVGG_B2g4(pretrained=False, use_ssld=False, **kwargs): ...@@ -367,6 +417,17 @@ def RepVGG_B2g4(pretrained=False, use_ssld=False, **kwargs):
return model return model
def RepVGG_B3(pretrained=False, use_ssld=False, **kwargs):
model = RepVGG(
num_blocks=[4, 6, 16, 1],
width_multiplier=[3, 3, 3, 5],
override_groups_map=None,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["RepVGG_B3"], use_ssld=use_ssld)
return model
def RepVGG_B3g4(pretrained=False, use_ssld=False, **kwargs): def RepVGG_B3g4(pretrained=False, use_ssld=False, **kwargs):
model = RepVGG( model = RepVGG(
num_blocks=[4, 6, 16, 1], num_blocks=[4, 6, 16, 1],
...@@ -376,3 +437,15 @@ def RepVGG_B3g4(pretrained=False, use_ssld=False, **kwargs): ...@@ -376,3 +437,15 @@ def RepVGG_B3g4(pretrained=False, use_ssld=False, **kwargs):
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["RepVGG_B3g4"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["RepVGG_B3g4"], use_ssld=use_ssld)
return model return model
def RepVGG_D2se(pretrained=False, use_ssld=False, **kwargs):
model = RepVGG(
num_blocks=[8, 14, 24, 1],
width_multiplier=[2.5, 2.5, 2.5, 5],
override_groups_map=None,
use_se=True,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["RepVGG_D2se"], use_ssld=use_ssld)
return model
ppcls/arch/backbone/model_zoo/resnest.py
浏览文件 @ fdaf24ee
...@@ -39,6 +39,10 @@ MODEL_URLS = { ...@@ -39,6 +39,10 @@ MODEL_URLS = {
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt50_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt50_pretrained.pdparams",
"ResNeSt101": "ResNeSt101":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt101_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt101_pretrained.pdparams",
"ResNeSt200":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt200_pretrained.pdparams",
"ResNeSt269":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNeSt269_pretrained.pdparams"
} }
__all__ = list(MODEL_URLS.keys()) __all__ = list(MODEL_URLS.keys())
...@@ -160,8 +164,7 @@ class SplatConv(nn.Layer): ...@@ -160,8 +164,7 @@ class SplatConv(nn.Layer):
padding=0, padding=0,
groups=groups, groups=groups,
weight_attr=ParamAttr( weight_attr=ParamAttr(
name=name + "_weights", initializer=KaimingNormal()), name=name + "_weights", initializer=KaimingNormal()))
bias_attr=False)
self.rsoftmax = rSoftmax(radix=radix, cardinality=groups) self.rsoftmax = rSoftmax(radix=radix, cardinality=groups)
...@@ -739,3 +742,39 @@ def ResNeSt101(pretrained=False, use_ssld=False, **kwargs): ...@@ -739,3 +742,39 @@ def ResNeSt101(pretrained=False, use_ssld=False, **kwargs):
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["ResNeSt101"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["ResNeSt101"], use_ssld=use_ssld)
return model return model
def ResNeSt200(pretrained=False, use_ssld=False, **kwargs):
model = ResNeSt(
layers=[3, 24, 36, 3],
radix=2,
groups=1,
bottleneck_width=64,
deep_stem=True,
stem_width=64,
avg_down=True,
avd=True,
avd_first=False,
final_drop=0.0,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["ResNeSt200"], use_ssld=use_ssld)
return model
def ResNeSt269(pretrained=False, use_ssld=False, **kwargs):
model = ResNeSt(
layers=[3, 30, 48, 8],
radix=2,
groups=1,
bottleneck_width=64,
deep_stem=True,
stem_width=64,
avg_down=True,
avd=True,
avd_first=False,
final_drop=0.0,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["ResNeSt269"], use_ssld=use_ssld)
return model
ppcls/arch/backbone/model_zoo/tnt.py
浏览文件 @ fdaf24ee
...@@ -20,7 +20,6 @@ import numpy as np ...@@ -20,7 +20,6 @@ import numpy as np
import paddle import paddle
import paddle.nn as nn import paddle.nn as nn
from paddle.nn.initializer import TruncatedNormal, Constant from paddle.nn.initializer import TruncatedNormal, Constant
from ..base.theseus_layer import Identity from ..base.theseus_layer import Identity
...@@ -28,7 +27,9 @@ from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_fro ...@@ -28,7 +27,9 @@ from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_fro
MODEL_URLS = { MODEL_URLS = {
"TNT_small": "TNT_small":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/TNT_small_pretrained.pdparams" "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/TNT_small_pretrained.pdparams",
"TNT_base":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/TNT_base_pretrained.pdparams"
} }
__all__ = MODEL_URLS.keys() __all__ = MODEL_URLS.keys()
...@@ -38,6 +39,14 @@ zeros_ = Constant(value=0.) ...@@ -38,6 +39,14 @@ zeros_ = Constant(value=0.)
ones_ = Constant(value=1.) ones_ = Constant(value=1.)
class Identity(nn.Layer):
def __init__(self):
super(Identity, self).__init__()
def forward(self, inputs):
return inputs
def drop_path(x, drop_prob=0., training=False): def drop_path(x, drop_prob=0., training=False):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper... the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
...@@ -165,8 +174,10 @@ class Block(nn.Layer): ...@@ -165,8 +174,10 @@ class Block(nn.Layer):
act_layer=act_layer, act_layer=act_layer,
drop=drop) drop=drop)
self.norm1_proj = norm_layer(in_dim) self.norm1_proj = norm_layer(in_dim * num_pixel)
self.proj = nn.Linear(in_dim * num_pixel, dim) self.proj = nn.Linear(in_dim * num_pixel, dim, bias_attr=False)
self.norm2_proj = norm_layer(in_dim * num_pixel)
# Outer transformer # Outer transformer
self.norm_out = norm_layer(dim) self.norm_out = norm_layer(dim)
self.attn_out = Attention( self.attn_out = Attention(
...@@ -196,11 +207,10 @@ class Block(nn.Layer): ...@@ -196,11 +207,10 @@ class Block(nn.Layer):
self.drop_path(self.mlp_in(self.norm_mlp_in(pixel_embed)))) self.drop_path(self.mlp_in(self.norm_mlp_in(pixel_embed))))
# outer # outer
B, N, C = patch_embed.shape B, N, C = patch_embed.shape
norm1_proj = self.norm1_proj(pixel_embed) norm1_proj = pixel_embed.reshape(shape=[B, N - 1, C])
norm1_proj = norm1_proj.reshape( norm1_proj = self.norm1_proj(norm1_proj)
(B, N - 1, norm1_proj.shape[1] * norm1_proj.shape[2])) patch_embed[:, 1:] = paddle.add(
patch_embed[:, 1:] = paddle.add(patch_embed[:, 1:], patch_embed[:, 1:], self.norm2_proj(self.proj(norm1_proj)))
self.proj(norm1_proj))
patch_embed = paddle.add( patch_embed = paddle.add(
patch_embed, patch_embed,
self.drop_path(self.attn_out(self.norm_out(patch_embed)))) self.drop_path(self.attn_out(self.norm_out(patch_embed))))
...@@ -217,6 +227,7 @@ class PixelEmbed(nn.Layer): ...@@ -217,6 +227,7 @@ class PixelEmbed(nn.Layer):
in_dim=48, in_dim=48,
stride=4): stride=4):
super().__init__() super().__init__()
self.patch_size = patch_size
num_patches = (img_size // patch_size)**2 num_patches = (img_size // patch_size)**2
self.img_size = img_size self.img_size = img_size
self.num_patches = num_patches self.num_patches = num_patches
...@@ -230,14 +241,12 @@ class PixelEmbed(nn.Layer): ...@@ -230,14 +241,12 @@ class PixelEmbed(nn.Layer):
def forward(self, x, pixel_pos): def forward(self, x, pixel_pos):
B, C, H, W = x.shape B, C, H, W = x.shape
assert H == self.img_size and W == self.img_size, f"Input image size ({H}*{W}) doesn't match model ({self.img_size}*{self.img_size})." assert H == self.img_size and W == self.img_size, f"Input image size ({H}*{W}) doesn't match model ({self.img_size}*{self.img_size})."
x = nn.functional.unfold(x, self.patch_size, self.patch_size)
x = self.proj(x)
x = nn.functional.unfold(x, self.new_patch_size, self.new_patch_size)
x = x.transpose((0, 2, 1)).reshape( x = x.transpose((0, 2, 1)).reshape(
(-1, self.in_dim, self.new_patch_size, self.new_patch_size)) (-1, C, self.patch_size, self.patch_size))
x = self.proj(x)
x = x.reshape((-1, self.in_dim, self.patch_size)).transpose((0, 2, 1))
x = x + pixel_pos x = x + pixel_pos
x = x.reshape((-1, self.in_dim, self.new_patch_size *
self.new_patch_size)).transpose((0, 2, 1))
return x return x
...@@ -288,8 +297,7 @@ class TNT(nn.Layer): ...@@ -288,8 +297,7 @@ class TNT(nn.Layer):
self.add_parameter("patch_pos", self.patch_pos) self.add_parameter("patch_pos", self.patch_pos)
self.pixel_pos = self.create_parameter( self.pixel_pos = self.create_parameter(
shape=(1, in_dim, new_patch_size, new_patch_size), shape=(1, patch_size, in_dim), default_initializer=zeros_)
default_initializer=zeros_)
self.add_parameter("pixel_pos", self.pixel_pos) self.add_parameter("pixel_pos", self.pixel_pos)
self.pos_drop = nn.Dropout(p=drop_rate) self.pos_drop = nn.Dropout(p=drop_rate)
...@@ -345,7 +353,6 @@ class TNT(nn.Layer): ...@@ -345,7 +353,6 @@ class TNT(nn.Layer):
(self.cls_token.expand((B, -1, -1)), patch_embed), axis=1) (self.cls_token.expand((B, -1, -1)), patch_embed), axis=1)
patch_embed = patch_embed + self.patch_pos patch_embed = patch_embed + self.patch_pos
patch_embed = self.pos_drop(patch_embed) patch_embed = self.pos_drop(patch_embed)
for blk in self.blocks: for blk in self.blocks:
pixel_embed, patch_embed = blk(pixel_embed, patch_embed) pixel_embed, patch_embed = blk(pixel_embed, patch_embed)
...@@ -385,3 +392,17 @@ def TNT_small(pretrained=False, use_ssld=False, **kwargs): ...@@ -385,3 +392,17 @@ def TNT_small(pretrained=False, use_ssld=False, **kwargs):
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["TNT_small"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["TNT_small"], use_ssld=use_ssld)
return model return model
def TNT_base(pretrained=False, use_ssld=False, **kwargs):
model = TNT(patch_size=16,
embed_dim=640,
in_dim=40,
depth=12,
num_heads=10,
in_num_head=4,
qkv_bias=False,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["TNT_base"], use_ssld=use_ssld)
return model
ppcls/arch/backbone/model_zoo/van.py
浏览文件 @ fdaf24ee
...@@ -26,6 +26,12 @@ from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_fro ...@@ -26,6 +26,12 @@ from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_fro
MODEL_URLS = { MODEL_URLS = {
"VAN_B0": "VAN_B0":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B0_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B0_pretrained.pdparams",
"VAN_B1":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B1_pretrained.pdparams",
"VAN_B2":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B2_pretrained.pdparams",
"VAN_B3":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/VAN_B3_pretrained.pdparams"
} }
__all__ = list(MODEL_URLS.keys()) __all__ = list(MODEL_URLS.keys())
...@@ -269,6 +275,7 @@ class VAN(nn.Layer): ...@@ -269,6 +275,7 @@ class VAN(nn.Layer):
x, H, W = patch_embed(x) x, H, W = patch_embed(x)
for blk in block: for blk in block:
x = blk(x) x = blk(x)
x = x.flatten(2) x = x.flatten(2)
x = swapdim(x, 1, 2) x = swapdim(x, 1, 2)
x = norm(x) x = norm(x)
...@@ -317,3 +324,39 @@ def VAN_B0(pretrained=False, use_ssld=False, **kwargs): ...@@ -317,3 +324,39 @@ def VAN_B0(pretrained=False, use_ssld=False, **kwargs):
_load_pretrained( _load_pretrained(
pretrained, model, MODEL_URLS["VAN_B0"], use_ssld=use_ssld) pretrained, model, MODEL_URLS["VAN_B0"], use_ssld=use_ssld)
return model return model
def VAN_B1(pretrained=False, use_ssld=False, **kwargs):
model = VAN(embed_dims=[64, 128, 320, 512],
mlp_ratios=[8, 8, 4, 4],
norm_layer=partial(
nn.LayerNorm, epsilon=1e-6),
depths=[2, 2, 4, 2],
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["VAN_B1"], use_ssld=use_ssld)
return model
def VAN_B2(pretrained=False, use_ssld=False, **kwargs):
model = VAN(embed_dims=[64, 128, 320, 512],
mlp_ratios=[8, 8, 4, 4],
norm_layer=partial(
nn.LayerNorm, epsilon=1e-6),
depths=[3, 3, 12, 3],
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["VAN_B2"], use_ssld=use_ssld)
return model
def VAN_B3(pretrained=False, use_ssld=False, **kwargs):
model = VAN(embed_dims=[64, 128, 320, 512],
mlp_ratios=[8, 8, 4, 4],
norm_layer=partial(
nn.LayerNorm, epsilon=1e-6),
depths=[3, 5, 27, 3],
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["VAN_B3"], use_ssld=use_ssld)
return model
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或