Merge pull request #1014 from PaddlePaddle/develop

Develop

deploy/configs/build_cartoon.yaml

@@ -31,6 +31,7 @@ IndexProcess:
   index_path: "./recognition_demo_data_v1.0/gallery_cartoon/index/"
   image_root: "./recognition_demo_data_v1.0/gallery_cartoon/"
   data_file:  "./recognition_demo_data_v1.0/gallery_cartoon/data_file.txt"
+  append_index: False
   delimiter: "\t"
   dist_type: "IP"
   pq_size: 100

deploy/configs/build_logo.yaml

@@ -29,6 +29,7 @@ IndexProcess:
   index_path: "./recognition_demo_data_v1.0/gallery_logo/index/"
   image_root: "./recognition_demo_data_v1.0/gallery_logo/"
   data_file:  "./recognition_demo_data_v1.0/gallery_logo/data_file.txt"
+  append_index: False
   delimiter: "\t"
   dist_type: "IP"
   pq_size: 100

deploy/configs/build_product.yaml

@@ -29,6 +29,7 @@ IndexProcess:
   index_path: "./recognition_demo_data_v1.0/gallery_product/index"
   image_root: "./recognition_demo_data_v1.0/gallery_product/"
   data_file:  "./recognition_demo_data_v1.0/gallery_product/data_file.txt"
+  append_index: False
   delimiter: "\t"
   dist_type: "IP"
   pq_size: 100

deploy/configs/build_vehicle.yaml

@@ -29,6 +29,7 @@ IndexProcess:
   index_path: "./recognition_demo_data_v1.0/gallery_vehicle/index/"
   image_root: "./recognition_demo_data_v1.0/gallery_vehicle/"
   data_file:  "./recognition_demo_data_v1.0/gallery_vehicle/data_file.txt"
+  append_index: False
   delimiter: "\t"
   dist_type: "IP"
   pq_size: 100

deploy/python/build_gallery.py

@@ -86,7 +86,8 @@ class GalleryBuilder(object):
             gallery_vectors=gallery_features,
             gallery_docs=gallery_docs,
             pq_size=config['pq_size'],
-            index_path=config['index_path'])
+            index_path=config['index_path'],
+            append_index=config["append_index"])
 
 
 def main(config):

deploy/vector_search/interface.py

@@ -132,7 +132,8 @@ class Graph_Index(object):
               gallery_vectors,
               gallery_docs=[],
               pq_size=100,
-              index_path='graph_index/'):
+              index_path='graph_index/',
+              append_index=False):
         """
         build index 
         """
@@ -181,7 +182,25 @@ class Graph_Index(object):
         self.gallery_doc_dict["dist_type"] = self.dist_type
         self.gallery_doc_dict["with_attr"] = self.with_attr
 
-        with open(index_path + "/info.json", "w") as f:
+        output_path = os.path.join(index_path, "info.json")
+        if append_index is True and os.path.exists(output_path):
+            with open(output_path, "r") as fin:
+                lines = fin.readlines()[0]
+                ori_gallery_doc_dict = json.loads(lines)
+            assert ori_gallery_doc_dict["dist_type"] == self.gallery_doc_dict[
+                "dist_type"]
+            assert ori_gallery_doc_dict["dim"] == self.gallery_doc_dict["dim"]
+            assert ori_gallery_doc_dict["with_attr"] == self.gallery_doc_dict[
+                "with_attr"]
+            offset = ori_gallery_doc_dict["total_num"]
+            for i in range(0, self.gallery_doc_dict["total_num"]):
+                ori_gallery_doc_dict[str(i + offset)] = self.gallery_doc_dict[
+                    str(i)]
+
+            ori_gallery_doc_dict["total_num"] += self.gallery_doc_dict[
+                "total_num"]
+            self.gallery_doc_dict = ori_gallery_doc_dict
+        with open(output_path, "w") as f:
             json.dump(self.gallery_doc_dict, f)
 
         print("finished creating index ...")

docs/en/application/vehicle_recognition_en.md

@@ -58,10 +58,7 @@ This method is used in VERI-Wild dataset. This dataset was captured in a large C
 |   GLAMOR(Resnet50+PGN)[3]    |   77.15   |   92.13   |   97.43   |
 |      PVEN(Resnet50)[4]       |   79.8    |   94.01   |   98.06   |
 |    SAVER(VAE+Resnet50)[5]    |   80.9    |   93.78   |   97.93   |
-|    PaddleClas  baseline1     |   65.6    |   92.37   |   97.23   |
-|    PaddleClas  baseline2     |   80.09   | **93.81** | **98.26** |
-
- Baseline1 is the released, and baseline2 will be released soon.
+|     PaddleClas  baseline     |   80.57   | **93.81** | **98.06** |
 
 ### 2.2 Vehicle Fine-grained Classification
 
@@ -79,7 +76,7 @@ The images in the dataset mainly come from the network and monitoring  data. The
 |      Fine-Tuning DARTS[7]       |   95.9%    |
 |       Resnet50 + COOC[8]        |   95.6%    |
 |             A3M[9]              |   95.4%    |
-| PaddleClas  baseline (ResNet50) | **97.36**% |
+| PaddleClas  baseline (ResNet50) | **97.37**% |
 
 ## 3 References
 

docs/en/tutorials/quick_start_recognition_en.md

@@ -41,7 +41,7 @@ The detection model with the recognition inference model for the 4 directions (L
 | Cartoon Face Recognition Model| Cartoon Face Scenario  | [Model Download Link](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/cartoon_rec_ResNet50_iCartoon_v1.0_infer.tar) | [inference_cartoon.yaml](../../../deploy/configs/inference_cartoon.yaml) | [build_cartoon.yaml](../../../deploy/configs/build_cartoon.yaml) |
 | Vehicle Fine-Grained Classfication Model | Vehicle Scenario  |   [Model Download Link](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/vehicle_cls_ResNet50_CompCars_v1.0_infer.tar) | [inference_vehicle.yaml](../../../deploy/configs/inference_vehicle.yaml) | [build_vehicle.yaml](../../../deploy/configs/build_vehicle.yaml) |
 | Product Recignition Model | Product Scenario  |  [Model Download Link](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/product_ResNet50_vd_Inshop_v1.0_infer.tar) | [inference_product.yaml](../../../deploy/configs/inference_product.yaml) | [build_product.yaml](../../../deploy/configs/build_product.yaml) |
-| Vehicle ReID Model | Vehicle ReID Scenario | [Model Download Link](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/vehicle_reid_ResNet50_VERI_Wild_v1.0_infer.tar) | - | - |
+| Vehicle ReID Model | Vehicle ReID Scenario | [Model Download Link](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/vehicle_reid_ResNet50_VERIWild_v1.0_infer.tar) | - | - |
 
 
 Demo data in this tutorial can be downloaded here: [download link](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/data/recognition_demo_data_en_v1.0.tar).

docs/zh_CN/application/vehicle_recognition.md

@@ -57,10 +57,7 @@ ReID，也就是 Re-identification，其定义是利用算法，在图像库中
 |   GLAMOR(Resnet50+PGN)[3]    |   77.15   |   92.13   |   97.43   |
 |      PVEN(Resnet50)[4]       |   79.8    |   94.01   |   98.06   |
 |    SAVER(VAE+Resnet50)[5]    |   80.9    |   93.78   |   97.93   |
-|    PaddleClas  baseline1     |   65.6    |   92.37   |   97.23   |
-|    PaddleClas  baseline2     |   80.09   | **93.81** | **98.26** |
-
-注：baseline1 为目前的开源模型，baseline2即将开源
+|     PaddleClas  baseline     |   80.57   | **93.81** | **98.06** |
 
 ### 2.2 车辆细分类
 

docs/zh_CN/tutorials/quick_start_new_user.md

@@ -40,7 +40,7 @@
 * 进入PaddleClas目录。
 
 ```bash
-## linux or mac， $path_to_PaddleClas表示PaddleClas的根目录，用户需要根据自己的真实目录修改
+# linux or mac， $path_to_PaddleClas表示PaddleClas的根目录，用户需要根据自己的真实目录修改
 cd $path_to_PaddleClas
 ```
 
@@ -61,7 +61,9 @@ unzip flowers102.zip
 
 此时flowers102数据集存放在**dataset/flowers102/jpg** 文件夹中，图像示例如下：
 
-<img src="../../images/quick_start/Examples-Flower-102.png" style="zoom:50%;" />
+<div align="center">
+<img src="../../images/quick_start/Examples-Flower-102.png" width = "800" />
+</div>
 
 * 返回`PaddleClas`根目录
 
@@ -73,18 +75,14 @@ cd ../../
 
 ## 四、模型训练
 
-### 预训练模型下载
-
-### 训练模型
-
-#### 使用CPU进行模型训练
+### 使用CPU进行模型训练
 
 由于使用CPU来进行模型训练，计算速度较慢，因此，此处以ShuffleNetV2_x0_25为例。此模型计算量较小，在CPU上计算速度较快。但是也因为模型较小，训练好的模型精度也不会太高。
 
-##### 不使用预训练模型
+#### 不使用预训练模型
 
 ```shell
-#windows在cmd中进入PaddleClas根目录，执行此命令
+# windows在cmd中进入PaddleClas根目录，执行此命令
 python tools/train.py -c ./ppcls/configs/quick_start/new_user/ShuffleNetV2_x0_25.yaml
 ```
 
@@ -92,16 +90,17 @@ python tools/train.py -c ./ppcls/configs/quick_start/new_user/ShuffleNetV2_x0_25
 - `yaml`文`Global.device` 参数设置为`cpu`，即使用CPU进行训练（若不设置，此参数默认为`True`）
 - `yaml`文件中`epochs`参数设置为20，说明对整个数据集进行20个epoch迭代，预计训练20分钟左右（不同CPU，训练时间略有不同），此时训练模型不充分。若提高训练模型精度，请将此参数设大，如**40**，训练时间也会相应延长
 
-##### 使用预训练模型
+#### 使用预训练模型
 
 ```shell
 python tools/train.py -c ./ppcls/configs/quick_start/new_user/ShuffleNetV2_x0_25.yaml  -o Arch.pretrained=True
+```
 
 - `-o` 参数可以选择为True或False，也可以是预训练模型存放路径，当选择为True时，预训练权重会自动下载到本地。注意：若为预训练模型路径，则不要加上：`.pdparams`
 
 可以使用将使用与不使用预训练模型训练进行对比，观察loss的下降情况。
 
-#### 使用GPU进行模型训练
+### 使用GPU进行模型训练
 
 由于GPU训练速度更快，可以使用更复杂模型，因此以ResNet50_vd为例。与ShuffleNetV2_x0_25相比，此模型计算量较大， 训练好的模型精度也会更高。
 
@@ -119,7 +118,7 @@ python tools/train.py -c ./ppcls/configs/quick_start/new_user/ShuffleNetV2_x0_25
   set CUDA_VISIBLE_DEVICES=0
   ```
 
-##### 不使用预训练模型
+#### 不使用预训练模型
 
 ```shell
 python3 tools/train.py -c ./ppcls/configs/quick_start/ResNet50_vd.yaml
@@ -127,9 +126,11 @@ python3 tools/train.py -c ./ppcls/configs/quick_start/ResNet50_vd.yaml
 
 训练完成后，验证集的`Top1 Acc`曲线如下所示，最高准确率为0.2735。训练精度曲线下图所示
 
-![image-20210329103510936](../../images/quick_start/r50_vd_acc.png)
+<div align="center">
+<img src="../../images/quick_start/r50_vd_acc.png"  width = "800" />
+</div>
 
-##### 使用预训练模型进行训练
+#### 使用预训练模型进行训练
 
 基于ImageNet1k分类预训练模型进行微调，训练脚本如下所示
 
@@ -143,7 +144,9 @@ python3 tools/train.py -c ./ppcls/configs/quick_start/ResNet50_vd.yaml -o Arch.p
 
 验证集的`Top1 Acc`曲线如下所示，最高准确率为0.9402，加载预训练模型之后，flowers102数据集精度大幅提升，绝对精度涨幅超过65%。
 
-![image-20210329105947361](../../images/quick_start/r50_vd_pretrained_acc.png)
+<div align="center">
+<img src="../../images/quick_start/r50_vd_pretrained_acc.png"  width = "800" />
+</div>
 
 ## 五、模型预测
 

docs/zh_CN/tutorials/quick_start_recognition.md

@@ -41,7 +41,7 @@
 | 动漫人物识别模型 | 动漫人物场景  | [模型下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/cartoon_rec_ResNet50_iCartoon_v1.0_infer.tar) | [inference_cartoon.yaml](../../../deploy/configs/inference_cartoon.yaml) | [build_cartoon.yaml](../../../deploy/configs/build_cartoon.yaml) |
 | 车辆细分类模型 | 车辆场景  |  [模型下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/vehicle_cls_ResNet50_CompCars_v1.0_infer.tar) | [inference_vehicle.yaml](../../../deploy/configs/inference_vehicle.yaml) | [build_vehicle.yaml](../../../deploy/configs/build_vehicle.yaml) |
 | 商品识别模型 | 商品场景  |  [模型下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/product_ResNet50_vd_aliproduct_v1.0_infer.tar) | [inference_product.yaml](../../../deploy/configs/inference_product.yaml) | [build_product.yaml](../../../deploy/configs/build_product.yaml) |
-| 车辆ReID模型 | 车辆ReID场景 | [模型下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/vehicle_reid_ResNet50_VERI_Wild_v1.0_infer.tar) | - | - |
+| 车辆ReID模型 | 车辆ReID场景 | [模型下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/vehicle_reid_ResNet50_VERIWild_v1.0_infer.tar) | - | - |
 
 
 本章节demo数据下载地址如下: [数据下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/data/recognition_demo_data_v1.0.tar)。

ppcls/arch/backbone/model_zoo/levit.py

@@ -80,7 +80,10 @@ class Linear_BN(nn.Sequential):
         super().__init__()
         self.add_sublayer('c', nn.Linear(a, b, bias_attr=False))
         bn = nn.BatchNorm1D(b)
-        ones_(bn.weight)
+        if bn_weight_init == 0:
+            zeros_(bn.weight)
+        else:
+            ones_(bn.weight)
         zeros_(bn.bias)
         self.add_sublayer('bn', bn)
 

ppcls/arch/backbone/model_zoo/tnt.py

@@ -44,7 +44,7 @@ def drop_path(x, drop_prob=0., training=False):
         return x
     keep_prob = paddle.to_tensor(1 - drop_prob)
     shape = (paddle.shape(x)[0], ) + (1, ) * (x.ndim - 1)
-    random_tensor = keep_prob + paddle.rand(shape, dtype=x.dtype)
+    random_tensor = paddle.add(keep_prob, paddle.rand(shape, dtype=x.dtype))
     random_tensor = paddle.floor(random_tensor)  # binarize
     output = x.divide(keep_prob) * random_tensor
     return output
@@ -114,14 +114,17 @@ class Attention(nn.Layer):
                 (2, 0, 3, 1, 4))
 
         q, k = qk[0], qk[1]
-        v = self.v(x).reshape((B, N, self.num_heads, -1)).transpose(
-            (0, 2, 1, 3))
+        v = self.v(x).reshape(
+            (B, N, self.num_heads, x.shape[-1] // self.num_heads)).transpose(
+                (0, 2, 1, 3))
 
-        attn = (q @k.transpose((0, 1, 3, 2))) * self.scale
+        attn = paddle.matmul(q, k.transpose((0, 1, 3, 2))) * self.scale
         attn = nn.functional.softmax(attn, axis=-1)
         attn = self.attn_drop(attn)
 
-        x = (attn @v).transpose((0, 2, 1, 3)).reshape((B, N, -1))
+        x = paddle.matmul(attn, v)
+        x = x.transpose((0, 2, 1, 3)).reshape(
+            (B, N, x.shape[-1] * x.shape[-3]))
         x = self.proj(x)
         x = self.proj_drop(x)
         return x
@@ -182,18 +185,22 @@ class Block(nn.Layer):
 
     def forward(self, pixel_embed, patch_embed):
         # inner
-        pixel_embed = pixel_embed + self.drop_path(
-            self.attn_in(self.norm_in(pixel_embed)))
-        pixel_embed = pixel_embed + self.drop_path(
-            self.mlp_in(self.norm_mlp_in(pixel_embed)))
+        pixel_embed = paddle.add(
+            pixel_embed,
+            self.drop_path(self.attn_in(self.norm_in(pixel_embed))))
+        pixel_embed = paddle.add(
+            pixel_embed,
+            self.drop_path(self.mlp_in(self.norm_mlp_in(pixel_embed))))
         # outer
         B, N, C = patch_embed.shape
-        patch_embed[:, 1:] = patch_embed[:, 1:] + self.proj(
-            self.norm1_proj(pixel_embed).reshape((B, N - 1, -1)))
-        patch_embed = patch_embed + self.drop_path(
-            self.attn_out(self.norm_out(patch_embed)))
-        patch_embed = patch_embed + self.drop_path(
-            self.mlp(self.norm_mlp(patch_embed)))
+        patch_embed[:, 1:] = paddle.add(
+            patch_embed[:, 1:],
+            self.proj(self.norm1_proj(pixel_embed).reshape((B, N - 1, -1))))
+        patch_embed = paddle.add(
+            patch_embed,
+            self.drop_path(self.attn_out(self.norm_out(patch_embed))))
+        patch_embed = paddle.add(
+            patch_embed, self.drop_path(self.mlp(self.norm_mlp(patch_embed))))
         return pixel_embed, patch_embed
 
 
@@ -222,11 +229,10 @@ class PixelEmbed(nn.Layer):
         x = self.proj(x)
         x = nn.functional.unfold(x, self.new_patch_size, self.new_patch_size)
         x = x.transpose((0, 2, 1)).reshape(
-            (B * self.num_patches, self.in_dim, self.new_patch_size,
-             self.new_patch_size))
+            (-1, self.in_dim, self.new_patch_size, self.new_patch_size))
         x = x + pixel_pos
-        x = x.reshape((B * self.num_patches, self.in_dim, -1)).transpose(
-            (0, 2, 1))
+        x = x.reshape((-1, self.in_dim, self.new_patch_size *
+                       self.new_patch_size)).transpose((0, 2, 1))
         return x
 
 
@@ -328,7 +334,8 @@ class TNT(nn.Layer):
         patch_embed = self.norm2_proj(
             self.proj(
                 self.norm1_proj(
-                    pixel_embed.reshape((B, self.num_patches, -1)))))
+                    pixel_embed.reshape((-1, self.num_patches, pixel_embed.
+                                         shape[-1] * pixel_embed.shape[-2])))))
         patch_embed = paddle.concat(
             (self.cls_token.expand((B, -1, -1)), patch_embed), axis=1)
         patch_embed = patch_embed + self.patch_pos

ppcls/arch/backbone/variant_models/resnet_variant.py

 from paddle.nn import Conv2D
-from ppcls.arch.backbone.legendary_models.resnet import ResNet50
+from ppcls.arch.backbone.legendary_models.resnet import ResNet50, MODEL_URLS, _load_pretrained
 
 __all__ = ["ResNet50_last_stage_stride1"]
 
@@ -17,6 +17,7 @@ def ResNet50_last_stage_stride1(pretrained=False, use_ssld=False, **kwargs):
         return new_conv
 
     match_re = "conv2d_4[4|6]"
-    model = ResNet50(pretrained=pretrained, use_ssld=use_ssld, **kwargs)
+    model = ResNet50(pretrained=False, use_ssld=use_ssld, **kwargs)
     model.replace_sub(match_re, replace_function, True)
+    _load_pretrained(pretrained, model, MODEL_URLS["ResNet50"], use_ssld)
     return model

ppcls/configs/Vehicle/ResNet50_ReID.yaml

@@ -52,11 +52,8 @@ Optimizer:
   name: Momentum
   momentum: 0.9
   lr:
-    name: MultiStepDecay
+    name: Cosine
     learning_rate: 0.01
-    milestones: [30, 60, 70, 80, 90, 100, 120, 140]
-    gamma: 0.5
-    verbose: False
     last_epoch: -1
   regularizer:
     name: 'L2'