Merge branch 'develop' of https://github.com/PaddlePaddle/PaddleClas into fix_cpp_serving_bug

README_ch.md

@@ -37,7 +37,7 @@ Res2Net200_vd预训练模型Top-1精度高达85.1%。
 * 您可以扫描下面的微信群二维码， 加入PaddleClas 微信交流群。获得更高效的问题答疑，与各行各业开发者充分交流，期待您的加入。
 
 <div align="center">
-<img src="./docs/images/wx_group.png"  width = "200" />
+<img src="https://user-images.githubusercontent.com/12560511/150500411-fdb27d17-0c50-4ac1-a484-fb4a9c2454b3.jpg"  width = "200" />
 </div>
 
 ## 快速体验
@@ -71,7 +71,7 @@ PP-ShiTu图像识别快速体验：[点击这里](./docs/zh_CN/quick_start/quick
     - [模型导出](./docs/zh_CN/inference_deployment/export_model.md)
     - Python/C++ 预测引擎
       - [基于Python预测引擎预测推理](./docs/zh_CN/inference_deployment/python_deploy.md)
-      - [基于C++预测引擎预测推理](./docs/zh_CN/inference_deployment/cpp_deploy.md)(当前只支持图像分类任务，图像识别更新中)
+      - [基于C++分类预测引擎预测推理](./docs/zh_CN/inference_deployment/cpp_deploy.md)、[基于C++的PP-ShiTu预测引擎预测推理](deploy/cpp_shitu/readme.md)
     - 服务化部署
       - [Paddle Serving服务化部署(推荐)](./docs/zh_CN/inference_deployment/paddle_serving_deploy.md)
       - [Hub serving服务化部署](./docs/zh_CN/inference_deployment/paddle_hub_serving_deploy.md)

README_en.md

@@ -41,7 +41,7 @@ Four sample solutions are provided, including product recognition, vehicle recog
 * You can also scan the QR code below to join the PaddleClas WeChat group to get more efficient answers to your questions and to communicate with developers from all walks of life. We look forward to hearing from you.
 
 <div align="center">
-<img src="./docs/images/wx_group.png"  width = "200" />
+<img src="https://user-images.githubusercontent.com/12560511/150500411-fdb27d17-0c50-4ac1-a484-fb4a9c2454b3.jpg"  width = "200" />
 </div>
 
 ## Quick Start
@@ -68,7 +68,7 @@ Quick experience of image recognition：[Link](./docs/en/tutorials/quick_start_r
     - [Feature Learning](./docs/en/tutorials/getting_started_retrieval_en.md)
 - Inference Model Prediction
     - [Python Inference](./docs/en/inference.md)
-    - [C++ Inference](./deploy/cpp/readme_en.md)(only support classification for now, recognition coming soon)
+    - [C++ Classfication Inference](./deploy/cpp/readme_en.md)， [C++ PP-ShiTu Inference](deploy/cpp_shitu/readme_en.md)
 - Model Deploy (only support classification for now, recognition coming soon)
     - [Hub Serving Deployment](./deploy/hubserving/readme_en.md)
     - [Mobile Deployment](./deploy/lite/readme_en.md)

deploy/utils/config.py

@@ -122,7 +122,7 @@ def override(dl, ks, v):
         if len(ks) == 1:
             # assert ks[0] in dl, ('{} is not exist in {}'.format(ks[0], dl))
             if not ks[0] in dl:
-                logger.warning('A new filed ({}) detected!'.format(ks[0], dl))
+                logger.warning('A new filed ({}) detected!'.format(ks[0]))
             dl[ks[0]] = str2num(v)
         else:
             override(dl[ks[0]], ks[1:], v)

docs/zh_CN/advanced_tutorials/theseus_layer.md

@@ -187,7 +187,7 @@ print("The result returned by update_res(): ", res)
 output = net(pd_input)
 print("The output's keys of processed net: ", output.keys())
 # The output's keys of net:  dict_keys(['output', 'blocks[0]', 'blocks[2]', 'blocks[4]', 'blocks[10]'])
-# 网络前向输出 output 为 dict 类型对象，其中，output["key"] 为网络最终输出，output["blocks[0]"] 等为网络中间层输出结果
+# 网络前向输出 output 为 dict 类型对象，其中，output["output"] 为网络最终输出，output["blocks[0]"] 等为网络中间层输出结果
 ```
 
 除了通过调用方法 `update_res()` 的方式之外，也同样可以在实例化网络对象时，通过指定参数 `return_patterns` 实现相同效果：

ppcls/configs/ImageNet/ESNet/ESNet_x0_25.yaml

@@ -4,7 +4,6 @@ Global:
   pretrained_model: null
   output_dir: ./output/
   device: gpu
-  class_num: 1000
   save_interval: 1
   eval_during_train: True
   eval_interval: 1
@@ -17,6 +16,7 @@ Global:
 # model architecture
 Arch:
   name: ESNet_x0_25
+  class_num: 1000
  
 # loss function config for traing/eval process
 Loss:

ppcls/configs/ImageNet/ESNet/ESNet_x0_5.yaml

@@ -4,7 +4,6 @@ Global:
   pretrained_model: null
   output_dir: ./output/
   device: gpu
-  class_num: 1000
   save_interval: 1
   eval_during_train: True
   eval_interval: 1
@@ -17,6 +16,7 @@ Global:
 # model architecture
 Arch:
   name: ESNet_x0_5
+  class_num: 1000
  
 # loss function config for traing/eval process
 Loss:

ppcls/configs/ImageNet/ESNet/ESNet_x0_75.yaml

@@ -4,7 +4,6 @@ Global:
   pretrained_model: null
   output_dir: ./output/
   device: gpu
-  class_num: 1000
   save_interval: 1
   eval_during_train: True
   eval_interval: 1
@@ -17,6 +16,7 @@ Global:
 # model architecture
 Arch:
   name: ESNet_x0_75
+  class_num: 1000
  
 # loss function config for traing/eval process
 Loss:

ppcls/configs/ImageNet/ESNet/ESNet_x1_0.yaml

@@ -4,7 +4,6 @@ Global:
   pretrained_model: null
   output_dir: ./output/
   device: gpu
-  class_num: 1000
   save_interval: 1
   eval_during_train: True
   eval_interval: 1
@@ -17,6 +16,7 @@ Global:
 # model architecture
 Arch:
   name: ESNet_x1_0
+  class_num: 1000
  
 # loss function config for traing/eval process
 Loss:

ppcls/configs/ImageNet/MobileNetV1/MobileNetV1.yaml

@@ -71,7 +71,7 @@ DataLoader:
       drop_last: False
       shuffle: True
     loader:
-      num_workers: 4
+      num_workers: 8
       use_shared_memory: True
 
   Eval:

ppcls/configs/ImageNet/MobileNetV2/MobileNetV2.yaml

@@ -69,7 +69,7 @@ DataLoader:
       drop_last: False
       shuffle: True
     loader:
-      num_workers: 4
+      num_workers: 8
       use_shared_memory: True
 
   Eval:

ppcls/configs/ImageNet/MobileNetV3/MobileNetV3_large_x1_0.yaml

@@ -70,7 +70,7 @@ DataLoader:
       drop_last: False
       shuffle: True
     loader:
-      num_workers: 4
+      num_workers: 8
       use_shared_memory: True
 
   Eval:

ppcls/configs/ImageNet/ResNet/ResNet50_fp16_dygraph.yaml → ppcls/configs/ImageNet/ResNet/ResNet50_amp_O1.yaml

@@ -22,7 +22,8 @@ Global:
 AMP:
   scale_loss: 128.0
   use_dynamic_loss_scaling: True
-  use_pure_fp16: &use_pure_fp16 False
+  # O1: mixed fp16
+  level: O1
 
 # model architecture
 Arch:
@@ -44,6 +45,7 @@ Loss:
 Optimizer:
   name: Momentum
   momentum: 0.9
+  multi_precision: True
   lr:
     name: Piecewise
     learning_rate: 0.1
@@ -74,12 +76,11 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
-            output_fp16: *use_pure_fp16
             channel_num: *image_channel
 
     sampler:
       name: DistributedBatchSampler
-      batch_size: 256
+      batch_size: 64
       drop_last: False
       shuffle: True
     loader:
@@ -104,7 +105,6 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
-            output_fp16: *use_pure_fp16
             channel_num: *image_channel
     sampler:
       name: DistributedBatchSampler
@@ -131,7 +131,6 @@ Infer:
         mean: [0.485, 0.456, 0.406]
         std: [0.229, 0.224, 0.225]
         order: ''
-        output_fp16: *use_pure_fp16
         channel_num: *image_channel
     - ToCHWImage:
   PostProcess:

ppcls/configs/ImageNet/ResNet/ResNet50_fp16.yaml → ppcls/configs/ImageNet/ResNet/ResNet50_amp_O2.yaml

@@ -10,8 +10,8 @@ Global:
   epochs: 120
   print_batch_step: 10
   use_visualdl: False
-  # used for static mode and model export
   image_channel: &image_channel 4
+  # used for static mode and model export
   image_shape: [*image_channel, 224, 224]
   save_inference_dir: ./inference
   # training model under @to_static
@@ -22,7 +22,8 @@ Global:
 AMP:
   scale_loss: 128.0
   use_dynamic_loss_scaling: True
-  use_pure_fp16: &use_pure_fp16 True
+  # O2: pure fp16
+  level: O2
 
 # model architecture
 Arch:
@@ -43,7 +44,7 @@ Loss:
 Optimizer:
   name: Momentum
   momentum: 0.9
-  multi_precision: *use_pure_fp16
+  multi_precision: True
   lr:
     name: Piecewise
     learning_rate: 0.1
@@ -74,7 +75,7 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
-            output_fp16: *use_pure_fp16
+            output_fp16: True
             channel_num: *image_channel
 
     sampler:
@@ -104,7 +105,7 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
-            output_fp16: *use_pure_fp16
+            output_fp16: True
             channel_num: *image_channel
     sampler:
       name: DistributedBatchSampler
@@ -131,7 +132,7 @@ Infer:
         mean: [0.485, 0.456, 0.406]
         std: [0.229, 0.224, 0.225]
         order: ''
-        output_fp16: *use_pure_fp16
+        output_fp16: True
         channel_num: *image_channel
     - ToCHWImage:
   PostProcess:

ppcls/configs/ImageNet/SENet/SE_ResNeXt101_32x4d_fp16.yaml → ppcls/configs/ImageNet/SENet/SE_ResNeXt101_32x4d_amp_O2.yaml

@@ -20,6 +20,7 @@ Arch:
   name: SE_ResNeXt101_32x4d
   class_num: 1000
   input_image_channel: *image_channel
+  data_format: "NHWC"
  
 # loss function config for traing/eval process
 Loss:
@@ -35,11 +36,13 @@ Loss:
 AMP:
     scale_loss: 128.0
     use_dynamic_loss_scaling: True
-    use_pure_fp16: &use_pure_fp16 True
+    # O2: pure fp16
+    level: O2
 
 Optimizer:
   name: Momentum
   momentum: 0.9
+  multi_precision: True
   lr:
     name: Cosine
     learning_rate: 0.1
@@ -67,7 +70,7 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
-            output_fp16: *use_pure_fp16
+            output_fp16: True
             channel_num: *image_channel
     sampler:
       name: DistributedBatchSampler
@@ -96,7 +99,7 @@ DataLoader:
             mean: [0.485, 0.456, 0.406]
             std: [0.229, 0.224, 0.225]
             order: ''
-            output_fp16: *use_pure_fp16
+            output_fp16: True
             channel_num: *image_channel
     sampler:
       name: BatchSampler
@@ -123,7 +126,7 @@ Infer:
         mean: [0.485, 0.456, 0.406]
         std: [0.229, 0.224, 0.225]
         order: ''
-        output_fp16: *use_pure_fp16
+        output_fp16: True
         channel_num: *image_channel
     - ToCHWImage:
   PostProcess:

ppcls/configs/ImageNet/ShuffleNet/ShuffleNetV2_x1_0.yaml

@@ -68,7 +68,7 @@ DataLoader:
       drop_last: False
       shuffle: True
     loader:
-      num_workers: 4
+      num_workers: 8
       use_shared_memory: True
 
   Eval:

ppcls/engine/engine.py

@@ -84,7 +84,8 @@ class Engine(object):
 
         # for visualdl
         self.vdl_writer = None
-        if self.config['Global']['use_visualdl'] and mode == "train":
+        if self.config['Global'][
+                'use_visualdl'] and mode == "train" and dist.get_rank() == 0:
             vdl_writer_path = os.path.join(self.output_dir, "vdl")
             if not os.path.exists(vdl_writer_path):
                 os.makedirs(vdl_writer_path)
@@ -97,7 +98,7 @@ class Engine(object):
             paddle.__version__, self.device))
 
         # AMP training
-        self.amp = True if "AMP" in self.config else False
+        self.amp = True if "AMP" in self.config and self.mode == "train" else False
         if self.amp and self.config["AMP"] is not None:
             self.scale_loss = self.config["AMP"].get("scale_loss", 1.0)
             self.use_dynamic_loss_scaling = self.config["AMP"].get(
@@ -112,6 +113,14 @@ class Engine(object):
             }
             paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING)
 
+        if "class_num" in config["Global"]:
+            global_class_num = config["Global"]["class_num"]
+            if "class_num" not in config["Arch"]:
+                config["Arch"]["class_num"] = global_class_num
+                msg = f"The Global.class_num will be deprecated. Please use Arch.class_num instead. Arch.class_num has been set to {global_class_num}."
+            else:
+                msg = "The Global.class_num will be deprecated. Please use Arch.class_num instead. The Global.class_num has been ignored."
+            logger.warning(msg)
         #TODO(gaotingquan): support rec
         class_num = config["Arch"].get("class_num", None)
         self.config["DataLoader"].update({"class_num": class_num})
@@ -217,15 +226,26 @@ class Engine(object):
             self.scaler = paddle.amp.GradScaler(
                 init_loss_scaling=self.scale_loss,
                 use_dynamic_loss_scaling=self.use_dynamic_loss_scaling)
-            if self.config['AMP']['use_pure_fp16'] is True:
-                self.model = paddle.amp.decorate(models=self.model, level='O2', save_dtype='float32')
+            amp_level = self.config['AMP'].get("level", "O1")
+            if amp_level not in ["O1", "O2"]:
+                msg = "[Parameter Error]: The optimize level of AMP only support 'O1' and 'O2'. The level has been set 'O1'."
+                logger.warning(msg)
+                self.config['AMP']["level"] = "O1"
+                amp_level = "O1"
+            self.model, self.optimizer = paddle.amp.decorate(
+                models=self.model,
+                optimizers=self.optimizer,
+                level=amp_level,
+                save_dtype='float32')
 
         # for distributed
-        self.config["Global"][
-            "distributed"] = paddle.distributed.get_world_size() != 1
+        world_size = dist.get_world_size()
+        self.config["Global"]["distributed"] = world_size != 1
+        if world_size != 4 and self.mode == "train":
+            msg = f"The training strategy in config files provided by PaddleClas is based on 4 gpus. But the number of gpus is {world_size} in current training. Please modify the stategy (learning rate, batch size and so on) if use config files in PaddleClas to train."
+            logger.warning(msg)
         if self.config["Global"]["distributed"]:
             dist.init_parallel_env()
-        if self.config["Global"]["distributed"]:
             self.model = paddle.DataParallel(self.model)
 
         # build postprocess for infer
@@ -336,8 +356,8 @@ class Engine(object):
     @paddle.no_grad()
     def infer(self):
         assert self.mode == "infer" and self.eval_mode == "classification"
-        total_trainer = paddle.distributed.get_world_size()
-        local_rank = paddle.distributed.get_rank()
+        total_trainer = dist.get_world_size()
+        local_rank = dist.get_rank()
         image_list = get_image_list(self.config["Infer"]["infer_imgs"])
         # data split
         image_list = image_list[local_rank::total_trainer]

ppcls/engine/evaluation/classification.py

@@ -59,10 +59,12 @@ def classification_eval(engine, epoch_id=0):
 
         # image input
         if engine.amp:
-            amp_level = 'O1'
-            if engine.config['AMP']['use_pure_fp16'] is True:
-                amp_level = 'O2'
-            with paddle.amp.auto_cast(custom_black_list={"flatten_contiguous_range", "greater_than"}, level=amp_level):
+            amp_level = engine.config['AMP'].get("level", "O1").upper()
+            with paddle.amp.auto_cast(
+                    custom_black_list={
+                        "flatten_contiguous_range", "greater_than"
+                    },
+                    level=amp_level):
                 out = engine.model(batch[0])
                 # calc loss
                 if engine.eval_loss_func is not None:
@@ -70,7 +72,8 @@ def classification_eval(engine, epoch_id=0):
                     for key in loss_dict:
                         if key not in output_info:
                             output_info[key] = AverageMeter(key, '7.5f')
-                        output_info[key].update(loss_dict[key].numpy()[0], batch_size)
+                        output_info[key].update(loss_dict[key].numpy()[0],
+                                                batch_size)
         else:
             out = engine.model(batch[0])
             # calc loss
@@ -79,7 +82,8 @@ def classification_eval(engine, epoch_id=0):
                 for key in loss_dict:
                     if key not in output_info:
                         output_info[key] = AverageMeter(key, '7.5f')
-                    output_info[key].update(loss_dict[key].numpy()[0], batch_size)
+                    output_info[key].update(loss_dict[key].numpy()[0],
+                                            batch_size)
 
         # just for DistributedBatchSampler issue: repeat sampling
         current_samples = batch_size * paddle.distributed.get_world_size()

ppcls/engine/train/train.py

@@ -42,10 +42,12 @@ def train_epoch(engine, epoch_id, print_batch_step):
 
         # image input
         if engine.amp:
-            amp_level = 'O1'
-            if engine.config['AMP']['use_pure_fp16'] is True:
-                amp_level = 'O2'
-            with paddle.amp.auto_cast(custom_black_list={"flatten_contiguous_range", "greater_than"}, level=amp_level):
+            amp_level = engine.config['AMP'].get("level", "O1").upper()
+            with paddle.amp.auto_cast(
+                    custom_black_list={
+                        "flatten_contiguous_range", "greater_than"
+                    },
+                    level=amp_level):
                 out = forward(engine, batch)
                 loss_dict = engine.train_loss_func(out, batch[1])
         else:

ppcls/static/program.py

@@ -158,7 +158,7 @@ def create_strategy(config):
     exec_strategy.num_threads = 1
     exec_strategy.num_iteration_per_drop_scope = (
         10000
-        if 'AMP' in config and config.AMP.get("use_pure_fp16", False) else 10)
+        if 'AMP' in config and config.AMP.get("level", "O1") == "O2" else 10)
 
     fuse_op = True if 'AMP' in config else False
 
@@ -206,7 +206,7 @@ def mixed_precision_optimizer(config, optimizer):
         scale_loss = amp_cfg.get('scale_loss', 1.0)
         use_dynamic_loss_scaling = amp_cfg.get('use_dynamic_loss_scaling',
                                                False)
-        use_pure_fp16 = amp_cfg.get('use_pure_fp16', False)
+        use_pure_fp16 = amp_cfg.get("level", "O1") == "O2"
         optimizer = paddle.static.amp.decorate(
             optimizer,
             init_loss_scaling=scale_loss,

ppcls/static/run_dali.sh

 #!/usr/bin/env bash
 
-export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
-export FLAGS_fraction_of_gpu_memory_to_use=0.80
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
 
 python3.7 -m paddle.distributed.launch \
-    --gpus="0,1,2,3,4,5,6,7" \
+    --gpus="0,1,2,3" \
     ppcls/static/train.py \
-    -c ./ppcls/configs/ImageNet/ResNet/ResNet50_fp16.yaml \
-    -o Global.use_dali=True
-
+    -c ./ppcls/configs/ImageNet/ResNet/ResNet50_amp_O1.yaml

ppcls/static/train.py

@@ -158,7 +158,7 @@ def main(args):
     # load pretrained models or checkpoints
     init_model(global_config, train_prog, exe)
 
-    if 'AMP' in config and config.AMP.get("use_pure_fp16", False):
+    if 'AMP' in config and config.AMP.get("level", "O1") == "O2":
         optimizer.amp_init(
             device,
             scope=paddle.static.global_scope(),

test_tipc/config/ResNet/ResNet50_vd_train_linux_gpu_use_dali.txt

+===========================train_params===========================
+model_name:ResNet50_vd
+python:python3.7
+gpu_list:0|0,1
+-o Global.device:gpu
+-o Global.auto_cast:null|amp
+-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/ResNet/ResNet50_vd.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.use_dali=True
+pact_train:null
+fpgm_train:null
+distill_train:null
+null:null
+null:null
+##

test_tipc/prepare.sh

@@ -108,6 +108,11 @@ if [[ $FILENAME == *GeneralRecognition* ]];then
    exit 0
 fi
 
+if [[ $FILENAME == *use_dali* ]];then
+    python_name=$(func_parser_value "${lines[2]}")
+    ${python_name} -m pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly --upgrade nvidia-dali-nightly-cuda102
+fi
+
 if [ ${MODE} = "lite_train_lite_infer" ] || [ ${MODE} = "lite_train_whole_infer" ];then
     # pretrain lite train data
     cd dataset