add static training (#1037)

* add static training

* fix typo

* add se fp16

* rm note

* fix loader

* fix cfg

ppcls/arch/backbone/legendary_models/resnet.py

@@ -104,7 +104,8 @@ class ConvBNLayer(TheseusLayer):
                  groups=1,
                  is_vd_mode=False,
                  act=None,
-                 lr_mult=1.0):
+                 lr_mult=1.0,
+                 data_format="NCHW"):
         super().__init__()
         self.is_vd_mode = is_vd_mode
         self.act = act
@@ -118,11 +119,13 @@ class ConvBNLayer(TheseusLayer):
             padding=(filter_size - 1) // 2,
             groups=groups,
             weight_attr=ParamAttr(learning_rate=lr_mult),
-            bias_attr=False)
+            bias_attr=False,
+            data_format=data_format)
         self.bn = BatchNorm(
             num_filters,
             param_attr=ParamAttr(learning_rate=lr_mult),
-            bias_attr=ParamAttr(learning_rate=lr_mult))
+            bias_attr=ParamAttr(learning_rate=lr_mult),
+            data_layout=data_format)
         self.relu = nn.ReLU()
 
     def forward(self, x):
@@ -136,14 +139,14 @@ class ConvBNLayer(TheseusLayer):
 
 
 class BottleneckBlock(TheseusLayer):
-    def __init__(
-            self,
-            num_channels,
-            num_filters,
-            stride,
-            shortcut=True,
-            if_first=False,
-            lr_mult=1.0, ):
+    def __init__(self,
+                 num_channels,
+                 num_filters,
+                 stride,
+                 shortcut=True,
+                 if_first=False,
+                 lr_mult=1.0,
+                 data_format="NCHW"):
         super().__init__()
 
         self.conv0 = ConvBNLayer(
@@ -151,20 +154,23 @@ class BottleneckBlock(TheseusLayer):
             num_filters=num_filters,
             filter_size=1,
             act="relu",
-            lr_mult=lr_mult)
+            lr_mult=lr_mult,
+            data_format=data_format)
         self.conv1 = ConvBNLayer(
             num_channels=num_filters,
             num_filters=num_filters,
             filter_size=3,
             stride=stride,
             act="relu",
-            lr_mult=lr_mult)
+            lr_mult=lr_mult,
+            data_format=data_format)
         self.conv2 = ConvBNLayer(
             num_channels=num_filters,
             num_filters=num_filters * 4,
             filter_size=1,
             act=None,
-            lr_mult=lr_mult)
+            lr_mult=lr_mult,
+            data_format=data_format)
 
         if not shortcut:
             self.short = ConvBNLayer(
@@ -173,7 +179,8 @@ class BottleneckBlock(TheseusLayer):
                 filter_size=1,
                 stride=stride if if_first else 1,
                 is_vd_mode=False if if_first else True,
-                lr_mult=lr_mult)
+                lr_mult=lr_mult,
+                data_format=data_format)
         self.relu = nn.ReLU()
         self.shortcut = shortcut
 
@@ -199,7 +206,8 @@ class BasicBlock(TheseusLayer):
                  stride,
                  shortcut=True,
                  if_first=False,
-                 lr_mult=1.0):
+                 lr_mult=1.0,
+                 data_format="NCHW"):
         super().__init__()
 
         self.stride = stride
@@ -209,13 +217,15 @@ class BasicBlock(TheseusLayer):
             filter_size=3,
             stride=stride,
             act="relu",
-            lr_mult=lr_mult)
+            lr_mult=lr_mult,
+            data_format=data_format)
         self.conv1 = ConvBNLayer(
             num_channels=num_filters,
             num_filters=num_filters,
             filter_size=3,
             act=None,
-            lr_mult=lr_mult)
+            lr_mult=lr_mult,
+            data_format=data_format)
         if not shortcut:
             self.short = ConvBNLayer(
                 num_channels=num_channels,
@@ -223,7 +233,8 @@ class BasicBlock(TheseusLayer):
                 filter_size=1,
                 stride=stride if if_first else 1,
                 is_vd_mode=False if if_first else True,
-                lr_mult=lr_mult)
+                lr_mult=lr_mult,
+                data_format=data_format)
         self.shortcut = shortcut
         self.relu = nn.ReLU()
 
@@ -256,7 +267,9 @@ class ResNet(TheseusLayer):
                  config,
                  version="vb",
                  class_num=1000,
-                 lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0]):
+                 lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0],
+                 data_format="NCHW",
+                 input_image_channel=3):
         super().__init__()
 
         self.cfg = config
@@ -279,22 +292,25 @@ class ResNet(TheseusLayer):
 
         self.stem_cfg = {
             #num_channels, num_filters, filter_size, stride
-            "vb": [[3, 64, 7, 2]],
-            "vd": [[3, 32, 3, 2], [32, 32, 3, 1], [32, 64, 3, 1]]
+            "vb": [[input_image_channel, 64, 7, 2]],
+            "vd":
+            [[input_image_channel, 32, 3, 2], [32, 32, 3, 1], [32, 64, 3, 1]]
         }
 
-        self.stem = nn.Sequential(*[
+        self.stem = nn.Sequential(* [
             ConvBNLayer(
                 num_channels=in_c,
                 num_filters=out_c,
                 filter_size=k,
                 stride=s,
                 act="relu",
-                lr_mult=self.lr_mult_list[0])
+                lr_mult=self.lr_mult_list[0],
+                data_format=data_format)
             for in_c, out_c, k, s in self.stem_cfg[version]
         ])
 
-        self.max_pool = MaxPool2D(kernel_size=3, stride=2, padding=1)
+        self.max_pool = MaxPool2D(
+            kernel_size=3, stride=2, padding=1, data_format=data_format)
         block_list = []
         for block_idx in range(len(self.block_depth)):
             shortcut = False
@@ -306,11 +322,12 @@ class ResNet(TheseusLayer):
                     stride=2 if i == 0 and block_idx != 0 else 1,
                     shortcut=shortcut,
                     if_first=block_idx == i == 0 if version == "vd" else True,
-                    lr_mult=self.lr_mult_list[block_idx + 1]))
+                    lr_mult=self.lr_mult_list[block_idx + 1],
+                    data_format=data_format))
                 shortcut = True
         self.blocks = nn.Sequential(*block_list)
 
-        self.avg_pool = AdaptiveAvgPool2D(1)
+        self.avg_pool = AdaptiveAvgPool2D(1, data_format=data_format)
         self.flatten = nn.Flatten()
         self.avg_pool_channels = self.num_channels[-1] * 2
         stdv = 1.0 / math.sqrt(self.avg_pool_channels * 1.0)
@@ -319,13 +336,19 @@ class ResNet(TheseusLayer):
             self.class_num,
             weight_attr=ParamAttr(initializer=Uniform(-stdv, stdv)))
 
+        self.data_format = data_format
+
     def forward(self, x):
-        x = self.stem(x)
-        x = self.max_pool(x)
-        x = self.blocks(x)
-        x = self.avg_pool(x)
-        x = self.flatten(x)
-        x = self.fc(x)
+        with paddle.static.amp.fp16_guard():
+            if self.data_format == "NHWC":
+                x = paddle.transpose(x, [0, 2, 3, 1])
+                x.stop_gradient = True
+            x = self.stem(x)
+            x = self.max_pool(x)
+            x = self.blocks(x)
+            x = self.avg_pool(x)
+            x = self.flatten(x)
+            x = self.fc(x)
         return x
 
 

ppcls/configs/ImageNet/ResNet/ResNet50_fp16.yaml

@@ -0,0 +1,145 @@
+# global configs
+Global:
+  checkpoints: null
+  pretrained_model: null
+  output_dir: ./output/
+  device: gpu
+  save_interval: 1
+  eval_during_train: True
+  eval_interval: 1
+  epochs: 120
+  print_batch_step: 10
+  use_visualdl: False
+  # used for static mode and model export
+  image_channel: &image_channel 4
+  image_shape: [*image_channel, 224, 224]
+  save_inference_dir: ./inference
+  # training model under @to_static
+  to_static: False
+
+# mixed precision training
+AMP:
+  scale_loss: 128.0
+  use_dynamic_loss_scaling: True
+  use_pure_fp16: &use_pure_fp16 True
+
+# model architecture
+Arch:
+  name: ResNet50
+  class_num: 1000
+ 
+# loss function config for traing/eval process
+Loss:
+  Train:
+    - CELoss:
+        weight: 1.0
+  Eval:
+    - CELoss:
+        weight: 1.0
+
+Optimizer:
+  name: Momentum
+  momentum: 0.9
+  multi_precision: False # *use_pure_fp16
+  lr:
+    name: Piecewise
+    learning_rate: 0.1
+    decay_epochs: [30, 60, 90]
+    values: [0.1, 0.01, 0.001, 0.0001]
+  regularizer:
+    name: 'L2'
+    coeff: 0.0001
+
+
+# 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
+        - RandCropImage:
+            size: 224
+        - RandFlipImage:
+            flip_code: 1
+        - NormalizeImage:
+            scale: 1.0/255.0
+            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: 32
+      drop_last: False
+      shuffle: 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: 256
+        - CropImage:
+            size: 224
+        - NormalizeImage:
+            scale: 1.0/255.0
+            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: 64
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+Infer:
+  infer_imgs: docs/images/whl/demo.jpg
+  batch_size: 10
+  transforms:
+    - DecodeImage:
+        to_rgb: True
+        channel_first: False
+    - ResizeImage:
+        resize_short: 256
+    - CropImage:
+        size: 224
+    - NormalizeImage:
+        scale: 1.0/255.0
+        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:
+    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/SENet/SE_ResNeXt101_32x4d_fp16.yaml

@@ -0,0 +1,139 @@
+# global configs
+Global:
+  checkpoints: null
+  pretrained_model: null
+  output_dir: ./output/
+  device: gpu
+  save_interval: 1
+  eval_during_train: True
+  eval_interval: 1
+  epochs: 200
+  print_batch_step: 10
+  use_visualdl: False
+  # used for static mode and model export
+  image_channel: &image_channel 4
+  image_shape: [*image_channel, 224, 224]
+  save_inference_dir: ./inference
+
+# model architecture
+Arch:
+  name: SE_ResNeXt101_32x4d
+  class_num: 1000
+ 
+# loss function config for traing/eval process
+Loss:
+  Train:
+    - CELoss:
+        weight: 1.0
+        epsilon: 0.1
+  Eval:
+    - CELoss:
+        weight: 1.0
+
+# mixed precision training
+AMP:
+    scale_loss: 128.0
+    use_dynamic_loss_scaling: True
+    use_pure_fp16: &use_pure_fp16 True
+
+Optimizer:
+  name: Momentum
+  momentum: 0.9
+  lr:
+    name: Cosine
+    learning_rate: 0.1
+  regularizer:
+    name: 'L2'
+    coeff: 0.00007
+
+# 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
+        - RandCropImage:
+            size: 224
+        - RandFlipImage:
+            flip_code: 1
+        - NormalizeImage:
+            scale: 1.0/255.0
+            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: 64
+      drop_last: False
+      shuffle: 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: 256
+        - CropImage:
+            size: 224
+        - NormalizeImage:
+            scale: 1.0/255.0
+            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: BatchSampler
+      batch_size: 64
+      drop_last: False
+      shuffle: False
+    loader:
+      num_workers: 4
+      use_shared_memory: True
+
+Infer:
+  infer_imgs: docs/images/whl/demo.jpg
+  batch_size: 10
+  transforms:
+    - DecodeImage:
+        to_rgb: True
+        channel_first: False
+    - ResizeImage:
+        resize_short: 256
+    - CropImage:
+        size: 224
+    - NormalizeImage:
+        scale: 1.0/255.0
+        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:
+    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/data/__init__.py

@@ -60,6 +60,7 @@ def build_dataloader(config, mode, device, use_dali=False, seed=None):
     if use_dali:
         from ppcls.data.dataloader.dali import dali_dataloader
         return dali_dataloader(config, mode, paddle.device.get_device(), seed)
+
     config_dataset = config[mode]['dataset']
     config_dataset = copy.deepcopy(config_dataset)
     dataset_name = config_dataset.pop('name')
@@ -74,10 +75,6 @@ def build_dataloader(config, mode, device, use_dali=False, seed=None):
 
     # build sampler
     config_sampler = config[mode]['sampler']
-    #config_sampler["batch_size"] = config_sampler[
-    #    "batch_size"] // paddle.distributed.get_world_size()
-    #assert config_sampler[
-    #    "batch_size"] >= 1, "The batch_size should be larger than gpu number."
     if "name" not in config_sampler:
         batch_sampler = None
         batch_size = config_sampler["batch_size"]

ppcls/data/dataloader/dali.py

@@ -148,7 +148,6 @@ def dali_dataloader(config, mode, device, seed=None):
     assert "gpu" in device, "gpu training is required for DALI"
     device_id = int(device.split(':')[1])
     config_dataloader = config[mode]
-    # mode = 'train' if mode.lower() == 'train' else 'eval'
     seed = 42 if seed is None else seed
     ops = [
         list(x.keys())[0]
@@ -160,6 +159,7 @@ def dali_dataloader(config, mode, device, seed=None):
     support_ops_eval = [
         "DecodeImage", "ResizeImage", "CropImage", "NormalizeImage"
     ]
+
     if mode.lower() == 'train':
         assert set(ops) == set(
             support_ops_train
@@ -171,6 +171,14 @@ def dali_dataloader(config, mode, device, seed=None):
         ), "The supported trasform_ops for eval_dataset in dali is : {}".format(
             ",".join(support_ops_eval))
 
+    normalize_ops = [
+        op for op in config_dataloader["dataset"]["transform_ops"]
+        if "NormalizeImage" in op
+    ][0]["NormalizeImage"]
+    channel_num = normalize_ops.get("channel_num", 3)
+    output_dtype = types.FLOAT16 if normalize_ops.get("output_fp16",
+                                                      False) else types.FLOAT
+
     env = os.environ
     #  assert float(env.get('FLAGS_fraction_of_gpu_memory_to_use', 0.92)) < 0.9, \
     #      "Please leave enough GPU memory for DALI workspace, e.g., by setting" \
@@ -179,9 +187,6 @@ def dali_dataloader(config, mode, device, seed=None):
     gpu_num = paddle.distributed.get_world_size()
 
     batch_size = config_dataloader["sampler"]["batch_size"]
-    #  assert batch_size % gpu_num == 0, \
-    #  "batch size must be multiple of number of devices"
-    #  batch_size = batch_size // gpu_num
 
     file_root = config_dataloader["dataset"]["image_root"]
     file_list = config_dataloader["dataset"]["cls_label_path"]
@@ -195,15 +200,9 @@ def dali_dataloader(config, mode, device, seed=None):
         INTERP_LANCZOS3,  # XXX use LANCZOS3 for cv2.INTER_LANCZOS4
     }
 
-    output_dtype = (types.FLOAT16 if 'AMP' in config and
-                    config.AMP.get("use_pure_fp16", False) else types.FLOAT)
-
     assert interp in interp_map, "interpolation method not supported by DALI"
     interp = interp_map[interp]
-    pad_output = False
-    image_shape = config.get("image_shape", None)
-    if image_shape and image_shape[0] == 4:
-        pad_output = True
+    pad_output = channel_num == 4
 
     transforms = {
         k: v
@@ -218,6 +217,10 @@ def dali_dataloader(config, mode, device, seed=None):
     mean = [v / scale for v in mean]
     std = [v / scale for v in std]
 
+    sampler_name = config_dataloader["sampler"].get("name",
+                                                    "DistributedBatchSampler")
+    assert sampler_name in ["DistributedBatchSampler", "BatchSampler"]
+
     if mode.lower() == "train":
         resize_shorter = 256
         crop = transforms["RandCropImage"]["size"]
@@ -279,10 +282,11 @@ def dali_dataloader(config, mode, device, seed=None):
     else:
         resize_shorter = transforms["ResizeImage"].get("resize_short", 256)
         crop = transforms["CropImage"]["size"]
-        if 'PADDLE_TRAINER_ID' in env and 'PADDLE_TRAINERS_NUM' in env:
+        if 'PADDLE_TRAINER_ID' in env and 'PADDLE_TRAINERS_NUM' in env and sampler_name == "DistributedBatchSampler":
             shard_id = int(env['PADDLE_TRAINER_ID'])
             num_shards = int(env['PADDLE_TRAINERS_NUM'])
             device_id = int(env['FLAGS_selected_gpus'])
+
             pipe = HybridValPipe(
                 file_root,
                 file_list,

ppcls/data/preprocess/ops/operators.py

@@ -197,14 +197,26 @@ class NormalizeImage(object):
     """ normalize image such as substract mean, divide std
     """
 
-    def __init__(self, scale=None, mean=None, std=None, order='chw'):
+    def __init__(self,
+                 scale=None,
+                 mean=None,
+                 std=None,
+                 order='chw',
+                 output_fp16=False,
+                 channel_num=3):
         if isinstance(scale, str):
             scale = eval(scale)
+        assert channel_num in [
+            3, 4
+        ], "channel number of input image should be set to 3 or 4."
+        self.channel_num = channel_num
+        self.output_dtype = 'float16' if output_fp16 else 'float32'
         self.scale = np.float32(scale if scale is not None else 1.0 / 255.0)
+        self.order = order
         mean = mean if mean is not None else [0.485, 0.456, 0.406]
         std = std if std is not None else [0.229, 0.224, 0.225]
 
-        shape = (3, 1, 1) if order == 'chw' else (1, 1, 3)
+        shape = (3, 1, 1) if self.order == 'chw' else (1, 1, 3)
         self.mean = np.array(mean).reshape(shape).astype('float32')
         self.std = np.array(std).reshape(shape).astype('float32')
 
@@ -215,7 +227,20 @@ class NormalizeImage(object):
 
         assert isinstance(img,
                           np.ndarray), "invalid input 'img' in NormalizeImage"
-        return (img.astype('float32') * self.scale - self.mean) / self.std
+
+        img = (img.astype('float32') * self.scale - self.mean) / self.std
+
+        if self.channel_num == 4:
+            img_h = img.shape[1] if self.order == 'chw' else img.shape[0]
+            img_w = img.shape[2] if self.order == 'chw' else img.shape[1]
+            pad_zeros = np.zeros(
+                (1, img_h, img_w)) if self.order == 'chw' else np.zeros(
+                    (img_h, img_w, 1))
+            img = (np.concatenate(
+                (img, pad_zeros), axis=0)
+                   if self.order == 'chw' else np.concatenate(
+                       (img, pad_zeros), axis=2))
+        return img.astype(self.output_dtype)
 
 
 class ToCHWImage(object):

ppcls/optimizer/__init__.py

@@ -41,7 +41,7 @@ def build_lr_scheduler(lr_config, epochs, step_each_epoch):
     return lr
 
 
-def build_optimizer(config, epochs, step_each_epoch, parameters):
+def build_optimizer(config, epochs, step_each_epoch, parameters=None):
     config = copy.deepcopy(config)
     # step1 build lr
     lr = build_lr_scheduler(config.pop('lr'), epochs, step_each_epoch)

ppcls/optimizer/optimizer.py

@@ -33,12 +33,14 @@ class Momentum(object):
                  learning_rate,
                  momentum,
                  weight_decay=None,
-                 grad_clip=None):
+                 grad_clip=None,
+                 multi_precision=False):
         super(Momentum, self).__init__()
         self.learning_rate = learning_rate
         self.momentum = momentum
         self.weight_decay = weight_decay
         self.grad_clip = grad_clip
+        self.multi_precision = multi_precision
 
     def __call__(self, parameters):
         opt = optim.Momentum(
@@ -46,6 +48,7 @@ class Momentum(object):
             momentum=self.momentum,
             weight_decay=self.weight_decay,
             grad_clip=self.grad_clip,
+            multi_precision=self.multi_precision,
             parameters=parameters)
         return opt
 
@@ -60,7 +63,8 @@ class Adam(object):
                  weight_decay=None,
                  grad_clip=None,
                  name=None,
-                 lazy_mode=False):
+                 lazy_mode=False,
+                 multi_precision=False):
         self.learning_rate = learning_rate
         self.beta1 = beta1
         self.beta2 = beta2
@@ -71,6 +75,7 @@ class Adam(object):
         self.grad_clip = grad_clip
         self.name = name
         self.lazy_mode = lazy_mode
+        self.multi_precision = multi_precision
 
     def __call__(self, parameters):
         opt = optim.Adam(
@@ -82,6 +87,7 @@ class Adam(object):
             grad_clip=self.grad_clip,
             name=self.name,
             lazy_mode=self.lazy_mode,
+            multi_precision=self.multi_precision,
             parameters=parameters)
         return opt
 
@@ -104,7 +110,8 @@ class RMSProp(object):
                  rho=0.95,
                  epsilon=1e-6,
                  weight_decay=None,
-                 grad_clip=None):
+                 grad_clip=None,
+                 multi_precision=False):
         super(RMSProp, self).__init__()
         self.learning_rate = learning_rate
         self.momentum = momentum
@@ -112,6 +119,7 @@ class RMSProp(object):
         self.epsilon = epsilon
         self.weight_decay = weight_decay
         self.grad_clip = grad_clip
+        self.multi_precision = multi_precision
 
     def __call__(self, parameters):
         opt = optim.RMSProp(
@@ -121,5 +129,6 @@ class RMSProp(object):
             epsilon=self.epsilon,
             weight_decay=self.weight_decay,
             grad_clip=self.grad_clip,
+            multi_precision=self.multi_precision,
             parameters=parameters)
-        return opt
+        return opt

ppcls/static/program.py

@@ -0,0 +1,456 @@
+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import time
+import numpy as np
+
+from collections import OrderedDict
+
+import paddle
+import paddle.nn.functional as F
+
+from paddle.distributed import fleet
+from paddle.distributed.fleet import DistributedStrategy
+
+# from ppcls.optimizer import OptimizerBuilder
+# from ppcls.optimizer.learning_rate import LearningRateBuilder
+
+from ppcls.arch import build_model
+from ppcls.loss import build_loss
+from ppcls.metric import build_metrics
+from ppcls.optimizer import build_optimizer
+from ppcls.optimizer import build_lr_scheduler
+
+from ppcls.utils.misc import AverageMeter
+from ppcls.utils import logger
+
+
+def create_feeds(image_shape, use_mix=None, dtype="float32"):
+    """
+    Create feeds as model input
+
+    Args:
+        image_shape(list[int]): model input shape, such as [3, 224, 224]
+        use_mix(bool): whether to use mix(include mixup, cutmix, fmix)
+
+    Returns:
+        feeds(dict): dict of model input variables
+    """
+    feeds = OrderedDict()
+    feeds['data'] = paddle.static.data(
+        name="data", shape=[None] + image_shape, dtype=dtype)
+    if use_mix:
+        feeds['y_a'] = paddle.static.data(
+            name="y_a", shape=[None, 1], dtype="int64")
+        feeds['y_b'] = paddle.static.data(
+            name="y_b", shape=[None, 1], dtype="int64")
+        feeds['lam'] = paddle.static.data(
+            name="lam", shape=[None, 1], dtype=dtype)
+    else:
+        feeds['label'] = paddle.static.data(
+            name="label", shape=[None, 1], dtype="int64")
+
+    return feeds
+
+
+def create_fetchs(out,
+                  feeds,
+                  architecture,
+                  topk=5,
+                  epsilon=None,
+                  use_mix=False,
+                  config=None,
+                  mode="Train"):
+    """
+    Create fetchs as model outputs(included loss and measures),
+    will call create_loss and create_metric(if use_mix).
+    Args:
+        out(variable): model output variable
+        feeds(dict): dict of model input variables.
+            If use mix_up, it will not include label.
+        architecture(dict): architecture information,
+            name(such as ResNet50) is needed
+        topk(int): usually top5
+        epsilon(float): parameter for label smoothing, 0.0 <= epsilon <= 1.0
+        use_mix(bool): whether to use mix(include mixup, cutmix, fmix)
+        config(dict): model config
+
+    Returns:
+        fetchs(dict): dict of model outputs(included loss and measures)
+    """
+    fetchs = OrderedDict()
+    # build loss
+    # TODO(littletomatodonkey): support mix training
+    if use_mix:
+        y_a = paddle.reshape(feeds['y_a'], [-1, 1])
+        y_b = paddle.reshape(feeds['y_b'], [-1, 1])
+        lam = paddle.reshape(feeds['lam'], [-1, 1])
+    else:
+        target = paddle.reshape(feeds['label'], [-1, 1])
+
+    loss_func = build_loss(config["Loss"][mode])
+
+    # TODO: support mix training
+    loss_dict = loss_func(out, target)
+
+    loss_out = loss_dict["loss"]
+    # if "AMP" in config and config.AMP.get("use_pure_fp16", False):
+    # loss_out = loss_out.astype("float16")
+
+    # if use_mix:
+    #     return loss_func(out, feed_y_a, feed_y_b, feed_lam)
+    # else:
+    #     return loss_func(out, target)
+
+    fetchs['loss'] = (loss_out, AverageMeter('loss', '7.4f', need_avg=True))
+
+    assert use_mix is False
+
+    # build metric
+    if not use_mix:
+        metric_func = build_metrics(config["Metric"][mode])
+
+        metric_dict = metric_func(out, target)
+
+        for key in metric_dict:
+            if mode != "Train" and paddle.distributed.get_world_size() > 1:
+                paddle.distributed.all_reduce(
+                    metric_dict[key], op=paddle.distributed.ReduceOp.SUM)
+                metric_dict[key] = metric_dict[
+                    key] / paddle.distributed.get_world_size()
+
+            fetchs[key] = (metric_dict[key], AverageMeter(
+                key, '7.4f', need_avg=True))
+
+    return fetchs
+
+
+def create_optimizer(config, step_each_epoch):
+    # create learning_rate instance
+    optimizer, lr_sch = build_optimizer(
+        config["Optimizer"], config["Global"]["epochs"], step_each_epoch)
+    return optimizer, lr_sch
+
+
+def create_strategy(config):
+    """
+    Create build strategy and exec strategy.
+
+    Args:
+        config(dict): config
+
+    Returns:
+        build_strategy: build strategy
+        exec_strategy: exec strategy
+    """
+    build_strategy = paddle.static.BuildStrategy()
+    exec_strategy = paddle.static.ExecutionStrategy()
+
+    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)
+
+    fuse_op = True if 'AMP' in config else False
+
+    fuse_bn_act_ops = config.get('fuse_bn_act_ops', fuse_op)
+    fuse_elewise_add_act_ops = config.get('fuse_elewise_add_act_ops', fuse_op)
+    fuse_bn_add_act_ops = config.get('fuse_bn_add_act_ops', fuse_op)
+    enable_addto = config.get('enable_addto', fuse_op)
+
+    build_strategy.fuse_bn_act_ops = fuse_bn_act_ops
+    build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops
+    build_strategy.fuse_bn_add_act_ops = fuse_bn_add_act_ops
+    build_strategy.enable_addto = enable_addto
+
+    return build_strategy, exec_strategy
+
+
+def dist_optimizer(config, optimizer):
+    """
+    Create a distributed optimizer based on a normal optimizer
+
+    Args:
+        config(dict):
+        optimizer(): a normal optimizer
+
+    Returns:
+        optimizer: a distributed optimizer
+    """
+    build_strategy, exec_strategy = create_strategy(config)
+
+    dist_strategy = DistributedStrategy()
+    dist_strategy.execution_strategy = exec_strategy
+    dist_strategy.build_strategy = build_strategy
+
+    dist_strategy.nccl_comm_num = 1
+    dist_strategy.fuse_all_reduce_ops = True
+    dist_strategy.fuse_grad_size_in_MB = 16
+    optimizer = fleet.distributed_optimizer(optimizer, strategy=dist_strategy)
+
+    return optimizer
+
+
+def mixed_precision_optimizer(config, optimizer):
+    if 'AMP' in config:
+        amp_cfg = config.AMP if config.AMP else dict()
+        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)
+        optimizer = paddle.static.amp.decorate(
+            optimizer,
+            init_loss_scaling=scale_loss,
+            use_dynamic_loss_scaling=use_dynamic_loss_scaling,
+            use_pure_fp16=use_pure_fp16,
+            use_fp16_guard=True)
+
+    return optimizer
+
+
+def build(config,
+          main_prog,
+          startup_prog,
+          step_each_epoch=100,
+          is_train=True,
+          is_distributed=True):
+    """
+    Build a program using a model and an optimizer
+        1. create feeds
+        2. create a dataloader
+        3. create a model
+        4. create fetchs
+        5. create an optimizer
+
+    Args:
+        config(dict): config
+        main_prog(): main program
+        startup_prog(): startup program
+        is_train(bool): train or eval
+        is_distributed(bool): whether to use distributed training method
+
+    Returns:
+        dataloader(): a bridge between the model and the data
+        fetchs(dict): dict of model outputs(included loss and measures)
+    """
+    with paddle.static.program_guard(main_prog, startup_prog):
+        with paddle.utils.unique_name.guard():
+            mode = "Train" if is_train else "Eval"
+            use_mix = "batch_transform_ops" in config["DataLoader"][mode][
+                "dataset"]
+            use_dali = config["Global"].get('use_dali', False)
+            feeds = create_feeds(
+                config["Global"]["image_shape"],
+                use_mix=use_mix,
+                dtype="float32")
+
+            # build model
+            # data_format should be assigned in arch-dict
+            input_image_channel = config["Global"]["image_shape"][
+                0]  # default as [3, 224, 224]
+            if input_image_channel != 3:
+                logger.warning(
+                    "Input image channel is changed to {}, maybe for better speed-up".
+                    format(input_image_channel))
+                config["Arch"]["input_image_channel"] = input_image_channel
+            model = build_model(config["Arch"])
+            out = model(feeds["data"])
+            # end of build model
+
+            fetchs = create_fetchs(
+                out,
+                feeds,
+                config["Arch"],
+                epsilon=config.get('ls_epsilon'),
+                use_mix=use_mix,
+                config=config,
+                mode=mode)
+            lr_scheduler = None
+            optimizer = None
+            if is_train:
+                optimizer, lr_scheduler = build_optimizer(
+                    config["Optimizer"], config["Global"]["epochs"],
+                    step_each_epoch)
+                optimizer = mixed_precision_optimizer(config, optimizer)
+                if is_distributed:
+                    optimizer = dist_optimizer(config, optimizer)
+                optimizer.minimize(fetchs['loss'][0])
+    return fetchs, lr_scheduler, feeds, optimizer
+
+
+def compile(config, program, loss_name=None, share_prog=None):
+    """
+    Compile the program
+
+    Args:
+        config(dict): config
+        program(): the program which is wrapped by
+        loss_name(str): loss name
+        share_prog(): the shared program, used for evaluation during training
+
+    Returns:
+        compiled_program(): a compiled program
+    """
+    build_strategy, exec_strategy = create_strategy(config)
+
+    compiled_program = paddle.static.CompiledProgram(
+        program).with_data_parallel(
+            share_vars_from=share_prog,
+            loss_name=loss_name,
+            build_strategy=build_strategy,
+            exec_strategy=exec_strategy)
+
+    return compiled_program
+
+
+total_step = 0
+
+
+def run(dataloader,
+        exe,
+        program,
+        feeds,
+        fetchs,
+        epoch=0,
+        mode='train',
+        config=None,
+        vdl_writer=None,
+        lr_scheduler=None):
+    """
+    Feed data to the model and fetch the measures and loss
+
+    Args:
+        dataloader(paddle io dataloader):
+        exe():
+        program():
+        fetchs(dict): dict of measures and the loss
+        epoch(int): epoch of training or evaluation
+        model(str): log only
+
+    Returns:
+    """
+    fetch_list = [f[0] for f in fetchs.values()]
+    metric_dict = OrderedDict([("lr", AverageMeter(
+        'lr', 'f', postfix=",", need_avg=False))])
+
+    for k in fetchs:
+        metric_dict[k] = fetchs[k][1]
+
+    metric_dict["batch_time"] = AverageMeter(
+        'batch_cost', '.5f', postfix=" s,")
+    metric_dict["reader_time"] = AverageMeter(
+        'reader_cost', '.5f', postfix=" s,")
+
+    for m in metric_dict.values():
+        m.reset()
+
+    use_dali = config["Global"].get('use_dali', False)
+    tic = time.time()
+
+    if not use_dali:
+        dataloader = dataloader()
+
+    idx = 0
+    batch_size = None
+    while True:
+        # The DALI maybe raise RuntimeError for some particular images, such as ImageNet1k/n04418357_26036.JPEG
+        try:
+            batch = next(dataloader)
+        except StopIteration:
+            break
+        except RuntimeError:
+            logger.warning(
+                "Except RuntimeError when reading data from dataloader, try to read once again..."
+            )
+            continue
+        idx += 1
+        # ignore the warmup iters
+        if idx == 5:
+            metric_dict["batch_time"].reset()
+            metric_dict["reader_time"].reset()
+
+        metric_dict['reader_time'].update(time.time() - tic)
+
+        if use_dali:
+            batch_size = batch[0]["data"].shape()[0]
+            feed_dict = batch[0]
+        else:
+            batch_size = batch[0].shape()[0]
+            feed_dict = {
+                key.name: batch[idx]
+                for idx, key in enumerate(feeds.values())
+            }
+
+        metrics = exe.run(program=program,
+                          feed=feed_dict,
+                          fetch_list=fetch_list)
+
+        for name, m in zip(fetchs.keys(), metrics):
+            metric_dict[name].update(np.mean(m), batch_size)
+        metric_dict["batch_time"].update(time.time() - tic)
+        if mode == "train":
+            metric_dict['lr'].update(lr_scheduler.get_lr())
+
+        fetchs_str = ' '.join([
+            str(metric_dict[key].mean)
+            if "time" in key else str(metric_dict[key].value)
+            for key in metric_dict
+        ])
+        ips_info = " ips: {:.5f} images/sec.".format(
+            batch_size / metric_dict["batch_time"].avg)
+        fetchs_str += ips_info
+
+        if lr_scheduler is not None:
+            lr_scheduler.step()
+
+        if vdl_writer:
+            global total_step
+            logger.scaler('loss', metrics[0][0], total_step, vdl_writer)
+            total_step += 1
+        if mode == 'eval':
+            if idx % config.get('print_interval', 10) == 0:
+                logger.info("{:s} step:{:<4d} {:s}".format(mode, idx,
+                                                           fetchs_str))
+        else:
+            epoch_str = "epoch:{:<3d}".format(epoch)
+            step_str = "{:s} step:{:<4d}".format(mode, idx)
+
+            if idx % config.get('print_interval', 10) == 0:
+                logger.info("{:s} {:s} {:s}".format(epoch_str, step_str,
+                                                    fetchs_str))
+
+        tic = time.time()
+
+    end_str = ' '.join([str(m.mean) for m in metric_dict.values()] +
+                       [metric_dict["batch_time"].total])
+    ips_info = "ips: {:.5f} images/sec.".format(
+        batch_size * metric_dict["batch_time"].count /
+        metric_dict["batch_time"].sum)
+    if mode == 'eval':
+        logger.info("END {:s} {:s} {:s}".format(mode, end_str, ips_info))
+    else:
+        end_epoch_str = "END epoch:{:<3d}".format(epoch)
+        logger.info("{:s} {:s} {:s} {:s}".format(end_epoch_str, mode, end_str,
+                                                 ips_info))
+    if use_dali:
+        dataloader.reset()
+
+    # return top1_acc in order to save the best model
+    if mode == 'eval':
+        return fetchs["top1"][1].avg

ppcls/static/run_dali.sh

@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+export FLAGS_fraction_of_gpu_memory_to_use=0.80
+
+python3.7 -m paddle.distributed.launch \
+    --gpus="0,1,2,3" \
+    ppcls/static//train.py \
+    -c ./ppcls/configs/ImageNet/ResNet/ResNet50_fp16.yaml \
+    -o Global.use_dali=True
+

ppcls/static/save_load.py

@@ -0,0 +1,139 @@
+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import errno
+import os
+import re
+import shutil
+import tempfile
+
+import paddle
+
+from ppcls.utils import logger
+
+__all__ = ['init_model', 'save_model']
+
+
+def _mkdir_if_not_exist(path):
+    """
+    mkdir if not exists, ignore the exception when multiprocess mkdir together
+    """
+    if not os.path.exists(path):
+        try:
+            os.makedirs(path)
+        except OSError as e:
+            if e.errno == errno.EEXIST and os.path.isdir(path):
+                logger.warning(
+                    'be happy if some process has already created {}'.format(
+                        path))
+            else:
+                raise OSError('Failed to mkdir {}'.format(path))
+
+
+def _load_state(path):
+    if os.path.exists(path + '.pdopt'):
+        # XXX another hack to ignore the optimizer state
+        tmp = tempfile.mkdtemp()
+        dst = os.path.join(tmp, os.path.basename(os.path.normpath(path)))
+        shutil.copy(path + '.pdparams', dst + '.pdparams')
+        state = paddle.static.load_program_state(dst)
+        shutil.rmtree(tmp)
+    else:
+        state = paddle.static.load_program_state(path)
+    return state
+
+
+def load_params(exe, prog, path, ignore_params=None):
+    """
+    Load model from the given path.
+    Args:
+        exe (fluid.Executor): The fluid.Executor object.
+        prog (fluid.Program): load weight to which Program object.
+        path (string): URL string or loca model path.
+        ignore_params (list): ignore variable to load when finetuning.
+            It can be specified by finetune_exclude_pretrained_params
+            and the usage can refer to the document
+            docs/advanced_tutorials/TRANSFER_LEARNING.md
+    """
+    if not (os.path.isdir(path) or os.path.exists(path + '.pdparams')):
+        raise ValueError("Model pretrain path {} does not "
+                         "exists.".format(path))
+
+    logger.info("Loading parameters from {}...".format(path))
+
+    ignore_set = set()
+    state = _load_state(path)
+
+    # ignore the parameter which mismatch the shape
+    # between the model and pretrain weight.
+    all_var_shape = {}
+    for block in prog.blocks:
+        for param in block.all_parameters():
+            all_var_shape[param.name] = param.shape
+    ignore_set.update([
+        name for name, shape in all_var_shape.items()
+        if name in state and shape != state[name].shape
+    ])
+
+    if ignore_params:
+        all_var_names = [var.name for var in prog.list_vars()]
+        ignore_list = filter(
+            lambda var: any([re.match(name, var) for name in ignore_params]),
+            all_var_names)
+        ignore_set.update(list(ignore_list))
+
+    if len(ignore_set) > 0:
+        for k in ignore_set:
+            if k in state:
+                logger.warning(
+                    'variable {} is already excluded automatically'.format(k))
+                del state[k]
+
+    paddle.static.set_program_state(prog, state)
+
+
+def init_model(config, program, exe):
+    """
+    load model from checkpoint or pretrained_model
+    """
+    checkpoints = config.get('checkpoints')
+    if checkpoints:
+        paddle.static.load(program, checkpoints, exe)
+        logger.info("Finish initing model from {}".format(checkpoints))
+        return
+
+    pretrained_model = config.get('pretrained_model')
+    if pretrained_model:
+        if not isinstance(pretrained_model, list):
+            pretrained_model = [pretrained_model]
+        for pretrain in pretrained_model:
+            load_params(exe, program, pretrain)
+        logger.info("Finish initing model from {}".format(pretrained_model))
+
+
+def save_model(program, model_path, epoch_id, prefix='ppcls'):
+    """
+    save model to the target path
+    """
+    if paddle.distributed.get_rank() != 0:
+        return
+    model_path = os.path.join(model_path, str(epoch_id))
+    _mkdir_if_not_exist(model_path)
+    model_prefix = os.path.join(model_path, prefix)
+    paddle.static.save(program, model_prefix)
+    logger.info("Already save model in {}".format(model_path))

ppcls/static/train.py

@@ -0,0 +1,197 @@
+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import os
+import sys
+__dir__ = os.path.dirname(os.path.abspath(__file__))
+sys.path.append(__dir__)
+sys.path.append(os.path.abspath(os.path.join(__dir__, '../../')))
+
+import paddle
+from paddle.distributed import fleet
+from visualdl import LogWriter
+
+from ppcls.data import build_dataloader
+from ppcls.utils.config import get_config, print_config
+from ppcls.utils import logger
+from ppcls.utils.logger import init_logger
+from ppcls.static.save_load import init_model, save_model
+from ppcls.static import program
+
+
+def parse_args():
+    parser = argparse.ArgumentParser("PaddleClas train script")
+    parser.add_argument(
+        '-c',
+        '--config',
+        type=str,
+        default='configs/ResNet/ResNet50.yaml',
+        help='config file path')
+    parser.add_argument(
+        '-o',
+        '--override',
+        action='append',
+        default=[],
+        help='config options to be overridden')
+    args = parser.parse_args()
+    return args
+
+
+def main(args):
+    """
+    all the config of training paradigm should be in config["Global"]
+    """
+    config = get_config(args.config, overrides=args.override, show=False)
+    global_config = config["Global"]
+
+    mode = "train"
+
+    log_file = os.path.join(global_config['output_dir'],
+                            config["Arch"]["name"], f"{mode}.log")
+    init_logger(name='root', log_file=log_file)
+    print_config(config)
+
+    if global_config.get("is_distributed", True):
+        fleet.init(is_collective=True)
+    # assign the device
+    use_gpu = global_config.get("use_gpu", True)
+    # amp related config
+    if 'AMP' in config:
+        AMP_RELATED_FLAGS_SETTING = {
+            'FLAGS_cudnn_exhaustive_search': "1",
+            'FLAGS_conv_workspace_size_limit': "1500",
+            'FLAGS_cudnn_batchnorm_spatial_persistent': "1",
+            'FLAGS_max_indevice_grad_add': "8",
+            "FLAGS_cudnn_batchnorm_spatial_persistent": "1",
+        }
+        for k in AMP_RELATED_FLAGS_SETTING:
+            os.environ[k] = AMP_RELATED_FLAGS_SETTING[k]
+
+    use_xpu = global_config.get("use_xpu", False)
+    assert (
+        use_gpu and use_xpu
+    ) is not True, "gpu and xpu can not be true in the same time in static mode!"
+
+    if use_gpu:
+        device = paddle.set_device('gpu')
+    elif use_xpu:
+        device = paddle.set_device('xpu')
+    else:
+        device = paddle.set_device('cpu')
+
+    # visualDL
+    vdl_writer = None
+    if global_config["use_visualdl"]:
+        vdl_dir = os.path.join(global_config["output_dir"], "vdl")
+        vdl_writer = LogWriter(vdl_dir)
+
+    # build dataloader
+    eval_dataloader = None
+    use_dali = global_config.get('use_dali', False)
+
+    train_dataloader = build_dataloader(
+        config["DataLoader"], "Train", device=device, use_dali=use_dali)
+    if global_config["eval_during_train"]:
+        eval_dataloader = build_dataloader(
+            config["DataLoader"], "Eval", device=device, use_dali=use_dali)
+
+    step_each_epoch = len(train_dataloader)
+
+    # startup_prog is used to do some parameter init work,
+    # and train prog is used to hold the network
+    startup_prog = paddle.static.Program()
+    train_prog = paddle.static.Program()
+
+    best_top1_acc = 0.0  # best top1 acc record
+
+    train_fetchs, lr_scheduler, train_feeds, optimizer = program.build(
+        config,
+        train_prog,
+        startup_prog,
+        step_each_epoch=step_each_epoch,
+        is_train=True,
+        is_distributed=global_config.get("is_distributed", True))
+
+    if global_config["eval_during_train"]:
+        eval_prog = paddle.static.Program()
+        eval_fetchs, _, eval_feeds, _ = program.build(
+            config,
+            eval_prog,
+            startup_prog,
+            is_train=False,
+            is_distributed=global_config.get("is_distributed", True))
+        # clone to prune some content which is irrelevant in eval_prog
+        eval_prog = eval_prog.clone(for_test=True)
+
+    # create the "Executor" with the statement of which device
+    exe = paddle.static.Executor(device)
+    # Parameter initialization
+    exe.run(startup_prog)
+    # load pretrained models or checkpoints
+    init_model(global_config, train_prog, exe)
+
+    if 'AMP' in config and config.AMP.get("use_pure_fp16", False):
+        optimizer.amp_init(
+            device,
+            scope=paddle.static.global_scope(),
+            test_program=eval_prog
+            if global_config["eval_during_train"] else None)
+
+    if not global_config.get("is_distributed", True):
+        compiled_train_prog = program.compile(
+            config, train_prog, loss_name=train_fetchs["loss"][0].name)
+    else:
+        compiled_train_prog = train_prog
+
+    if eval_dataloader is not None:
+        compiled_eval_prog = program.compile(config, eval_prog)
+
+    for epoch_id in range(global_config["epochs"]):
+        # 1. train with train dataset
+        program.run(train_dataloader, exe, compiled_train_prog, train_feeds,
+                    train_fetchs, epoch_id, 'train', config, vdl_writer,
+                    lr_scheduler)
+        # 2. evaate with eval dataset
+        if global_config["eval_during_train"] and epoch_id % global_config[
+                "eval_interval"] == 0:
+            top1_acc = program.run(eval_dataloader, exe, compiled_eval_prog,
+                                   eval_feeds, eval_fetchs, epoch_id, "eval",
+                                   config)
+            if top1_acc > best_top1_acc:
+                best_top1_acc = top1_acc
+                message = "The best top1 acc {:.5f}, in epoch: {:d}".format(
+                    best_top1_acc, epoch_id)
+                logger.info(message)
+                if epoch_id % global_config["save_interval"] == 0:
+
+                    model_path = os.path.join(global_config["output_dir"],
+                                              config["Arch"]["name"])
+                    save_model(train_prog, model_path, "best_model")
+
+        # 3. save the persistable model
+        if epoch_id % global_config["save_interval"] == 0:
+            model_path = os.path.join(global_config["output_dir"],
+                                      config["Arch"]["name"])
+            save_model(train_prog, model_path, epoch_id)
+
+
+if __name__ == '__main__':
+    paddle.enable_static()
+    args = parse_args()
+    main(args)