add device num for xpu (#581)

* add device num for xpu
* fix vgg

ppcls/data/reader.py

@@ -25,7 +25,7 @@ from . import imaug
 from .imaug import transform
 from ppcls.utils import logger
 
-trainers_num = int(os.environ.get('PADDLE_TRAINERS_NUM', 0))
+trainers_num = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
 trainer_id = int(os.environ.get("PADDLE_TRAINER_ID", 0))
 
 
@@ -257,6 +257,9 @@ class Reader:
             raise ModeException(mode=mode)
 
         self.use_gpu = config.get("use_gpu", True)
+        self.use_xpu = config.get("use_xpu", False)
+        self.is_distributed = config.get("is_distributed", True)
+
         use_mix = config.get('use_mix')
         self.params['mode'] = mode
         if seed is not None:
@@ -265,14 +268,20 @@ class Reader:
         if use_mix and mode == "train":
             self.batch_ops = create_operators(self.params['mix'])
 
+    def get_device_num(self):
+        if self.is_distributed:
+            device_num = trainers_num
+        elif self.use_gpu:
+            device_num = fluid.core.get_cuda_device_count()
+        elif self.use_xpu:
+            xpus = os.environ.get("FLAGS_selected_xpus", '1')
+            device_num = len(xpus.split(','))
+        else:
+            device_num = int(os.environ.get('CPU_NUM', 1))
+        return device_num
+
     def __call__(self):
-        device_num = trainers_num
-        # non-distributed launch
-        if trainers_num <= 0:
-            if self.use_gpu:
-                device_num = fluid.core.get_cuda_device_count()
-            else:
-                device_num = int(os.environ.get('CPU_NUM', 1))
+        device_num = self.get_device_num()
         batch_size = int(self.params['batch_size']) // device_num
 
         def wrapper():

ppcls/modeling/architectures/vgg.py

@@ -23,8 +23,9 @@ __all__ = ["VGGNet", "VGG11", "VGG13", "VGG16", "VGG19"]
 
 
 class VGGNet():
-    def __init__(self, layers=16):
+    def __init__(self, layers=16, stop_grad_layers=0, **args):
         self.layers = layers
+        self.stop_grad_layers = stop_grad_layers
 
     def net(self, input, class_dim=1000):
         layers = self.layers
@@ -44,6 +45,10 @@ class VGGNet():
         conv4 = self.conv_block(conv3, 512, nums[3], name="conv4_")
         conv5 = self.conv_block(conv4, 512, nums[4], name="conv5_")
 
+        for idx, conv in enumerate([conv1, conv2, conv3, conv4, conv5]):
+            if self.stop_grad_layers >= idx + 1:
+                conv.stop_gradient = True
+
         fc_dim = 4096
         fc_name = ["fc6", "fc7", "fc8"]
         fc1 = fluid.layers.fc(
@@ -88,21 +93,21 @@ class VGGNet():
             input=conv, pool_size=2, pool_type='max', pool_stride=2)
 
 
-def VGG11():
-    model = VGGNet(layers=11)
+def VGG11(stop_grad_layers=0, **args):
+    model = VGGNet(layers=11, stop_grad_layers=stop_grad_layers, **args)
     return model
 
 
-def VGG13():
-    model = VGGNet(layers=13)
+def VGG13(stop_grad_layers=0, **args):
+    model = VGGNet(layers=13, stop_grad_layers=stop_grad_layers, **args)
     return model
 
 
-def VGG16():
-    model = VGGNet(layers=16)
+def VGG16(stop_grad_layers=0, **args):
+    model = VGGNet(layers=16, stop_grad_layers=stop_grad_layers, **args)
     return model
 
 
-def VGG19():
-    model = VGGNet(layers=19)
+def VGG19(stop_grad_layers=0, **args):
+    model = VGGNet(layers=19, stop_grad_layers=stop_grad_layers, **args)
     return model

tools/program.py

@@ -282,6 +282,7 @@ def create_optimizer(config):
     # create optimizer instance
     opt_config = config['OPTIMIZER']
     opt = OptimizerBuilder(**opt_config)
+
     return opt(lr)
 
 
@@ -305,7 +306,6 @@ def dist_optimizer(config, optimizer):
     dist_strategy.fuse_all_reduce_ops = True
     dist_strategy.exec_strategy = exec_strategy
     optimizer = fleet.distributed_optimizer(optimizer, strategy=dist_strategy)
-
     return optimizer