!1588 GPU upadate resnet50 script in example

Merge pull request !1588 from VectorSL/r0.3

example/resnet50_cifar10/README.md

@@ -123,3 +123,15 @@ Inference result will be stored in the example path, whose folder name is "infer
 ```
 result: {'acc': 0.91446314102564111} ckpt=~/resnet50_cifar10/train_parallel0/resnet-90_195.ckpt
 ```
+
+### Running on GPU
+```
+# distributed training example
+mpirun -n 8 python train.py --dataset_path=~/cifar-10-batches-bin --device_target="GPU" --run_distribute=True
+
+# standalone training example
+python train.py --dataset_path=~/cifar-10-batches-bin --device_target="GPU"
+
+# infer example
+python eval.py --dataset_path=~/cifar10-10-verify-bin --device_target="GPU" --checkpoint_path=resnet-90_195.ckpt
+```
\ No newline at end of file

example/resnet50_cifar10/dataset.py

@@ -20,10 +20,11 @@ import mindspore.common.dtype as mstype
 import mindspore.dataset.engine as de
 import mindspore.dataset.transforms.vision.c_transforms as C
 import mindspore.dataset.transforms.c_transforms as C2
+from mindspore.communication.management import get_rank, get_group_size
 from config import config
 
 
-def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32):
+def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend"):
     """
     create a train or eval dataset
 
@@ -32,12 +33,17 @@ def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32):
         do_train(bool): whether dataset is used for train or eval.
         repeat_num(int): the repeat times of dataset. Default: 1
         batch_size(int): the batch size of dataset. Default: 32
+        target(str): the device target. Default: Ascend
 
     Returns:
         dataset
     """
-    device_num = int(os.getenv("DEVICE_NUM"))
-    rank_id = int(os.getenv("RANK_ID"))
+    if target == "Ascend":
+        device_num = int(os.getenv("DEVICE_NUM"))
+        rank_id = int(os.getenv("RANK_ID"))
+    else:
+        rank_id = get_rank()
+        device_num = get_group_size()
 
     if device_num == 1:
         ds = de.Cifar10Dataset(dataset_path, num_parallel_workers=8, shuffle=True)

example/resnet50_cifar10/eval.py

@@ -25,7 +25,7 @@ from mindspore.parallel._auto_parallel_context import auto_parallel_context
 from mindspore.nn.loss import SoftmaxCrossEntropyWithLogits
 from mindspore.train.model import Model, ParallelMode
 from mindspore.train.serialization import load_checkpoint, load_param_into_net
-from mindspore.communication.management import init
+from mindspore.communication.management import init, get_group_size
 
 parser = argparse.ArgumentParser(description='Image classification')
 parser.add_argument('--run_distribute', type=bool, default=False, help='Run distribute')
@@ -34,26 +34,32 @@ parser.add_argument('--do_train', type=bool, default=False, help='Do train or no
 parser.add_argument('--do_eval', type=bool, default=True, help='Do eval or not.')
 parser.add_argument('--checkpoint_path', type=str, default=None, help='Checkpoint file path')
 parser.add_argument('--dataset_path', type=str, default=None, help='Dataset path')
+parser.add_argument('--device_target', type=str, default='Ascend', help='Device target')
 args_opt = parser.parse_args()
 
-device_id = int(os.getenv('DEVICE_ID'))
-
-context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False)
-context.set_context(device_id=device_id)
-
 if __name__ == '__main__':
+    target = args_opt.device_target
+    context.set_context(mode=context.GRAPH_MODE, device_target=target, save_graphs=False)
     if not args_opt.do_eval and args_opt.run_distribute:
-        context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
-                                          mirror_mean=True)
-        auto_parallel_context().set_all_reduce_fusion_split_indices([140])
-        init()
+        if target == "Ascend":
+            device_id = int(os.getenv('DEVICE_ID'))
+            context.set_context(device_id=device_id)
+            context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
+                                              mirror_mean=True)
+            auto_parallel_context().set_all_reduce_fusion_split_indices([140])
+            init()
+        elif target == "GPU":
+            init("nccl")
+            context.set_auto_parallel_context(device_num=get_group_size(), parallel_mode=ParallelMode.DATA_PARALLEL,
+                                              mirror_mean=True)
 
     epoch_size = config.epoch_size
     net = resnet50(class_num=config.class_num)
     loss = SoftmaxCrossEntropyWithLogits(sparse=True)
 
     if args_opt.do_eval:
-        dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=False, batch_size=config.batch_size)
+        dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=False, batch_size=config.batch_size,
+                                 target=target)
         step_size = dataset.get_dataset_size()
 
         if args_opt.checkpoint_path:

example/resnet50_cifar10/train.py

@@ -29,7 +29,7 @@ from mindspore.train.model import Model, ParallelMode
 
 from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor
 from mindspore.train.loss_scale_manager import FixedLossScaleManager
-from mindspore.communication.management import init
+from mindspore.communication.management import init, get_rank, get_group_size
 
 parser = argparse.ArgumentParser(description='Image classification')
 parser.add_argument('--run_distribute', type=bool, default=False, help='Run distribute')
@@ -37,28 +37,37 @@ parser.add_argument('--device_num', type=int, default=1, help='Device num.')
 parser.add_argument('--do_train', type=bool, default=True, help='Do train or not.')
 parser.add_argument('--do_eval', type=bool, default=False, help='Do eval or not.')
 parser.add_argument('--dataset_path', type=str, default=None, help='Dataset path')
+parser.add_argument('--device_target', type=str, default='Ascend', help='Device target')
 args_opt = parser.parse_args()
 
-device_id = int(os.getenv('DEVICE_ID'))
-
-context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id,
-                    enable_auto_mixed_precision=True)
 
 if __name__ == '__main__':
+    target = args_opt.device_target
     if not args_opt.do_eval and args_opt.run_distribute:
-        context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
-                                          mirror_mean=True)
-        auto_parallel_context().set_all_reduce_fusion_split_indices([107, 160])
-        init()
+        if target == "Ascend":
+            device_id = int(os.getenv('DEVICE_ID'))
+            context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id,
+                                enable_auto_mixed_precision=True)
+            init()
+            context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
+                                              mirror_mean=True)
+            auto_parallel_context().set_all_reduce_fusion_split_indices([107, 160])
+            ckpt_save_dir = config.save_checkpoint_path
+            loss = SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
+        elif target == "GPU":
+            context.set_context(mode=context.GRAPH_MODE, device_target="GPU", save_graphs=False)
+            init("nccl")
+            context.set_auto_parallel_context(device_num=get_group_size(), parallel_mode=ParallelMode.DATA_PARALLEL,
+                                              mirror_mean=True)
+            ckpt_save_dir = config.save_checkpoint_path + "ckpt_" + str(get_rank()) + "/"
+            loss = SoftmaxCrossEntropyWithLogits(sparse=True, is_grad=False, reduction='mean')
 
     epoch_size = config.epoch_size
     net = resnet50(class_num=config.class_num)
-    loss = SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
-
 
     if args_opt.do_train:
         dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=True,
-                                 repeat_num=epoch_size, batch_size=config.batch_size)
+                                 repeat_num=epoch_size, batch_size=config.batch_size, target=target)
         step_size = dataset.get_dataset_size()
 
         loss_scale = FixedLossScaleManager(config.loss_scale, drop_overflow_update=False)
@@ -67,9 +76,11 @@ if __name__ == '__main__':
                            lr_decay_mode='poly'))
         opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr, config.momentum,
                        config.weight_decay, config.loss_scale)
-
-        model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'}, amp_level="O2",
-                      keep_batchnorm_fp32=False)
+        if target == 'GPU':
+            model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'})
+        else:
+            model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'},
+                          amp_level="O2", keep_batchnorm_fp32=True)
 
         time_cb = TimeMonitor(data_size=step_size)
         loss_cb = LossMonitor()
@@ -77,6 +88,6 @@ if __name__ == '__main__':
         if config.save_checkpoint:
             config_ck = CheckpointConfig(save_checkpoint_steps=config.save_checkpoint_steps,
                                          keep_checkpoint_max=config.keep_checkpoint_max)
-            ckpt_cb = ModelCheckpoint(prefix="resnet", directory=config.save_checkpoint_path, config=config_ck)
+            ckpt_cb = ModelCheckpoint(prefix="resnet", directory=ckpt_save_dir, config=config_ck)
             cb += [ckpt_cb]
         model.train(epoch_size, dataset, callbacks=cb)

example/resnet50_imagenet2012/README.md

@@ -133,3 +133,18 @@ Inference result will be stored in the example path, whose folder name is "infer
 ```
 result: {'acc': 0.7671054737516005} ckpt=train_parallel0/resnet-90_5004.ckpt
 ```
+
+### Running on GPU
+```
+# distributed training example
+mpirun -n 8 python train.py --dataset_path=dataset/ilsvrc/train --device_target="GPU" --run_distribute=True
+
+# standalone training example
+python train.py --dataset_path=dataset/ilsvrc/train --device_target="GPU"
+
+# standalone training example with pretrained checkpoint
+python train.py --dataset_path=dataset/ilsvrc/train --device_target="GPU" --pre_trained=pretrained.ckpt
+
+# infer example
+python eval.py --dataset_path=dataset/ilsvrc/val --device_target="GPU" --checkpoint_path=resnet-90_5004ss.ckpt
+```
\ No newline at end of file

example/resnet50_imagenet2012/dataset.py

@@ -20,9 +20,9 @@ import mindspore.common.dtype as mstype
 import mindspore.dataset.engine as de
 import mindspore.dataset.transforms.vision.c_transforms as C
 import mindspore.dataset.transforms.c_transforms as C2
+from mindspore.communication.management import get_rank, get_group_size
 
-
-def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32):
+def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32, target="Ascend"):
     """
     create a train or eval dataset
 
@@ -31,12 +31,17 @@ def create_dataset(dataset_path, do_train, repeat_num=1, batch_size=32):
         do_train(bool): whether dataset is used for train or eval.
         repeat_num(int): the repeat times of dataset. Default: 1
         batch_size(int): the batch size of dataset. Default: 32
+        target(str): the device target. Default: Ascend
 
     Returns:
         dataset
     """
-    device_num = int(os.getenv("DEVICE_NUM"))
-    rank_id = int(os.getenv("RANK_ID"))
+    if target == "Ascend":
+        device_num = int(os.getenv("DEVICE_NUM"))
+        rank_id = int(os.getenv("RANK_ID"))
+    else:
+        rank_id = get_rank()
+        device_num = get_group_size()
 
     if device_num == 1:
         ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True)

example/resnet50_imagenet2012/eval.py

@@ -32,12 +32,13 @@ parser.add_argument('--do_train', type=bool, default=False, help='Do train or no
 parser.add_argument('--do_eval', type=bool, default=True, help='Do eval or not.')
 parser.add_argument('--checkpoint_path', type=str, default=None, help='Checkpoint file path')
 parser.add_argument('--dataset_path', type=str, default=None, help='Dataset path')
+parser.add_argument('--device_target', type=str, default='Ascend', help='Device target')
 args_opt = parser.parse_args()
-
-device_id = int(os.getenv('DEVICE_ID'))
-
-context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False)
-context.set_context(device_id=device_id)
+target = args_opt.device_target
+context.set_context(mode=context.GRAPH_MODE, device_target=target, save_graphs=False)
+if target == "Ascend":
+    device_id = int(os.getenv('DEVICE_ID'))
+    context.set_context(device_id=device_id)
 
 if __name__ == '__main__':
 
@@ -47,7 +48,8 @@ if __name__ == '__main__':
     loss = CrossEntropy(smooth_factor=config.label_smooth_factor, num_classes=config.class_num)
 
     if args_opt.do_eval:
-        dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=False, batch_size=config.batch_size)
+        dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=False, batch_size=config.batch_size,
+                                 target=target)
         step_size = dataset.get_dataset_size()
 
         if args_opt.checkpoint_path:

example/resnet50_imagenet2012/train.py

@@ -29,7 +29,7 @@ from mindspore.train.model import Model, ParallelMode
 from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor
 from mindspore.train.loss_scale_manager import FixedLossScaleManager
 from mindspore.train.serialization import load_checkpoint, load_param_into_net
-from mindspore.communication.management import init
+from mindspore.communication.management import init, get_rank, get_group_size
 import mindspore.nn as nn
 import mindspore.common.initializer as weight_init
 from crossentropy import CrossEntropy
@@ -40,21 +40,28 @@ parser.add_argument('--device_num', type=int, default=1, help='Device num.')
 parser.add_argument('--do_train', type=bool, default=True, help='Do train or not.')
 parser.add_argument('--do_eval', type=bool, default=False, help='Do eval or not.')
 parser.add_argument('--dataset_path', type=str, default=None, help='Dataset path')
+parser.add_argument('--device_target', type=str, default='Ascend', help='Device target')
 parser.add_argument('--pre_trained', type=str, default=None, help='Pretrained checkpoint path')
 args_opt = parser.parse_args()
 
-device_id = int(os.getenv('DEVICE_ID'))
-
-
-context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id,
-                    enable_auto_mixed_precision=True)
-
 if __name__ == '__main__':
+    target = args_opt.device_target
     if not args_opt.do_eval and args_opt.run_distribute:
-        context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
-                                          mirror_mean=True, parameter_broadcast=True)
-        auto_parallel_context().set_all_reduce_fusion_split_indices([107, 160])
-        init()
+        if target == "Ascend":
+            device_id = int(os.getenv('DEVICE_ID'))
+            context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id,
+                                enable_auto_mixed_precision=True)
+            init()
+            context.set_auto_parallel_context(device_num=args_opt.device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
+                                              mirror_mean=True)
+            auto_parallel_context().set_all_reduce_fusion_split_indices([107, 160])
+            ckpt_save_dir = config.save_checkpoint_path
+        elif target == "GPU":
+            context.set_context(mode=context.GRAPH_MODE, device_target="GPU", save_graphs=False)
+            init("nccl")
+            context.set_auto_parallel_context(device_num=get_group_size(), parallel_mode=ParallelMode.DATA_PARALLEL,
+                                              mirror_mean=True)
+            ckpt_save_dir = config.save_checkpoint_path + "ckpt_" + str(get_rank()) + "/"
 
     epoch_size = config.epoch_size
     net = resnet50(class_num=config.class_num)
@@ -81,7 +88,7 @@ if __name__ == '__main__':
 
     if args_opt.do_train:
         dataset = create_dataset(dataset_path=args_opt.dataset_path, do_train=True,
-                                 repeat_num=epoch_size, batch_size=config.batch_size)
+                                 repeat_num=epoch_size, batch_size=config.batch_size, target=target)
         step_size = dataset.get_dataset_size()
 
         loss_scale = FixedLossScaleManager(config.loss_scale, drop_overflow_update=False)
@@ -93,9 +100,11 @@ if __name__ == '__main__':
 
         opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr, config.momentum,
                        config.weight_decay, config.loss_scale)
-
-        model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'}, amp_level="O2",
-                      keep_batchnorm_fp32=False)
+        if target == "Ascend":
+            model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'},
+                          amp_level="O2", keep_batchnorm_fp32=False)
+        elif target == "GPU":
+            model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics={'acc'})
 
 
         time_cb = TimeMonitor(data_size=step_size)
@@ -104,6 +113,6 @@ if __name__ == '__main__':
         if config.save_checkpoint:
             config_ck = CheckpointConfig(save_checkpoint_steps=config.save_checkpoint_epochs*step_size,
                                          keep_checkpoint_max=config.keep_checkpoint_max)
-            ckpt_cb = ModelCheckpoint(prefix="resnet", directory=config.save_checkpoint_path, config=config_ck)
+            ckpt_cb = ModelCheckpoint(prefix="resnet", directory=ckpt_save_dir, config=config_ck)
             cb += [ckpt_cb]
         model.train(epoch_size, dataset, callbacks=cb)