update resnet acc_mode scripts.

mindspore/nn/acc/acc.py

@@ -40,6 +40,7 @@ class AutoAcc:
     def __init__(self, level, kwargs):
         if level not in _acc_config_level.keys():
             level = 'O0'
+        self.level = level
         acc_config = _acc_config_level[level]
         self._acc_config = acc_config
         self._fn_flag = True

mindspore/nn/acc/base.py

@@ -73,16 +73,16 @@ class OptimizerProcess:
                     params_gc_value.append(param)
                 else:
                     params_value.append(param)
-                if params_gc_value:
+            if params_gc_value:
-                    new_group_param = copy.deepcopy(group_param)
+                new_group_param = copy.deepcopy(group_param)
-                    new_group_param['params'] = params_gc_value
+                new_group_param['params'] = params_gc_value
-                    new_group_param['grad_centralization'] = True
+                new_group_param['grad_centralization'] = True
-                    group_params.append(new_group_param)
+                group_params.append(new_group_param)
-                if params_value:
+            if params_value:
-                    new_group_param = copy.deepcopy(group_param)
+                new_group_param = copy.deepcopy(group_param)
-                    new_group_param['params'] = params_value
+                new_group_param['params'] = params_value
-                    group_params.append(new_group_param)
+                group_params.append(new_group_param)
-            self.origin_params = group_params
+        self.origin_params = group_params
 
     def generate_new_optimizer(self):
         """Generate new optimizer."""
@@ -135,6 +135,7 @@ class ParameterProcess:
         else:
             group_params = []
             params_name = [param.name for param in parameters]
+            new_params_count = copy.deepcopy(params_name)
             for group_param in origin_params_copy:
                 if 'order_params' in group_param.keys():
                     new_group_param = copy.deepcopy(group_param)
@@ -146,9 +147,16 @@ class ParameterProcess:
                     if param.name in params_name:
                         index = params_name.index(param.name)
                         params_value.append(parameters[index])
+                        new_params_count.remove(param.name)
                 new_group_param = copy.deepcopy(group_param)
                 new_group_param['params'] = params_value
                 group_params.append(new_group_param)
+            if new_params_count:
+                params_value = []
+                for param in new_params_count:
+                    index = params_name.index(param)
+                    params_value.append(parameters[index])
+                group_params.append({"params": params_value})
         return group_params
 
 _gradient_accumulation_op = C.MultitypeFuncGraph("gradient_accumulation_op")

mindspore/train/model.py

@@ -139,6 +139,7 @@ class Model:
         self._device_number = _get_device_num()
         self._global_rank = _get_global_rank()
         self._parameter_broadcast = _get_parameter_broadcast()
+        self._metrics = metrics
 
         self._check_amp_level_arg(optimizer, amp_level)
         self._check_for_graph_cell(kwargs)
@@ -175,7 +176,7 @@ class Model:
 
     def _check_kwargs(self, kwargs):
         for arg in kwargs:
-            if arg not in ['loss_scale_manager', 'keep_batchnorm_fp32', 'total_steps']:
+            if arg not in ['loss_scale_manager', 'keep_batchnorm_fp32']:
                 raise ValueError(f"Unsupported arg '{arg}'")
 
     def _check_reuse_dataset(self, dataset):
@@ -187,15 +188,18 @@ class Model:
     def _build_acc_network(self, kwargs):
         """Build the acc network."""
         processor = acc.AutoAcc(self._acc_level, kwargs)
+        if processor.level not in ["O1", "O2"]:
+            return
         if self._optimizer is None:
             logger.warning("In acc mode, the optimizer must be defined.")
             return
-        if self._eval_network is None:
+        if self._eval_network is None and self._metrics is None:
-            logger.warning("In acc mode, the eval_network must be defined.")
+            logger.warning("In acc mode, the eval_network and metrics cannot be undefined at the same time.")
             return
 
         self._network, self._optimizer = processor.network_auto_process_train(self._network, self._optimizer)
-        self._eval_network = processor.network_auto_process_eval(self._eval_network)
+        if self._eval_network is not None:
+            self._eval_network = processor.network_auto_process_eval(self._eval_network)
 
     def _build_train_network(self):
         """Build train network"""

model_zoo/official/cv/resnet/resnet101_imagenet2012_config.yaml

@@ -48,6 +48,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/resnet18_cifar10_config.yaml

@@ -48,6 +48,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/resnet18_imagenet2012_config.yaml

@@ -50,6 +50,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/resnet34_imagenet2012_config.yaml

@@ -50,6 +50,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/resnet50_cifar10_config.yaml

@@ -48,6 +48,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/resnet50_imagenet2012_Acc_config.yaml

@@ -0,0 +1,78 @@
+# Builtin Configurations(DO NOT CHANGE THESE CONFIGURATIONS unless you know exactly what you are doing)
+enable_modelarts: False
+# Url for modelarts
+data_url: ""
+train_url: ""
+checkpoint_url: ""
+# Path for local
+run_distribute: False
+enable_profiling: False
+data_path: "/cache/data"
+output_path: "/cache/train"
+load_path: "/cache/checkpoint_path/"
+device_target: "Ascend"
+checkpoint_path: "./checkpoint/"
+checkpoint_file_path: ""
+
+# ==============================================================================
+# Training options
+optimizer: "Momentum"
+infer_label: ""
+class_num: 1001
+batch_size: 256
+loss_scale: 1024
+momentum: 0.9
+weight_decay: 0.0001
+epoch_size: 90
+pretrain_epoch_size: 0
+save_checkpoint: True
+save_checkpoint_epochs: 5
+keep_checkpoint_max: 10
+warmup_epochs: 0
+lr_decay_mode: "linear"
+use_label_smooth: True
+label_smooth_factor: 0.1
+lr_init: 0
+lr_max: 0.8
+lr_end: 0.0
+
+net_name: "resnet50"
+dataset: "imagenet2012"
+device_num: 1
+pre_trained: ""
+run_eval: False
+eval_dataset_path: ""
+parameter_server: False
+filter_weight: False
+save_best_ckpt: True
+eval_start_epoch: 40
+eval_interval: 1
+enable_cache: False
+cache_session_id: ""
+mode_name: "GRAPH"
+acc_mode: "O1"
+
+# Export options
+device_id: 0
+width: 224
+height: 224
+file_name: "resnet50"
+file_format: "AIR"
+ckpt_file: ""
+network_dataset: "resnet50_imagenet2012"
+
+---
+# Help description for each configuration
+enable_modelarts: "Whether training on modelarts, default: False"
+data_url: "Dataset url for obs"
+checkpoint_url: "The location of checkpoint for obs"
+data_path: "Dataset path for local"
+output_path: "Training output path for local"
+load_path: "The location of checkpoint for obs"
+device_target: "Target device type, available: [Ascend, GPU, CPU]"
+enable_profiling: "Whether enable profiling while training, default: False"
+num_classes: "Class for dataset"
+batch_size: "Batch size for training and evaluation"
+epoch_size: "Total training epochs."
+checkpoint_path: "The location of the checkpoint file."
+checkpoint_file_path: "The location of the checkpoint file."

model_zoo/official/cv/resnet/resnet50_imagenet2012_GPU_config.yaml

@@ -51,6 +51,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/resnet50_imagenet2012_Thor_config.yaml

@@ -51,6 +51,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/resnet50_imagenet2012_config.yaml

@@ -50,6 +50,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/resnet_benchmark_GPU.yaml

@@ -25,6 +25,7 @@ eval: False
 save_ckpt: False
 mode_name: "GRAPH"
 dtype: "fp16"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/se-resnet50_imagenet2012_config.yaml

@@ -51,6 +51,7 @@ eval_interval: 1
 enable_cache: False
 cache_session_id: ""
 mode_name: "GRAPH"
+acc_mode: "O0"
 
 # Export options
 device_id: 0

model_zoo/official/cv/resnet/train.py

@@ -105,7 +105,8 @@ def set_parameter():
                                               gradients_mean=True)
             set_algo_parameters(elementwise_op_strategy_follow=True)
             if config.net_name == "resnet50" or config.net_name == "se-resnet50":
-                context.set_auto_parallel_context(all_reduce_fusion_config=[85, 160])
+                if config.acc_mode not in ["O1", "O2"]:
+                    context.set_auto_parallel_context(all_reduce_fusion_config=[85, 160])
             elif config.net_name == "resnet101":
                 context.set_auto_parallel_context(all_reduce_fusion_config=[80, 210, 313])
             init()
@@ -228,7 +229,8 @@ def train_net():
         model = Model(net, loss_fn=loss, optimizer=opt, metrics=metrics, eval_network=dist_eval_network)
     else:
         model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics=metrics,
-                      amp_level="O2", keep_batchnorm_fp32=False, eval_network=dist_eval_network)
+                      amp_level="O2", acc_level=config.acc_mode, keep_batchnorm_fp32=False,
+                      eval_network=dist_eval_network)
 
     if config.optimizer == "Thor" and config.dataset == "imagenet2012":
         from src.lr_generator import get_thor_damping