mindspore.Profiler

class mindspore.Profiler(**kwargs)[source]

This class to enable the profiling of MindSpore neural networks. MindSpore users can import the mindspore.Profiler, initialize the Profiler object to start profiling, and use Profiler.analyse() to stop profiling and analyse the results. Users can visualize the results using the MindInsight tool. Now, Profiler supports AICORE operator, AICPU operator, HostCPU operator, memory, correspondence, cluster, etc data analysis.

Parameters
  • output_path (str, optional) – Output data path. Default: “./data”.

  • op_time (bool, optional) – (Ascend/GPU) Whether to collect operators performance data. Default value: True.

  • profile_communication (bool, optional) – (Ascend only) Whether to collect communication performance data in a multi devices training,collect when True. Setting this parameter has no effect during single device training. When using this parameter, op_time must be set to True. Default: False.

  • profile_memory (bool, optional) – (Ascend only) Whether to collect tensor memory data, collect when True. When using this parameter, op_time must be set to True. Default: False.

  • parallel_strategy (bool, optional) – (Ascend only) Whether to collect parallel policy performance data. Default value: true.

  • start_profile (bool, optional) – The start_profile parameter controls whether to enable or disable performance data collection based on conditions. Default: True.

  • aicore_metrics (int, optional) –

    (Ascend only) Types of AICORE performance data collected, when using this parameter, op_time must be set to True, and the value must be in [-1, 0, 1, 2, 3, 4, 5], Default: 0, the data items contained in each metric are as follows:

    • -1: Does not collect AICORE data.

    • 0: ArithmeticUtilization contains mac_fp16/int8_ratio, vec_fp32/fp16/int32_ratio, vec_misc_ratio etc.

    • 1: PipeUtilization contains vec_ratio, mac_ratio, scalar_ratio, mte1/mte2/mte3_ratio, icache_miss_rate etc.

    • 2: Memory contains ub_read/write_bw, l1_read/write_bw, l2_read/write_bw, main_mem_read/write_bw etc.

    • 3: MemoryL0 contains l0a_read/write_bw, l0b_read/write_bw, l0c_read/write_bw etc.

    • 4: ResourceConflictRatio contains vec_bankgroup/bank/resc_cflt_ratio etc.

    • 5: MemoryUB contains ub_read/write_bw_mte, ub_read/write_bw_vector, ub_/write_bw_scalar etc.

  • l2_cache (bool, optional) – (Ascend only) Whether to collect l2 cache data, collect when True. Default: False.

  • sync_enable (bool, optional) –

    (GPU only) Whether the profiler collects operators in a synchronous way. Default: True.

    • True: The synchronous way. Before sending the operator to the GPU, the CPU records the start timestamp. Then the operator is returned to the CPU after execution, and the end timestamp is recorded, The duration of the operator is the difference between the two timestamps.

    • False: The asynchronous way. The duration of the operator is that of sending from the CPU to the GPU. This method can reduce the impact of adding profiler on overall training time.

  • data_process (bool, optional) – (Ascend/GPU) Whether to collect data to prepare performance data. Default value: True.

  • timeline_limit (int, optional) – Set the maximum storage size of the timeline file (unit M). When using this parameter, op_time must be set to True. Default value: 500.

Raises

RuntimeError – When the version of CANN does not match the version of MindSpore, MindSpore cannot parse the generated ascend_job_id directory structure.

Supported Platforms:

Ascend GPU

Examples

>>> import numpy as np
>>> import mindspore as ms
>>> from mindspore import nn
>>> import mindspore.dataset as ds
>>> from mindspore import Profiler
>>>
>>>
>>> class Net(nn.Cell):
...     def __init__(self):
...         super(Net, self).__init__()
...         self.fc = nn.Dense(2,2)
...     def construct(self, x):
...         return self.fc(x)
>>>
>>> def generator():
...     for i in range(2):
...         yield (np.ones([2, 2]).astype(np.float32), np.ones([2]).astype(np.int32))
>>>
>>> def train(net):
...     optimizer = nn.Momentum(net.trainable_params(), 1, 0.9)
...     loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True)
...     data = ds.GeneratorDataset(generator, ["data", "label"])
...     model = ms.Model(net, loss, optimizer)
...     model.train(1, data)
>>>
>>> if __name__ == '__main__':
...     # If the device_target is GPU, set the device_target to "GPU"
...     ms.set_context(mode=ms.GRAPH_MODE, device_target="Ascend")
...
...     # Init Profiler
...     # Note that the Profiler should be initialized before model.train
...     profiler = Profiler()
...
...     # Train Model
...     net = Net()
...     train(net)
...
...     # Profiler end
...     profiler.analyse()
analyse()[source]

Collect and analyze training performance data, support calls during and after training. The example shows above.

op_analyse(op_name, device_id=None)[source]

Profiler users can use this interface to obtain operator performance data.

Parameters
  • op_name (str or list) – The primitive operator name to query.

  • device_id (int, optional) – ID of the target device. This parameter is optional during network training or inference, and users can use device_id parameter to specify which card operator performance data to parse. If this interface is used for offline data parsing, Default: 0.

Raises
  • TypeError – If the op_name parameter type is incorrect.

  • TypeError – If the device_id parameter type is incorrect.

  • RuntimeError – If MindSpore runs on Ascend, this interface cannot be used.

Supported Platforms:

GPU CPU

Examples

>>> from mindspore import Profiler
>>>
...     # Profiler init.
...     profiler = Profiler()
...
...     # Train Model or eval Model, taking LeNet5 as an example
...     # Refer to https://gitee.com/mindspore/docs/blob/r2.0/docs/mindspore/code/lenet.py
...     net = LeNet5()
...     train(net)
...
...     # Profiler end
...     profiler.analyse()
...
...     profiler.op_analyse(op_name=["BiasAdd", "Conv2D"])
...
>>> from mindspore import Profiler
>>>
...     # Profiler init.
...     profiler = Profiler(output_path="my_profiler_path")
...     profiler.op_analyse(op_name="Conv2D")
start()[source]

Used for Ascend, GPU, start profiling. Profiling can be turned on based on step and epoch.

Raises
  • RuntimeError – If the profiler has already started.

  • RuntimeError – If MD profiling has stopped, repeated start action is not supported.

  • RuntimeError – If the start_profile parameter is not set or is set to True.

Examples

>>> class StopAtStep(Callback):
...     def __init__(self, start_step, stop_step):
...         super(StopAtStep, self).__init__()
...         self.start_step = start_step
...         self.stop_step = stop_step
...         self.profiler = Profiler(start_profile=False)
...
...     def step_begin(self, run_context):
...         cb_params = run_context.original_args()
...         step_num = cb_params.cur_step_num
...         if step_num == self.start_step:
...             self.profiler.start()
...
...     def step_end(self, run_context):
...         cb_params = run_context.original_args()
...         step_num = cb_params.cur_step_num
...         if step_num == self.stop_step:
...             self.profiler.stop()
...
...     def end(self, run_context):
...         self.profiler.analyse()
stop()[source]

Used for Ascend, GPU, stop profiling. Profiling can be turned off based on step and epoch.

Raises

RuntimeError – If the profiler has not started, this function is disabled.

Examples

>>> class StopAtEpoch(Callback):
...     def __init__(self, start_epoch, stop_epoch):
...         super(StopAtEpoch, self).__init__()
...         self.start_epoch = start_epoch
...         self.stop_epoch = stop_epoch
...         self.profiler = Profiler(start_profile=False)
...
...     def epoch_begin(self, run_context):
...         cb_params = run_context.original_args()
...         epoch_num = cb_params.cur_epoch_num
...         if epoch_num == self.start_epoch:
...             self.profiler.start()
...
...     def epoch_end(self, run_context):
...         cb_params = run_context.original_args()
...         epoch_num = cb_params.cur_epoch_num
...         if epoch_num == self.stop_epoch:
...             self.profiler.stop()
...
...     def end(self, run_context):
...         self.profiler.analyse()