Customized Debugging Information


This section describes how to use the customized capabilities provided by MindSpore, such as callback, metrics,Print operator and log printing, to help you quickly debug the training network.

Introduction to Callback

Callback here is not a function but a class. You can use callback to observe the internal status and related information of the network during training or perform specific actions in a specific period. For example, you can monitor the loss, save model parameters, dynamically adjust parameters, and terminate training tasks in advance.

Callback Capabilities of MindSpore

MindSpore provides the callback capabilities to allow users to insert customized operations in a specific phase of training or inference, including:

  • Callback functions such as ModelCheckpoint, LossMonitor, and SummaryStep provided by the MindSpore framework

  • Custom callback functions

Usage: Transfer the callback object in the model.train method. The callback object can be a list, for example:

ckpt_cb = ModelCheckpoint()                                                            
loss_cb = LossMonitor()
summary_cb = SummaryStep()
model.train(epoch, dataset, callbacks=[ckpt_cb, loss_cb, summary_cb])

ModelCheckpoint can save model parameters for retraining or inference. LossMonitor can output loss information in logs for users to view. In addition, LossMonitor monitors the loss value change during training. When the loss value is Nan or Inf, the training terminates. SummaryStep can save the training information to a file for later use. During the training process, the callback list will execute the callback function in the defined order. Therefore, in the definition process, the dependency between callbacks needs to be considered.

Custom Callback

You can customize callback based on the callback base class as required.

The callback base class is defined as follows:

class Callback():
    """Callback base class""" 
    def begin(self, run_context):
        """Called once before the network executing."""

    def epoch_begin(self, run_context):
        """Called before each epoch beginning."""

    def epoch_end(self, run_context):
        """Called after each epoch finished.""" 

    def step_begin(self, run_context):
        """Called before each epoch beginning.""" 

    def step_end(self, run_context):
        """Called after each step finished."""

    def end(self, run_context):
        """Called once after network training."""

The callback can record important information during training and transfer the information to the callback object through a dictionary variable cb_params, You can obtain related attributes from each custom callback and perform customized operations. You can also customize other variables and transfer them to the cb_params object.

The main attributes of cb_params are as follows:

  • loss_fn: Loss function

  • optimizer: Optimizer

  • train_dataset: Training dataset

  • cur_epoch_num: Number of current epochs

  • cur_step_num: Number of current steps

  • batch_num: Number of steps in an epoch

You can inherit the callback base class to customize a callback object.

The following example describes how to use a custom callback function.

class StopAtTime(Callback):
    def __init__(self, run_time):
        super(StopAtTime, self).__init__()
        self.run_time = run_time*60

    def begin(self, run_context):
        cb_params = run_context.original_args()
        cb_params.init_time = time.time()
    def step_end(self, run_context):
        cb_params = run_context.original_args()
        epoch_num = cb_params.cur_epoch_num
        step_num = cb_params.cur_step_num
        loss = cb_params.net_outputs
	cur_time = time.time()
	if (cur_time - cb_params.init_time) > self.run_time:
            print("epoch: ", epoch_num, " step: ", step_num, " loss: ", loss)

stop_cb = StopAtTime(run_time=10)
model.train(100, dataset, callbacks=stop_cb)

The output is as follows:

epoch: 20 step: 32 loss: 2.298344373703003

This callback function is used to terminate the training within a specified period. You can use the run_context.original_args method to obtain the cb_params dictionary, which contains the main attribute information described above. In addition, you can modify and add values in the dictionary. In the preceding example, an init_time object is defined in begin and transferred to the cb_params dictionary. A decision is made at each step_end. When the training time is greater than the configured time threshold, a training termination signal will be sent to the run_context to terminate the training in advance and the current values of epoch, step, and loss will be printed.

MindSpore Metrics

After the training is complete, you can use metrics to evaluate the training result.

MindSpore provides multiple metrics, such as accuracy, loss, tolerance, recall, and F1.

You can define a metrics dictionary object that contains multiple metrics and transfer them to the model.eval interface to verify the training precision.

metrics = {
    'accuracy': nn.Accuracy(),
    'loss': nn.Loss(),
    'precision': nn.Precision(),
    'recall': nn.Recall(),
    'f1_score': nn.F1()
net = ResNet()
loss = CrossEntropyLoss()
opt = Momentum()
model = Model(net, loss_fn=loss, optimizer=opt, metrics=metrics, callbacks=TimeMonitor())
ds_eval = create_dataset()
output = model.eval(ds_eval)

The model.eval method returns a dictionary that contains the metrics and results transferred to the metrics.

The callback function can also be used in the eval process, and the user can call the related API or customize the callback method to achieve the desired function.

You can also define your own metrics class by inheriting the Metric base class and rewriting the clear, update, and eval methods.

The accuracy operator is used as an example to describe the internal implementation principle.

The accuracy inherits the EvaluationBase base class and rewrites the preceding three methods. The clear method initializes related calculation parameters in the class. The update method accepts the predicted value and tag value and updates the internal variables of accuracy. The eval method calculates related indicators and returns the calculation result. By invoking the eval method of accuracy, you will obtain the calculation result.

You can understand how accuracy runs by using the following code:

x = Tensor(np.array([[0.2, 0.5], [0.3, 0.1], [0.9, 0.6]]))
y = Tensor(np.array([1, 0, 1]))
metric = Accuracy()
metric.update(x, y)
accuracy = metric.eval()
print('Accuracy is ', accuracy)

The output is as follows:

Accuracy is 0.6667

MindSpore Print Operator

MindSpore-developed Print operator is used to print the tensors or character strings input by users. Multiple strings, multiple tensors, and a combination of tensors and strings are supported, which are separated by comma (,). The use method of MindSpore Print operator is the same that of other operators. You need to assert MindSpore Print operator in __init__ and invoke using construct. The following is an example.

import numpy as np
from mindspore import Tensor
from mindspore.ops import operations as P
import mindspore.nn as nn
import mindspore.context as context


class PrintDemo(nn.Cell):
    def __init__(self):
        super(PrintDemo, self).__init__()
        self.print = P.Print()

    def construct(self, x, y):
        self.print('print Tensor x and Tensor y:', x, y)
        return x

x = Tensor(np.ones([2, 1]).astype(np.int32))
y = Tensor(np.ones([2, 2]).astype(np.int32))
net = PrintDemo()
output = net(x, y)

The output is as follows:

print Tensor x and Tensor y:
Tensor shape:[[const vector][2, 1]]Int32
Tensor shape:[[const vector][2, 2]]Int32
val:[[1 1]
[1 1]]

Asynchronous Data Dump

When the training result deviates from the expectation on Ascend, the input and output of the operator can be dumped for debugging through Asynchronous Data Dump.

comm_ops operators are not supported by Asynchronous Data Dump. comm_ops can be found in Operator List.

  1. Turn on the switch to save graph IR: context.set_context(save_graphs=True).

  2. Execute training script.

  3. Open hwopt_d_end_graph_{graph id}.ir in the directory you execute the script and find the name of the operators you want to Dump.

  4. Configure json file: data_dump.json.

        "DumpSettings": {
            "net_name": "ResNet50",
            "dump_mode": 0,
            "op_debug_mode": 0,
            "iteration": 0,
            "kernels": ["Default/Conv2D-op2", "Default/TensorAdd-op10"]
        "DumpSettingsSpec": {
            "net_name": "net name eg:ResNet50",
            "dump_mode": "0: dump all kernels, 1: dump kernels in kernels list",
            "op_debug_mode": "0: close debug, 1: debug ai-core overflow, 2: debug atomic overflow, 3: debug all overflow",
            "iteration": "specified iteration",
            "kernels": "op's full scope name which need to be dump"
    • Iteration should be set to 0 when dataset_sink_mode is False and data of every iteration will be dumped.

    • Iteration should increase by 1 when dataset_sink_mode is True. For example, data of GetNext will be dumped in iteration 0 and data of compute graph will be dumped in iteration 1.

  5. Set environment variables.

    export ENABLE_DATA_DUMP=1
    export DATA_DUMP_PATH=/test
    export DATA_DUMP_CONFIG_PATH=data_dump.json
    • Set the environment variables before executing the training script. Setting environment variables during training will not take effect.

    • Dump environment variables need to be configured before calling

  6. Execute the training script again.

  7. Parse the Dump file.

    Change directory to /var/log/npu/ide_daemon/dump/ after training and execute the following commands to parse Dump data file:

    python /usr/local/Ascend/toolkit/tools/operator_cmp/compare/dump_data_conversion.pyc -type offline -target numpy -i ./{Dump file path}} -o ./{output file path}