Collecting Summary Record

Linux Ascend GPU CPU Model Optimization Intermediate Expert

Overview

Scalars, images, computational graphs, and model hyperparameters during training are recorded in files and can be viewed on the web page.

Operation Process

Prepare a training script, specify scalars, images, computational graphs, and model hyperparameters in the training script, record them in the summary log file, and run the training script.
Start MindInsight and specify the summary log file directory using startup parameters. After MindInsight is started, access the visualization page based on the IP address and port number. The default access IP address is http://127.0.0.1:8080.
During the training, when data is written into the summary log file, you can view the data on the web page.

Preparing The Training Script

Currently, MindSpore supports to save scalars, images, computational graph, and model hyperparameters to summary log file and display them on the web page.

MindSpore currently supports multiple ways to record data into summary log files.

Method one: Automatically collected through SummaryCollector

The Callback mechanism in MindSpore provides a quick and easy way to collect common information, including the calculational graph, loss value, learning rate, parameter weights, etc. It is named ‘SummaryCollector’.

When you write a training script, you just instantiate the SummaryCollector and apply it to either model.train or model.eval. You can automatically collect some common summary data. The detailed usage of SummaryCollector can refer to the API document mindspore.train.callback.SummaryCollector.

The sample code is as follows:

import mindspore
import mindspore.nn as nn
from mindspore import ops
from mindspore import context, Tensor, Model
from mindspore.nn.metrics import Accuracy
from mindspore.train.callback import SummaryCollector


def conv(in_channels, out_channels, kernel_size, stride=1, padding=0, pad_mode="valid", has_bias=True):
    return nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding,
                     has_bias=has_bias, pad_mode=pad_mode)


def fc_with_initialize(input_channels, out_channels, has_bias=True):
    return nn.Dense(input_channels, out_channels, has_bias=has_bias)


class AlexNet(nn.Cell):
    """AlexNet"""
    def __init__(self, num_classes=10, channel=3, phase='train', include_top=True):
        super(AlexNet, self).__init__()
        self.conv1 = conv(channel, 64, 11, stride=4, pad_mode="same", has_bias=True)
        self.conv2 = conv(64, 128, 5, pad_mode="same", has_bias=True)
        self.conv3 = conv(128, 192, 3, pad_mode="same", has_bias=True)
        self.conv4 = conv(192, 256, 3, pad_mode="same", has_bias=True)
        self.conv5 = conv(256, 256, 3, pad_mode="same", has_bias=True)
        self.relu = ops.ReLU()
        self.max_pool2d = nn.MaxPool2d(kernel_size=3, stride=2, pad_mode='valid')
        self.include_top = include_top
        if self.include_top:
            dropout_ratio = 0.65
            if phase == 'test':
                dropout_ratio = 1.0
            self.flatten = nn.Flatten()
            self.fc1 = fc_with_initialize(6 * 6 * 256, 4096)
            self.fc2 = fc_with_initialize(4096, 4096)
            self.fc3 = fc_with_initialize(4096, num_classes)
            self.dropout = nn.Dropout(dropout_ratio)

    def construct(self, x):
        """define network"""
        x = self.conv1(x)
        x = self.relu(x)
        x = self.max_pool2d(x)
        x = self.conv2(x)
        x = self.relu(x)
        x = self.max_pool2d(x)
        x = self.conv3(x)
        x = self.relu(x)
        x = self.conv4(x)
        x = self.relu(x)
        x = self.conv5(x)
        x = self.relu(x)
        x = self.max_pool2d(x)
        if not self.include_top:
            return x
        x = self.flatten(x)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc2(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc3(x)
        return x


context.set_context(mode=context.GRAPH_MODE)

network = AlexNet(num_classes=10)
loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
lr = Tensor(0.5, mindspore.float32)
opt = nn.Momentum(network.trainable_params(), lr, momentum=0.9)
model = Model(network, loss, opt, metrics={"Accuracy": Accuracy()})

# How to create a valid dataset instance,
# for detail, see the https://www.mindspore.cn/tutorial/training/en/r1.1/quick_start/quick_start.html document.
ds_train = create_dataset('./dataset_path')

# Init a SummaryCollector callback instance, and use it in model.train or model.eval
summary_collector = SummaryCollector(summary_dir='./summary_dir', collect_freq=1)

# Note: dataset_sink_mode should be set to False, else you should modify collect freq in SummaryCollector
model.train(epoch=1, train_dataset=ds_train, callbacks=[summary_collector], dataset_sink_mode=False)

ds_eval = create_dataset('./dataset_path')
model.eval(ds_eval, callbacks=[summary_collector])

When using summary, it is recommended that you set dataset_sink_mode argument of model.train() to False. Please see notices for more information.

Method two: Custom collection of network data with summary operators and SummaryCollector

In addition to providing the SummaryCollector that automatically collects some summary data, MindSpore provides summary operators that enable customized collection of other data on the network, such as the input of each convolutional layer, or the loss value in the loss function, etc.

The following summary operators are currently supported:

ScalarSummary: Record a scalar data.
TensorSummary: Record a tensor data.
ImageSummary: Record a image data.
HistogramSummary: Convert tensor data into histogram data records.

The recording method is shown in the following steps.

Step 1: Call the summary operator in the construct function of the derived class that inherits nn.Cell to collect image or scalar data.

For example, when a network is defined, image data is recorded in construct of the network. When the loss function is defined, the loss value is recorded in construct of the loss function.

Record the dynamic learning rate in construct of the optimizer when defining the optimizer.

The sample code is as follows:

import mindspore
import mindspore.ops as ops
from mindspore import Tensor, nn
from mindspore.nn import Optimizer


class CrossEntropyLoss(nn.Cell):
    """Loss function definition."""
    def __init__(self):
        super(CrossEntropyLoss, self).__init__()
        self.cross_entropy = ops.SoftmaxCrossEntropyWithLogits()
        self.mean = ops.ReduceMean()
        self.one_hot = ops.OneHot()
        self.on_value = Tensor(1.0, mindspore.float32)
        self.off_value = Tensor(0.0, mindspore.float32)

        # Init ScalarSummary
        self.scalar_summary = ops.ScalarSummary()

    def construct(self, logits, label):
        label = self.one_hot(label, ops.shape(logits)[1], self.on_value, self.off_value)
        loss = self.cross_entropy(logits, label)[0]
        loss = self.mean(loss, (-1,))

        # Record loss
        self.scalar_summary("loss", loss)
        return loss


class MyOptimizer(Optimizer):
    """Optimizer definition."""
    def __init__(self, learning_rate, params, ...):
        ...
        # Initialize ScalarSummary
        self.scalar_summary = ops.ScalarSummary()
        self.histogram_summary = ops.HistogramSummary()
        self.weight_names = [param.name for param in self.parameters]

    def construct(self, grads):
        ...
        # Record learning rate here
        self.scalar_summary("learning_rate", learning_rate)

        # Record weight
        self.histogram_summary(self.weight_names[0], self.parameters[0])
        # Record gradient
        self.histogram_summary(self.weight_names[0] + ".gradient", grads[0])

        ...

class Net(nn.Cell):
    """Net definition."""
    def __init__(self):
        super(Net, self).__init__()
        ...

        # Init ImageSummary
        self.image_summary = ops.ImageSummary()
        # Init TensorSummary
        self.tensor_summary = ops.TensorSummary()

    def construct(self, data):
        # Record image by Summary operator
        self.image_summary("image", data)
        # Record tensor by Summary operator
        self.tensor_summary("tensor", data)
        ...
        return out

In the same Summary operator, the name given to the data must not be repeated, otherwise the data collection and presentation will have unexpected behavior. For example, if two ScalarSummary operators are used to collect scalar data, two scalars cannot be given the same name.

Step 2: In the training script, instantiate the SummaryCollector and apply it to model.train.

The sample code is as follows:

from mindspore import Model, nn, context
from mindspore.train.callback import SummaryCollector
...

context.set_context(mode=context.GRAPH_MODE)
network = Net()
loss_fn = CrossEntropyLoss()
optim = MyOptimizer(learning_rate=0.01, params=network.trainable_params())
model = Model(network, loss_fn=loss_fn, optimizer=optim, metrics={"Accuracy": Accuracy()})

ds_train = create_dataset('./dataset_path')

summary_collector = SummaryCollector(summary_dir='./summary_dir', collect_freq=1)
model.train(epoch=2, train_dataset=ds_train, callbacks=[summary_collector])

Method three: Custom callback recording data

MindSpore supports customized callback and supports to record data into summary log file in custom callback, and display the data by the web page.

The following pseudocode is shown in the CNN network, where developers can use the network output with the original tag and the prediction tag to generate the image of the confusion matrix. It is then recorded into the summary log file through the SummaryRecord module. The detailed usage of SummaryRecord can refer to the API document mindspore.train.summary.SummaryRecord.

The sample code is as follows:

from mindspore.train.callback import Callback
from mindspore.train.summary import SummaryRecord

class ConfusionMatrixCallback(Callback):
    def __init__(self, summary_dir):
        self._summary_dir = summary_dir

    def __enter__(self):
        # init you summary record in here, when the train script run, it will be inited before training
        self.summary_record = SummaryRecord(self._summary_dir)

    def __exit__(self, *exc_args):
        # Note: you must close the summary record, it will release the process pool resource
        # else your training script will not exit from training.
        self.summary_record.close()
        return self

    def step_end(self, run_context):
        cb_params = run_context.run_context.original_args()

        # create a confusion matric image, and record it to summary file
        confusion_matrix = create_confusion_matrix(cb_params)
        self.summary_record.add_value('image', 'confusion_matrix', confusion_matrix)
        self.summary_record.record(cb_params.cur_step)

# init you train script
...

confusion_matrix = ConfusionMatrixCallback(summary_dir='./summary_dir')
model.train(network, train_dataset=ds_train, callbacks=[confusion_matrix])

The above three ways support the record computational graph, loss value and other data. In addition, MindSpore also supports the saving of computational graph for other phases of training, through the save_graphs option of context.set_context in the training script is set to True to record computational graphs of other phases, including the computational graph after operator fusion.

In the saved files, ms_output_after_hwopt.pb is the computational graph after operator fusion, which can be viewed on the web page.

Method four: Advanced usage, custom training cycle

If you are not using the Model interface provided by MindSpore, you can implement a method by imitating train method of Model interface to control the number of iterations. You can imitate the SummaryCollector and record the summary operator data in the following manner. For a detailed custom training cycle tutorial, please refer to the tutorial on the official website.

The following example demonstrates how to record data in a custom training cycle using the summary operator and the add_value interface of SummaryRecord. For more tutorials about SummaryRecord, refer to the Python API documentation.

from mindspore import nn
from mindspore.train.summary import SummaryRecord
import mindspore.ops as ops

class LeNet5(nn.Cell):
    def __init__(self, num_class=10):
        super(LeNet5, self).__init__()
        self.num_class = num_class
        self.batch_size = 32
        self.conv1 = conv(1, 6, 5)
        ...

        self.image_summary = ops.ImageSummary()
        self.tensor_summary = ops.TensorSummary()

    def construct(self, x):
        self.image_summary('x1', x)
        x = self.conv1(x)
        self.tensor_summary('after_conv1', x)
        x = self.relu(x)
        ...
        return x

...

def train():
    epochs = 10
    net = LeNet5()
    with SummaryRecord('./summary_dir') as summary_record:
        for epoch in range(epochs):
            step = 1
            for inputs in dataset_helper:
                output = net(*inputs)
                current_step = epoch * len(dataset_helper) + step
                print("step: {0}, losses: {1}".format(current_step, output.asnumpy()))

                # Note1: The output should be a scalar, and use 'add_value' method to record loss.
                # Note2: You must use the 'record(step)' method to record the data of this step.
                summary_record.add_value('scalar', 'loss', output)
                summary_record.record(current_step)

                step += 1

if __name__ == '__main__':
    train()

Tip: Recording gradients

There is a tip for recording gradients with summary in addition to the above methods. Please note that the tip should be used with one of the above methods.

Recording gradients is possible by inheriting your original optimizer and inserting calls to summary operator. An example of code is as follows:

import mindspore.nn as nn
import mindspore.ops as ops
...

# Define a new optimizer class by inheriting your original optimizer.
class MyOptimizer(nn.Momentum):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._original_construct = super().construct
        self.histogram_summary = ops.HistogramSummary()
        self.gradient_names = [param.name + ".gradient" for param in self.parameters]

    def construct(self, grads):
        # Record gradient.
        self.histogram_summary(self.gradient_names[0], grads[0])
        return self._original_construct(grads)

...

# Initialize your model with the newly defined optimizer.
model = Model(network, loss_fn=loss_fn, optimizer=MyOptimizer(arg1=arg1value))

Run MindInsight

After completing the data collection in the tutorial above, you can start MindInsight to visualize the collected data. When start MindInsight, you need to specify the summary log file directory with the --summary-base-dir parameter.

The specified summary log file directory can be the output directory of a training or the parent directory of the output directory of multiple training.

The output directory structure for a training is as follows

└─summary_dir
    events.out.events.summary.1596869898.hostname_MS
    events.out.events.summary.1596869898.hostname_lineage

Execute command:

mindinsight start --summary-base-dir ./summary_dir

The output directory structure of multiple training is as follows:

└─summary
    ├─summary_dir1
    │      events.out.events.summary.1596869898.hostname_MS
    │      events.out.events.summary.1596869898.hostname_lineage
    │
    └─summary_dir2
            events.out.events.summary.1596869998.hostname_MS
            events.out.events.summary.1596869998.hostname_lineage

Execute command:

mindinsight start --summary-base-dir ./summary

After successful startup, the visual page can be viewed by visiting the http://127.0.0.1:8080 address through the browser.

Stop MindInsight command:

mindinsight stop

For more parameter Settings, see the MindInsight related commands page.

Notices

To limit time of listing summaries, MindInsight lists at most 999 summary items.

Multiple SummaryRecord instances can not be used at the same time. (SummaryRecord is used in SummaryCollector)

If you use two or more instances of SummaryCollector in the callback list of ‘model.train’ or ‘model.eval’, it is seen as using multiple SummaryRecord instances at the same time, and it may cause recoding data failure.

If the customized callback uses SummaryRecord, it can not be used with SummaryCollector at the same time.

Correct code:

...
summary_collector = SummaryCollector('./summary_dir')
model.train(2, train_dataset, callbacks=[summary_collector])
...
model.eval(dataset, callbacks=[summary_collector])

Wrong code:

...
summary_collector1 = SummaryCollector('./summary_dir1')
summary_collector2 = SummaryCollector('./summary_dir2')
model.train(2, train_dataset, callbacks=[summary_collector1, summary_collector2])

Wrong code:

...
# Note: the 'ConfusionMatrixCallback' is user-defined, and it uses SummaryRecord to record data.
confusion_callback = ConfusionMatrixCallback('./summary_dir1')
summary_collector = SummaryCollector('./summary_dir2')
model.train(2, train_dataset, callbacks=[confusion_callback, summary_collector])

In each Summary log file directory, only one training data should be placed. If a summary log directory contains summary data from multiple training, MindInsight will overlay the summary data from these training when visualizing the data, which may not be consistent with the expected visualizations.
Currently, SummaryCollector and SummaryRecord do not support scenarios with GPU multi-card running.
When using summary, it is recommended that you set dataset_sink_mode argument of model.train() to False, so that the unit of collect_freq is step. When dataset_sink_mode was True, the unit of collect_freq would be epoch and it is recommend you set collect_freq manually.
The maximum amount of data saved per step is 2147483647Bytes. If this limit is exceeded, data for the step cannot be recorded and an error occurs.
In PyNative mode, the SummaryCollector can be used properly, but the computational graph can not be recorded and the summary operator can not be used.