Collecting Summary Record

Overview

Scalars, images, computational graphs, training optimization process, 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, training optimization process, and model hyperparameters in the training script, record them in the summary log file, and run the training script.
Start MindSpore Insight and specify the summary log file directory using startup parameters. After MindSpore Insight 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 visualized data in Viewing Dashboard on the web page.

To view visualized data in ModelArts, see managing visualized Jobs on ModelArts.

Preparing The Training Script

Currently, MindSpore supports to save scalars, images, computational graph, training optimization process, and model hyperparameters to summary log file and display them on the web page. The computational graph can only be recorded in the graph mode. The detailed process of data collection and landscape drawing in the training optimization process can be referred to Training Optimization Process Visualization.

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.SummaryCollector.

The sample code snippet is shown as follows. The whole script is put on gitee.

def train(ds_train):
    ...
    model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})

    # Init a SummaryCollector callback instance, and use it in model.train or model.eval
    specified = {"collect_metric": True, "histogram_regular": "^conv1.*|^conv2.*", "collect_graph": True,
                 "collect_dataset_graph": True}

    summary_collector = SummaryCollector(summary_dir="./summary_dir/summary_01", collect_specified_data=specified,
                                         collect_freq=1, keep_default_action=False, collect_tensor_freq=200)

    print("============== Starting Training ==============")
    model.train(epoch=1, train_dataset=ds_train, callbacks=[time_cb, LossMonitor(), summary_collector],
                dataset_sink_mode=False)

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 APIs and SummaryCollector

In addition to providing the SummaryCollector that automatically collects some summary data, MindSpore provides summary APIs 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 APIs 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. The whole script is put on gitee.

Step 1: Call the summary API 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:

class AlexNet(nn.Cell):
    """
    Alexnet
    """
    def __init__(self, num_classes=10, channel=3):
        super(AlexNet, self).__init__()
        self.conv1 = conv(channel, 96, 11, stride=4)
        ...
        # Init TensorSummary
        self.tensor_summary = ops.TensorSummary()
        # Init ImageSummary
        self.image_summary = ops.ImageSummary()

    def construct(self, x):
        # Record image by Summary API
        self.image_summary("Image", x)
        x = self.conv1(x)
        # Record tensor by Summary API
        self.tensor_summary("Tensor", x)
        ...
        return x

In the same Summary API, 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 APIs are used to collect scalar data, two scalars cannot be given the same name.

Summary API needs to be used in construct of nn.Cell.

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

The sample code is as follows:

def train(ds_train):
    ...
    # Init a SummaryCollector callback instance, and use it in model.train or model.eval
    specified = {"collect_metric": True, "histogram_regular": "^conv1.*|^conv2.*", "collect_graph": True,
                 "collect_dataset_graph": True}

    summary_collector = SummaryCollector(summary_dir="./summary_dir/summary_02", collect_specified_data=specified,
                                         collect_freq=1, keep_default_action=False, collect_tensor_freq=200)

    print("============== Starting Training ==============")
    model.train(epoch=1, train_dataset=ds_train, callbacks=[time_cb, LossMonitor(), summary_collector],
                dataset_sink_mode=False)

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.SummaryRecord.

The sample code snippet is as follows. The whole script is put on gitee.

class MyCallback(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)
        return self

    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()

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

        # create a confusion matric image, and record it to summary file
        self.summary_record.add_value('image', 'image0', cb_params.train_dataset_element[0])
        self.summary_record.add_value('scalar', 'loss', cb_params.net_outputs)
        self.summary_record.record(cb_params.cur_step_num)


def train(ds_train):
    ...
    model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})

    # Init a specified callback instance, and use it in model.train or model.eval
    specified_callback = MyCallback(summary_dir='./summary_dir/summary_03')

    print("============== Starting Training ==============")
    model.train(epoch=1, train_dataset=ds_train, callbacks=[time_cb, LossMonitor(), specified_callback],
                dataset_sink_mode=False)

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 set_context in the training script is set to True to record computational graphs of other phases, including the computational graph after API fusion.

In the saved files, ms_output_after_hwopt.pb is the computational graph after API 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 API data in the following manner.

The following code snippet demonstrates how to record data in a custom training cycle using the summary API and the add_value interface of SummaryRecord. The whole script is put on gitee.

For more tutorials about SummaryRecord, refer to the Python API documentation. Please note that SummaryRecord will not record computational graph automatically. If you need to record the computational graph, please manually pass the instance of network that inherits from Cell. The recorded computational graph only includes the code and functions used in the construct method.

def train(ds_train):
    ...
    summary_collect_frequency = 200
    # Note1: An instance of the network should be passed to SummaryRecord if you want to record
    # computational graph.
    with SummaryRecord('./summary_dir/summary_04', network=train_net) as summary_record:
        for epoch in range(epochs):
            step = 1
            for inputs in ds_train:
                output = train_net(*inputs)
                current_step = epoch * ds_train.get_dataset_size() + step
                print("step: {0}, losses: {1}".format(current_step, output.asnumpy()))

                # Note2: The output should be a scalar, and use 'add_value' method to record loss.
                # Note3: You must use the 'record(step)' method to record the data of this step.
                if current_step % summary_collect_frequency == 0:
                    summary_record.add_value('scalar', 'loss', output)
                    summary_record.record(current_step)

                step += 1

Distributed Training Scene

The SummaryCollector and the SummaryRecord are not multi-process safe when writing data, so in a single-machine multi-card scenario, you need to make sure that each card stores data in a different directory. In a distributed scenario, we set the summary directory with the ‘get_rank’ function.

from mindspore.communication import get_rank
summary_dir = "summary_dir" + str(get_rank())

The sample code is as follows:

from mindspore.communication import get_rank

...

network = ResNet50(num_classes=10)

# Init a SummaryCollector callback instance, and use it in model.train or model.eval
summary_dir = "summary_dir" + str(get_rank())
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)

model.eval(ds_eval, callbacks=[summary_collector])

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 API. An example of code snippet is as follows. The whole script is put on gitee.

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.
        l = len(self.gradient_names)
        for i in range(l):
            self.histogram_summary(self.gradient_names[i], grads[i])
        return self._original_construct(grads)

...

def train(ds_train):
    device_target = "GPU"
    set_context(mode=GRAPH_MODE, device_target=device_target)
    network = AlexNet(num_classes=10)
    net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
    lr = Tensor(get_lr(0, 0.002, 10, ds_train.get_dataset_size()))
    net_opt = MyOptimizer(network.trainable_params(), learning_rate=lr, momentum=0.9)
    time_cb = TimeMonitor(data_size=ds_train.get_dataset_size())
    model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})

    # Init a SummaryCollector callback instance, and use it in model.train or model.eval
    summary_collector = SummaryCollector(summary_dir="./summary_dir/summary_gradients",
                                         collect_freq=200, keep_default_action=False, collect_tensor_freq=200)

    print("============== Starting Training ==============")
    model.train(epoch=1, train_dataset=ds_train, callbacks=[time_cb, LossMonitor(), summary_collector],
                dataset_sink_mode=False)

Running MindSpore Insight

After completing the data collection in the tutorial above, you can start MindSpore Insight to visualize the collected data. When start MindSpore Insight, 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 MindSpore Insight command:

mindinsight stop

For more parameter Settings, see the MindSpore Insight related commands page.

Notices

To limit time of listing summaries, MindSpore Insight 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, MindSpore Insight will overlay the summary data from these training when visualizing the data, which may not be consistent with the expected visualizations.
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 recommended that you set collect_freq manually. The default value of the collect_freq parameter is 10.
The maximum amount of data saved per step is 2147483647 Bytes. 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.