中文

VisualDL Guide

Overview

VisualDL is a visualization tool designed for Deep Learning. VisualDL provides a variety of charts to show the trends of parameters. It enables users to understand the training process and model structures of Deep Learning models more clearly and intuitively so as to optimize models efficiently.

Currently, VisualDL provides Fifteen Components: scalar, image, audio, text, graph(dynamic, static), histogram, pr curve, ROC curve, high dimensional and hyperparameters, profiler, x2paddle, fastdeployserver, fastdeployclient. VisualDL iterates rapidly and new functions will be continuously added.

At the same time, VisualDL provides VDL.service , which allows developers to easily save, track and share visualization results of experiments with anyone for free.

Scalar--Line Chart

Introduction

The data type of the input is scalar values. Scalar is used to present the training parameters in the form of a line chart. By using Scalar to record loss and accuracy, developers are able to track the trend of changes easily through line charts.

Record Interface

The interface of the Scalar is shown as follows:

add_scalar(tag, value, step, walltime=None)

The interface parameters are described as follows:

*Note that the rules of specifying tags (e.g.train/acc) are:

1. The tag before the first / is the parent tag and serves as the tag of the same raw
2. The tag after the first / is a child tag, the charts with child tag will be displayed under the parent tag. The data of the same parent tag but different child tags will be displayed in the same column, but not in the same picture.
3. Users can use multiple /, but the tag of a raw is the parent tag--the tag before the first /

Here are three examples:

• When 'train' is created as the parent tag and 'acc' and 'loss' are created as child tags：train/acc、 train/loss，the tag of a raw is 'train' , which includes two sub charts--'acc' and 'loss':

• When 'train' is created as the parent tag, and 'test/acc' and 'test/loss' are created as child tags：train/test/acc、 train/test/loss, the tag of a raw is 'train', which includes two sub charts--'test/acc' and 'test/loss':

• When two parent tags are created：acc、 loss， two rows of charts are named as 'acc' and 'loss' respectively.

Demo

• Fundamental Methods

The following shows an example of using Scalar to record data, and the script can be found in Scalar Demo

from visualdl import LogWriter

if __name__ == '__main__':
    value = [i/1000.0 for i in range(1000)]
    # initialize a recorder
    with LogWriter(logdir="./log/scalar_test/train") as writer:
        for step in range(1000):
            # add accuracy with tag of 'acc' to the recorder
            writer.add_scalar(tag="acc", step=step, value=value[step])
            # add loss with tag of 'loss' to the recorder
            writer.add_scalar(tag="loss", step=step, value=1/(value[step] + 1))

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the address: http://127.0.0.1:8080to view line charts:

• Advanced Usage--Comparison of Multiple Experiments

The following shows the comparison of multiple sets of experiments using Scalar.

There are two steps to achieve this function:

1. Create sub-log files to store the parameter data of each group of experiments
2. When recording data to the scalar component，developers can compare the same type of parameters for different experiments by using the same tag. Note that the log files you want to display must be placed in different directories because only one log file in a directory is valid and displayed.

from visualdl import LogWriter

if __name__ == '__main__':
    value = [i/1000.0 for i in range(1000)]
    # Step 1: Create a parent folder: log and a child folder: scalar_test
    with LogWriter(logdir="./log/scalar_test") as writer:
        for step in range(1000):
            # Step 2: Add data with tag train/acc to the recorder
            writer.add_scalar(tag="train/acc", step=step, value=value[step])
            # Step 2: Add data with tag train/loss to the recorder
            writer.add_scalar(tag="train/loss", step=step, value=1/(value[step] + 1))
    # Step 1: Create a second child folder: scalar_test2    
    value = [i/500.0 for i in range(1000)]
    with LogWriter(logdir="./log/scalar_test2") as writer:
        for step in range(1000):
            # Step 2: Add the accuracy data of scalar_test2 under the same name `train/acc`
            writer.add_scalar(tag="train/acc", step=step, value=value[step])
            # Step 2: Add the loss data of scalar_test2 under the same name as `train/loss`
            writer.add_scalar(tag="train/loss", step=step, value=1/(value[step] + 1))

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the address: http://127.0.0.1:8080 to view line charts:

Functional Instruction

• Developers are allowed to zoom in, restore, transform of the coordinate axis (y-axis logarithmic coordinates), download the line chart.

• Details can be shown by hovering on specific data points.

• Developers can find target scalar charts by searching corresponded tags.

• Specific runs can be selected by searching for the corresponded experiment tags.

• Display the global extrema

• Only display smoothed data

• There are three measurement scales of X axis

1. Step: number of iterations
2. Walltime: absolute training time
3. Relative: training time

• The smoothness of the curve can be adjusted to better show the change of the overall trend.

Image--Image Visualization

Introduction

The Image is used to present the change of image data during training. Developers can view images in different training stages by adding few lines of codes to record images in a log file.

Record Interface

The interface of the Image is shown as follows:

add_image(tag, img, step, walltime=None, dataformats="HWC")

The interface parameters are described as follows:

Demo

The following shows an example of using Image to record data, and the script can be found in Image Demo.

import numpy as np
from PIL import Image
from visualdl import LogWriter


def random_crop(img):
    """get random 100x100 slices of image
    """
    img = Image.open(img)
    w, h = img.size
    random_w = np.random.randint(0, w - 100)
    random_h = np.random.randint(0, h - 100)
    r = img.crop((random_w, random_h, random_w + 100, random_h + 100))
    return np.asarray(r)


if __name__ == '__main__':
    # initialize a recorder
    with LogWriter(logdir="./log/image_test/train") as writer:
        for step in range(6):
            # add image data
            writer.add_image(tag="eye",
                             img=random_crop("../../docs/images/eye.jpg"),
                             step=step)

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the address: http://127.0.0.1:8080to view:

Functional Instructions

• Developers can find target images by searching corresponded tags.

• Developers are allowed to view image data under different iterations by scrolling the Step/iteration slider.

Audio--Audio Play

Introduction

Audio aims to allow developers to listen to the audio in real-time during the training process, helping developers to monitor the process of speech recognition and text-to-speech.

Record Interface

The interface of the Image is shown as follows:

add_audio(tag, audio_array, step, sample_rate)

The interface parameters are described as follows:

Demo

The following shows an example of using Audio to record data, and the script can be found in Audio Demo.

from visualdl import LogWriter
from scipy.io import wavfile


if __name__ == '__main__':
    with LogWriter(logdir="./log/audio_test/train") as writer:
        sample_rate, audio_data = wavfile.read('./test.wav')
        writer.add_audio(tag="audio_tag",
                         audio_array=audio_data,
                         step=0,
                         sample_rate=sample_rate)

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the address: http://127.0.0.1:8080to view:

Functional Instructions

• Developers can find the target audio by searching corresponded tags.

• Developers are allowed to listen to the audio under different iterations by scrolling the Step/iteration slider.

• Play/Pause the audio

• Adjust the volume

• Download the audio

Text

Introduction

visualizes the text output of NLP models within any stage, aiding developers to compare the changes of outputs so as to deeply understand the training process and simply evaluate the performance of the model.

Record Interface

The interface of the Text is shown as follows:

add_text(tag, text_string, step=None, walltime=None)

The interface parameters are described as follows:

Demo

The following shows an example of how to use Text component, and script can be found in Text Demo

from visualdl import LogWriter
if __name__ == '__main__':
    texts = [
        '上联: 众 佛 群 灵 光 圣 地	下联: 众 生 一 念 证 菩 提',
        '上联: 乡 愁 何 处 解	下联: 故 事 几 时 休',
        '上联: 清 池 荷 试 墨	下联: 碧 水 柳 含 情',
        '上联: 既 近 浅 流 安 笔 砚	下联: 欲 将 直 气 定 乾 坤',
        '上联: 日 丽 萱 闱 祝 无 量 寿	下联: 月 明 桂 殿 祝 有 余 龄',
        '上联: 一 地 残 红 风 拾 起	下联: 半 窗 疏 影 月 窥 来'
    ]
    with LogWriter(logdir="./log/text_test/train") as writer:
        for step in range(len(texts)):
            writer.add_text(tag="output", step=step, text_string=texts[step])

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the addresshttp://127.0.0.1:8080 to view:

Functional Instrucions

• Developers can find the target text by searching corresponded tags.

  

• Developers can find the target runs by searching corresponded tags.

  

• Developers can fold the tab of text.

Graph--Network Structure

Introduction

Graph can visualize the network structure of the model by one click. It enables developers to view the model attributes, node information, searching node and so on. These functions help developers analyze model structures and understand the directions of data flow quickly.

Record Interface

The interface of the Graph is shown as follows:

add_graph(model, input_spec, verbose=False):

The interface parameters are described as follows:

Note

If you want to use add_graph interface, paddle package is required. Please refer to website of PaddlePaddle。

Demo

The following shows an example of how to use Graph component, and script can be found in Graph Demo There are two methods to launch this component:

import paddle
import paddle.nn as nn
import paddle.nn.functional as F

from visualdl import LogWriter


class MyNet(nn.Layer):
    def __init__(self):
        super(MyNet, self).__init__()
        self.conv1 = nn.Conv2D(
            in_channels=1, out_channels=20, kernel_size=5, stride=1, padding=2)
        self.max_pool1 = nn.MaxPool2D(kernel_size=2, stride=2)
        self.conv2 = nn.Conv2D(
            in_channels=20,
            out_channels=20,
            kernel_size=5,
            stride=1,
            padding=2)
        self.max_pool2 = nn.MaxPool2D(kernel_size=2, stride=2)
        self.fc = nn.Linear(in_features=980, out_features=10)

    def forward(self, inputs):
        x = self.conv1(inputs)
        x = F.relu(x)
        x = self.max_pool1(x)
        x = self.conv2(x)
        x = F.relu(x)
        x = self.max_pool2(x)
        x = paddle.reshape(x, [x.shape[0], -1])
        x = self.fc(x)
        return x


net = MyNet()
with LogWriter(logdir="./log/graph_test/") as writer:
    writer.add_graph(
        model=net,
        input_spec=[paddle.static.InputSpec([-1, 1, 28, 28], 'float32')],
        verbose=True)

After running the above program, developers can launch the panel by:

visualdl --logdir ./log/graph_test/ --port 8080

Then, open the browser and enter the addresshttp://127.0.0.1:8080 to view:

Note

We provide option --model to specify model structure file in previous versions, and this option is still supported now. You can specify model exported by add_graph interface ("vdlgraph" contained in filename), which will be shown in dynamic graph page, and we use string "manual_input_model" in the page to denote the model you specify by this option. Other supported file formats are presented in static graph page.

For example

visualdl --model ./log/model.pdmodel --port 8080

which will be shown in static graph page. And

visualdl --model ./log/vdlgraph.1655783158.log --port 8080

shown in dynamic graph page.

Functional Instructions

Graph page is divided into dynamic and static version currently. Dynamic version is used to visualize dynamic model of paddle, which is exported by add_graph interface. The other is used to visualize static model of paddle, which is exported by paddle.jit.save interface and other supported formats.

Common functions

• Developers are allowed to drag the model up and down，left and right，zoom in and zoom out.

• Search to locate the specific node

• Click to view the model properties

• Display the model information by selecting corresponded attributes

• Files can be ex as PNG or SVG format

• Click nodes to view attribute information

• Switch the model by one-click

Specific feature in dynamic version

• Fold and unfold one node

• Fold and unfold all nodes

• Link api specification of paddle

  If you use paddle.nn components to construct your network model, you can use alt+click mouse to direct to corresponding api specification.

Specific feature in static version

• Upload the model file by one-click
  • Supported model：PaddlePaddle、ONNX、Keras、Core ML、Caffe、Caffe2、Darknet、MXNet、ncnn、TensorFlow Lite
  • Experimental supported model：TorchScript、PyTorch、Torch、 ArmNN、BigDL、Chainer、CNTK、Deeplearning4j、MediaPipe、ML.NET、MNN、OpenVINO、Scikit-learn、Tengine、TensorFlow.js、TensorFlow

Histogram--Distribution of Tensors

Introduction

Histogram displays how the trend of tensors (weight, bias, gradient, etc.) changes during the training process in the form of histogram. Developers can adjust the model structures accurately by having an in-depth understanding of the effect of each layer.

Record Interface

The interface of the Histogram is shown as follows:

add_histogram(tag, values, step, walltime=None, buckets=10)

The interface parameters are described as follows:

Demo

The following shows an example of using Histogram to record data, and the script can be found in Histogram Demo

from visualdl import LogWriter
import numpy as np


if __name__ == '__main__':
    values = np.arange(0, 1000)
    with LogWriter(logdir="./log/histogram_test/train") as writer:
        for index in range(1, 101):
            interval_start = 1 + 2 * index / 100.0
            interval_end = 6 - 2 * index / 100.0
            data = np.random.uniform(interval_start, interval_end, size=(10000))
            writer.add_histogram(tag='default tag',
                                 values=data,
                                 step=index,
                                 buckets=10)

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the address: http://127.0.0.1:8080to view the histogram.

Functional Instructions

• Developers are allowed to zoom in and download the histogram.

  

• Provide two modes: Offset and Overlay.

  

  • Offset mode

  

  • Overlay mode

  

• Display the parameters、training steps and frequency by hovering on specific data points.

  • In the 240th training step, the weight is -0.0031and the frequency is 2734

  

• Developers can find target histogram by searching corresponded tags.

  

• Search tags to show the histograms generated by corresponded experiments.

  

PR Curve

Introduction

PR Curve presents precision-recall curves in line charts, describing the tradeoff relationship between precision and recall in order to choose a best threshold.

Record Interface

The interface of the PR Curve is shown as follows:

add_pr_curve(tag, labels, predictions, step=None, num_thresholds=10)

The interface parameters are described as follows:

Demo

The following shows an example of how to use PR Curve component, and script can be found in PR Curve Demo

from visualdl import LogWriter
import numpy as np

with LogWriter("./log/pr_curve_test/train") as writer:
    for step in range(3):
        labels = np.random.randint(2, size=100)
        predictions = np.random.rand(100)
        writer.add_pr_curve(tag='pr_curve',
                            labels=labels,
                            predictions=predictions,
                            step=step,
                            num_thresholds=5)

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the addresshttp://127.0.0.1:8080 to view:

Functional Instrucions

• Developers can zoom in, restore, and download PR Curves

  

• Developers hover on the specific data point to learn about the detailed information: TP, TN, FP, FN and the corresponded thresholds

  

• The targeted PR Curves can be displayed by searching tags

  

• Developers can find specific labels by searching tags or view the all labels

• Developers is able to observe the changes of PR Curves across training steps

  

• There are three measurement scales of X axis

  1. Step: number of iterations
  2. Walltime: absolute training time
  3. Relative: training time

  

ROC Curve

Introduction

ROC Curve shows the performance of a classification model at all classification thresholds; the larger the area under the curve, the better the model performs, aiding developers to evaluate the model performance and choose an appropriate threshold.

Record Interface

The interface of the PR Curve is shown as follows:

add_roc_curve(tag, labels, predictions, step=None, num_thresholds=10)

The interface parameters are described as follows:

Demo

The following shows an example of how to use ROC curve component, and script can be found in ROC Curve Demo

from visualdl import LogWriter
import numpy as np

with LogWriter("./log/roc_curve_test/train") as writer:
    for step in range(3):
        labels = np.random.randint(2, size=100)
        predictions = np.random.rand(100)
        writer.add_roc_curve(tag='roc_curve',
                             labels=labels,
                             predictions=predictions,
                             step=step,
                             num_thresholds=5)

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the addresshttp://127.0.0.1:8080 to view:

*Note: the use of ROC Curve in the frontend is the same as that of PR Curve, please refer to the instructions in PR Curve section if needed.

High Dimensional--Data Dimensionality Reduction

Introduction

High Dimensional projects high-dimensional data into a low dimensional space, aiding users to have an in-depth analysis of the relationship between high-dimensional data. Three dimensionality reduction algorithms are supported:

• PCA : Principle Component Analysis
• t-SNE : t-distributed Stochastic Neighbor Embedding
• umap: Uniform Manifold Approximation and Projection

Record Interface

The interface of the High Dimensional is shown as follows:

add_embeddings(tag, labels, hot_vectors, walltime=None)

The interface parameters are described as follows:

Demo

The following shows an example of how to use High Dimensional component, and script can be found in High Dimensional Demo

from visualdl import LogWriter


if __name__ == '__main__':
    hot_vectors = [
        [1.3561076367500755, 1.3116267195134017, 1.6785401875616097],
        [1.1039614644440658, 1.8891609992484688, 1.32030488587171],
        [1.9924524852447711, 1.9358920727142739, 1.2124401279391606],
        [1.4129542689796446, 1.7372166387197474, 1.7317806077076527],
        [1.3913371800587777, 1.4684674577930312, 1.5214136352476377]]

    labels = ["label_1", "label_2", "label_3", "label_4", "label_5"]
    # initialize a recorder
    with LogWriter(logdir="./log/high_dimensional_test/train") as writer:
        # recorde a set of labels and corresponding hot_vectors to the recorder 
        writer.add_embeddings(tag='default',
                              labels=labels,
                              hot_vectors=hot_vectors)

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the addresshttp://127.0.0.1:8080 to view:

Functional Instrucions

• Developers are allowed to select specific runs of data or certain labels of data to display

  

• TSNE

  

• PCA

  

• UMAP

  

HyperParameters--HyperParameter Visualization

Introduction

HyperParameters visualize the relationship between hyperparameters and model metrics (such as accuracy and loss) in a rich view, helping you identify the best hyperparameters in an efficient way.

Record Interface

The interface of the HyperParameters is slightly different from other components'. Firstly, you need to use the add_hparams to record the hyperparameter data(hparams_dict) and specify the name of the metrics(metrics_list). Then, for the metrics you just added, you need to record those metrics values by using add_scalar. In this way you can get all data for HpyerParameters Visualization.

add_hparams(hparam_dict, metric_list, walltime=None):

The interface parameters are described as follows:

Demo

The following shows an example of how to use HyperParameters component, and script can be found in HyperParameters Demo

from visualdl import LogWriter

# This demo demonstrates the hyperparameter records of two experiments. Take the first
# experiment data as an example, First, record the data of the hyperparameter `hparams`
# in the `add_hparams` interface. Then specify the name of `metrics` to be recorded later.
# Finally, use `add_scalar` to specifically record the data of `metrics`. Note that the
# `metrics_list` parameter in the `add_hparams` interface needs to include the `tag`
# parameter of the `add_scalar` interface.
if __name__ == '__main__':
    # Record the data of the first experiment
    with LogWriter('./log/hparams_test/train/run1') as writer:
        # Record the value of `hparams` and the name of `metrics`
        writer.add_hparams(hparams_dict={'lr': 0.1, 'bsize': 1, 'opt': 'sgd'},
                           metrics_list=['hparam/accuracy', 'hparam/loss'])
        # Record the metrics values ​​of different steps in an experiment by matching
        # the `tag` in the `add_scalar` interface with `metrics_list` in `add_hparams` interface.
        for i in range(10):
            writer.add_scalar(tag='hparam/accuracy', value=i, step=i)
            writer.add_scalar(tag='hparam/loss', value=2*i, step=i)

    # Record the data of the second experiment
    with LogWriter('./log/hparams_test/train/run2') as writer:
        # Record the value of `hparams` and the name of `metrics`
        writer.add_hparams(hparams_dict={'lr': 0.2, 'bsize': 2, 'opt': 'relu'},
                           metrics_list=['hparam/accuracy', 'hparam/loss'])
        # Record the metrics values ​​of different steps in an experiment by matching
        # the `tag` in the `add_scalar` interface with `metrics_list` in `add_hparams` interface.
        for i in range(10):
            writer.add_scalar(tag='hparam/accuracy', value=1.0/(i+1), step=i)
            writer.add_scalar(tag='hparam/loss', value=5*i, step=i)

After running the above program, developers can launch the panel by:

visualdl --logdir ./log --port 8080

Then, open the browser and enter the addresshttp://127.0.0.1:8080 to view:

Functional Instrucions

• Table View

  • The table view can be displayed in a sorted order.
  • Trial ID represents a specific experiment name, the column name displayed in other normal fonts is the hyperparameter name, and the column displayed in bold font is the metric name.
  • The position of hyperparameters and metrics can be customized by dragging.
  • The column width of the table view can be adjusted by dragging.
  • You can click to expand to view the scalar of the metrics.

  

• Parallel Coordinates View

  • The specific values ​​of hyperparameters and metrics in a certain set of experiments can be displayed by hovering.
  • Scalar of the metrics in this group of experiments can be displayed by selecting a certain curve.

  

• Scatter Plot Matrix View

  • The specific values ​​of hyperparameters and metrics in a certain set of experiments can be displayed by hovering.
  • Scalar of the metrics in this group of experiments can be displayed by selecting a certain point.

  

• Scalar of Metrics

  • Can be viewed in table view, parallel coordinates view and scatter plot matrix view.
  • Scalar of the metrics viewed here can also be viewed under the SCALARS board.

  

• Hyperparameter/metric range selection

  • Display part of the data by selecting the range of hyperparameters or metrics.

  

• download data

  • Two formats can be selected, CSV or TSV.

  

Profiler--profiling data visualization

Introduction

VisualDL supports to visualize profiling data exported by paddle and helps you identify program bottlenecks and optimize performance. Please refer to VisualDL Profiler Guide.

X2Paddle--model format transformation

Introduction

The X2Paddle component is used to read the onnx model, display the network structure of the onnx model, and help users convert the onnx model into a paddle model. Users can compare the original onnx model and the converted paddle model network, and obtain the converted model for use.

Usage

Launch the panel by:

visualdl --port 8080

Then, open the browser and enter the addresshttp://127.0.0.1:8080 to use X2Paddle component.

Functional Instrucions

• Convert onnx model and download

  

• Reload a new model

  

• Compare model network between conversions

  

Note: If failed to convert an onnx model to paddle, you can copy the error message of the model conversion to X2Paddle issue to help us improve this tool.

FastDeployServer--fastdeploy serving deployment visualization

Introduction

The FastDeployServer component assists users to use fastdeployserver to deploy service conveniently based on FastDeploy project. It mainly provides the functions of loading and editing the model repository, service management and monitoring, and providing the client to test service. Please refer to use VisualDL for fastdeploy serving deployment management.

FastDeployClient--fastdeploy client for request visualization

Introduction

The FastDeployClient component is mainly used to quickly access the fastdeployserver service based on FastDeploy project, to help users visualize prediction requests and results, and make quick verification of deployed services. Please refer to use VisualDL as fastdeploy client for request visualization.

VDL.service

Introduction

VDL.service enables developers to easily save, track and share visualization results with anyone for free.

Usage Steps

1. Make sure that your get the lastest version of VisualDL, if not, please update by:

pip install visualdl --upgrade

2. Upload log/model to save, track and share the visualization results.

visualdl service upload --logdir ./log \
                        --model ./__model__

3. An unique URL will be given. Then you can view the visualization results by simply copying and pasting the URL to the browser.