Getting started

Neptune

• 比较和调试ML实验和模型
• 监测实验结果
• 保存模型、参数
• 监测GPU利用状态

Install Neptune client library

$ pip install neptune-client
# or
$ conda install -c conda-forge neptune-client

Copy and export your API token

在~/.bashrc 中添加

export NEPTUNE_API_TOKEN="your API token"

执行

$ source ~\.bashrc

测试

新建python文件，执行

import neptune

# The init() function called this way assumes that
# NEPTUNE_API_TOKEN environment variable is defined.

neptune.init('ybb-ybb/sandbox')
neptune.create_experiment(name='minimal_example')

# log some metrics

for i in range(100):
    neptune.log_metric('loss', 0.95**i)

neptune.log_metric('AUC', 0.96)

$ bash main.py

使用

create an experiment

定义参数运行新实验：

# Define parameters

PARAMS = {'decay_factor' : 0.5,
          'n_iterations' : 117}

# Create experiment with defined parameters

neptune.create_experiment (name='example_with_parameters',
                          params=PARAMS)

记录图像：

# Log image data

import numpy as np

array = np.random.rand(10, 10, 3)*255
array = np.repeat(array, 30, 0)
array = np.repeat(array, 30, 1)
neptune.log_image('mosaics', array)

记录文本

# Log image data

import numpy as np

array = np.random.rand(10, 10, 3)*255
array = np.repeat(array, 30, 0)
array = np.repeat(array, 30, 1)
neptune.log_image('mosaics', array)

保存模型

# log some file

# replace this file with your own file from local machine
neptune.log_artifact('model_weights.pkl')

# log file to some specific directory (see second parameter below)

# replace this file with your own file from local machine
neptune.log_artifact('model_checkpoints/checkpoint_3.pkl', 'training/model_checkpoints/checkpoint_3.pkl')

上传代码

# Upload source code

# replace these two source files with your own files.
neptune.create_experiment(upload_source_files=['main.py', 'model.py'])

标签

# add tag when experiment is created
neptune.create_experiment(tags=['training'])

# add single tag
neptune.append_tag('transformer')

# add few tags at once
neptune.append_tags('BERT', 'ELMO', 'ideas-exploration')