pymlog 0.0.26

A minimal logging utility for machine learning experiments

MLog

A minimal logging utility for machine learning experiments.

Installation

> pip install pymlog

Logging

import mlog
import random

CONFIG = {'num_epochs': 100}

# Create a new run with an associated configuration
run = mlog.start(run='run_name', config=CONFIG, save='*.py')

# Log seamlessly
for epoch in range(CONFIG['num_epochs']):
    loss = random.random() * (1.05 ** (- epoch))
    run.log(epoch=epoch, loss=loss)
    metric = random.random()
    run.log(epoch=epoch, metric=metric)

Quick preview

> mlog plot epoch loss
> mlog plot epoch loss --aggregate median
> mlog plot epoch loss --aggregate median --intervals max
> mlog plot loss metric --scatter

Manage runs

> mlog list
        _name  num_epochs  learning_rate  batch_size
_run_id
1         run         100          0.001          32
2         run         100          0.001          32
3         run         100          0.001          32
4         run         100          0.001          32
5         run         100          0.001          32
6         run         100          0.001          32
7         run         100          0.001          32
8         run         100          0.001          32
9         run         100          0.001          32
10        run         100          0.001          32

This command starts an interactive interface where you can use commands like:

• hjkl to navigate left, down, up and right,
• gG to go up and down,
• d to delete run,
• space to preview plot,
• q to exit.

Plotting

import mlog
import pandas as pd
import matplotlib.pyplot as plt

# Retrieve data
df = mlog.get('epoch', 'loss')
df = df.groupby('epoch').aggregate(['mean', 'min', 'max'])

# Plot data
fig, ax = plt.subplots()
ax.plot(df.index, df.loss['mean'])
ax.fill_between(df.index, df.loss['min'], df.loss['max'], alpha=0.4)
plt.show()