Plotting Research results

`plot_results_by_config` example for synthetic dataset



In [1]:

    
import pandas as pd
import numpy as np

import random

import sys
sys.path.append('../../..')

from batchflow.utils import plot_results_by_config

%matplotlib inline

plot_results_by_config allows to draw plots of specified variables for all configs in given results of Research.run(). Generally results can be any pandas.DataFrame that follows certain structure. Such dataframe is constructed below.



In [2]:

    
# constructing test dataframe
d1 = pd.DataFrame(data={'iteration': np.arange(10), 'val2': np.sin(np.arange(10))})
d1['name'] = 'train'

d2 = pd.DataFrame(data={'iteration': np.arange(10), 'val1': np.arange(10)})
d2['name'] = 'test'

d3 = pd.concat([d1, d2], sort=False)

dfs = []
for rep in range(2):
    dft = d3.copy()
    dft['val1'] += 0.1 * rep
    dft['val2'] += 0.1 * rep
    dft['repetition'] = rep
    dfs.append(dft)
    
d4 = pd.concat(dfs, sort=False)

dfs = []
for i, config in enumerate(('conf1', 'conf2', 'conf3', 'conf4')):
    for cv_split in range(3):
        dft = d4.copy()
        dft['val1'] += (0.1 + i * 0.5) * np.random.randn(len(dft))
        dft['val2'] += (0.1 + i * 0.5) * np.random.randn(len(dft))
        dft['config'] = config
        dft['cv_split'] = cv_split
        dft['update'] = 0
        dft['sample_index'] = str(random.getrandbits(32))
        dfs.append(dft)
    
res = pd.concat(dfs, sort=False)

res.info()









    



<class 'pandas.core.frame.DataFrame'>
Int64Index: 480 entries, 0 to 9
Data columns (total 9 columns):
iteration       480 non-null int64
val2            240 non-null float64
name            480 non-null object
val1            240 non-null float64
repetition      480 non-null int64
config          480 non-null object
cv_split        480 non-null int64
update          480 non-null int64
sample_index    480 non-null object
dtypes: float64(2), int64(4), object(3)
memory usage: 37.5+ KB

Plotting results for several configs and several variables. If figure size is not specified, it is set to (x, y), where x = (5 * number of variables), y = (5 * number of configs in results)

Plotting results of research with cross-validation



In [3]:

    
plot_results_by_config(res, (('train', 'val2'), ('test', 'val1')))

Plotting results of research without cross-validation



In [4]:

    
plot_results_by_config(res[res.cv_split == 0].drop(columns='cv_split'), (('train', 'val2'), ('test', 'val1')))

By default if only one variable is specified the plots for different configs are arranged vertically



In [5]:

    
plot_results_by_config(res, ('train', 'val2'))

This can be changed by setting layout='flat'



In [6]:

    
plot_results_by_config(res, ('train', 'val2'), layout='flat')

Or layout='square'



In [7]:

    
plot_results_by_config(res, ('train', 'val2'), layout='square')

Variables to plot and their sources can be passed via dict, custom figure size can be specified



In [8]:

    
plot_results_by_config(res, {'train': 'val2', 'test': 'val1'}, figsize=(20, 20))

Following 2 code cells test plots layouts in marginal situations



In [9]:

    
plot_results_by_config(res[res.config == 'conf1'], {'train': 'val2', 'test': 'val1'})



In [10]:

    
plot_results_by_config(res[res.config == 'conf1'], ('train', 'val2'))

Plotting Research results

plot_results_by_config example for synthetic dataset

`plot_results_by_config` example for synthetic dataset