notebook.community

Edit and run



In [1]:

    
%matplotlib inline
from io import StringIO
import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

mpl.rc('savefig', dpi=150)



In [2]:

    
# https://docs.google.com/spreadsheets/d/1c5yGydEANMzHjBufTzph0w-WGwJyiwPMRYz3yBZatb4/edit#gid=0
# tf_cnn_benchmarks, our measurements, inception3/resnet50
table = StringIO("""machine	GPUs	model	images/sec	speedup	efficiency
az-2x-m60	1	inception3	47.49	1.00	100.00
az-2x-m60	2	inception3	95.43	2.01	100.47
7gforce	1	inception3	78.44	1.00	100.00
7gforce	2	inception3	118.32	1.51	75.42
7gforce	4	inception3	236.86	3.02	75.49
7gforce	6	inception3	351.98	4.49	74.79
az-2x-m60	1	resnet50	77.94	1.00	100.00
az-2x-m60	2	resnet50	151.65	1.95	97.29
7gforce	1	resnet50	119.8	1.00	100.00
7gforce	2	resnet50	196	1.64	81.80
7gforce	4	resnet50	364.32	3.04	76.03
7gforce	6	resnet50	571.9	4.77	79.56
""")
df = pd.read_csv(table, sep='\t')



In [3]:

    
df



In [4]:

    
df_inception3 = df[df['model'] == 'inception3']
df_resnet50 = df[df['model'] == 'resnet50']



In [5]:

    
def select(model, machine):
    return df[(df['model']==model) & (df['machine']==machine)]



In [10]:

    
def plot_table(model, metric):
    fig = plt.figure()
    title = 'tf_cnn_benchmark - %s/imagenet-synth - %s' % (model, metric)
    plt.suptitle(title)
    plt.xlabel('number of GPUs')
    plt.ylabel(metric)
    gpu_count = df['GPUs'].max()
    x_ideal = 1+np.arange(gpu_count)
    if metric == 'speedup':
        y_ideal = 1+np.arange(gpu_count)
    elif metric == 'efficiency':
        y_ideal = 100 * np.ones(gpu_count) # %
    plt.bar(x_ideal, y_ideal, alpha=0.25, label='ideal')
    d = select(model, 'az-2x-m60')
    plt.plot(d['GPUs'], d[metric], 'o-', label='az-2x-m60')
    d = select(model, '7gforce')
    plt.plot(d['GPUs'], d[metric], 'o-', label='7gforce')
    plt.grid(True)
    plt.legend()
    plt.savefig('tf_cnn_benchmark_%s_%s.png' % (model, metric))



In [11]:

    
for model in ['inception3', 'resnet50']:
    for metric in ['speedup', 'efficiency']:
        plot_table(model, metric)



In [ ]:

	machine	GPUs	model	images/sec	speedup	efficiency
0	az-2x-m60	1	inception3	47.49	1.00	100.00
1	az-2x-m60	2	inception3	95.43	2.01	100.47
2	7gforce	1	inception3	78.44	1.00	100.00
3	7gforce	2	inception3	118.32	1.51	75.42
4	7gforce	4	inception3	236.86	3.02	75.49
5	7gforce	6	inception3	351.98	4.49	74.79
6	az-2x-m60	1	resnet50	77.94	1.00	100.00
7	az-2x-m60	2	resnet50	151.65	1.95	97.29
8	7gforce	1	resnet50	119.80	1.00	100.00
9	7gforce	2	resnet50	196.00	1.64	81.80
10	7gforce	4	resnet50	364.32	3.04	76.03
11	7gforce	6	resnet50	571.90	4.77	79.56