-
Notifications
You must be signed in to change notification settings - Fork 88
/
Copy pathprobe.py
231 lines (205 loc) · 9.29 KB
/
probe.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
import os
import tensorwatch as tw
import argparse
from itertools import cycle
from collections import OrderedDict
import pandas as pd
from bokeh.io import output_file, save, show
from bokeh.layouts import column
from bokeh.plotting import figure as bk_figure
import numpy as np
parser = argparse.ArgumentParser(
description='Probe experiments done with convNet.pytorch using tensorwatch')
parser.add_argument('experiments', metavar='N', type=str, nargs='+',
help='pathes to multiple experiments')
parser.add_argument('--legend', type=str, nargs='+',
help='legend to multiple experiments')
parser.add_argument('--metrics', type=str, nargs='+', default=['loss', 'prec1', 'prec5', 'grad'],
help='metrics (train/val) to extract')
parser.add_argument('--additional', type=str, nargs='+', default=['lr'],
help='additional metrics (standalone) to extract')
parser.add_argument('--x-axis', type=str, default='steps',
help='x axis to experiments')
parser.add_argument('--colors', type=str, nargs='+',
default=['green', 'red', 'blue', 'orange',
'black', 'purple', 'brown'],
help='color for each experiment')
parser.add_argument('--all', action='store_true', default=False,
help='show data for every step (not only epochs)')
parser.add_argument('--paper', action='store_true', default=False,
help='publish ready plots')
def multi_line_opts(num,
dashes=None,
colors=['red', 'green', 'blue', 'orange',
'black', 'purple', 'brown'],
**defaults):
defaults.setdefault('line_width', 2)
options = []
colors = cycle(colors)
if dashes is not None:
dashes = cycle(dashes)
for _ in range(num):
if dashes is not None:
defaults.update(line_dash=next(dashes))
options.append({'line_color': next(colors), **defaults})
return options
def reduce_multiple(result):
x_values = [x for x, _ in result]
x_unq = sorted(set(x_values), key=x_values.index)
if len(x_unq) != len(x_values):
new_result = []
for curr_x in x_unq:
y_values = [y for x, y in result if curr_x == x]
aggr_y = sum(y_values) / len(y_values)
new_result.append((curr_x, aggr_y))
result = new_result
return result
def plot_comparison(experiments,
figure=None,
line_options=None,
title=None,
x_axis_label=None,
y_axis_label=None,
x_axis_type='linear',
y_axis_type='linear',
x_range=None,
y_range=None,
width=800,
height=400,
legend_text_font_size=None,
tools='pan,box_zoom,wheel_zoom,box_select,hover,reset,save',
figure_fn=bk_figure):
line_options = line_options or multi_line_opts(len(experiments))
if len(line_options) < len(experiments):
line_options += multi_line_opts(len(experiments) -
len(line_options))
if figure is None:
figure = figure_fn(title=title, tools=tools,
width=width, height=height,
x_axis_label=x_axis_label,
y_axis_label=y_axis_label,
x_axis_type=x_axis_type,
y_axis_type=y_axis_type,
x_range=x_range,
y_range=y_range)
plotted = False
for i, (name, result) in enumerate(experiments.items()):
if result is None or len(result) == 0:
continue
result_x, result_y = zip(*result)
figure.line(result_x, result_y, legend=name, **line_options[i])
plotted = True
if plotted:
figure.legend.click_policy = "hide"
if legend_text_font_size is not None:
figure.legend.label_text_font_size = legend_text_font_size
else:
figure = None
return figure
def dump_stream(client, stream_name):
stream = client.open_stream(name=stream_name)
output = [getattr(entry, 'value') for entry in stream.read_all()]
return reduce_multiple(output)
def reduce_ranges(results, x_values):
# assumes x_values are sorted unique values
x_values = np.array(x_values)
new_results = []
for x, y in results:
idx = x_values.searchsorted(x)
if idx < len(x_values):
x = x_values[idx]
new_results.append((x, y))
return reduce_multiple(new_results)
def annotate_indices(results, x_values, interpolate=False):
x_values = np.array(x_values)
new_results = []
for x, y in results:
idx = x_values.searchsorted(x)
if interpolate and idx < len(x_values) - 1:
length = float(x_values[idx+1] - x_values[idx])
idx = float(idx) + (x - x_values[idx]) / length
new_results.append((idx, y))
return reduce_multiple(new_results)
if __name__ == '__main__':
args = parser.parse_args()
metrics = args.metrics
additional = args.additional
if args.paper:
def figure_fn(*kargs, **kwargs):
fig = bk_figure(*kargs, **kwargs)
fig.title.text_font_size = '14pt'
fig.xaxis.axis_label_text_font_size = '24pt'
fig.yaxis.axis_label_text_font_size = '24pt'
fig.xaxis.major_label_text_font_size = '24pt'
fig.yaxis.major_label_text_font_size = '24pt'
return fig
line_width = 3
plot_defaults = dict(figure_fn=figure_fn, legend_text_font_size='20pt',
width=1000, height=500)
else:
line_width = 2
plot_defaults = dict(figure_fn=bk_figure,
width=800, height=400)
defaults_streams = ['train_' + n for n in metrics] +\
['eval_' + n for n in metrics]
if len(args.experiments) > 1:
exp_names = args.legend or [exp.replace(
'./results', '') for exp in args.experiments]
else:
exp_names = ['']
results = OrderedDict()
for exp_name, exp in zip(exp_names, args.experiments):
client = tw.WatcherClient(
filename=os.path.join(exp, 'tensorwatch.log'))
epoch_steps = [x for x, _ in dump_stream(client, 'epoch')]
prefix = exp_name + ' - ' if exp_name != '' else ''
for metric_name in metrics:
results.setdefault(metric_name, {})
train_metrics = dump_stream(client, 'train_%s' % metric_name)
val_metrics = dump_stream(client, 'eval_%s' % metric_name)
if not args.all:
train_metrics = reduce_ranges(train_metrics, epoch_steps)
val_metrics = reduce_ranges(val_metrics, epoch_steps)
if args.x_axis == 'epochs':
train_metrics = annotate_indices(
train_metrics, epoch_steps, interpolate=args.all)
val_metrics = annotate_indices(
val_metrics, epoch_steps, interpolate=args.all)
results[metric_name].update({'%svalidation %s' % (prefix, metric_name): val_metrics,
'%straining %s' % (prefix, metric_name): train_metrics})
for name in additional:
results.setdefault(name, {})
metric = dump_stream(client, name)
if not args.all:
metric = reduce_ranges(metric, epoch_steps)
if args.x_axis == 'epochs':
metric = annotate_indices(metric, epoch_steps)
results[name].update({'%s%s' % (prefix, name): metric})
line_options = OrderedDict()
if len(args.experiments) == 1:
# Training / val in different color
line_options.update({metric_name: multi_line_opts(2, line_width=line_width)
for metric_name in metrics})
line_options.update({name: multi_line_opts(1, line_width=line_width)
for name in additional})
else:
# Training / val in different line style, experiments in different color
line_options_exp = multi_line_opts(
len(args.experiments), line_width=line_width)
for i in range(len(args.experiments)):
train_ops = {'line_dash': [6, 3], **line_options_exp[i]}
val_ops = line_options_exp[i]
for metric_name in metrics:
line_options.setdefault(metric_name, [])
line_options[metric_name] += [val_ops, train_ops]
for name in additional:
line_options.setdefault(name, [])
line_options[name] += [line_options_exp[i]]
figures = []
for name, result in results.items():
fig = plot_comparison(result, title=name, line_options=line_options[name],
x_axis_label=args.x_axis, y_axis_label=name, **plot_defaults)
if fig is not None:
figures.append(fig)
plots = column(*figures)
save(plots)