-
Notifications
You must be signed in to change notification settings - Fork 0
/
parse_tensorboard_efficient.py
196 lines (173 loc) · 7.18 KB
/
parse_tensorboard_efficient.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
import gzip
import json
from turtle import update
import os
import math
def find_all_matches(event, d):
min_idx = 0
max_idx = len(d) - 1
cur_idx = int(max_idx / 2)
target_time = event['ts'] + event['dur'] / 2
oscil_count = 0
while min_idx < max_idx:
if max_idx - min_idx == 1:
oscil_count += 1
if oscil_count > 5:
break
# print(min_idx, max_idx, cur_idx)
if target_time < d[cur_idx]['ts'] - 0.1:
# print("small")
max_idx = cur_idx
cur_idx = int(min_idx + math.floor((max_idx - min_idx)/2))
elif d[cur_idx]['ts'] + d[cur_idx]['dur'] + 0.1< target_time:
# print("big")
min_idx = cur_idx
cur_idx = int(min_idx + math.ceil((max_idx - min_idx)/2))
else:
return d[cur_idx]
return None
def parse(path):
print("----------------------------------------")
print(path[:path.find("2022")].split("/")[1])
gzf = gzip.open(path)
d = json.load(gzf)
xla_ops_list = []
tf_ops_list = []
tf_name_list = {}
tf_name_list['0'] = []
tf_name_list['1'] = []
tf_name_list['2'] = []
tf_name_list['3'] = []
forward = 0
backward = 0
computing_norms = 0
clipping_grads = 0
second_backprop = 0
reduce = 0
add_noise = 0
update_params = 0
elses = 0
cpu = 0
tpu = 0
for event in d['traceEvents']:
if 'pid' in event and event['pid'] == 3:
# Event is on TPU.
tpu += 1
if 'tid' in event:
if event['tid'] == 5:
# XLA Ops
if 'ts' in event:
xla_ops_list += [event]
elif event['tid'] == 3:
# TF Ops
if 'ts' in event:
tf_ops_list += [event]
elif event['tid'] == 2:
if 'ts' in event and 'args' in event and 'l' in event['args']:
if event['args']['l'] in ['0', '1', '2', '3']:
tf_name_list[event['args']['l']] += [event]
else:
cpu += 1
for event in xla_ops_list:
anyone = False
if 'ts' in event and 'while' not in event['name']:
anyone = False
events = []
matched_tf_name = []
for l in ['0', '1', '2', '3']:
matched_tf_name += [find_all_matches(event, tf_name_list[l])]
matched_tf_ops = []
matched_tf_ops += [find_all_matches(event, tf_ops_list)]
for tf_name in matched_tf_name:
if not tf_name:
continue
if 'ts' in tf_name:
# print(tf_name)
if (tf_name['ts'] <= event['ts'] + 0.1) and (event['ts'] + event['dur'] - 0.1 <= tf_name['ts'] + tf_name['dur']):
# print(event['ts'], event['dur'], tf_name['ts'], tf_name['dur'])
if 'tf_name' not in event:
event['tf_name'] = [tf_name]
else:
event['tf_name'] += [tf_name]
if "_SGD_" not in path:
if tf_name['args']['l'] == '1' and tf_name['name'] == "computing_grads":
forward += event['dur']
anyone=True
elif tf_name['args']['l'] == '2' and tf_name['name'] == "computing_grads":
backward += event['dur']
anyone=True
elif tf_name['args']['l'] in ['0', '1', '2'] and tf_name['name'] == "second_backprop":
second_backprop += event['dur']
anyone=True
else:
if tf_name['args']['l'] == '0' and tf_name['name'] == "computing_grads":
forward += event['dur']
anyone=True
elif tf_name['args']['l'] == '1' and tf_name['name'] == "computing_grads":
backward += event['dur']
anyone=True
elif tf_name['args']['l'] in ['0', '1', '2'] and tf_name['name'] == "second_backprop":
second_backprop += event['dur']
anyone=True
if tf_name['args']['l'] == '1' and tf_name['name'] == "computing_norms":
computing_norms += event['dur']
anyone=True
if tf_name['args']['l'] == '1' and tf_name['name'] == "compute_clipping_factor":
computing_norms += event['dur']
anyone=True
elif tf_name['args']['l'] == '1' and tf_name['name'] == "clipping_grads":
clipping_grads += event['dur']
anyone=True
elif tf_name['args']['l'] == '0' and tf_name['name'] == "add_noise_and_reduce":
if 'reduce' in event['args']['hlo_category']:
reduce += event['dur']
else:
add_noise += event['dur']
anyone=True
# print(event)
if anyone==True:
break
# print(event)
for tf_op in matched_tf_ops:
if not tf_op:
continue
if 'ts' in tf_op:
# print(tf_op)
if (tf_op['ts'] <= event['ts'] + 0.1) and (event['ts'] + event['dur'] - 0.1 <= tf_op['ts'] + tf_op['dur']):
if 'tf_op' not in event:
event['tf_op'] = [tf_op]
else:
event['tf_op'] += [tf_op]
if tf_op['name'] == "ResourceApplyGradientDescent":
update_params += event['dur']
anyone=True
if not anyone:
# print(event)
elses += event['dur']
else:
anyone = True
# if anyone:
# print(event)
# print("AAAAAAAAA")
print("forward,backward,computing_norms,clipping_grads_and_reduce,second_backprop,add_noise,update_params,elses")
print(forward/1000, end=",")
print(backward/1000, end=",")
print(computing_norms/1000, end=",")
print(clipping_grads/1000 + reduce/1000, end=",")
print(second_backprop/1000, end=",")
print(add_noise/1000, end=",")
print(update_params/1000, end=",")
print(elses/1000)
for root, subdirs, files in os.walk("logs"):
# if "large_batch" in root:
# continue
# if "mobilenet" in root:
# continue
# if "squeezenet" in root:
# continue
# if "smallnet" in root:
# continue
# print(root, subdirs, files)
for file in files:
if "trace.json.gz" in file:
parse(os.path.join(root, file))