|
1
|
|
|
from .contextlibbackport import ContextDecorator |
|
2
|
|
|
from .internal import ContextMethodDecorator |
|
3
|
|
|
import sacred.optional as opt |
|
4
|
|
|
if opt.has_tensorflow: |
|
5
|
|
|
import tensorflow |
|
6
|
|
|
else: |
|
7
|
|
|
tensorflow = None |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
class LogFileWriter(ContextDecorator, ContextMethodDecorator): |
|
11
|
|
|
""" |
|
12
|
|
|
Intercept ``logdir`` each time a new ``FileWriter`` instance is created. |
|
13
|
|
|
|
|
14
|
|
|
:param experiment: Tensorflow experiment. |
|
15
|
|
|
|
|
16
|
|
|
The state of the experiment must be running when entering the annotated |
|
17
|
|
|
function / the context manager. |
|
18
|
|
|
|
|
19
|
|
|
When creating ``FileWriters`` in Tensorflow, you might want to |
|
20
|
|
|
store the path to the produced log files in the sacred database. |
|
21
|
|
|
|
|
22
|
|
|
In the scope of ``LogFileWriter``, the corresponding log directory path |
|
23
|
|
|
is appended to a list in experiment.info["tensorflow"]["logdirs"]. |
|
24
|
|
|
|
|
25
|
|
|
``LogFileWriter`` can be used both as a context manager or as |
|
26
|
|
|
an annotation (decorator) on a function. |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
Example usage as decorator:: |
|
30
|
|
|
|
|
31
|
|
|
ex = Experiment("my experiment") |
|
32
|
|
|
@LogFileWriter(ex) |
|
33
|
|
|
def run_experiment(_run): |
|
34
|
|
|
with tf.Session() as s: |
|
35
|
|
|
swr = tf.summary.FileWriter("/tmp/1", s.graph) |
|
36
|
|
|
# _run.info["tensorflow"]["logdirs"] == ["/tmp/1"] |
|
37
|
|
|
swr2 tf.summary.FileWriter("./test", s.graph) |
|
38
|
|
|
#_run.info["tensorflow"]["logdirs"] == ["/tmp/1", "./test"] |
|
39
|
|
|
|
|
40
|
|
|
|
|
41
|
|
|
Example usage as context manager:: |
|
42
|
|
|
|
|
43
|
|
|
ex = Experiment("my experiment") |
|
44
|
|
|
def run_experiment(_run): |
|
45
|
|
|
with tf.Session() as s: |
|
46
|
|
|
with LogFileWriter(ex): |
|
47
|
|
|
swr = tf.summary.FileWriter("/tmp/1", s.graph) |
|
48
|
|
|
# _run.info["tensorflow"]["logdirs"] == ["/tmp/1"] |
|
49
|
|
|
swr3 = tf.summary.FileWriter("./test", s.graph) |
|
50
|
|
|
#_run.info["tensorflow"]["logdirs"] == ["/tmp/1", "./test"] |
|
51
|
|
|
# This is called outside the scope and won't be captured |
|
52
|
|
|
swr3 = tf.summary.FileWriter("./nothing", s.graph) |
|
53
|
|
|
# Nothing has changed: |
|
54
|
|
|
#_run.info["tensorflow"]["logdirs"] == ["/tmp/1", "./test"] |
|
55
|
|
|
|
|
56
|
|
|
""" |
|
57
|
|
|
|
|
58
|
|
|
def __init__(self, experiment): |
|
59
|
|
|
self.experiment = experiment |
|
60
|
|
|
|
|
61
|
|
|
def log_writer_decorator(instance, original_method, original_args, |
|
62
|
|
|
original_kwargs): |
|
63
|
|
|
result = original_method(instance, *original_args, |
|
64
|
|
|
**original_kwargs) |
|
65
|
|
|
if "logdir" in original_kwargs: |
|
66
|
|
|
logdir = original_kwargs["logdir"] |
|
67
|
|
|
else: |
|
68
|
|
|
logdir = original_args[0] |
|
69
|
|
|
self.experiment.info.setdefault("tensorflow", {}).setdefault( |
|
70
|
|
|
"logdirs", []).append(logdir) |
|
71
|
|
|
return result |
|
72
|
|
|
|
|
73
|
|
|
ContextMethodDecorator.__init__(self, |
|
74
|
|
|
tensorflow.summary.FileWriter, |
|
75
|
|
|
"__init__", |
|
76
|
|
|
log_writer_decorator) |
|
77
|
|
|
|
IDSIA /
sacred
