|
360
|
1 #!/usr/bin/python
|
|
|
2 # coding: utf-8
|
|
|
3
|
|
|
4 from __future__ import with_statement
|
|
|
5
|
|
|
6 import jobman
|
|
|
7 from jobman import DD
|
|
|
8
|
|
|
9 from pylearn.io.seriestables import *
|
|
|
10 import tables
|
|
|
11
|
|
|
12
|
|
|
13
|
|
|
14 # from pylearn codebase
|
|
|
15 # useful in __init__(param1, param2, etc.) to save
|
|
|
16 # values in self.param1, self.param2... just call
|
|
|
17 # update_locals(self, locals())
|
|
|
18 def update_locals(obj, dct):
|
|
|
19 if 'self' in dct:
|
|
|
20 del dct['self']
|
|
|
21 obj.__dict__.update(dct)
|
|
|
22
|
|
|
23 # from a dictionary of possible values for hyperparameters, e.g.
|
|
|
24 # hp_values = {'learning_rate':[0.1, 0.01], 'num_layers': [1,2]}
|
|
|
25 # create a list of other dictionaries representing all the possible
|
|
|
26 # combinations, thus in this example creating:
|
|
|
27 # [{'learning_rate': 0.1, 'num_layers': 1}, ...]
|
|
|
28 # (similarly for combinations (0.1, 2), (0.01, 1), (0.01, 2))
|
|
|
29 def produit_cartesien_jobs(val_dict):
|
|
|
30 job_list = [DD()]
|
|
|
31 all_keys = val_dict.keys()
|
|
|
32
|
|
|
33 for key in all_keys:
|
|
|
34 possible_values = val_dict[key]
|
|
|
35 new_job_list = []
|
|
|
36 for val in possible_values:
|
|
|
37 for job in job_list:
|
|
|
38 to_insert = job.copy()
|
|
|
39 to_insert.update({key: val})
|
|
|
40 new_job_list.append(to_insert)
|
|
|
41 job_list = new_job_list
|
|
|
42
|
|
|
43 return job_list
|
|
|
44
|
|
|
45 def jobs_from_reinsert_list(cols, job_vals):
|
|
|
46 job_list = []
|
|
|
47 for vals in job_vals:
|
|
|
48 job = DD()
|
|
|
49 for i, col in enumerate(cols):
|
|
|
50 job[col] = vals[i]
|
|
|
51 job_list.append(job)
|
|
|
52
|
|
|
53 return job_list
|
|
|
54
|
|
|
55 def save_params(all_params, filename):
|
|
|
56 import pickle
|
|
|
57 with open(filename, 'wb') as f:
|
|
|
58 values = [p.value for p in all_params]
|
|
|
59
|
|
|
60 # -1 for HIGHEST_PROTOCOL
|
|
|
61 pickle.dump(values, f, -1)
|
|
|
62
|
|
|
63 # Perform insertion into the Postgre DB based on combination
|
|
|
64 # of hyperparameter values above
|
|
|
65 # (see comment for produit_cartesien_jobs() to know how it works)
|
|
|
66 def jobman_insert_job_vals(job_db, experiment_path, job_vals):
|
|
|
67 jobs = produit_cartesien_jobs(job_vals)
|
|
|
68
|
|
|
69 db = jobman.sql.db(job_db)
|
|
|
70 for job in jobs:
|
|
|
71 job.update({jobman.sql.EXPERIMENT: experiment_path})
|
|
|
72 jobman.sql.insert_dict(job, db)
|
|
|
73
|
|
|
74 def jobman_insert_specific_jobs(job_db, experiment_path,
|
|
|
75 insert_cols, insert_vals):
|
|
|
76 jobs = jobs_from_reinsert_list(insert_cols, insert_vals)
|
|
|
77
|
|
|
78 db = jobman.sql.db(job_db)
|
|
|
79 for job in jobs:
|
|
|
80 job.update({jobman.sql.EXPERIMENT: experiment_path})
|
|
|
81 jobman.sql.insert_dict(job, db)
|
|
|
82
|
|
|
83 # Just a shortcut for a common case where we need a few
|
|
|
84 # related Error (float) series
|
|
|
85 def get_accumulator_series_array( \
|
|
|
86 hdf5_file, group_name, series_names,
|
|
|
87 reduce_every,
|
|
|
88 index_names=('epoch','minibatch'),
|
|
|
89 stdout_too=True,
|
|
|
90 skip_hdf5_append=False):
|
|
|
91 all_series = []
|
|
|
92
|
|
|
93 new_group = hdf5_file.createGroup('/', group_name)
|
|
|
94
|
|
|
95 other_targets = []
|
|
|
96 if stdout_too:
|
|
|
97 other_targets = [StdoutAppendTarget()]
|
|
|
98
|
|
|
99 for sn in series_names:
|
|
|
100 series_base = \
|
|
|
101 ErrorSeries(error_name=sn,
|
|
|
102 table_name=sn,
|
|
|
103 hdf5_file=hdf5_file,
|
|
|
104 hdf5_group=new_group._v_pathname,
|
|
|
105 index_names=index_names,
|
|
|
106 other_targets=other_targets,
|
|
|
107 skip_hdf5_append=skip_hdf5_append)
|
|
|
108
|
|
|
109 all_series.append( \
|
|
|
110 AccumulatorSeriesWrapper( \
|
|
|
111 base_series=series_base,
|
|
|
112 reduce_every=reduce_every))
|
|
|
113
|
|
|
114 ret_wrapper = SeriesArrayWrapper(all_series)
|
|
|
115
|
|
|
116 return ret_wrapper
|
|
|
117
|
|
|
118
|
