-
Notifications
You must be signed in to change notification settings - Fork 1
/
experiments.py
285 lines (224 loc) · 16 KB
/
experiments.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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
import os
import numpy as np
from sklearn.metrics import accuracy_score, explained_variance_score
from algorithms import inner_solver_selector
from losses import AbsoluteLoss, HingeLoss
from plots import plot_resultsList
from train import EXP_FOLDER, Results, val, meta_val
from utils import make_exp_dir, save_exp_parameters
from data import data_generator, data_load
def exp_itl_only(exp_str='exp1', seed = 0, lambdas = np.logspace(-6, 3, num=10),
gamma = None, n_processes=30, w_bar = 4, y_snr = 100, task_std = 1, n_tasks = 100, n_train = 100, n_dims = 30,
n_tasks_test = 200, n_test = 100, val_perc = 0.0, h_start = None,
exp_dir = EXP_FOLDER,
inner_solver_test_str = ('ssubgd', 'fista'), show_plot=False, verbose =0):
loss_class, tasks_gen, inner_exp_name,\
metric_dict, val_metric = select_exp(exp_str, seed=seed, task_std=task_std, y_snr=y_snr,
n_train_tasks=n_tasks, n_val_tasks=n_tasks_test, n_dims=n_dims,
val_perc=val_perc, w_bar=w_bar)
exp_name = 'grid_searchITL' + inner_exp_name + 'seed' + str(seed) + 'vm' + val_metric + 'is' \
+ 'ist' + str(inner_solver_test_str) + 'n' + str(n_train) + 'val_perc' + str(val_perc)
exp_parameters = locals()
print('parameters ' + exp_name, exp_parameters)
n_tasks_val = n_tasks_test//2 # use half the tasks for testing for the validation (with synthetic datasets)
data_train, oracle_train = tasks_gen(n_tasks=n_tasks, n_train=n_train, n_test=n_test, sel='train')
data_valid, oracle_valid = tasks_gen(n_tasks=n_tasks_val, n_train=n_train, n_test=n_test, sel='val')
data_test, oracle_test = tasks_gen(n_tasks=n_tasks_test, n_train=n_train, n_test=n_test, sel='test')
print('n_tasks train, val ,test', len(data_train['X_train']), len(data_valid['X_train']),
len(data_test['X_train']))
exp_dir_path = make_exp_dir(os.path.join(exp_dir, exp_name))
save_exp_parameters(exp_parameters, exp_dir_path)
res_dict = {}
for ts in inner_solver_test_str:
inner_solver_test_class = inner_solver_selector(ts)
# metaval for ITL
results = Results(save_dir=exp_dir_path, do_plot=False, show_plot=show_plot, name='ITL-ts' + ts)
if h_start is None:
h = np.zeros(tasks_gen.n_dims)
else:
h = np.copy(h_start)
itl_res = val(h, val_metric, lambdas, gamma, inner_solver_test_class, loss_class,
data_valid, data_test, metric_dict, results, n_processes=n_processes, verbose=verbose)
res_dict[itl_res.name] = itl_res
plot_resultsList(n_tasks+1, res_dict, save_dir_path=exp_dir_path, show_plot=show_plot, filename='ltl_plots')
return res_dict
def exp(exp_str = 'exp1', seed=0, lambdas = np.logspace(-6, 3, num=10), alphas = np.logspace(-6, 3, num=10),
gamma = None, n_processes = 30, w_bar = 4, y_snr=100, task_std = 1, n_tasks = 100, n_train = 100, n_dims = 30,
n_tasks_test = 200, n_test = 100, val_perc = 0.0,
exp_dir = EXP_FOLDER, inner_solver_str = ('ssubgd', 'fista'),
inner_solver_test_str = ('ssubgd', 'fista'), show_plot=False, verbose =0):
loss_class, tasks_gen, inner_exp_name,\
metric_dict, val_metric = select_exp(exp_str, seed=seed, task_std=task_std, y_snr=y_snr,
n_train_tasks=n_tasks, n_val_tasks=n_tasks_test, n_dims=n_dims,
val_perc=val_perc, w_bar=w_bar)
exp_name = 'grid_search' + inner_exp_name + 'seed' + str(seed) + 'vm' + val_metric + 'is' \
+ str(inner_solver_str) + 'ist' + str(inner_solver_test_str) + 'n' + str(n_train) + 'val_perc' + str(val_perc)
n_tasks_val = n_tasks_test//2 # use half the tasks for testing for the validation (with synthetic datasets)
exp_parameters = locals()
print('parameters ' + exp_name, exp_parameters)
exp_dir_path = make_exp_dir(os.path.join(exp_dir, exp_name))
save_exp_parameters(exp_parameters, exp_dir_path)
with open(os.path.join(exp_dir_path, tasks_gen.desc+".txt"), "w") as text_file:
text_file.write(tasks_gen.desc)
data_train, oracle_train = tasks_gen(n_tasks=n_tasks, n_train=n_train, n_test=n_test, sel='train')
data_valid, oracle_valid = tasks_gen(n_tasks=n_tasks_val, n_train=n_train, n_test=n_test, sel='val')
data_test, oracle_test = tasks_gen(n_tasks=n_tasks_test, n_train=n_train, n_test=n_test, sel='test')
res_dict = {}
for ts in inner_solver_test_str:
inner_solver_test_class = inner_solver_selector(ts)
# metaval for ITL
results = Results(save_dir=exp_dir_path, do_plot=False, show_plot=show_plot, name='ITL-ts' + ts)
h = np.zeros(tasks_gen.n_dims)
itl_res = val(h, val_metric, lambdas, gamma, inner_solver_test_class, loss_class,
data_valid, data_test, metric_dict, results, n_processes=n_processes, verbose=verbose)
res_dict[itl_res.name] = itl_res
# metaval for MEAN
if oracle_valid is not None:
results = Results(save_dir=exp_dir_path, do_plot=False, show_plot=show_plot, name='MEAN-ts' + ts)
h = oracle_valid['w_bar']
oracle_res = val(h, val_metric, lambdas, gamma, inner_solver_test_class, loss_class,
data_valid, data_test, metric_dict, results, n_processes=n_processes, verbose=verbose)
res_dict[oracle_res.name] = oracle_res
for s in inner_solver_str:
inner_solver_class = inner_solver_selector(s)
results = Results(save_dir=exp_dir_path, do_plot=False, show_plot=show_plot, name='LTL-tr'+s +'ts'+ts)
h0 = np.zeros(tasks_gen.n_dims)
ltl_res = meta_val(val_metric=val_metric, h0=h0, alphas=alphas, lambdas=lambdas, gamma=gamma,
inner_solver_class=inner_solver_class, inner_solver_test_class=inner_solver_test_class,
loss_class=loss_class, data_train=data_train, data_valid=data_valid, data_test=data_test,
metric_dict=metric_dict, results=results, n_processes=n_processes, verbose=verbose)
res_dict[ltl_res.name] = ltl_res
plot_resultsList(n_tasks+1, res_dict, save_dir_path=exp_dir_path, show_plot=show_plot, filename='plots')
return res_dict
def multi_seed(exp_str='exp1', seeds=list(range(10)), lambdas=np.logspace(-6, 3, num=10), alphas=np.logspace(-6, 3,
num=10),
gamma=None, n_processes=30, w_bar=4, y_snr=100, task_std=1, n_tasks=100, n_train=100, n_dims=30,
n_tasks_test=200, n_test=100, val_perc=0.0, search_oracle=False,
exp_dir=EXP_FOLDER, inner_solver_str=('ssubgd', 'subgd'),
use_hyper_bounds=False, inner_solver_test_str=('ssubgd', 'subgd'), show_plot=True, save_res=True, verbose=1):
loss_class, tasks_gen, inner_exp_name,\
metric_dict, val_metric = select_exp(exp_str, seed=seeds[0], task_std=task_std, y_snr=y_snr,
n_train_tasks=n_tasks, n_val_tasks=n_tasks_test, n_dims=n_dims,
val_perc=val_perc, w_bar=w_bar)
exp_name = 'grid_search' + inner_exp_name + 'over' + str(len(seeds)) + 'seeds' + 'is' \
+ str(inner_solver_str) + 'ist' + str(inner_solver_test_str) + 'n' + str(n_train) + 'val_perc' + str(val_perc)
exp_parameters = locals()
print('parameters ' + exp_str, exp_parameters)
exp_dir_path = make_exp_dir(os.path.join(exp_dir, exp_name))
res_list = []
avg_res_dict = {}
for s in seeds:
r = exp(exp_str=exp_str, seed=s, lambdas=lambdas, alphas=alphas, gamma=gamma, n_processes=n_processes,
w_bar=w_bar, y_snr=y_snr, task_std=task_std, n_tasks=n_tasks, n_train=n_train, n_dims=n_dims,
n_tasks_test=n_tasks_test, n_test=n_test, exp_dir=exp_dir_path, val_perc=val_perc,
inner_solver_str=inner_solver_str, inner_solver_test_str=inner_solver_test_str,
show_plot=show_plot, verbose=verbose)
res_list.append(r)
for name in list(res_list[0].keys()):
avg_res_dict[name] = Results(save_dir=exp_dir_path, do_plot=False, show_plot=False, name=name)
metric_dict = {}
for m_name in list(res_list[0][name].metrics.keys()):
metric_dict[m_name] = [res[name].metrics[m_name] for res in res_list]
last_axes = len(metric_dict[m_name][0].shape)-1
metric_dict[m_name] = np.concatenate([np.expand_dims(np.mean(o, axis=last_axes), last_axes)
for o in metric_dict[m_name]], axis=last_axes)
avg_res_dict[name].add_metrics(metric_dict)
plot_resultsList(n_tasks + 1, avg_res_dict, save_dir_path=exp_dir_path, show_plot=show_plot, filename='plots',
title=str(len(res_list))+' runs')
def lenk_multi_seed(seeds=list(range(10)), lambdas=np.logspace(-3, 3, num=30), alphas=np.logspace(-3, 3, num=30), reg=False,
n_train=None,gamma=None, n_processes=30, n_tasks=100, n_val_tasks=40, exp_dir=EXP_FOLDER, inner_solver_str=('ssubgd', 'subgd'),
use_hyper_bounds=False, inner_solver_test_str=('ssubgd', 'fista'), show_plot=True, save_res=True, verbose=1):
reg_str = '' if not reg else 'Reg'
return multi_seed(exp_str='lenk'+reg_str, seeds=seeds, lambdas=lambdas, alphas=alphas,
gamma=gamma, n_processes=n_processes, w_bar=0, y_snr=0, task_std=0, n_tasks=n_tasks, n_train=n_train, n_dims=0,
n_tasks_test=n_val_tasks, n_test=0, exp_dir=exp_dir, inner_solver_str=inner_solver_str,
use_hyper_bounds=use_hyper_bounds, inner_solver_test_str=inner_solver_test_str, show_plot=show_plot,
save_res=save_res, verbose=verbose)
def mulan_multi_seed(data_name='Corel5k', seeds=list(range(10)), lambdas=np.logspace(-3, 3, num=5), alphas=np.logspace(-3, 3, num=5),
n_train=None, gamma=None, n_processes=30, n_tasks=200, n_val_tasks=73, exp_dir=EXP_FOLDER, inner_solver_str=('ssubgd'),
use_hyper_bounds=False, inner_solver_test_str=('ssubgd'), show_plot=True, save_res=True, verbose=1):
return multi_seed(exp_str=data_name, seeds=seeds, lambdas=lambdas, alphas=alphas,
gamma=gamma, n_processes=n_processes, w_bar=0, y_snr=0, task_std=0, n_tasks=n_tasks, n_train=n_train, n_dims=0,
n_tasks_test=n_val_tasks, n_test=0, exp_dir=exp_dir, inner_solver_str=inner_solver_str,
use_hyper_bounds=use_hyper_bounds, inner_solver_test_str=inner_solver_test_str, show_plot=show_plot,
save_res=save_res, verbose=verbose)
def school_multi_seed(seeds=list(range(10)), lambdas=np.logspace(-3, 3, num=30), alphas=np.logspace(-1, 6, num=30),
gamma=None, n_processes=30, n_tasks=75, n_val_tasks=25, exp_dir=EXP_FOLDER, inner_solver_str=('ssubgd', 'subgd'),
use_hyper_bounds=False, inner_solver_test_str='ssubgd', show_plot=True, save_res=True, verbose=1):
return multi_seed(exp_str='school', seeds=seeds, lambdas=lambdas, alphas=alphas,
gamma=gamma, n_processes=n_processes, w_bar=0, y_snr=0, task_std=0, n_tasks=n_tasks, n_train=None, n_dims=0,
n_tasks_test=n_val_tasks, n_test=0, val_perc=0.5, exp_dir=exp_dir, inner_solver_str=inner_solver_str,
use_hyper_bounds=use_hyper_bounds, inner_solver_test_str=inner_solver_test_str, show_plot=show_plot,
save_res=save_res, verbose=verbose)
def select_exp(exp_str, seed=0, task_std=1, y_snr=10, val_perc=0.5, w_bar=4, n_dims=30,
n_train_tasks=0, n_val_tasks=0):
if exp_str == 'exp1':
tasks_gen = data_generator.TasksGenerator(seed=seed, task_std=task_std, y_snr=y_snr, val_perc=val_perc, n_dims=n_dims,
tasks_generation='exp1', w_bar=w_bar)
exp_name = exp_str + 'w_bar' + str(w_bar) + 'taskstd' + str(task_std) + 'y_snr' + str(y_snr) + \
'dim' + str(n_dims)
loss = AbsoluteLoss
metric_dict = {}
val_metric = 'loss'
elif exp_str == 'exp2':
tasks_gen = data_generator.TasksGenerator(seed=seed, task_std=task_std, y_snr=y_snr, val_perc=val_perc, n_dims=n_dims,
tasks_generation='expclass', w_bar=w_bar)
exp_name = exp_str + 'w_bar' + str(w_bar) + 'taskstd' + str(task_std) + 'y_snr' + str(y_snr) + \
'dim' + str(n_dims) + 'y_dist' + str(tasks_gen.y_dist)
loss = HingeLoss
val_metric = 'loss'
metric_dict = {'accuracy': accuracy}
elif exp_str == 'school':
tasks_gen = data_load.RealDatasetGenerator(gen_f=data_load.schools_data_gen, seed=seed,
n_train_tasks=n_train_tasks,
n_val_tasks=n_val_tasks,
val_perc=val_perc)
exp_name = 'expSchool' + 'n_tasks_val' + str(n_val_tasks) + 'n_tasks' + str(tasks_gen.n_tasks) \
+ 'dim' + str(tasks_gen.n_dims)
loss = AbsoluteLoss
val_metric = 'loss'
metric_dict = {'negexpvar': ne_exp_var}
elif exp_str == 'lenk':
tasks_gen = data_load.RealDatasetGenerator(gen_f=data_load.computer_data_gen,
seed=seed, n_train_tasks=n_train_tasks,
n_val_tasks=n_val_tasks,
val_perc=val_perc)
exp_name = 'expLenk' + 'n_tasks_train' + str(n_train_tasks) + 'n_tasks_val' + str(n_val_tasks) \
+ 'n_tasks' + str(tasks_gen.n_tasks) + 'dim' + str(tasks_gen.n_dims)
loss = HingeLoss
val_metric = 'loss'
metric_dict = {'accuracy': accuracy}
elif exp_str == 'lenkReg':
tasks_gen = data_load.RealDatasetGenerator(gen_f=data_load.computer_data_ge_reg,
seed=seed, n_train_tasks=n_train_tasks,
n_val_tasks=n_val_tasks,
val_perc=val_perc)
exp_name = 'expLenkReg' + 'n_tasks_train' + str(n_train_tasks) + 'n_tasks_val' + str(n_val_tasks) \
+ 'n_tasks' + str(tasks_gen.n_tasks) + 'dim' + str(tasks_gen.n_dims)
loss = AbsoluteLoss
val_metric = 'loss'
metric_dict = {}
elif exp_str in data_load.mulan_dict.keys():
tasks_gen = data_load.RealDatasetGenerator(gen_f=data_load.get_mulan_loader(exp_str),
seed=seed, n_train_tasks=n_train_tasks,
n_val_tasks=n_val_tasks,
val_perc=val_perc)
exp_name = 'exp' + exp_str + 'n_tasks-tot-tr-val-tst' + str(tasks_gen.n_tasks) + '-' + str(tasks_gen.n_train_tasks) \
+ '-' + str(tasks_gen.n_val_tasks) + '-' + str(tasks_gen.n_test_tasks) \
+ 'dim' + str(tasks_gen.n_dims)
loss = HingeLoss
val_metric = 'loss'
metric_dict = {'accuracy': accuracy}
else:
raise NotImplementedError('exp: {} not implemented'.format(exp_str))
return loss, tasks_gen, exp_name, metric_dict, val_metric
def dg(d, k): return None if d is None else d[k]
def accuracy(y_true, y_pred):
yp = np.ones_like(y_true)
yp[y_pred < 0] = -1
return accuracy_score(y_true, yp)
def ne_exp_var(y_true, y_pred):
return - explained_variance_score(y_true, y_pred)
if __name__ == '__main__':
lenk_multi_seed(lambdas=[0.1, 10], alphas=[0.1, 10], reg=True,
inner_solver_test_str=['ssubgd', 'fista'])