-
Notifications
You must be signed in to change notification settings - Fork 605
/
main.py
348 lines (287 loc) · 17 KB
/
main.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
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
"""
Usage Instructions:
10-shot sinusoid:
python main.py --datasource=sinusoid --logdir=logs/sine/ --metatrain_iterations=70000 --norm=None --update_batch_size=10
10-shot sinusoid baselines:
python main.py --datasource=sinusoid --logdir=logs/sine/ --pretrain_iterations=70000 --metatrain_iterations=0 --norm=None --update_batch_size=10 --baseline=oracle
python main.py --datasource=sinusoid --logdir=logs/sine/ --pretrain_iterations=70000 --metatrain_iterations=0 --norm=None --update_batch_size=10
5-way, 1-shot omniglot:
python main.py --datasource=omniglot --metatrain_iterations=60000 --meta_batch_size=32 --update_batch_size=1 --update_lr=0.4 --num_updates=1 --logdir=logs/omniglot5way/
20-way, 1-shot omniglot:
python main.py --datasource=omniglot --metatrain_iterations=60000 --meta_batch_size=16 --update_batch_size=1 --num_classes=20 --update_lr=0.1 --num_updates=5 --logdir=logs/omniglot20way/
5-way 1-shot mini imagenet:
python main.py --datasource=miniimagenet --metatrain_iterations=60000 --meta_batch_size=4 --update_batch_size=1 --update_lr=0.01 --num_updates=5 --num_classes=5 --logdir=logs/miniimagenet1shot/ --num_filters=32 --max_pool=True
5-way 5-shot mini imagenet:
python main.py --datasource=miniimagenet --metatrain_iterations=60000 --meta_batch_size=4 --update_batch_size=5 --update_lr=0.01 --num_updates=5 --num_classes=5 --logdir=logs/miniimagenet5shot/ --num_filters=32 --max_pool=True
To run evaluation, use the '--train=False' flag and the '--test_set=True' flag to use the test set.
For omniglot and miniimagenet training, acquire the dataset online, put it in the correspoding data directory, and see the python script instructions in that directory to preprocess the data.
Note that better sinusoid results can be achieved by using a larger network.
"""
import csv
import numpy as np
import pickle
import random
import tensorflow as tf
from data_generator import DataGenerator
from maml import MAML
from tensorflow.python.platform import flags
FLAGS = flags.FLAGS
## Dataset/method options
flags.DEFINE_string('datasource', 'sinusoid', 'sinusoid or omniglot or miniimagenet')
flags.DEFINE_integer('num_classes', 5, 'number of classes used in classification (e.g. 5-way classification).')
# oracle means task id is input (only suitable for sinusoid)
flags.DEFINE_string('baseline', None, 'oracle, or None')
## Training options
flags.DEFINE_integer('pretrain_iterations', 0, 'number of pre-training iterations.')
flags.DEFINE_integer('metatrain_iterations', 15000, 'number of metatraining iterations.') # 15k for omniglot, 50k for sinusoid
flags.DEFINE_integer('meta_batch_size', 25, 'number of tasks sampled per meta-update')
flags.DEFINE_float('meta_lr', 0.001, 'the base learning rate of the generator')
flags.DEFINE_integer('update_batch_size', 5, 'number of examples used for inner gradient update (K for K-shot learning).')
flags.DEFINE_float('update_lr', 1e-3, 'step size alpha for inner gradient update.') # 0.1 for omniglot
flags.DEFINE_integer('num_updates', 1, 'number of inner gradient updates during training.')
## Model options
flags.DEFINE_string('norm', 'batch_norm', 'batch_norm, layer_norm, or None')
flags.DEFINE_integer('num_filters', 64, 'number of filters for conv nets -- 32 for miniimagenet, 64 for omiglot.')
flags.DEFINE_bool('conv', True, 'whether or not to use a convolutional network, only applicable in some cases')
flags.DEFINE_bool('max_pool', False, 'Whether or not to use max pooling rather than strided convolutions')
flags.DEFINE_bool('stop_grad', False, 'if True, do not use second derivatives in meta-optimization (for speed)')
## Logging, saving, and testing options
flags.DEFINE_bool('log', True, 'if false, do not log summaries, for debugging code.')
flags.DEFINE_string('logdir', '/tmp/data', 'directory for summaries and checkpoints.')
flags.DEFINE_bool('resume', True, 'resume training if there is a model available')
flags.DEFINE_bool('train', True, 'True to train, False to test.')
flags.DEFINE_integer('test_iter', -1, 'iteration to load model (-1 for latest model)')
flags.DEFINE_bool('test_set', False, 'Set to true to test on the the test set, False for the validation set.')
flags.DEFINE_integer('train_update_batch_size', -1, 'number of examples used for gradient update during training (use if you want to test with a different number).')
flags.DEFINE_float('train_update_lr', -1, 'value of inner gradient step step during training. (use if you want to test with a different value)') # 0.1 for omniglot
def train(model, saver, sess, exp_string, data_generator, resume_itr=0):
SUMMARY_INTERVAL = 100
SAVE_INTERVAL = 1000
if FLAGS.datasource == 'sinusoid':
PRINT_INTERVAL = 1000
TEST_PRINT_INTERVAL = PRINT_INTERVAL*5
else:
PRINT_INTERVAL = 100
TEST_PRINT_INTERVAL = PRINT_INTERVAL*5
if FLAGS.log:
train_writer = tf.summary.FileWriter(FLAGS.logdir + '/' + exp_string, sess.graph)
print('Done initializing, starting training.')
prelosses, postlosses = [], []
num_classes = data_generator.num_classes # for classification, 1 otherwise
multitask_weights, reg_weights = [], []
for itr in range(resume_itr, FLAGS.pretrain_iterations + FLAGS.metatrain_iterations):
feed_dict = {}
if 'generate' in dir(data_generator):
batch_x, batch_y, amp, phase = data_generator.generate()
if FLAGS.baseline == 'oracle':
batch_x = np.concatenate([batch_x, np.zeros([batch_x.shape[0], batch_x.shape[1], 2])], 2)
for i in range(FLAGS.meta_batch_size):
batch_x[i, :, 1] = amp[i]
batch_x[i, :, 2] = phase[i]
inputa = batch_x[:, :num_classes*FLAGS.update_batch_size, :]
labela = batch_y[:, :num_classes*FLAGS.update_batch_size, :]
inputb = batch_x[:, num_classes*FLAGS.update_batch_size:, :] # b used for testing
labelb = batch_y[:, num_classes*FLAGS.update_batch_size:, :]
feed_dict = {model.inputa: inputa, model.inputb: inputb, model.labela: labela, model.labelb: labelb}
if itr < FLAGS.pretrain_iterations:
input_tensors = [model.pretrain_op]
else:
input_tensors = [model.metatrain_op]
if (itr % SUMMARY_INTERVAL == 0 or itr % PRINT_INTERVAL == 0):
input_tensors.extend([model.summ_op, model.total_loss1, model.total_losses2[FLAGS.num_updates-1]])
if model.classification:
input_tensors.extend([model.total_accuracy1, model.total_accuracies2[FLAGS.num_updates-1]])
result = sess.run(input_tensors, feed_dict)
if itr % SUMMARY_INTERVAL == 0:
prelosses.append(result[-2])
if FLAGS.log:
train_writer.add_summary(result[1], itr)
postlosses.append(result[-1])
if (itr!=0) and itr % PRINT_INTERVAL == 0:
if itr < FLAGS.pretrain_iterations:
print_str = 'Pretrain Iteration ' + str(itr)
else:
print_str = 'Iteration ' + str(itr - FLAGS.pretrain_iterations)
print_str += ': ' + str(np.mean(prelosses)) + ', ' + str(np.mean(postlosses))
print(print_str)
prelosses, postlosses = [], []
if (itr!=0) and itr % SAVE_INTERVAL == 0:
saver.save(sess, FLAGS.logdir + '/' + exp_string + '/model' + str(itr))
# sinusoid is infinite data, so no need to test on meta-validation set.
if (itr!=0) and itr % TEST_PRINT_INTERVAL == 0 and FLAGS.datasource !='sinusoid':
if 'generate' not in dir(data_generator):
feed_dict = {}
if model.classification:
input_tensors = [model.metaval_total_accuracy1, model.metaval_total_accuracies2[FLAGS.num_updates-1], model.summ_op]
else:
input_tensors = [model.metaval_total_loss1, model.metaval_total_losses2[FLAGS.num_updates-1], model.summ_op]
else:
batch_x, batch_y, amp, phase = data_generator.generate(train=False)
inputa = batch_x[:, :num_classes*FLAGS.update_batch_size, :]
inputb = batch_x[:, num_classes*FLAGS.update_batch_size:, :]
labela = batch_y[:, :num_classes*FLAGS.update_batch_size, :]
labelb = batch_y[:, num_classes*FLAGS.update_batch_size:, :]
feed_dict = {model.inputa: inputa, model.inputb: inputb, model.labela: labela, model.labelb: labelb, model.meta_lr: 0.0}
if model.classification:
input_tensors = [model.total_accuracy1, model.total_accuracies2[FLAGS.num_updates-1]]
else:
input_tensors = [model.total_loss1, model.total_losses2[FLAGS.num_updates-1]]
result = sess.run(input_tensors, feed_dict)
print('Validation results: ' + str(result[0]) + ', ' + str(result[1]))
saver.save(sess, FLAGS.logdir + '/' + exp_string + '/model' + str(itr))
# calculated for omniglot
NUM_TEST_POINTS = 600
def test(model, saver, sess, exp_string, data_generator, test_num_updates=None):
num_classes = data_generator.num_classes # for classification, 1 otherwise
np.random.seed(1)
random.seed(1)
metaval_accuracies = []
for _ in range(NUM_TEST_POINTS):
if 'generate' not in dir(data_generator):
feed_dict = {}
feed_dict = {model.meta_lr : 0.0}
else:
batch_x, batch_y, amp, phase = data_generator.generate(train=False)
if FLAGS.baseline == 'oracle': # NOTE - this flag is specific to sinusoid
batch_x = np.concatenate([batch_x, np.zeros([batch_x.shape[0], batch_x.shape[1], 2])], 2)
batch_x[0, :, 1] = amp[0]
batch_x[0, :, 2] = phase[0]
inputa = batch_x[:, :num_classes*FLAGS.update_batch_size, :]
inputb = batch_x[:,num_classes*FLAGS.update_batch_size:, :]
labela = batch_y[:, :num_classes*FLAGS.update_batch_size, :]
labelb = batch_y[:,num_classes*FLAGS.update_batch_size:, :]
feed_dict = {model.inputa: inputa, model.inputb: inputb, model.labela: labela, model.labelb: labelb, model.meta_lr: 0.0}
if model.classification:
result = sess.run([model.metaval_total_accuracy1] + model.metaval_total_accuracies2, feed_dict)
else: # this is for sinusoid
result = sess.run([model.total_loss1] + model.total_losses2, feed_dict)
metaval_accuracies.append(result)
metaval_accuracies = np.array(metaval_accuracies)
means = np.mean(metaval_accuracies, 0)
stds = np.std(metaval_accuracies, 0)
ci95 = 1.96*stds/np.sqrt(NUM_TEST_POINTS)
print('Mean validation accuracy/loss, stddev, and confidence intervals')
print((means, stds, ci95))
out_filename = FLAGS.logdir +'/'+ exp_string + '/' + 'test_ubs' + str(FLAGS.update_batch_size) + '_stepsize' + str(FLAGS.update_lr) + '.csv'
out_pkl = FLAGS.logdir +'/'+ exp_string + '/' + 'test_ubs' + str(FLAGS.update_batch_size) + '_stepsize' + str(FLAGS.update_lr) + '.pkl'
with open(out_pkl, 'wb') as f:
pickle.dump({'mses': metaval_accuracies}, f)
with open(out_filename, 'w') as f:
writer = csv.writer(f, delimiter=',')
writer.writerow(['update'+str(i) for i in range(len(means))])
writer.writerow(means)
writer.writerow(stds)
writer.writerow(ci95)
def main():
if FLAGS.datasource == 'sinusoid':
if FLAGS.train:
test_num_updates = 5
else:
test_num_updates = 10
else:
if FLAGS.datasource == 'miniimagenet':
if FLAGS.train == True:
test_num_updates = 1 # eval on at least one update during training
else:
test_num_updates = 10
else:
test_num_updates = 10
if FLAGS.train == False:
orig_meta_batch_size = FLAGS.meta_batch_size
# always use meta batch size of 1 when testing.
FLAGS.meta_batch_size = 1
if FLAGS.datasource == 'sinusoid':
data_generator = DataGenerator(FLAGS.update_batch_size*2, FLAGS.meta_batch_size)
else:
if FLAGS.metatrain_iterations == 0 and FLAGS.datasource == 'miniimagenet':
assert FLAGS.meta_batch_size == 1
assert FLAGS.update_batch_size == 1
data_generator = DataGenerator(1, FLAGS.meta_batch_size) # only use one datapoint,
else:
if FLAGS.datasource == 'miniimagenet': # TODO - use 15 val examples for imagenet?
if FLAGS.train:
data_generator = DataGenerator(FLAGS.update_batch_size+15, FLAGS.meta_batch_size) # only use one datapoint for testing to save memory
else:
data_generator = DataGenerator(FLAGS.update_batch_size*2, FLAGS.meta_batch_size) # only use one datapoint for testing to save memory
else:
data_generator = DataGenerator(FLAGS.update_batch_size*2, FLAGS.meta_batch_size) # only use one datapoint for testing to save memory
dim_output = data_generator.dim_output
if FLAGS.baseline == 'oracle':
assert FLAGS.datasource == 'sinusoid'
dim_input = 3
FLAGS.pretrain_iterations += FLAGS.metatrain_iterations
FLAGS.metatrain_iterations = 0
else:
dim_input = data_generator.dim_input
if FLAGS.datasource == 'miniimagenet' or FLAGS.datasource == 'omniglot':
tf_data_load = True
num_classes = data_generator.num_classes
if FLAGS.train: # only construct training model if needed
random.seed(5)
image_tensor, label_tensor = data_generator.make_data_tensor()
inputa = tf.slice(image_tensor, [0,0,0], [-1,num_classes*FLAGS.update_batch_size, -1])
inputb = tf.slice(image_tensor, [0,num_classes*FLAGS.update_batch_size, 0], [-1,-1,-1])
labela = tf.slice(label_tensor, [0,0,0], [-1,num_classes*FLAGS.update_batch_size, -1])
labelb = tf.slice(label_tensor, [0,num_classes*FLAGS.update_batch_size, 0], [-1,-1,-1])
input_tensors = {'inputa': inputa, 'inputb': inputb, 'labela': labela, 'labelb': labelb}
random.seed(6)
image_tensor, label_tensor = data_generator.make_data_tensor(train=False)
inputa = tf.slice(image_tensor, [0,0,0], [-1,num_classes*FLAGS.update_batch_size, -1])
inputb = tf.slice(image_tensor, [0,num_classes*FLAGS.update_batch_size, 0], [-1,-1,-1])
labela = tf.slice(label_tensor, [0,0,0], [-1,num_classes*FLAGS.update_batch_size, -1])
labelb = tf.slice(label_tensor, [0,num_classes*FLAGS.update_batch_size, 0], [-1,-1,-1])
metaval_input_tensors = {'inputa': inputa, 'inputb': inputb, 'labela': labela, 'labelb': labelb}
else:
tf_data_load = False
input_tensors = None
model = MAML(dim_input, dim_output, test_num_updates=test_num_updates)
if FLAGS.train or not tf_data_load:
model.construct_model(input_tensors=input_tensors, prefix='metatrain_')
if tf_data_load:
model.construct_model(input_tensors=metaval_input_tensors, prefix='metaval_')
model.summ_op = tf.summary.merge_all()
saver = loader = tf.train.Saver(tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES), max_to_keep=10)
sess = tf.InteractiveSession()
if FLAGS.train == False:
# change to original meta batch size when loading model.
FLAGS.meta_batch_size = orig_meta_batch_size
if FLAGS.train_update_batch_size == -1:
FLAGS.train_update_batch_size = FLAGS.update_batch_size
if FLAGS.train_update_lr == -1:
FLAGS.train_update_lr = FLAGS.update_lr
exp_string = 'cls_'+str(FLAGS.num_classes)+'.mbs_'+str(FLAGS.meta_batch_size) + '.ubs_' + str(FLAGS.train_update_batch_size) + '.numstep' + str(FLAGS.num_updates) + '.updatelr' + str(FLAGS.train_update_lr)
if FLAGS.num_filters != 64:
exp_string += 'hidden' + str(FLAGS.num_filters)
if FLAGS.max_pool:
exp_string += 'maxpool'
if FLAGS.stop_grad:
exp_string += 'stopgrad'
if FLAGS.baseline:
exp_string += FLAGS.baseline
if FLAGS.norm == 'batch_norm':
exp_string += 'batchnorm'
elif FLAGS.norm == 'layer_norm':
exp_string += 'layernorm'
elif FLAGS.norm == 'None':
exp_string += 'nonorm'
else:
print('Norm setting not recognized.')
resume_itr = 0
model_file = None
tf.global_variables_initializer().run()
tf.train.start_queue_runners()
if FLAGS.resume or not FLAGS.train:
model_file = tf.train.latest_checkpoint(FLAGS.logdir + '/' + exp_string)
if FLAGS.test_iter > 0:
model_file = model_file[:model_file.index('model')] + 'model' + str(FLAGS.test_iter)
if model_file:
ind1 = model_file.index('model')
resume_itr = int(model_file[ind1+5:])
print("Restoring model weights from " + model_file)
saver.restore(sess, model_file)
if FLAGS.train:
train(model, saver, sess, exp_string, data_generator, resume_itr)
else:
test(model, saver, sess, exp_string, data_generator, test_num_updates)
if __name__ == "__main__":
main()