-
Notifications
You must be signed in to change notification settings - Fork 8
/
meta_LSM.py
252 lines (215 loc) · 12.9 KB
/
meta_LSM.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
import numpy as np
import tensorflow as tf
import pandas as pd
from modeling import MAML
from scene_segmentation import SLICProcessor, TaskSampling
from tensorflow.python.platform import flags
from utils import tasksbatch_generator, batch_generator, meta_train_test1, save_tasks, \
read_tasks, savepts_fortask, cal_measure
from unsupervised_pretraining.DAS_pretraining_v2 import Unsupervise_pretrain
from sklearn.metrics import accuracy_score
from sklearn.metrics import cohen_kappa_score
from sklearn.neural_network import MLPClassifier
from comparison import SHAP_
import warnings
import os
warnings.filterwarnings("ignore")
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
FLAGS = flags.FLAGS
"""for task sampling"""
flags.DEFINE_float('M', 500, 'determine how distance influence the segmentation')
flags.DEFINE_integer('K', 512, 'number of superpixels')
flags.DEFINE_integer('loop', 5, 'number of SLIC iterations')
flags.DEFINE_string('str_region', 'HK', 'the study area')
flags.DEFINE_string('sample_pts', './src_data/samples_HK.csv', 'path to (non)/landslide samples')
flags.DEFINE_string('Ts_pts', './src_data/Ts_HK.csv', 'path to Ts samples')
"""for meta-train"""
flags.DEFINE_string('basemodel', 'DAS', 'MLP: no unsupervised pretraining; DAS: pretraining with DAS')
flags.DEFINE_string('norm', 'batch_norm', 'batch_norm, layer_norm, or None')
flags.DEFINE_string('log', './tmp/data', 'batch_norm, layer_norm, or None')
flags.DEFINE_string('logdir', './checkpoint_dir', 'directory for summaries and checkpoints.')
flags.DEFINE_integer('dim_input', 14, 'dim of input data')
flags.DEFINE_integer('dim_output', 2, 'dim of output data')
flags.DEFINE_integer('meta_batch_size', 16, 'number of tasks sampled per meta-update, not nums tasks')
flags.DEFINE_integer('num_samples_each_task', 16,
'number of samples sampling from each task when training, inner_batch_size')
flags.DEFINE_integer('test_update_batch_size', 8,
'number of examples used for gradient update during adapting (K=1,3,5 in experiment, K-shot); -1: M.')
flags.DEFINE_integer('metatrain_iterations', 5001, 'number of meta-training iterations.')
flags.DEFINE_integer('num_updates', 5, 'number of inner gradient updates during training.')
flags.DEFINE_integer('pretrain_iterations', 0, 'number of pre-training iterations.')
# flags.DEFINE_integer('num_samples', 18469, 'total number of samples in HK, see samples_HK.')
flags.DEFINE_float('update_lr', 1e-2, 'learning rate of single task objective (inner)') # le-2 is the best
flags.DEFINE_float('meta_lr', 1e-3, 'the base learning rate of meta objective (outer)') # le-2 or le-3
flags.DEFINE_bool('stop_grad', False, 'if True, do not use second derivatives in meta-optimization (for speed)')
flags.DEFINE_bool('resume', True, 'resume training if there is a model available')
def train(model, saver, sess, exp_string, tasks, resume_itr):
SUMMARY_INTERVAL = 100
SAVE_INTERVAL = 1000
PRINT_INTERVAL = 1000
print('Done model initializing, starting training...')
prelosses, postlosses = [], []
if resume_itr != FLAGS.pretrain_iterations + FLAGS.metatrain_iterations - 1:
if FLAGS.log:
train_writer = tf.compat.v1.summary.FileWriter(FLAGS.logdir + '/' + exp_string, sess.graph)
for itr in range(resume_itr, FLAGS.pretrain_iterations + FLAGS.metatrain_iterations):
batch_x, batch_y, cnt_sample = tasksbatch_generator(tasks, FLAGS.meta_batch_size
, FLAGS.num_samples_each_task,
FLAGS.dim_input,
FLAGS.dim_output) # task_batch[i]: (x, y, features)
# batch_y = _transform_labels_to_network_format(batch_y, FLAGS.num_classes)
inputa = batch_x[:, :int(FLAGS.num_samples_each_task / 2), :] # a used for training
labela = batch_y[:, :int(FLAGS.num_samples_each_task / 2), :]
inputb = batch_x[:, int(FLAGS.num_samples_each_task / 2):, :] # b used for testing
labelb = batch_y[:, int(FLAGS.num_samples_each_task / 2):, :]
# # when deal with few-shot problem
# inputa = batch_x[:, :int(len(batch_x[0]) / 2), :] # a used for training
# labela = batch_y[:, :int(len(batch_y[0]) / 2), :]
# inputb = batch_x[:, int(len(batch_x[0]) / 2):, :] # b used for testing
# labelb = batch_y[:, int(len(batch_y[0]) / 2):, :]
feed_dict = {model.inputa: inputa, model.inputb: inputb, model.labela: labela,
model.labelb: labelb, model.cnt_sample: cnt_sample}
if itr < FLAGS.pretrain_iterations:
input_tensors = [model.pretrain_op] # for comparison
else:
input_tensors = [model.metatrain_op] # meta_train
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]])
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) # add sum_op
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)
print('inner lr:', sess.run(model.update_lr))
prelosses, postlosses = [], []
# save model
if (itr != 0) and itr % SAVE_INTERVAL == 0:
saver.save(sess, FLAGS.logdir + '/' + exp_string + '/model' + str(itr))
saver.save(sess, FLAGS.logdir + '/' + exp_string + '/model' + str(itr))
def test(model, saver, sess, exp_string, tasks, num_updates=5):
print('start evaluation...')
print(exp_string)
total_Ytest, total_Ypred, total_Ytest1, total_Ypred1, sum_accuracies, sum_accuracies1 = [], [], [], [], [], []
for i in range(len(tasks)):
np.random.shuffle(tasks[i])
train_ = tasks[i][:int(len(tasks[i]) / 2)]
test_ = tasks[i][int(len(tasks[i]) / 2):] # test_ samples account 25%
"""few-steps tuning (不用op跑是因为采用的batch_size(input shape)不一致,且不想更新model.weight)"""
with tf.compat.v1.variable_scope('model', reuse=True): # np.normalize()里Variable重用
fast_weights = model.weights
for j in range(num_updates):
inputa, labela = batch_generator(train_, FLAGS.dim_input, FLAGS.dim_output,
FLAGS.test_update_batch_size)
loss = model.loss_func(model.forward(inputa, fast_weights, reuse=True),
labela) # fast_weight和grads(stopped)有关系,但不影响这里的梯度计算
grads = tf.gradients(ys=loss, xs=list(fast_weights.values()))
gradients = dict(zip(fast_weights.keys(), grads))
fast_weights = dict(zip(fast_weights.keys(),
[fast_weights[key] - model.update_lr * gradients[key] for key in
fast_weights.keys()]))
"""Single task test accuracy"""
inputb, labelb = batch_generator(test_, FLAGS.dim_input, FLAGS.dim_output, len(test_))
Y_array = sess.run(tf.nn.softmax(model.forward(inputb, fast_weights, reuse=True))) # pred_prob
total_Ypred1.extend(Y_array) # pred_prob_test
total_Ytest1.extend(labelb) # label
Y_test = [] # for single task test
for j in range(len(labelb)):
Y_test.append(labelb[j][0])
total_Ytest.append(labelb[j][0])
Y_pred = [] # for single task test
for j in range(len(labelb)):
if Y_array[j][0] > Y_array[j][1]:
Y_pred.append(1)
total_Ypred.append(1) # total_Ypred: 1d-array label
else:
Y_pred.append(0)
total_Ypred.append(0)
accuracy = accuracy_score(Y_test, Y_pred)
sum_accuracies.append(accuracy)
# print('Test_Accuracy: %f' % accuracy)
# print('SHAP...')
# SHAP_() # TODO: SHAP for proposed
"""Overall evaluation (test data)"""
total_Ypred = np.array(total_Ypred).reshape(len(total_Ypred), )
total_Ytest = np.array(total_Ytest)
total_acc = accuracy_score(total_Ytest, total_Ypred)
print('Test_Accuracy: %f' % total_acc)
cal_measure(total_Ypred, total_Ytest)
kappa_value = cohen_kappa_score(total_Ypred, total_Ytest)
print('Cohen_Kappa: %f' % kappa_value)
# save prediction for test samples, which can be used in calculating statistical measure such as AUROC
pred_prob = np.array(total_Ypred1)
label_bi = np.array(total_Ytest1)
savearr = np.hstack((pred_prob, label_bi))
writer = pd.ExcelWriter('proposed_test.xlsx')
data_df = pd.DataFrame(savearr)
data_df.to_excel(writer)
writer.close()
sess.close()
def main():
"""1.Unsupervised pretraining; 2.segmentation and meta-task sampling; 3.meta-training and -testing"""
"""Unsupervised pretraining"""
# TODO: if it's necessary to mimic batch normalization in Pretraining?
if not os.path.exists('./unsupervised_pretraining/model_init/savedmodel.npz'):
print("start unsupervised pretraining")
tmp = np.loadtxt(FLAGS.sample_pts, dtype=str, delimiter=",", encoding='UTF-8')
tmp_feature = tmp[1:, :].astype(np.float32)
np.random.shuffle(tmp_feature)
Unsupervise_pretrain(tmp_feature)
"""meta task sampling"""
tasks_path = './metatask_sampling/' + FLAGS.str_region + '_tasks_K' + str(FLAGS.K) + '.xlsx'
if not os.path.exists(
'./metatask_sampling/' + FLAGS.str_region + '_SLIC_M{m}_K{k}_loop{loop}.tif'.format(loop=0, m=FLAGS.M,
k=FLAGS.K)):
print('start scene segmentation using SLIC algorithm:')
p = SLICProcessor('./src_data/' + FLAGS.str_region + '/composite.tif', FLAGS.K, FLAGS.M)
p.iterate_times(loop=FLAGS.loop)
print('start meta-task sampling:')
t = TaskSampling(p.clusters)
tasks = t.sampling(p.im_geotrans, FLAGS.sample_pts)
save_tasks(tasks, tasks_path) # save each meta-task samples into respective sheet in a .xlsx file
savepts_fortask(p.clusters, './metatask_sampling/' + FLAGS.str_region + 'pts_tasks_K' + str(FLAGS.K) + '.xlsx')
print('produce meta training and testing datasets...')
HK_tasks = read_tasks(tasks_path)
tasks_train, tasks_test = meta_train_test1(HK_tasks)
"""meta-training and -testing"""
print('model construction...')
model = MAML(FLAGS.dim_input, FLAGS.dim_output, test_num_updates=5)
input_tensors_input = (FLAGS.meta_batch_size, int(FLAGS.num_samples_each_task / 2), FLAGS.dim_input)
input_tensors_label = (FLAGS.meta_batch_size, int(FLAGS.num_samples_each_task / 2), FLAGS.dim_output)
model.construct_model(input_tensors_input=input_tensors_input, input_tensors_label=input_tensors_label,
prefix='metatrain_')
model.summ_op = tf.compat.v1.summary.merge_all()
saver = tf.compat.v1.train.Saver(tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES),
max_to_keep=10)
sess = tf.compat.v1.InteractiveSession()
init = tf.compat.v1.global_variables() # optimizer里会有额外variable需要初始化
sess.run(tf.compat.v1.variables_initializer(var_list=init))
exp_string = '.mbs' + str(FLAGS.meta_batch_size) + '.ubs_' + \
str(FLAGS.num_samples_each_task) + '.numstep' + str(FLAGS.num_updates) + \
'.updatelr' + str(FLAGS.update_lr) + '.meta_lr' + str(FLAGS.meta_lr)
resume_itr = 0
# 续点训练
if FLAGS.resume:
model_file = tf.train.latest_checkpoint(FLAGS.logdir + '/' + exp_string)
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) # 以model_file初始化sess中图
train(model, saver, sess, exp_string, tasks_train, resume_itr)
test(model, saver, sess, exp_string, tasks_test, num_updates=FLAGS.num_updates)
# TODO: use tf.estimator
if __name__ == "__main__":
# device=tf.config.list_physical_devices('GPU')
tf.compat.v1.disable_eager_execution()
main()
print('finished!')