forked from UW-EXP/GLOBEM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdl_mldg.py
171 lines (137 loc) · 8.73 KB
/
dl_mldg.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
"""
Implementation of Meta-Learning for Domain Generalization (MLDG) algorithm for depression detection
reference:
Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. 2017.
Learning to Generalize: Meta-Learning for Domain Generalization.
arXiv:1710.03463 [cs] (Oct. 2017). http://arxiv.org/abs/1710.03463 arXiv: 1710.03463.
"""
import os, sys
sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
from utils.common_settings import *
from algorithm.dl_erm import DepressionDetectionAlgorithm_DL_erm, DepressionDetectionClassifier_DL_erm
from data_loader.data_loader_ml import DataRepo
from utils import network
import warnings
class DepressionDetectionClassifier_DL_mldg(DepressionDetectionClassifier_DL_erm):
""" MLDG network classifier, extended from ERM classifier """
def __init__(self, config):
super().__init__(config = config)
if (self.model_params["num_domain"] == 1):
warnings.warn("Only have one domain. Degrade to basic ERM")
def fit(self, X, y):
tf.keras.utils.set_random_seed(42)
self.__assert__(X)
training_params_inner = deepcopy(self.training_params)
training_params_inner['optimizer'] = training_params_inner['inner_optimizer']
training_params_inner['learning_rate'] = training_params_inner['inner_learning_rate']
inner_model_optimizer = network.prep_model_optimizer(training_params_inner)
training_params_outer = deepcopy(self.training_params)
training_params_outer['optimizer'] = training_params_outer['outer_optimizer']
training_params_outer['learning_rate'] = training_params_outer['outer_learning_rate']
outer_model_optimizer = network.prep_model_optimizer(training_params_outer)
self.prep_callbacks(X)
num_meta_train = int(self.model_params["num_domain"] * self.model_params["meta_training_proportion"])
batch_size = self.config["data_loader"]["step_batch_size"]
if (self.training_params.get("skip_training", False) == False):
train_epoch_counter = 0
pbar = tqdm(total = self.training_params["epochs"])
train_step_counter = 0
train_loss_record = []
log_history = []
for data, label in X["train"] if self.flag_X_dict else X:
task_list = np.random.permutation(self.model_params["num_domain"])
if (self.model_params["num_domain"] > 1):
meta_train_index_list_idx = task_list[:num_meta_train]
meta_test_index_list_idx = task_list[num_meta_train:]
meta_train_index_list = np.concatenate(
[np.arange(idx * batch_size, (idx + 1) * batch_size) for idx in meta_train_index_list_idx])
meta_test_index_list = np.concatenate(
[np.arange(idx * batch_size, (idx + 1) * batch_size) for idx in meta_test_index_list_idx])
data_metatrain = {}
data_metatest = {}
for k in data:
data_metatrain[k] = tf.gather(data[k], meta_train_index_list, axis = 0)
data_metatest[k] = tf.gather(data[k], meta_test_index_list, axis = 0)
old_vars = self.clf.get_weights()
with tf.GradientTape(persistent=True) as tape1:
pred_metatrain = self.clf(data_metatrain)
loss_metatrain = tf.reduce_mean(
tf.keras.metrics.categorical_crossentropy(y_pred=pred_metatrain, y_true=data_metatrain["input_y"]))
grads = tape1.gradient(target=loss_metatrain, sources=self.clf.trainable_variables)
inner_model_optimizer.apply_gradients(zip(grads, self.clf.trainable_variables))
with tf.GradientTape(persistent=False) as tape2:
pred_metatest = self.clf(data_metatest)
loss_metatest = tf.reduce_mean(
tf.keras.metrics.categorical_crossentropy(y_pred=pred_metatest, y_true=data_metatest["input_y"]))
loss_summary = loss_metatrain + self.model_params["meta_testing_weight"] * loss_metatest
train_loss_record.append([loss_summary, loss_metatrain, loss_metatest])
self.clf.set_weights(old_vars)
grads_summary = tape2.gradient(target=loss_summary, sources=self.clf.trainable_variables)
outer_model_optimizer.apply_gradients(zip(grads_summary, self.clf.trainable_variables))
del tape1
del tape2
else: # degrade to basic ERM
meta_train_index_list_idx = task_list
meta_train_index_list = np.concatenate(
[np.arange(idx * batch_size, (idx + 1) * batch_size) for idx in meta_train_index_list_idx])
data_metatrain = {}
for k in data:
data_metatrain[k] = tf.gather(data[k], meta_train_index_list, axis = 0)
old_vars = self.clf.get_weights()
with tf.GradientTape(persistent=False) as tape1:
pred_metatrain = self.clf(data_metatrain)
loss_metatrain = tf.reduce_mean(
tf.keras.metrics.categorical_crossentropy(y_pred=pred_metatrain, y_true=data_metatrain["input_y"]))
train_loss_record.append([loss_metatrain, loss_metatrain, 0])
grads_summary = tape1.gradient(target=loss_metatrain, sources=self.clf.trainable_variables)
outer_model_optimizer.apply_gradients(zip(grads_summary, self.clf.trainable_variables))
train_step_counter += 1
if (train_step_counter % self.training_params["steps_per_epoch"] == 0):
self.__assert__(X) # ensure the flag_X_dict is correct
train_epoch_counter += 1
pbar.update(1)
train_loss, train_loss_metatrain, train_loss_metatest = tf.reduce_mean(train_loss_record, axis = 0).numpy()
# calc val loss
for data_val, label_val in (X["val"] if self.flag_X_dict else X):
pred_val = self.clf.predict(data_val)
loss_val = tf.reduce_mean(
tf.keras.metrics.categorical_crossentropy(y_pred=pred_val, y_true=data_val["input_y"]))
logs = {"loss": train_loss, "loss_metatrain": train_loss_metatrain, "loss_metatest": train_loss_metatest, "val_loss":loss_val.numpy()}
if (self.training_params["verbose"] > 1):
print(f"Epoch: {train_epoch_counter}", end = " - ")
print(logs)
# manual callback
log_history.append(deepcopy(logs))
# save model
self.model_saver.model_repo_dict[train_epoch_counter] = deepcopy(self.clf.get_weights())
# calc metrics
self.eval_callback.on_epoch_end(epoch=train_epoch_counter, logs = logs)
if (train_epoch_counter >= self.training_params["epochs"]):
pbar.close()
break
self.log_history = log_history
best_epoch, df_results_record = self.find_best_epoch()
self.clf.set_weights(self.model_saver.model_repo_dict[best_epoch])
else:
df_results_record = self.fit_skip_training()
return df_results_record
class DepressionDetectionAlgorithm_DL_mldg(DepressionDetectionAlgorithm_DL_erm):
""" The MLDG algorithm. Extends the ERM algorithm """
def __init__(self, config_dict = None, config_name = "dl_mldg"):
super().__init__(config_dict, config_name)
assert self.config["data_loader"]["generate_by"] in ["across_person", "across_dataset"]
def prep_model(self, data_train: DataRepo, criteria: str = "balanced_acc") -> sklearn.base.ClassifierMixin:
self.config["model_params"].update(
{
"input_shape": self.input_shape,
"flag_return_embedding":False, "flag_embedding_norm":False,
"flag_input_dict":True
}
)
if (self.config["data_loader"]["generate_by"] == "across_person"):
num_domain = len(self.data_generator_whole.person_dict)
else:
num_domain = len(self.data_generator_whole.dataset_dict)
self.config["model_params"]["num_domain"] = num_domain
self.config["data_loader"]["step_batch_size"] = self.data_generator_whole.step_size
return DepressionDetectionClassifier_DL_mldg(config = self.config)