-
Notifications
You must be signed in to change notification settings - Fork 77
/
dl_csd.py
213 lines (171 loc) · 10.1 KB
/
dl_csd.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
"""
Implementation of Common Specific Decomposition (CSD) algorithm for depression detection
reference:
Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. 2020.
Efficient Domain Generalization via Common-Specific Low-Rank Decomposition.
arXiv:2003.12815 [cs, stat] (April 2020). http://arxiv.org/abs/2003.12815 arXiv: 2003.12815.
"""
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
class DepressionDetectionClassifier_DL_csd(DepressionDetectionClassifier_DL_erm):
""" CSD network classifier, extended from ERM classifier """
def __init__(self, config):
super().__init__(config=config)
self.clf = self.build_csd_model()
class csd_layer(layers.Layer):
def __init__(self, embedding_size, common_specific_weights = 0.5, num_domain = 3, low_rank_dim = 20):
super().__init__()
self.num_classes = 2
self.num_domain = num_domain
self.common_specific_weights = common_specific_weights
self.low_rank_dim = low_rank_dim
self.embedding_size = embedding_size
self.common_wt = tf.Variable(tf.ones([1]), name = "common_wt", trainable=False)
specialized_common_wt_init = tf.random_normal_initializer(.5, 1e-2)
self.specialized_common_wt = tf.Variable(name = "specialized_common_wt",
initial_value=specialized_common_wt_init(shape=[1], dtype=tf.float64),
trainable=True)
emb_matrix_init = tf.random_normal_initializer(0, 1e-4)
self.emb_matrix = tf.Variable(name = "emb_matrix",
initial_value=emb_matrix_init(shape=[self.num_domain, self.low_rank_dim], dtype=tf.float64),
trainable=True)
sms_init = tf.random_normal_initializer(0, 0.05)
self.sms = tf.Variable( name = "sms",
initial_value=sms_init(shape=[self.low_rank_dim+1, self.embedding_size, self.num_classes], dtype=tf.float64),
trainable=True)
sm_biases_init = tf.random_normal_initializer(0, 0.05)
self.sm_biases = tf.Variable(name = "sm_biases",
initial_value=sm_biases_init(shape=[self.low_rank_dim+1, self.num_classes], dtype=tf.float64),
trainable=True)
def csd(self, embeds, label_placeholder, domain_placeholder):
"""CSD layer to be used as a replacement for the final classification layer
Modified from: https://gist.github.com/vihari/bad9868049ef62db783e0fc11b22bb5c
Args:
embeds (tensor): final layer representations of dim 2
label_placeholder (tensor): tf tensor with label index of dim 1
domain_placeholder (tensor): tf tensor with domain index of dim 1 -- set to all zeros when testing
Returns:
tuple of final loss, logits
"""
batch_size = tf.shape(embeds)[0]
common_cwt = tf.identity(tf.concat([self.common_wt, tf.zeros([self.low_rank_dim])], axis=0), name='common_cwt')
common_cwt = tf.cast(common_cwt, tf.float64)
# Batch size x self.low_rank_dim + 1
c_wts = tf.nn.embedding_lookup(self.emb_matrix, domain_placeholder)
c_wts = tf.concat([ tf.cast(tf.ones([batch_size, 1]), tf.float64) * self.specialized_common_wt, c_wts], axis=1)
c_wts = tf.reshape(c_wts, [batch_size, self.low_rank_dim+1])
specific_sms = tf.einsum("ij,jkl->ikl", c_wts, self.sms)
common_sm = tf.einsum("j,jkl->kl", common_cwt, self.sms)
specific_bias = tf.einsum("ij,jl->il", c_wts, self.sm_biases)
common_bias = tf.einsum("j,jl->l", common_cwt, self.sm_biases)
diag_tensor = tf.eye(self.low_rank_dim+1, batch_shape=[self.num_classes], dtype=tf.float64)
cps = tf.stack([tf.matmul(self.sms[:, :, _], self.sms[:, :, _], transpose_b=True) for _ in range(self.num_classes)])
orthn_loss = tf.reduce_mean((cps - diag_tensor)**2)
reg_loss = orthn_loss
logits1 = tf.einsum("ik,ikl->il", embeds, specific_sms) + specific_bias
logits2 = tf.matmul(embeds, common_sm) + common_bias
loss1 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits1, labels=label_placeholder))
loss2 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits2, labels=label_placeholder))
loss = (1-self.common_specific_weights)*loss1 + self.common_specific_weights*loss2 + reg_loss
# return common logits
return loss, tf.math.softmax(logits2)
def call(self, embeds, labels, domains):
loss, logits = self.csd(embeds,labels,domains)
self.add_loss(loss)
return logits
def build_csd_model(self):
input_X = Input(shape = self.model_params["input_shape"], name = "input_X")
if (self.model_params["flag_input_dict"]):
if (self.model_params.get("flag_y_vector", True)):
input_y = Input((2,), name="input_y")
else:
input_y = Input((), name="input_y")
input_dataset = Input((), name="input_dataset", dtype = tf.int64)
input_person = Input((), name="input_person", dtype = tf.int64)
if (self.model_params["arch"] == "1dCNN"):
feature_extractor = network.build_1dCNN(**self.model_params)
elif (self.model_params["arch"] == "2dCNN"):
feature_extractor = network.build_2dCNN(**self.model_params)
elif (self.model_params["arch"] == "LSTM"):
feature_extractor = network.build_LSTM(**self.model_params)
elif (self.model_params["arch"] == "Transformer"):
feature_extractor = network.build_Transformer(**self.model_params)
embeddings = feature_extractor({"input_X":input_X, "input_y": input_y, "input_dataset":input_dataset, "input_person":input_person})
csd_layer = self.csd_layer(embedding_size=self.model_params["embedding_size"],
common_specific_weights = self.model_params["common_specific_weights"],
num_domain = self.model_params["num_domain"],
low_rank_dim = self.model_params["low_rank_dim"]
)
if (self.model_params["domain_target"] == "person"):
preds = csd_layer(embeddings, input_y, input_person)
elif (self.model_params["domain_target"] == "dataset"):
preds = csd_layer(embeddings, input_y, input_dataset)
model = Model(inputs=[input_X, input_y, input_dataset, input_person], outputs=[preds])
return model
def fit(self, X, y):
tf.keras.utils.set_random_seed(42)
self.__assert__(X)
model_optimizer = network.prep_model_optimizer(self.training_params)
self.clf.compile(optimizer = model_optimizer, metrics="acc")
callbacks = self.prep_callbacks(X)
if (self.training_params.get("skip_training", False) == False):
history = self.clf.fit(x = X["train"] if self.flag_X_dict else X,
steps_per_epoch = self.training_params["steps_per_epoch"],
epochs = self.training_params["epochs"],
validation_data = X["val"] if self.flag_X_dict else X,
verbose = 1 if self.training_params["verbose"] > 1 else 0,
callbacks = callbacks
)
self.log_history = history.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
def predict(self, X, y=None):
self.__assert__(X)
if (self.flag_X_dict):
X_ = X["val_whole"] # only use the whole val set for eval
else:
X_ = X
for data, label in X_:
# zero out the domain
if (self.model_params["domain_target"] == "person"):
data["input_person"] = tf.zeros_like(data["input_person"])
elif (self.model_params["domain_target"] == "dataset"):
data["input_dataset"] = tf.zeros_like(data["input_dataset"])
return np.argmax(self.clf.predict(data), axis = 1)
def predict_proba(self, X, y=None):
self.__assert__(X)
if (self.flag_X_dict):
X_ = X["val_whole"] # only use the whole val set for eval
else:
X_ = X
for data, label in X_:
# zero out the domain
if (self.model_params["domain_target"] == "person"):
data["input_person"] = tf.zeros_like(data["input_person"])
elif (self.model_params["domain_target"] == "dataset"):
data["input_dataset"] = tf.zeros_like(data["input_dataset"])
return self.clf.predict(data)
class DepressionDetectionAlgorithm_DL_csd(DepressionDetectionAlgorithm_DL_erm):
""" The CSD algorithm. Extends the ERM algorithm """
def __init__(self, config_dict = None, config_name = "dl_csd"):
super().__init__(config_dict, config_name)
assert self.config["model_params"]["domain_target"] in ["person", "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":True, "flag_embedding_norm":False,
"flag_input_dict":True}
)
if (self.config["model_params"]["domain_target"] == "person"):
num_domain = len(self.data_generator_whole.person_dict)
elif (self.config["model_params"]["domain_target"] == "dataset"):
num_domain = len(self.data_generator_whole.dataset_dict)
self.config["model_params"]["num_domain"] = num_domain
return DepressionDetectionClassifier_DL_csd(config=self.config)