def run_rnn_grucell2(self):
train_input, train_output = self.generate_data()
DLogger.logger().debug("test and training data loaded")
data = tf.placeholder(Const.FLOAT, [None, 20, 3]) # Number of examples, number of input, dimension of each input
target = tf.placeholder(Const.FLOAT, [None, 21])
num_hidden = 24
n_samples = 5
cell = GRUCell2(num_hidden, n_samples)
n_bacthes = len(train_input)
dW1, db1, dW2, db2 = cell.get_weight_dims(3, num_hidden)
W1, b1, W2, b2 = tf.get_variable(name='w1', shape=[n_samples, n_bacthes] + dW1, initializer=tf.constant_initializer(0.1),
dtype=Const.FLOAT), \
tf.get_variable(name='b1', shape=[n_samples, n_bacthes] + db1, initializer=tf.constant_initializer(0.1),
dtype=Const.FLOAT), \
tf.get_variable(name='w2', shape=[n_samples, n_bacthes] + dW2, initializer=tf.constant_initializer(0.1),
dtype=Const.FLOAT), \
tf.get_variable(name='b2', shape=[n_samples, n_bacthes] + db2, initializer=tf.constant_initializer(0.1),
dtype=Const.FLOAT)
cell.set_weights(W1, b1, W2, b2)
step_size = tf.ones([tf.shape(data)[0]], dtype=tf.int32) * tf.shape(data)[1]
state_track, last_state = cell.dynamic_rnn(data, step_size)
state_track = tf.transpose(state_track[0], [1, 0, 2])
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
entropy = sess.run(state_track, {data: train_input, target: train_output})
DLogger.logger().debug('Entropy:' + str(entropy))
def opt_model_train_test(cls, model_path):
output = '../nongit/local/BD/opt/'
with LogFile(output, 'run.log'):
action, reward, state, ids, seq_lengths = BD.get_data()
tf.reset_default_graph()
with tf.device('/device:GPU:0'):
model = HYPMMD(enc_cells=20,
dec_cells=3,
a_size=2,
s_size=0,
latent_size=2,
n_T=action.shape[1],
static_loops=False,
mmd_coef=50)
ensure_dir(output)
actions_train, actions_test, rewards_train, rewards_test, seq_train, seq_test, id_train, id_test = \
stratified_train_test_split(action, reward, state, ids, seq_lengths)
DLogger.logger().debug("test points: " +
str(actions_test.shape[0]))
DLogger.logger().debug("train points: " +
str(actions_train.shape[0]))
def lr_schedule(t):
if t < 2000:
_lr = 0.001
elif t < 5000:
_lr = 0.0001
else:
_lr = 0.00001
return _lr
BD.opt_model_mddae(model,
actions_train,
rewards_train,
None,
seq_train,
actions_test,
rewards_test,
None,
seq_test,
output + '/model/',
model_path,
hessian_term=False,
lr_schedule=lr_schedule)
def predict_z(cls, model_path):
output = '../nongit/local/BD/opt/'
with LogFile(output, 'run.log'):
action, reward, state, ids, seq_lengths = BD.get_data()
DLogger.logger().debug("data points: " + str(action.shape[0]))
tf.reset_default_graph()
ensure_dir(output)
model = HYPMMD(enc_cells=20,
dec_cells=3,
a_size=2,
s_size=0,
latent_size=2,
n_T=action.shape[1],
static_loops=False)
BD.predict(model, action, reward, state, ids, seq_lengths,
'../nongit/local/BD/', model_path)
def get_data(cls):
data = DataReader.read_synth_normal()
# data = data.loc[data.diag == 'Healthy']
ids = data['id'].unique().tolist()
dftr = pd.DataFrame({'id': ids, 'train': 'train'})
tdftr = pd.DataFrame({'id': ids, 'train': 'test'})
train, test = DataProcess.train_test_between_subject(
data, pd.concat((dftr, tdftr)), [1], 2)
DLogger.logger().debug("total points: " + str(get_total_pionts(train)))
train = DataProcess.merge_data(train)
action = train['merged'][0]['action']
reward = train['merged'][0]['reward']
state = train['merged'][0]['state']
ids = train['merged'][0]['id']
seq_lengths = train['merged'][0]['seq_lengths']
return action, reward, state, ids, seq_lengths
def align_model(cls, model_path):
output = '../nongit/local/BD/align/'
with LogFile(output, 'run.log'):
action, reward, state, ids, seq_lengths = BD.get_data()
tf.reset_default_graph()
with tf.device('/device:GPU:0'):
model = HYPMMD(enc_cells=20,
dec_cells=3,
a_size=2,
s_size=0,
latent_size=2,
n_T=action.shape[1],
static_loops=True,
mmd_coef=2)
ensure_dir(output)
actions_train, actions_test, rewards_train, rewards_test, seq_train, seq_test, id_train, id_test = \
stratified_train_test_split(action, reward, state, ids, seq_lengths)
DLogger.logger().debug("test points: " +
str(actions_test.shape[0]))
DLogger.logger().debug("train points: " +
str(actions_train.shape[0]))
BD.opt_model_align(model,
actions_train,
rewards_train,
None,
seq_train,
actions_test,
rewards_test,
None,
seq_test,
output + '/model/',
model_path,
hessian_term=True,
_beta=0.5,
hessian_lr=0.0001,
_h=0.1)
def run_rnn_grucell(self):
train_input, train_output = self.generate_data()
DLogger.logger().debug("test and training data loaded")
data = tf.placeholder(Const.FLOAT,
[None, 20, 3]) # Number of examples, number of input, dimension of each input
target = tf.placeholder(Const.FLOAT, [None, 21])
num_hidden = 24
cell = tf.nn.rnn_cell.GRUCell(num_hidden,
kernel_initializer=tf.constant_initializer(0.1),
bias_initializer=tf.constant_initializer(0.1))
val, _ = tf.nn.dynamic_rnn(cell, data, dtype=Const.FLOAT)
val = tf.transpose(val, [1, 0, 2])
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
entropy = sess.run(val, {data: train_input, target: train_output})
DLogger.logger().debug('Entropy:' + str(entropy))
sess.close()