def CRNN1_1D(input_shape, n_classes):
X_input = Input(input_shape)
X = Lambda(lambda q: expand_dims(q, -1), name='expand_dims')(X_input)
X = Conv1D(16, 9, activation=relu, padding='valid')(X)
X = MaxPool1D(8)(X)
X = Conv1D(32, 9, activation=relu, padding='valid')(X)
X = MaxPool1D(8)(X)
X = Conv1D(32, 9, activation=relu, padding='valid')(X)
X = MaxPool1D(6)(X)
X = CuDNNGRU(32, return_sequences=True)(X)
X = Dropout(0.1)(X)
X = CuDNNGRU(32, return_sequences=True)(X)
X = Dropout(0.1)(X)
X = Flatten()(X)
X = Dense(64)(X)
X = Dropout(0.5)(X)
X = Activation(relu)(X)
X = Dense(n_classes, activation=softmax)(X)
model = Model(inputs=X_input, outputs=X)
return model
def model_lstm_atten(embedding_matrix):
inp = Input(shape=(maxlen, ))
x = Embedding(max_features,
embed_size,
weights=[embedding_matrix],
trainable=False)(inp)
x = SpatialDropout1D(0.1)(x)
x = Bidirectional(CuDNNLSTM(40, return_sequences=True))(x)
y = Bidirectional(CuDNNGRU(40, return_sequences=True))(x)
atten_1 = Attention(maxlen)(x) # skip connect
atten_2 = Attention(maxlen)(y)
avg_pool = GlobalAveragePooling1D()(y)
max_pool = GlobalMaxPooling1D()(y)
conc = concatenate([atten_1, atten_2, avg_pool, max_pool])
conc = Dense(16, activation="relu")(conc)
conc = Dropout(0.1)(conc)
outp = Dense(1, activation="sigmoid")(conc)
model = Model(inputs=inp, outputs=outp)
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=[f1])
return model
def build_model():
gpu_check = tf.test.is_gpu_available()
m = Sequential()
#potentially put in dense or embedding layer
if gpu_check:
m.add(
CuDNNGRU(
units=512,
input_shape=(train_data.shape[1], train_data.shape[2]
), # (batch_size, num_time_steps, num_features)
return_sequences=True))
else: # no GPU
m.add(
GRU(
units=512,
input_shape=(train_data.shape[1], train_data.shape[2]
), # (batch_size, num_time_steps, num_features)
return_sequences=True))
m.add(Flatten())
m.add(Dense(units=4))
# configure how model will be trained
# define loss function and optimizer choices
m.compile(loss=MSE, optimizer=Adam(), metrics=["accuracy"])
m.summary() # prints out architecture of network
return m
def ModelCreator(self, input_nodes):
input_ = Input(shape=(input_nodes, ))
embd_seq = Embedding(
len(self.dictionary) + 1,
output_dim=300, # 词向量维度
weights=[self.embedding_matrix],
input_length=config.sequence_max_len, # 文本或者句子截断长度
trainable=False,
mask_zero=False)(input_)
gru = Bidirectional(CuDNNGRU(256))(embd_seq)
# gru = Lambda(lambda x: expand_dims(x, axis=1))(gru)
outputs_list = []
for i in range(20):
# cap = Capsule(10,16,5,True)(gru)
# cap = Flatten()(cap)
out = Dense(64)(gru)
out = Dropout(0.5)(out)
out = Dense(4, activation='softmax')(out)
outputs_list.append(out)
model = Model(inputs=input_, outputs=outputs_list, name='base')
print(model.summary())
return model
def initialise_model():
global model, bottleneck_mdl
from tensorflow.keras.layers import CuDNNGRU, Dropout, Dense, Embedding, CuDNNLSTM, Input, add
#===========================================================#
#=========================[MODEL]===========================#
#===========================================================#
bottleneck_input = Input(shape=(2048, ))
fe1 = Dropout(0.5)(bottleneck_input)
fe2 = Dense(256, activation=tf.nn.selu)(fe1)
#Partial Caption Input
cap_inputs = Input(shape=(72, ))
#se1 is already pretrained on GLOVE model
se1 = Embedding(5411, 200)(cap_inputs)
se2 = Dropout(0.5)(se1)
se3 = CuDNNGRU(256, return_sequences=True)(se2)
se4 = CuDNNLSTM(256)(se3)
decoder1 = add([fe2, se4])
decoder2 = Dense(256, activation=tf.nn.selu)(decoder1)
outputs = Dense(5411, activation=tf.nn.softmax)(decoder2)
model = Model(inputs=[bottleneck_input, cap_inputs], outputs=outputs)
model.load_weights('model_xeon.h5')
#===========================================================#
#=========================[MODEL]===========================#
#===========================================================#
incep_mdl = InceptionV3(weights="imagenet")
bottleneck_mdl = Model(incep_mdl.input, incep_mdl.layers[-2].output)
def build_model(feature_len,
num_classes,
gru_size=128,
classify_activation='softmax',
time_steps=None,
allow_cudnn=True):
"""Build a bidirectional GRU model with CuDNNGRU support.
CuDNNGRU implementation is claimed to give speed-up on GPU of 7x.
The function will build a model capable of running on GPU with
CuDNNGRU provided a) a GPU is present, b) the option has been
allowed by the `allow_cudnn` argument; otherwise a compatible
(but not CuDNNGRU accelerated model) is built.
:param feature_len: int, number of features for each pileup column.
:param num_classes: int, number of output class labels.
:param gru_size: int, size of each GRU layer.
:param classify_activation: str, activation to use in classification layer.
:param time_steps: int, number of pileup columns in a sample.
:param allow_cudnn: bool, opt-in to cudnn when using a GPU.
:returns: `keras.models.Sequential` object.
"""
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, GRU, CuDNNGRU, Bidirectional
# Determine whether to use CuDNNGRU or not
cudnn = False
if tf.test.is_gpu_available(cuda_only=True) and allow_cudnn:
cudnn = True
logger.info("Building model with cudnn optimization: {}".format(cudnn))
model = Sequential()
input_shape = (time_steps, feature_len)
for i in [1, 2]:
name = 'gru{}'.format(i)
# Options here are to be mutually compatible: train with CuDNNGRU
# but allow inference with GRU (on cpu).
# https://gist.github.com/bzamecnik/bd3786a074f8cb891bc2a397343070f1
if cudnn:
gru = CuDNNGRU(gru_size, return_sequences=True, name=name)
else:
gru = GRU(gru_size,
reset_after=True,
recurrent_activation='sigmoid',
return_sequences=True,
name=name)
model.add(Bidirectional(gru, input_shape=input_shape))
# see keras #10417 for why we specify input shape
model.add(
Dense(num_classes,
activation=classify_activation,
name='classify',
input_shape=(time_steps, 2 * gru_size)))
return model
def build_model(chunk_size,
feature_len,
num_classes,
gru_size=128,
classify_activation='softmax'):
"""Builds a bidirectional GRU model. Uses CuDNNGRU for additional
speed-up on GPU (claimed 7x).
:param chunk_size: int, number of pileup columns in a sample.
:param feature_len: int, number of features for each pileup column.
:param num_classes: int, number of output class labels.
:param gru_size: int, size of each GRU layer.
:param classify_activation: str, activation to use in classification layer.
:returns: `keras.models.Sequential` object.
"""
import tensorflow as tf
from tensorflow.keras import backend as K
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, GRU, CuDNNGRU, Bidirectional
# if we can see a gpu, use CuDNNGRU for speed
cudnn = False
if tf.test.is_gpu_available(cuda_only=True):
cudnn = True
logger.info("With cudnn: {}".format(cudnn))
model = Sequential()
input_shape = (chunk_size, feature_len)
for i in [1, 2]:
name = 'gru{}'.format(i)
# Options here are to be mutually compatible: train with CuDNNGRU
# but allow inference with GRU (on cpu).
# https://gist.github.com/bzamecnik/bd3786a074f8cb891bc2a397343070f1
if cudnn:
gru = CuDNNGRU(gru_size, return_sequences=True, name=name)
else:
gru = GRU(gru_size,
reset_after=True,
recurrent_activation='sigmoid',
return_sequences=True,
name=name)
model.add(Bidirectional(gru, input_shape=input_shape))
# see keras #10417 for why we specify input shape
model.add(
Dense(num_classes,
activation=classify_activation,
name='classify',
input_shape=(chunk_size, 2 * gru_size)))
return model
def gru(units):
if tf.test.is_gpu_available():
return CuDNNGRU(units,
return_sequences=False,
return_state=False,
recurrent_initializer='glorot_uniform')
else:
return GRU(units,
return_sequences=False,
return_state=False,
recurrent_activation='sigmoid',
recurrent_initializer='glorot_uniform')
def CRNN_v1(input_shape, n_classes):
X_input = Input(input_shape)
X = Lambda(lambda q: expand_dims(q, -1), name='expand_dims')(X_input)
X = Permute((2, 1, 3))(X)
X = Conv2D(32,
kernel_size=[5, 5],
strides=[2, 2],
activation=relu,
name='conv_1')(X)
X = Conv2D(1,
kernel_size=[1, 1],
strides=[1, 1],
activation=relu,
name='conv_1x1')(X)
X = Lambda(lambda q: squeeze(q, -1), name='squeeze_last_dim')(X)
X = CuDNNGRU(32, return_sequences=True)(X)
X = CuDNNGRU(32, return_sequences=True)(X)
X = Flatten()(X)
X = Dense(64)(X)
X = Dropout(0.5)(X)
X = Activation(relu)(X)
X = Dense(n_classes, activation=softmax)(X)
model = Model(inputs=X_input, outputs=X)
return model
def get_cudnngru(shape, dropout):
model = Sequential()
with tf.variable_scope("CuDNNGRU1", reuse=tf.AUTO_REUSE) as scope:
model.add(CuDNNGRU(64, input_shape=(shape), return_sequences=True))
model.add(BatchNormalization())
model.add(Dropout(dropout))
with tf.variable_scope("CuDNNGRU2", reuse=tf.AUTO_REUSE) as scope:
model.add(CuDNNGRU(64, return_sequences=True))
model.add(BatchNormalization())
model.add(Dropout(dropout))
with tf.variable_scope("CuDNNGRU3", reuse=tf.AUTO_REUSE) as scope:
model.add(CuDNNGRU(64, return_sequences=True))
model.add(BatchNormalization())
model.add(Dropout(dropout))
with tf.variable_scope("CuDNNGRU4", reuse=tf.AUTO_REUSE) as scope:
model.add(CuDNNGRU(64, return_sequences=False))
model.add(BatchNormalization())
model.add(Dropout(dropout))
with tf.variable_scope("DENSE1", reuse=tf.AUTO_REUSE) as scope:
model.add(Dense(128, activation='relu'))
model.add(BatchNormalization())
model.add(Dropout(dropout))
with tf.variable_scope("DENSE2", reuse=tf.AUTO_REUSE) as scope:
model.add(Dense(1, activation='sigmoid'))
opt = tf.keras.optimizers.Adam(lr=1e-2, decay=1e-3)
model.compile(loss='binary_crossentropy',
optimizer=opt,
metrics=['accuracy'])
model.summary()
return model
def get_model(self):
x = Input(shape=(187, 1))
layer = Convolution1D(16, kernel_size=3, activation=activations.relu, padding="valid")(x)
layer = Convolution1D(16, kernel_size=3, activation=activations.relu, padding="valid")(layer)
layer = MaxPool1D(pool_size=2)(layer)
layer = Dropout(rate=self.dropout)(layer)
layer = Convolution1D(32, kernel_size=3, activation=activations.relu, padding="valid")(layer)
layer = Convolution1D(32, kernel_size=3, activation=activations.relu, padding="valid")(layer)
layer = MaxPool1D(pool_size=2)(layer)
layer = Dropout(rate=self.dropout)(layer)
layer = BatchNormalization()(layer)
gru = Bidirectional(CuDNNGRU(self.hidden_size, name='rnn'), merge_mode='concat')(layer)
layer = Dense(self.dense_size, activation=activations.relu, name='dense')(gru)
y = Dense(self.output_classes, name='out_layer', activation=self.last_activation)(layer)
model = models.Model(inputs=x, outputs=y, name=self.name)
return model
def my_gru(name, units):
# If you have a GPU, we recommend using CuDNNGRU(provides a 3x speedup than GRU)
# the code automatically does that.
if tf.test.is_gpu_available():
print('### This is CuDNNGRU ###')
return CuDNNGRU(name=name,
units=units,
return_sequences=True,
return_state=True,
recurrent_initializer='glorot_uniform')
else:
return GRU(name=name,
units=units,
return_sequences=True,
return_state=True,
recurrent_initializer='glorot_uniform',
recurrent_activation='sigmoid')
def __init__(self, vocab_size, embedding_dim, num_RNN_units, batch_size):
super(Model, self).__init__()
self.num_RNN_units = num_RNN_units
self.batch_size = batch_size
# layer 1
self.embedding = Embedding(vocab_size, embedding_dim)
# layer 2
if tf.test.is_gpu_available():
self.gru = CuDNNGRU(self.num_RNN_units,
return_sequences=True,
return_state=True,
recurrent_initializer=glorot_uniform())
else:
self.gru = GRU(self.num_RNN_units,
return_state=True,
recurrent_activation=sigmoid,
recurrent_initializer=glorot_uniform())
# layer 3
self.full_connect = Dense(vocab_size)
import numpy as np
import pickle
from scipy.stats import zscore
import datetime
import pytz
np.random.seed(seed=11)
with open('series_34697_1000.pkl', 'rb') as f:
segments = pickle.load(f)
segments = zscore(segments).astype(np.float32) # standardize
deep_model = Sequential(name="LSTM-autoencoder")
deep_model.add(CuDNNGRU(80, input_shape=(1000, 1), return_sequences=False))
#deep_model.add(CuDNNGRU(100, return_sequences=False))
deep_model.add(Dense(20, activation=None))
deep_model.add(RepeatVector(1000))
#deep_model.add(CuDNNGRU(100, return_sequences=True))
deep_model.add(CuDNNGRU(80, return_sequences=True))
deep_model.add(TimeDistributed(Dense(1)))
deep_model.compile(optimizer=Adam(lr=5e-3, clipnorm=1.0), loss='mse')
#deep_model.load_weights("model_weights/lstm_autoencoder_2020-01-09_18-20-21.h5")
training_time_stamp = datetime.datetime.now(
tz=pytz.timezone('Europe/London')).strftime("%Y-%m-%d_%H-%M-%S")
CB = EarlyStopping(monitor='val_loss',
min_delta=1e-4,
def build_model():
in_id = Input(shape=(MAX_SEQ_LENGTH, ), name="input_ids")
in_mask = Input(shape=(MAX_SEQ_LENGTH, ), name="input_masks")
in_segment = Input(shape=(MAX_SEQ_LENGTH, ), name="segment_ids")
bert_inputs = [in_id, in_mask, in_segment]
bert_cls_output, bert_exp_output = BertLayer(n_fine_tune_layers=10,
name='bert')(bert_inputs)
outputs = []
model_cls, model_exp = None, None
if 'seq' not in dataset and not exp_only:
# Classifier output
dense = Dense(DIM_DENSE_CLS, activation='tanh',
name='cls_dense')(bert_cls_output)
cls_output = Dense(1, activation='sigmoid',
name='cls_output')(dense)
outputs.append(cls_output)
model_cls = Model(inputs=bert_inputs, outputs=cls_output)
optimizer = Adam(LEARNING_RATE)
model_cls.compile(loss=loss['cls_output'],
optimizer=optimizer,
metrics=[metrics['cls_output']])
else:
model_cls = None
if 'cls' not in dataset and exp_structure != 'none' and not cls_only:
# Explainer output
if exp_structure == 'gru':
gru = CuDNNGRU(NUM_GRU_UNITS_BERT_SEQ,
kernel_initializer='random_uniform',
return_sequences=True,
name='exp_gru_gru')(bert_exp_output)
exp = Dense(1, activation='sigmoid', name='exp_gru_dense')(gru)
output_mask = Reshape((MAX_SEQ_LENGTH, 1),
name='exp_gru_reshape')(in_mask)
exp_outputs = Multiply(name='exp_output')([output_mask, exp])
elif exp_structure == 'rnr':
M1 = Bidirectional(
layer=CuDNNLSTM(NUM_INTERVAL_LSTM_WIDTH,
return_sequences=True,
name='exp_rnr_lstm1'),
merge_mode='concat',
name='exp_rnr_bidirectional1')(bert_exp_output)
p_starts = Dense(1,
activation='sigmoid',
name='exp_rnr_starts')(Concatenate(axis=-1)(
[bert_exp_output, M1]))
start_mask = Reshape((MAX_SEQ_LENGTH, 1))(in_mask)
p_starts = Multiply()([p_starts, start_mask])
m1_tilde = Dot(axes=-2)([p_starts, M1])
M1_tilde = Lambda(
lambda x: tf.tile(x, (1, MAX_SEQ_LENGTH, 1)))(m1_tilde)
x = Multiply()([M1, M1_tilde])
M2 = Bidirectional(layer=CuDNNLSTM(NUM_INTERVAL_LSTM_WIDTH,
return_sequences=True,
name='exp_rnr_lstm2'),
merge_mode='concat',
name='exp_rnr_bidirecitonal2')(
Concatenate(axis=-1)(
[bert_exp_output, M1, M1_tilde, x]))
p_end_given_start = Dense(MAX_SEQ_LENGTH,
activation='linear',
name='exp_rnr_end')(Concatenate(
axis=-1)([bert_exp_output, M2]))
end_mask = Lambda(
lambda x: tf.tile(x, (1, MAX_SEQ_LENGTH, 1)))(Reshape(
(1, MAX_SEQ_LENGTH))(in_mask))
p_end_given_start = Multiply()([p_end_given_start, end_mask])
p_end_given_start = Lambda(
lambda x: tf.linalg.band_part(x, 0, -1))(p_end_given_start)
p_end_given_start = Softmax(axis=-1)(p_end_given_start)
exp_outputs = Concatenate(
axis=-1, name='exp_output')([p_starts, p_end_given_start])
outputs.append(exp_outputs)
model_exp = Model(inputs=bert_inputs, outputs=exp_outputs)
optimizer = Adam(LEARNING_RATE)
model_exp.compile(loss=loss['exp_output'],
optimizer=optimizer,
metrics=[metrics['exp_output']])
else:
model_exp = None
model = Model(inputs=bert_inputs, outputs=outputs)
optimizer = Adam(LEARNING_RATE)
model.compile(loss=loss,
loss_weights=loss_weights,
optimizer=optimizer,
metrics=metrics)
return model, model_cls, model_exp
np.random.shuffle(self.indexes)
# Training and Validation generators in a 95/5 split
training_generator = DataGenerator(
segments[:int(np.floor(len(segments) * 0.95))],
errors[:int(np.floor(len(errors) * 0.95))],
batch_size=384)
validation_generator = DataGenerator(
segments[int(np.floor(len(segments) * 0.95)):],
errors[int(np.floor(len(errors) * 0.95)):],
batch_size=384)
y_in = Input(shape=(1024, 1))
y_err = Input(shape=(1024, 1))
h_enc = CuDNNGRU(512, return_sequences=True)(y_in)
h_enc = CuDNNGRU(256, return_sequences=True)(h_enc)
h_enc = CuDNNGRU(128, return_sequences=False)(h_enc)
# h_enc = BatchNormalization()(h_enc)
h_enc = Dense(16, activation=None, name='bottleneck')(h_enc)
# h_enc = BatchNormalization()(h_enc)
h_dec = RepeatVector(1024)(h_enc)
h_dec = CuDNNGRU(64, return_sequences=True)(h_dec)
# h_dec = CuDNNGRU(256, return_sequences=True)(h_dec)
h_dec = TimeDistributed(Dense(1))(h_dec)
model = Model(inputs=[y_in, y_err], outputs=h_dec)
model.compile(optimizer=Adam(clipvalue=0.5), loss=chi2(y_err))
# model.load_weights("model_weights/lstm_autoencoder_2020-01-09_18-20-21.h5")
training_time_stamp = datetime.datetime.now(
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense, Embedding, SpatialDropout1D, Dropout, add, concatenate
from tensorflow.keras.layers import CuDNNGRU, Bidirectional, GlobalMaxPooling1D, GlobalAveragePooling1D, Conv1D
# In[ ]:
sequence_input = Input(shape=(max_seq_size, ), dtype='int32')
embedding_layer = Embedding(total_vocab,
embedding_size,
weights=[embedding_matrix],
input_length=max_seq_size,
trainable=False)
x_layer = embedding_layer(sequence_input)
x_layer = SpatialDropout1D(0.2)(x_layer)
x_layer = Bidirectional(CuDNNGRU(64, return_sequences=True))(x_layer)
x_layer = Conv1D(64,
kernel_size=2,
padding="valid",
kernel_initializer="he_uniform")(x_layer)
avg_pool1 = GlobalAveragePooling1D()(x_layer)
max_pool1 = GlobalMaxPooling1D()(x_layer)
x_layer = concatenate([avg_pool1, max_pool1])
preds = Dense(1, activation=sigmoid)(x_layer)
model = Model(sequence_input, preds)
model.summary()
maxlen = 20
batch_size = 64
dict = eval(open("dictionary.txt").read())
words_dict = list(dict.keys())
filenames = ['train1.tfrecords']
dataset = tf.data.TFRecordDataset(filenames).apply(
tf.contrib.data.map_and_batch(
_parse_function, 256, num_parallel_calls=12)).shuffle(4096).prefetch(1)
model = Sequential()
model.add(Embedding(len(words_dict), 50, input_length=maxlen))
# model.add(LSTM(256, return_sequences=True))
# model.add(Dropout(0.1))
model.add(CuDNNGRU(1000, return_sequences=False))
model.add(Dropout(rate=0.20))
model.add(Dense(100, activation='relu'))
model.add(Dense(len(words_dict), activation='softmax'))
optimizer = tf.train.AdamOptimizer(0.001)
model.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
iter = dataset.make_one_shot_iterator()
for i in range(10000):
print(i)
t = time.time()
print(model.train_on_batch(iter))
print(time.time() - t)
# Training and Validation generators in a 95/5 split
training_generator = DataGenerator(
segments[:int(np.floor(len(segments) * 0.95))],
errors[:int(np.floor(len(errors) * 0.95))],
batch_size=384)
validation_generator = DataGenerator(
segments[int(np.floor(len(segments) * 0.95)):],
errors[int(np.floor(len(errors) * 0.95)):],
batch_size=384)
y_in = Input(shape=(1024, 1))
y_err = Input(shape=(1024, 1))
h_enc = CuDNNGRU(64, return_sequences=False)(y_in)
h_enc = Dense(32, activation=None, name='bottleneck')(h_enc)
h_dec = RepeatVector(1024)(h_enc)
h_dec = CuDNNGRU(64, return_sequences=True)(h_dec)
h_dec = TimeDistributed(Dense(1))(h_dec)
model = Model(inputs=[y_in, y_err], outputs=h_dec)
model.compile(optimizer='adam', loss=chi2(y_err))
#model.load_weights("model_weights/lstm_autoencoder_2020-01-09_18-20-21.h5")
training_time_stamp = datetime.datetime.now(
tz=pytz.timezone('Europe/London')).strftime("%Y-%m-%d_%H-%M-%S")
CB = EarlyStopping(monitor='val_loss',
min_delta=1e-5,
patience=100,
model.add(
Conv1D(filters=filters,
kernel_size=3,
padding='same',
activation='relu'))
model.add(MaxPool1D(pool_size=2))
for l in range(conv - 1):
model.add(
Conv1D(filters=filters,
kernel_size=3,
padding='same',
activation='relu'))
model.add(MaxPool1D(pool_size=2))
model.add(CuDNNGRU(128, return_sequences=True))
model.add(Dropout(0.2))
model.add(CuDNNGRU(64, return_sequences=False))
model.add(Dropout(0.2))
model.add(Flatten())
for n in range(dense):
model.add(Dense(filters, activation="relu"))
model.add(Dense(1, activation="sigmoid"))
model.compile(loss="binary_crossentropy",
optimizer="adam",
metrics=["accuracy"])
loaded_model_json = json_file.read()
json_file.close()
model = model_from_json(loaded_model_json)
print("Loaded model " + model_name + " from disk")
# load latest weights into new model
list_of_files = glob.glob('./data/model/weights-' + model_name + "*.hdf5")
if len(list_of_files) > 0:
latest_file = max(list_of_files, key=os.path.getmtime)
model.load_weights(latest_file)
epoch_init = int(latest_file.split('epoch')[1].split("-")[0])
print("loading weights from file", latest_file, "starting from epoch",
epoch_init)
elif model_name == "basic_CuDNNGRU":
model = Sequential()
model.add(CuDNNGRU(10, input_shape=(4096, 1)))
model.add(Dense(1, activation='linear'))
elif model_name == "basic_CuDNNGRU_reg":
model = Sequential()
model.add(
CuDNNGRU(
10,
input_shape=(4096, 1),
kernel_regularizer=regularizers.l2(0.01),
))
model.add(Dense(1, activation='linear'))
elif model_name == "basic_big_CuDNNGRU_reg":
model = Sequential()
model.add(
def make_bi_gru(self, units): if tf.test.is_gpu_available() and not self.params['no_gpu']: return Bidirectional(CuDNNGRU(units, return_sequences=True), merge_mode='concat') else: return Bidirectional(GRU(units, return_sequences=True), merge_mode='concat')