def _create_model(self,
data_shape,
channel_size=5,
kernel_size=10,
activation='relu'):
if self._is_regression:
loss_func = 'mean_squared_error'
else:
loss_func = 'sparse_categorical_crossentropy'
conv1d_1 = Conv1D(filters=1,
kernel_size=channel_size,
padding='valid',
activation=activation,
input_shape=(data_shape[1], data_shape[2]))
conv1d_2 = Conv1D(filters=1,
kernel_size=5,
padding='same',
activation='tanh')
global_max_1d = GlobalMaxPooling1D()
if self._with_functional_api:
inputs = Input(name='layer_in',
shape=(data_shape[1], data_shape[2]))
x1 = conv1d_1(inputs)
x2 = conv1d_2(x1)
x3 = global_max_1d(x2)
if not self._is_regression:
layer_out = Dense(name='layer_out', units=2)
acti_out = Activation('softmax', name='acti_out')
outputs = acti_out(layer_out(x3))
else:
outputs = x3
model = Model(inputs=inputs,
outputs=outputs,
name='cnn_model_constructor')
else:
model = Sequential([conv1d_1, conv1d_2, global_max_1d],
name='cnn_seq_constructor')
if not self._is_regression:
model.add(Dense(name='layer_out', units=2))
model.add(Activation('softmax', name='acti_out'))
model.compile(optimizer='adam', loss=loss_func, metrics=['acc'])
return model
def _build_model(self,
input_size,
output_size,
hidden_size0=200,
hidden_size1=10):
if self.dueling_dqn:
inputs = Input(shape=(input_size, ))
x = Dense(hidden_size0, activation='relu')(inputs)
x = Dense(hidden_size1, activation='relu')(x)
value = Dense(3, activation='linear')(x)
a = Dense(3, activation='linear')(x)
mean = Lambda(lambda x: K.mean(x, axis=1, keepdims=True))(a)
advantage = Subtract()([a, mean])
q = Add()([value, advantage])
model = Model(inputs=inputs, outputs=q)
model.compile(loss='mse', optimizer=Adam(self.lr))
else:
model = Sequential()
model.add(
Dense(hidden_size0,
input_shape=(input_size, ),
activation='relu'))
model.add(Dense(hidden_size1, activation='relu'))
model.add(Dense(output_size, activation='linear'))
model.compile(loss='mse', optimizer=Adam(lr=self.lr))
return model
class FCN():
def __init__(self, method=1):
self.method = method
self._build()
def _build(self):
if self.method == 1:
self.model = Sequential([
Flatten(input_shape=(32, 32, 3)),
Dense(200, activation='relu'),
Dense(150, activation='relu'),
Dense(10, activation='softmax') # 10: the number of classes
])
elif self.method == 2:
self.model = Sequential()
self.model.add(Flatten(input_shape=(32, 32, 3)))
self.model.add(Dense(200, activation='relu'))
self.model.add(Dense(150, activation='relu'))
self.model.add(Dense(10, activation='softmax'))
elif self.method == 3:
input_layer = Input(shape=(32, 32, 3))
x = input_layer
x = Flatten()(x)
x = Dense(200, activation='relu')(x)
x = Dense(150, activation='relu')(x)
output_layer = Dense(10, activation='softmax')(x)
self.model = Model(input_layer, output_layer)
show_structure(self.model, __file__)
def compile(self, learning_rate):
opt = Adam(lr=learning_rate)
self.model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
def train(self, x_train, y_train, batch_size, epochs):
self.model.fit(x_train,
y_train,
batch_size=batch_size,
epochs=epochs,
shuffle=True)
def evaluate(self, x_test, y_test, batch_size):
self.model.evaluate(x_test, y_test, batch_size=batch_size)
def predict(self, x):
return self.model.predict(x)
def create_model(optimiser, learning_rate, c, transfer_learning, deepen_model,
base_model):
# transfer learning condition
model = copy_model(base_model, optimiser, transfer_learning)
model.pop()
if deepen_model:
if transfer_learning:
model.layers[0].trainable = False
start = Dense(120, activation='relu',
activity_regularizer=l2(c))(model.layers[-1].output)
end = Dense(250, activation='relu',
activity_regularizer=l2(c))(start)
out_r = Dense(1,
activation="sigmoid",
activity_regularizer=l2(c),
name='r_prop')(end)
final_model = Model(model.input, out_r)
else:
final_model = Sequential()
final_model.add(
Dense(120,
input_dim=51,
activation="relu",
activity_regularizer=l2(c)))
final_model.add(
Dense(250, activation='relu', activity_regularizer=l2(c)))
final_model.add(
Dense(1,
activation="sigmoid",
activity_regularizer=l2(c),
name='r_prop'))
else:
out_r = Dense(1,
activation="sigmoid",
activity_regularizer=l2(c),
name='r_prop')(model.layers[-1].output)
final_model = Model(model.input, out_r)
# Compile model
if optimiser == "nag":
opt = SGD(learning_rate=learning_rate, nesterov=True)
elif optimiser == "adam":
opt = Adam(learning_rate=learning_rate)
final_model.compile(loss=cross_entropy, optimizer=opt, metrics=["acc"])
return final_model
Out = TimeDistributed(Lambda(part))(mer)
layer_1 = -main_input
Out = TimeDistributed(Lambda(part2))([main_input, layer_1])
model = Model(inputs=[main_input, aux_input], outputs=Out)
model.compile(loss='mean_squared_error', optimizer='RMSprop')
model.predict([DATA, DATA2])
#%%
from tensorflow.keras.utils import plot_model
plot_model(model)
model = Sequential()
x = Input
model.add(InputLayer(input_shape=(3, 2)))
model.add(TimeDistributed(Lambda(lambda x: x[:, 0])))
model.compile(loss='mean_squared_error', optimizer='RMSprop')
out = model.predict(DATA)
model.predict()
# model.add(BatchNormalization())
for i in range(n_layer):
if i == n_layer - 1:
model.add(
LSTM(n_hidden,
dropout=dropout,
kernel_initializer=init,
return_sequences=False))
else:
model.add(
reture_value = LSTM(units=16,return_sequences=False)(counsellor_msg)
#reture_value = Dense(units=8)(counsellor_msg)
drop = Dropout(0.5)(reture_value)
out = Dense(1, activation='sigmoid')(drop)
optimizer = Adam(lr=0.001)
model = Model(sequence_input, outputs=out)
'''
model = Sequential()
model.add(
Embedding(num_words,
embedding_dim,
weights=[embedding_matrix],
input_length=max_tokens,
trainable=False))
model.add(Bidirectional(LSTM(units=32, return_sequences=True)))
model.add(LSTM(units=16, return_sequences=False))
model.add(Dropout(0.1))
model.add(Dense(1, activation='sigmoid'))
optimizer = Adam(lr=0.001)
print(model.summary())
model.compile(loss=[focal_loss(alpha=.75, gamma=2.)],
optimizer=optimizer,
metrics=['accuracy'])
def michael_model_2(input_shape=(600, 800, 3)):
base_model = ResNet50(weights='imagenet', include_top=False)
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(200, activation="softmax")(x)
model = Model(inputs=base_model.input, output=predictions)
for layer in base_model.layers:
layer.trainable = False
model.add(Flatten())
model.add(BatchNormalization())
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.5))
model.add(BatchNormalization())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(BatchNormalization())
model.add(Dense(64, activation='relu'))
model.add(Dropout(0.5))
model.add(BatchNormalization())
model.add(Dense(2, activation='relu'))
return model
#2. 모델링
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.layers import Conv1D, MaxPooling1D, Dense, Flatten, Dropout, Input
input1 = Input(shape=(x_train[1], 1))
conv1 = Conv1D(100, 2, padding='SAME')(input1)
maxp = MaxPooling1D(2)(conv1)
conv1 = Conv1D(100, 2, padding='SAME')(maxp)
drop = Dropout(0.2)
dense1 = Dense(50, activation='relu')(drop)
dense1 = Dense(50, activation='relu')(drop)
dense1 = Dense(42, activation='relu')(drop)
output1 = Dense(3, activation='relu')(dense1)
model = Model(Input=input1, outputs=output1)
# model.add(Dropout(0.2)),
model.add(Dense(1, activation='relu'))
#3. 컴파일, 훈련
model.compile(loss='mse', optimizer='adam', metrics=['mae'])
'''
EarlyStopping
'''
from tensorflow.keras.callbacks import EarlyStopping
early_stopping = EarlyStopping(monitor='loss', patience=20, mode='auto')
model.fit(x_train,
y_train,
epochs=1000,
batch_size=7,
validation_data=(x_val, y_val),
callbacks=[early_stopping])
def build_model(self):
if self.pretrained_embeddings:
self.vocab_size, self.emb_size = self.pretrained_embeddings.shape
emb_layer = Embedding(input_dim=self.vocab_size, output_dim=self.emb_size,
weights=[self.pretrained_embeddings],
input_length=self.max_seq_len, trainable=self.finetune_embeddings)
else:
emb_layer = Embedding(input_dim=self.vocab_size, output_dim=self.emb_size, input_length=self.max_seq_len)
if self.labels_num == 2:
output_layer = Dense(1, activation="sigmoid")
else:
output_layer = Dense(self.labels_num, activation="softmax")
# Functional API
if self.model_api == 'functional':
INPUT = Input(shape=(self.max_seq_len,))
EMB = emb_layer(INPUT)
x = Bidirectional(LSTM_LAYER(128, return_sequences=True))(EMB)
x = Bidirectional(LSTM_LAYER(64, return_sequences=True))(x)
x = AttentionWithContext()(x)
x = Dense(64, activation="relu")(x)
OUTPUT = output_layer(x)
model = Model(inputs=INPUT, outputs=OUTPUT)
# Sequential API
elif self.model_api == 'sequential':
model = Sequential()
model.add(emb_layer)
model.add(Bidirectional(LSTM_LAYER(128, return_sequences=True)))
model.add(Bidirectional(LSTM_LAYER(64, return_sequences=True)))
model.add(AttentionWithContext)
model.add(Dense(64, activation="relu"))
model.add(output_layer)
else:
logger.error(f"Model-api type `{self.model_api}` is not supported. Please choose either `functional` or `sequential`")
raise IOError(f"Model-api type `{self.model_api}` is not supported. Please choose either `functional` or `sequential`")
if self.labels_num == 2:
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=[f1_m, 'accuracy'])
else:
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=[f1_m, 'accuracy'])
return model
test_generator = data_generator('./dataset/test/')
if __name__ == '__main__':
os.mkdir('./models/')
print("Extracting features with ResNet50")
base_model = ResNet50(include_top=False, input_shape=(128, 128, 3))
output = base_model.layers[38].output
model = Model(inputs=base_model.input, outputs=output)
pretrain(model, 'train')
pretrain(model, 'test')
print()
print("Building CNN")
model = Sequential()
model.add(Conv2D(256, (1, 1), input_shape=(32, 32, 256),
activation='relu'))
model.add(Conv2D(256, (3, 3), activation='relu'))
model.add(MaxPooling2D())
model.add(Conv2D(512, (2, 2), activation='relu'))
model.add(MaxPooling2D())
model.add(Flatten())
model.add(Dropout(0.2))
model.add(Dense(196))
model.compile('adam', loss='mse', metrics=['accuracy'])
model.summary()
plot_model(model, to_file='./models/model.png', show_shapes=True)
print()
history = model.fit_generator(train_generator,
steps_per_epoch=STEPS_PER_EPOCH,
validation_data=test_generator,
class Models:
embedding_dim = 500
input_length = 100
lstm_units = 75
lstm_dropout = 0.4
recurrent_dropout = 0.4
spatial_dropout = 0.3
filters = 32
kernel_size = 3
max_sequence_length = 75
def build_Base_model(self, embedding_matrix):
self.sequence_input = Input(shape=(self.max_sequence_length, ))
embedding = Embedding(
input_dim=embedding_matrix.shape[0],
output_dim=embedding_matrix.shape[1],
weights=[embedding_matrix],
input_length=self.max_sequence_length,
trainable=False,
)(self.sequence_input)
base = SpatialDropout1D(self.spatial_dropout)(embedding)
return base
def build_GRU_model(self, base):
base = GRU(self.lstm_units,
dropout=self.lstm_dropout,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True)(base)
base = GRU(self.lstm_units, return_sequences=True)(base)
base = GRU(self.lstm_units, return_sequences=True)(base)
base = GRU(self.lstm_units, return_sequences=True)(base)
base = GRU(self.lstm_units, return_sequences=True)(base)
base = GRU(self.lstm_units,
dropout=self.lstm_dropout,
recurrent_dropout=self.recurrent_dropout)(base)
# base = BatchNormalization(name="batchy3")(base)
return base
def build_LSTM_model(self, embedding_matrix):
self.model = Sequential()
self.model.add(
Embedding(input_dim=embedding_matrix.shape[0],
output_dim=embedding_matrix.shape[1],
weights=[embedding_matrix],
input_length=75,
trainable=False))
self.model.add(SpatialDropout1D(0.5))
self.model.add(
Conv1D(self.filters,
kernel_size=self.kernel_size,
kernel_regularizer=regularizers.l2(0.00001),
padding='same'))
self.model.add(LeakyReLU(alpha=0.2))
self.model.add(MaxPooling1D(pool_size=2))
self.model.add(
Bidirectional(
LSTM(self.lstm_units,
dropout=0.5,
recurrent_dropout=0.5,
return_sequences=True)))
self.model.add(SpatialDropout1D(0.5))
self.model.add(
Conv1D(self.filters,
kernel_size=self.kernel_size,
kernel_regularizer=regularizers.l2(0.00001),
padding='same'))
self.model.add(LeakyReLU(alpha=0.2))
self.model.add(MaxPooling1D(pool_size=2))
self.model.add(
Bidirectional(
LSTM(self.lstm_units,
dropout=0.5,
recurrent_dropout=0.5,
return_sequences=True)))
self.model.add(SpatialDropout1D(0.5))
self.model.add(
Conv1D(self.filters,
kernel_size=self.kernel_size,
kernel_regularizer=regularizers.l2(0.00001),
padding='same'))
self.model.add(LeakyReLU(alpha=0.2))
self.model.add(MaxPooling1D(pool_size=2))
self.model.add(
Bidirectional(
LSTM(self.lstm_units, dropout=0.5, recurrent_dropout=0.5)))
self.model.add(Dense(5, activation='softmax'))
self.model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
def build_BLSTM_model(self, embedding_matrix):
self.model = Sequential()
self.model.add(
Embedding(input_dim=embedding_matrix.shape[0],
output_dim=embedding_matrix.shape[1],
weights=[embedding_matrix],
input_length=self.max_sequence_length,
trainable=False))
self.model.add(SpatialDropout1D(0.5))
self.model.add(
Bidirectional(LSTM(self.lstm_units, return_sequences=True)))
self.model.add(
Bidirectional(LSTM(self.lstm_units, return_sequences=True)))
self.model.add(
Bidirectional(
LSTM(self.lstm_units, dropout=0.5, recurrent_dropout=0.5)))
self.model.add(Dense(5, activation='softmax'))
self.model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
def build_BLTSM_model(self, base):
base = Bidirectional(
LSTM(self.lstm_units,
return_sequences=True,
dropout=self.lstm_dropout,
recurrent_dropout=self.recurrent_dropout))(base)
base = Bidirectional(
LSTM(self.lstm_units,
dropout=self.lstm_dropout,
return_sequences=True,
recurrent_dropout=self.recurrent_dropout))(base)
base = Bidirectional(
LSTM(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True))(base)
base = Bidirectional(
LSTM(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True))(base)
base = Bidirectional(
LSTM(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True))(base)
base = Bidirectional(
LSTM(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True))(base)
base = Bidirectional(
LSTM(self.lstm_units,
dropout=self.lstm_dropout,
recurrent_dropout=self.recurrent_dropout))(base)
# max = GlobalMaxPooling1D()(base)
# concat = concatenate([avg,max])
# base = BatchNormalization(name="batch_blstm")(base)
return base
def build_BGRU_model(self, base):
base = Bidirectional(
GRU(self.lstm_units,
return_sequences=True,
dropout=self.lstm_dropout,
recurrent_dropout=self.recurrent_dropout))(base)
base = Bidirectional(
GRU(self.lstm_units,
dropout=self.lstm_dropout,
return_sequences=True,
recurrent_dropout=self.recurrent_dropout))(base)
base = Bidirectional(
GRU(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True))(base)
base = Bidirectional(
GRU(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True))(base)
base = Bidirectional(
GRU(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True))(base)
base = Bidirectional(
GRU(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True))(base)
base = Bidirectional(
GRU(self.lstm_units,
dropout=self.lstm_dropout,
recurrent_dropout=self.recurrent_dropout))(base)
# max = GlobalMaxPooling1D()(base)
# concat = concatenate([avg,max])
# base = BatchNormalization(name="batch_blstm")(base)
return base
def build_LTSM_model(self, base):
base = LSTM(self.lstm_units,
return_sequences=True,
dropout=self.lstm_dropout,
recurrent_dropout=self.recurrent_dropout)(base)
base = LSTM(self.lstm_units,
dropout=self.lstm_dropout,
return_sequences=True,
recurrent_dropout=self.recurrent_dropout)(base)
base = LSTM(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True)(base)
base = LSTM(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True)(base)
base = LSTM(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True)(base)
base = LSTM(self.lstm_units,
recurrent_dropout=self.recurrent_dropout,
return_sequences=True)(base)
base = LSTM(self.lstm_units,
dropout=self.lstm_dropout,
recurrent_dropout=self.recurrent_dropout)(base)
# max = GlobalMaxPooling1D()(base)
# concat = concatenate([avg,max])
# base = BatchNormalization(name="batch_blstm")(base)
return base
def build_CNN_model(self, base):
base = Conv1D(self.filters,
kernel_size=self.kernel_size,
kernel_regularizer=regularizers.l2(0.0005),
padding='same')(base)
base = LeakyReLU(alpha=0.2)(base)
base = Conv1D(self.filters,
kernel_size=self.kernel_size,
kernel_regularizer=regularizers.l2(0.0005),
padding='same')(base)
base = LeakyReLU(alpha=0.2)(base)
base = Conv1D(self.filters,
kernel_size=self.kernel_size,
kernel_regularizer=regularizers.l2(0.0005),
padding='same')(base)
base = LeakyReLU(alpha=0.2)(base)
base = Conv1D(self.filters,
kernel_size=self.kernel_size,
kernel_regularizer=regularizers.l2(0.0005),
padding='same')(base)
base = LeakyReLU(alpha=0.2)(base)
base = Conv1D(self.filters,
kernel_size=self.kernel_size,
kernel_regularizer=regularizers.l2(0.0005),
padding='same')(base)
base = LeakyReLU(alpha=0.2)(base)
avg = GlobalAveragePooling1D()(base)
max = GlobalMaxPooling1D()(base)
concat = concatenate([avg, max])
base = BatchNormalization()(concat)
return base
def f1(self, y_true, y_pred):
def recall(y_true, y_pred):
"""Recall metric.
Only computes a batch-wise average of recall.
Computes the recall, a metric for multi-label classification of
how many relevant items are selected.
"""
true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))
possible_positives = K.sum(K.round(K.clip(y_true, 0, 1)))
recall = true_positives / (possible_positives + K.epsilon())
return recall
def precision(y_true, y_pred):
"""Precision metric.
Only computes a batch-wise average of precision.
Computes the precision, a metric for multi-label classification of
how many selected items are relevant.
"""
true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))
predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1)))
precision = true_positives / (predicted_positives + K.epsilon())
return precision
precision = precision(y_true, y_pred)
recall = recall(y_true, y_pred)
return 2 * ((precision * recall) / (precision + recall + K.epsilon()))
def build_myModel(self, text_embedding, model):
for layer in model.layers[2:-2]:
layer.trainable = False
# print("layer ",model.layers[2].name)
concat_cnn = model.layers[-2].output
base = model.layers[2].output
# base = self.build_Base_model(text_embedding)
# model.layers[3] = model.layers[3](base)
print("***** concat_cnn ", concat_cnn.shape)
concat_blstm = self.build_BLTSM_model(base)
# concat_lstm = self.build_LTSM_model(base)
print("***** concat_blstm ", concat_blstm.shape)
# concat_cnn = self.build_CNN_model(base)
concat_gru = self.build_GRU_model(base)
print("***** concat_gru ", concat_gru.shape)
concat_out = concatenate([concat_cnn, concat_blstm], name="concat1")
concat_out = concatenate([concat_out, concat_gru], name="concat2")
# concat_out = concatenate([concat_out, concat_lstm],name = "concat3")
# avg = GlobalAveragePooling1D()(base)
# out = BatchNormalization(name = "batchyend")(concat_out)
# out = Dense(300, activation='softmax',kernel_regularizer=regularizers.l2(0.00005))(concat_out)
pred = Dense(5, activation='softmax')(concat_out)
self.model = Model(model.input, pred)
weights = np.ones((5, ))
op = SGD(lr=0.0001)
self.model.compile(
optimizer='adam',
loss=self.weighted_categorical_crossentropy(weights),
metrics=['acc'])
def compile(self, model):
op = SGD(lr=4e-4, momentum=0.9)
model.compile(optimizer=op,
loss='categorical_crossentropy',
metrics=['acc'])
return model
def buil_pre_model(self, text_embedding):
base = self.build_Base_model(text_embedding)
# concat_cnn = self.build_CNN_model(base)
# concat_blstm = self.build_BLTSM_model(base)
# concat_gru = self.build_GRU_model(base)
# concat_lstm = self.build_LTSM_model(base)
concat_bgru = self.build_BGRU_model(base)
# out = BatchNormalization()(concat_cnn)
# out = Dense(75, activation='softmax',kernel_regularizer=regularizers.l2(0.00005))(out)
pred = Dense(5, activation='softmax')(concat_bgru)
self.model = Model(self.sequence_input, pred)
op = SGD(lr=0.0001)
self.model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['acc'])
def weighted_categorical_crossentropy(self, weights):
# A weighted version of keras.objectives.categorical_crossentropy
# Variables:
# weights: numpy array of shape (C,) where C is the number of classes
# Usage:
# weights = np.array([0.5,2,10]) # Class one at 0.5, class 2 twice the normal weights, class 3 10x.
# loss = weighted_categorical_crossentropy(weights)
# model.compile(loss=loss,optimizer='adam')
weights = K.variable(weights)
def loss(y_true, y_pred):
# scale predictions so that the class probas of each sample sum to 1
y_pred /= K.sum(y_pred, axis=-1, keepdims=True)
# clip to prevent NaN's and Inf's
y_pred = K.clip(y_pred, K.epsilon(), 1 - K.epsilon())
# calc
loss = y_true * K.log(y_pred) * weights
loss = -K.sum(loss, -1)
return loss
return loss
dense (Dense) (None, 5) 10 _________________________________________________________________ dense_1 (Dense) (None, 3) 18 _________________________________________________________________ dense_2 (Dense) (None, 4) 16 _________________________________________________________________ dense_3 (Dense) (None, 1) 5 ================================================================= Total params: 49 Trainable params: 49 Non-trainable params: 0 _________________________________________________________________ ''' #2.2 시퀀셜 모델 model= Sequential() model.add(Dense(10,input_dim=1)) #인풋레이어(칼럼)=1, 은닉층 첫번째 5 model.add(Dense(5)) model.add(Dense(3)) model.add(Dense(4)) model.add(Dense(1)) model.summary() ''' ''' Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= dense (Dense) (None, 10) 20 _________________________________________________________________ dense_2 (Dense) (None, 3) 18 _________________________________________________________________
## add more if you want x = Flatten()(mobilnet.output) prediction = Dense(len(folders), activation='softmax')(x) # create a model object model = Model(inputs=mobilnet.input, outputs=prediction) model.summary() from tensorflow.keras.layers import MaxPooling2D ### Create Model from scratch using CNN model=Sequential() model.add(Conv2D(filters=16,kernel_size=2,padding="same",activation="relu",input_shape=(224,224,3))) model.add(MaxPooling2D(pool_size=2)) model.add(Conv2D(filters=32,kernel_size=2,padding="same",activation ="relu")) model.add(MaxPooling2D(pool_size=2)) model.add(Conv2D(filters=64,kernel_size=2,padding="same",activation="relu")) model.add(MaxPooling2D(pool_size=2)) model.add(Flatten()) model.add(Dense(500,activation="relu")) model.add(Dense(2,activation="softmax")) model.summary() model.compile( loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'] )
headModel = Flatten(name="flatten")(headModel)
headModel = Dense(256, activation="relu")(headModel)
headModel = Dropout(0.5)(headModel)
headModel = Dense(5, activation="softmax")(headModel)
# place the head FC model on top of the base model (this will become
# the actual model we will train)
model = Model(inputs=baseModel.input, outputs=headModel)
# loop over all layers in the base model and freeze them so they will
# *not* be updated during the training process
for layer in baseModel.layers:
layer.trainable = False
# Construction of model
model = Sequential()
model.add(Conv2D(32, (5, 5), activation='relu', input_shape=(128, 128, 3)))
model.add(MaxPooling2D((2, 2)))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2)))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2)))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dense(5, activation='softmax'))
# Configures the model for training
model.compile(optimizer='adam', # Optimization routine, which tells the computer how to adjust the parameter values to minimize the loss function.
loss='sparse_categorical_crossentropy', # Loss function, which tells us how bad our predictions are.
metrics=['accuracy']) # List of metrics to be evaluated by the model during training and testing.
# Trains the model for a given number of epochs (iterations on a dataset) and validates it.
def cnn_model(X_train, X_test , y_train, y_test):
base_model = applications.VGG16(include_top=False, input_shape=X_train.shape[1:], weights='imagenet',classes=CLASSES)
# Freezing VGG16 layers
for layer in base_model.layers:
layer.trainable=False
last_layer = 'block5_pool'
model = Model(base_model.input, base_model.get_layer(last_layer).output)
model.layers[-1].output.shape
model = Sequential()
model.add(base_model) # Stack vgg16
model.add(Conv2D(128,(3,3),activation="relu", input_shape=model.layers[-1].output.shape, data_format='channels_first'))
model.add(MaxPooling2D(2,2))
# model.add(Conv2D(128,(3,3),activation="relu"))
# model.add(MaxPooling2D(2,2))
model.add(Flatten()) # Flatten the output
model.add(Dense(64, activation='relu'))
model.add(Dropout(0.2))
# Output layer
model.add(Dense(CLASSES, activation="sigmoid"))
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
model.fit(X_train, y_train, batch_size=bs, epochs=ep, validation_data = (X_test, y_test), callbacks=[tensorboard])
# Save model
model.save('cnn.model')
model.summary()
def get_model_architecture(dataset, is_dropout=False):
assert dataset.name in DATASETS, "dataset must be one of: mnist, mnist_fashion, cifar10, cifar100"
num_classes = dataset.num_classes
img_shape = dataset.img_shape
img_input = Input(shape=img_shape)
if dataset.name == 'mnist' or dataset.name == 'mnist_fashion':
# architecture from: https://keras.io/examples/mnist_cnn/
x = Conv2D(32, kernel_size=(3, 3))(img_input)
x = Activation('relu')(x)
x = Conv2D(64, (3, 3))(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(2, 2), name='pool1')(x)
if is_dropout:
x = Dropout(0.25)(x)
x = Flatten()(x)
x = Dense(128)(x)
x = Activation('relu')(x)
if is_dropout:
x = Dropout(0.5)(x)
x = Dense(num_classes, name='features')(x)
x = Activation('softmax')(x)
# Create model
model = Model(img_input, x)
elif dataset.name == 'cifar100' or dataset.name == 'cifar10':
# taken from: https://github.com/geifmany/cifar-vgg/tree/e7d4bd4807d15631177a2fafabb5497d0e4be3ba
model = Sequential()
weight_decay = 0.0005
model.add(
Conv2D(64, (3, 3),
padding='same',
input_shape=img_shape,
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.3))
model.add(
Conv2D(64, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(
Conv2D(128, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.4))
model.add(
Conv2D(128, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(
Conv2D(256, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.4))
model.add(
Conv2D(256, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.4))
model.add(
Conv2D(256, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(
Conv2D(512, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.4))
model.add(
Conv2D(512, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.4))
model.add(
Conv2D(512, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(
Conv2D(512, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.4))
model.add(
Conv2D(512, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.4))
model.add(
Conv2D(512, (3, 3),
padding='same',
kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2, 2)))
if is_dropout:
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(512, kernel_regularizer=regularizers.l2(weight_decay)))
model.add(Activation('relu'))
model.add(BatchNormalization())
if is_dropout:
model.add(Dropout(0.5))
model.add(Dense(num_classes, name='features'))
model.add(Activation('softmax'))
'''
x = Conv2D(32, (3, 3), padding='same')(img_input)
x = Activation('relu')(x)
x = Conv2D(32, (3, 3))(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
if is_dropout:
x = Dropout(0.25)(x)
x = Conv2D(64, (3, 3), padding='same')(x)
x = Activation('relu')(x)
x = Conv2D(64, (3, 3))(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
if is_dropout:
x = Dropout(0.25)(x)
x = Flatten()(x)
x = Dense(512)(x)
x = Activation('relu')(x)
if is_dropout:
x = Dropout(0.5)(x)
x = Dense(num_classes, name='features')(x)
x = Activation('softmax')(x)
# Create model
model = Model(img_input, x)
'''
#model.summary()
return model
def GenModel(data_objects, model_params):
# Clear the TF graph
K.clear_session()
# Load the data to check dimensionality
x_train, y_train = data_objects["x_train"], data_objects["y_train"]
print("Input data shape:", x_train.shape)
# Model-controlled parameters:
F_modeltype = data_objects["F_modeltype"]
#######################################################################
# Number of block layers
BLOCK_LAYERS = model_params["BLOCK_LAYERS"]
# Alternative block type setup:
BLOCK1_TYPE = model_params["BLOCK1_TYPE"]
BLOCK2_TYPE = model_params["BLOCK2_TYPE"]
BLOCK3_TYPE = model_params["BLOCK3_TYPE"]
BLOCK4_TYPE = model_params["BLOCK4_TYPE"]
FC_BLOCK1 = model_params["FC_BLOCK1"]
FC_BLOCK2 = model_params["FC_BLOCK2"]
FC_BLOCK3 = model_params["FC_BLOCK3"]
FC_BLOCK4 = model_params["FC_BLOCK4"]
#CNN related params
DROPOUT_RATE = model_params["DROPOUT_RATE"]
FULLY_CONNECTED = model_params["FULLY_CONNECTED"]
NUM_FILTERS = model_params["NUM_FILTERS"]
KERNEL_SIZE = model_params["KERNEL_SIZE"]
KERNEL_STRIDE = model_params["KERNEL_STRIDE"]
POOL_STRIDE = model_params["POOL_STRIDE"]
POOL_SIZE = model_params["POOL_SIZE"]
PADDING = 0 #used for analysis only
input_dim = (x_train.shape[1], x_train.shape[2])
# Print the configuration to be trained:
print("Hyper params:")
print("================================" * 3)
print("================================" * 3)
print("\nModel: ", F_modeltype)
print("\nTraining:")
print("- batch_size: ", (model_params["batch_size"]))
print("- optimizer: ", (model_params["optimizer"]))
print("- learningrate: ", (model_params["learningrate"]))
print("\nArchitecture::")
print("Blocks: ", model_params["BLOCK_LAYERS"])
print("")
print("Block types: ", model_params["BLOCK1_TYPE"],
model_params["BLOCK2_TYPE"], model_params["BLOCK3_TYPE"],
model_params["BLOCK4_TYPE"])
print("Hidden units: ", model_params["FC_BLOCK1"],
model_params["FC_BLOCK2"], model_params["FC_BLOCK3"],
model_params["FC_BLOCK4"])
print("Dropout: ", (model_params["DROPOUT_RATE"]))
print("\n")
print("NUM_FILTERS: ", NUM_FILTERS)
print("KERNEL_SIZE: ", KERNEL_SIZE)
print("KERNEL_STRIDE: ", KERNEL_STRIDE)
print("POOL_SIZE: ", POOL_SIZE)
print("")
print("================================" * 3)
print("================================" * 3)
#######################################################################
#################### Architecture #####################
#######################################################################
if F_modeltype == "LSTM":
K.clear_session()
# X matrix inputs
inputs = Input(shape=(input_dim))
# Network:
x = LSTM(FULLY_CONNECTED,
recurrent_dropout=DROPOUT_RATE,
return_sequences=False)(inputs)
x2 = Dense(64, activation='relu')(x)
predictions = Dense(1, activation='linear')(x2)
model = Model(inputs=inputs, outputs=predictions)
blocks_available = 1
if F_modeltype == "LSTM-0":
model = Sequential() ####### input_shape=input_dim
model.add(
LSTM(FULLY_CONNECTED,
implementation=2,
input_shape=input_dim,
recurrent_dropout=DROPOUT_RATE,
return_sequences=True))
model.add(BatchNormalization())
model.add(Dense(
1)) #, dtype='float32' #Only the softmax is adviced to be float32
blocks_available = 1
#################### TESTING #########################
"""
# Print model overview
print(model.summary())
#compiling the model, creating the callbacks
model.compile(loss='mae',
optimizer='Nadam',
metrics=['mae'])
trainhist = model.fit(x_train,
y_train,
validation_split=0.2,
epochs=10,
batch_size=(batch_size))
scores = model.evaluate(x_test, y_test, verbose=1, batch_size=512)
mae_test = scores[1]#/(24.0*3600)
mae_test = mae_test/(24.0*3600)
mae_test
"""
return model, blocks_available
def make_model(classes, lr_rate, height, width, model_size, rand_seed):
size = len(classes)
#check_file = os.path.join(output_dir, 'tmp.h5')
if model_size == 'L':
if height != 224 or width != 224:
Top = False
weights = None
layer_cut = -1
else:
Top = True
weights = 'imagenet'
layer_cut = -6
# mobile = keras.applications.mobilenet_v2.MobileNetV2(input_shape=input_shape)
#keras.applications.mobilenet.MobileNet(input_shape=None, alpha=1.0, depth_multiplier=1,
#dropout=1e-3, include_top=True, weights='imagenet', input_tensor=None, pooling=None, classes=1000)
mobile = tf.keras.applications.mobilenet.MobileNet(include_top=Top,
input_shape=(height,
width,
3),
pooling='avg',
weights=weights,
alpha=1,
depth_multiplier=1)
x = mobile.layers[layer_cut].output
x = Dense(128,
kernel_regularizer=regularizers.l2(l=0.015),
activation='relu')(x)
x = Dropout(rate=.4, seed=rand_seed)(x)
predictions = Dense(size, activation='softmax')(x)
model = Model(inputs=mobile.input, outputs=predictions)
for layer in model.layers:
layer.trainable = True
model.compile(Adam(lr=lr_rate),
loss='categorical_crossentropy',
metrics=['accuracy'])
else:
if model_size == 'M':
fm = 2
else:
fm = 1
model = Sequential()
model.add(
Conv2D(filters=4 * fm,
kernel_size=(3, 3),
activation='relu',
padding='same',
name='L11',
kernel_regularizer=regularizers.l2(l=0.015),
input_shape=(height, width, 3)))
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='L12'))
model.add(BatchNormalization(name='L13'))
model.add(
Conv2D(filters=8 * fm,
kernel_size=(3, 3),
activation='relu',
kernel_regularizer=regularizers.l2(l=0.015),
padding='same',
name='L21'))
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='L22'))
model.add(BatchNormalization(name='L23'))
model.add(
Conv2D(filters=16 * fm,
kernel_size=(3, 3),
activation='relu',
kernel_regularizer=regularizers.l2(l=0.015),
padding='same',
name='L31'))
model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='L32'))
model.add(BatchNormalization(name='L33'))
if fm == 2:
model.add(
Conv2D(filters=32 * fm,
kernel_size=(3, 3),
activation='relu',
kernel_regularizer=regularizers.l2(l=0.015),
padding='same',
name='L41'))
model.add(
MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='L42'))
model.add(BatchNormalization(name='L43'))
model.add(
Conv2D(filters=64 * fm,
kernel_size=(3, 3),
activation='relu',
kernel_regularizer=regularizers.l2(l=0.015),
padding='same',
name='L51'))
model.add(
MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='L52'))
model.add(BatchNormalization(name='L53'))
model.add(Flatten())
model.add(
Dense(256 * fm,
kernel_regularizer=regularizers.l2(l=0.015),
activation='relu',
name='Dn1'))
model.add(Dropout(rate=.5))
model.add(Dense(size, activation='softmax', name='predict'))
model.compile(Adam(lr=lr_rate, ),
loss='categorical_crossentropy',
metrics=['accuracy'])
#early_stop = EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=10, mode='min', verbose=1)
#checkpoint = ModelCheckpoint(check_file, monitor='val_loss', verbose=1, save_best_only=True, mode='min', period=1)
#lrck=keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=.8, patience=1,
# verbose=1, mode='min', min_delta=0.000001, cooldown=1, min_lr=1.0e-08)
#callbacks=[checkpoint,lrck, early_stop, ]
return model
class MixedModel:
def __init__(self,
mode,
model_params,
input_shapes,
output_shape,
output_folder,
neurons=None,
auto_build=False):
self.model = None
self.mode = mode
self.is_trained = False
self.output_folder = output_folder
self.input_shapes = input_shapes
self.output_shape = output_shape
self.batch_size = model_params['MINIBATCH_SIZE']
self.epochs = model_params['EPOCHS']
self.loss_func = model_params['LOSS']
self.optimizer = model_params['OPTIMIZER']
self.patience = model_params['PATIENCE']
self.verbose = model_params['VERBOSE_OUTPUT']
self.validation_percentage = model_params['VALIDATION_PERCENTAGE']
self.neurons = neurons
if auto_build:
self.build_model(self.mode)
def build_model(self, neurons=None, cnn_layers=None, auto_compile=True):
#if self.mode != "mixed" and len(self.input_shapes) > 1:
# raise ValueError("Cannot accept more than one input shape if mode is not mixed")
if self.mode == "mlp":
inputs, x = self.build_default_mlp(self.input_shapes[0])
x = Dense(self.output_shape[0][0], activation="softmax")(x)
self.model = Model(inputs, x)
elif self.mode == "cnn":
inputs, x = self.build_default_cnn(self.input_shapes[0])
x = Dense(self.output_shape[0][0], activation="softmax")(x)
self.model = Model(inputs, x)
elif self.mode == "mixed":
inputs1, x = self.build_default_cnn(self.input_shapes[0])
inputs2, y = self.build_default_mlp(self.input_shapes[1])
z = concatenate([x, y])
z = Dense(self.output_shape[0][0], activation="softmax")(z)
self.model = Model([inputs1, inputs2], z)
elif self.mode == "mlp-custom":
inputs, x = self.build_custom_mlp(self.input_shapes[0], neurons)
x = Dense(self.output_shape[0][0], activation="softmax")(x)
self.model = Model(inputs, x)
elif self.mode == "cnn-custom":
#input_size = self.input_shapes[0] + (1,)
self.model = tf.keras.Sequential()
self.model.add(Input(shape=self.input_shapes[0]))
self.build_custom_cnn(cnn_layers)
self.model.add(Flatten())
self.model.add(Dense(6, activation="relu"))
self.model.add(Dense(self.output_shape[0][0],
activation="softmax"))
elif self.mode == "mixed-searched":
# build model for image
inputs1 = Input(shape=self.input_shapes[0])
a = Conv2D(filters=16,
kernel_size=(3, 3),
padding="same",
activation="relu")(inputs1)
a = BatchNormalization(axis=-1)(a)
a = Dropout(0.25)(a)
a = Conv2D(filters=16,
kernel_size=(5, 5),
padding="same",
activation="relu")(a)
a = BatchNormalization(axis=-1)(a)
a = MaxPooling2D(pool_size=(2, 2))(a)
a = Flatten()(a)
a = Dense(6, activation="relu")(a)
# build model for mlp
inputs2 = Input(shape=self.input_shapes[1])
b = Dense(10, activation="relu")(inputs2)
b = Dropout(0.25)(b) # removing this yieled about 58% accuracy
b = Dense(4, activation="relu")(b)
b = Dense(10, activation="relu")(b)
# concatenate
z = concatenate([a, b])
# softmax activation to find output
z = Dense(self.output_shape[0][0], activation="softmax")(z)
# create model
self.model = Model([inputs1, inputs2], z)
elif self.mode == "full-mixed":
# build model for img1
inputs1 = Input(shape=self.input_shapes[0])
a = Conv2D(filters=16,
kernel_size=(3, 3),
padding="same",
activation="relu")(inputs1)
a = BatchNormalization(axis=-1)(a)
a = Dropout(0.25)(a)
a = Conv2D(filters=16,
kernel_size=(5, 5),
padding="same",
activation="relu")(a)
a = BatchNormalization(axis=-1)(a)
a = MaxPooling2D(pool_size=(2, 2))(a)
a = Flatten()(a)
a = Dense(6, activation="relu")(a)
# build model for img2
inputs2 = Input(shape=self.input_shapes[1])
b = Conv2D(filters=16,
kernel_size=(3, 3),
padding="same",
activation="relu")(inputs2)
b = BatchNormalization(axis=-1)(b)
b = Dropout(0.25)(b)
b = Conv2D(filters=16,
kernel_size=(5, 5),
padding="same",
activation="relu")(b)
b = BatchNormalization(axis=-1)(b)
b = MaxPooling2D(pool_size=(2, 2))(b)
b = Flatten()(b)
b = Dense(6, activation="relu")(b)
# build model for img3
inputs3 = Input(shape=self.input_shapes[2])
c = Conv2D(filters=16,
kernel_size=(3, 3),
padding="same",
activation="relu")(inputs3)
c = BatchNormalization(axis=-1)(c)
c = Dropout(0.25)(c)
c = Conv2D(filters=16,
kernel_size=(5, 5),
padding="same",
activation="relu")(c)
c = BatchNormalization(axis=-1)(c)
c = MaxPooling2D(pool_size=(2, 2))(c)
c = Flatten()(c)
c = Dense(6, activation="relu")(c)
# build model for img4
inputs4 = Input(shape=self.input_shapes[3])
d = Conv2D(filters=16,
kernel_size=(3, 3),
padding="same",
activation="relu")(inputs4)
d = BatchNormalization(axis=-1)(d)
d = Dropout(0.25)(d)
d = Conv2D(filters=16,
kernel_size=(5, 5),
padding="same",
activation="relu")(d)
d = BatchNormalization(axis=-1)(d)
d = MaxPooling2D(pool_size=(2, 2))(d)
d = Flatten()(d)
d = Dense(6, activation="relu")(d)
# build model for mlp
inputs5 = Input(shape=self.input_shapes[4])
e = Dense(25, activation="relu")(inputs5)
#e = Dropout(0.25)(e) # removing this yieled about 58% accuracy
#e = Dense(4, activation="relu")(e)
#e = Dense(10, activation="relu")(e)
# concatenate
z = concatenate([a, b, c, d, e])
# softmax activation to find output
z = Dense(self.output_shape[0][0], activation="softmax")(z)
# create model
self.model = Model([inputs1, inputs2, inputs3, inputs4, inputs5],
z)
else:
raise ValueError("Unrecognized mode for model generation")
if auto_compile:
self.compile_model()
def build_default_mlp(self, input_size):
inputs = Input(shape=input_size[0])
x = Dense(8, activation="relu")(inputs)
x = Dense(4, activation="relu")(x)
x = Dense(10, activation="relu")(x)
#x = Dense(self.output_shape[0][0], activation="softmax")(x)
return inputs, x
def build_default_cnn(self, input_size):
input_size = input_size + (
1,
) # our images are all b&w so we need to attach the single layer here
inputs = Input(shape=input_size)
x = Conv2D(16, (3, 3), padding="same", activation="relu")(inputs)
x = BatchNormalization(axis=-1)(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
x = Dropout(0.5)(x)
x = Conv2D(32, (3, 3), padding="same", activation="relu")(x)
x = BatchNormalization(axis=-1)(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
x = Dropout(0.25)(x)
x = Flatten()(x)
x = Dense(6, activation="relu")(x)
x = BatchNormalization(axis=-1)(x)
x = Dropout(0.25)(x)
return inputs, x
def build_custom_mlp(self, input_size, neurons):
inputs = Input(shape=input_size[0])
x = Dense(8, activation="relu")(inputs)
for n in neurons:
x = Dense(n, activation="relu")(x)
#x = Dense(self.output_shape[0][0], activation="softmax")(x)
return inputs, x
def build_custom_cnn(
self, layer_defs
): # This will only add layers, input/output should be defined in the build_model function
for l in layer_defs:
self.model.add(self.decode_cnn_layer(l))
self.model.add(BatchNormalization(axis=-1))
def decode_cnn_layer(self, op):
print(op)
k = op.keys()
if "conv" in k:
# Generate convolutional layer
f = op['filters']
s = op['conv']
return Conv2D(f, (s, s), padding='same', activation="relu")
if "pooling" in k:
# Generate pooling layer
s = op['pooling']
return MaxPooling2D((s, s))
def compile_model(self):
if self.model is not None:
self.model.compile(loss=self.loss_func,
optimizer=self.optimizer,
metrics='accuracy')
return True
return False
def train(self,
x,
y,
output_model_file,
save_at_checkpoints=True,
early_stopping=True,
use_existing=True):
if self.model is not None:
# TODO: Implement loading from existing model
callbacks = []
if early_stopping:
callbacks.append(
tf.keras.callbacks.EarlyStopping(monitor='val_loss',
patience=self.patience,
verbose=self.verbose))
# TODO: Implement saving at checkpoints
print(f"Training {self.mode} model...")
history = self.model.fit(
x,
y,
validation_split=self.validation_percentage,
epochs=self.epochs,
shuffle=True,
verbose=self.verbose,
callbacks=callbacks)
self.is_trained = True
return history
def print_model(self):
if self.model is not None:
print(self.model.summary())
else:
raise ValueError("Model is undefined, cannot print")
def save_model(self):
pass
def load_model(self):
pass
def test(self, x, y):
if not self.is_trained:
raise ValueError("Model is not yet trained")
results = self.model.evaluate(x, y, return_dict=True)
return results
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.layers import Dense, Input
input1=Input(shape=(5,))
dense1 = Dense(5, activation='relu')(input1)
dense2=Dense(3)(dense1)
dense3=Dense(4)(dense2)
outputs=Dense(2)(dense3)
#차례대로 앞에 있던게 맨 뒤로 온다. input을 뒤에 명시
model=Model(inputs=input1, outputs=outputs)
model.summary()
#인풋과 아웃풋 dim 맞추기
#이거 함수 input일
'''
model=Sequential()
#model.add(Dense(10, input_dim=5)) #행이 무시되고 열만 표시 3열
model.add(Dense(5,activation='relu', input_shape=(5,)))
#input_shape=(5,) 컬럼이 5개라는 말
model.add(Dense(3))
model.add(Dense(4))
model.add(Dense(2))
model.summary()
# output_dim=2 y도 3차원 이므로 Dense(2)
'''
#3. 컴파일 훈련
model.compile(loss='mse', optimizer='adam', metrics=['mae'])
model.fit(x_train,y_train,batch_size=1, epochs=50,
validation_split=0.2) #각 컬럼별 20%씩 쓰는 것
def build_ATN(architecture=1, input_shape=[28, 28, 1], num_classes=10):
if architecture == 0:
image = Input(shape=input_shape)
target = Input(shape=(num_classes, ))
#target_int = Lambda(lambda x:K.argmax(x,axis=-1))(target)
x1 = Flatten()(image)
#x2 = Embedding(10,20,input_length=1)(target_int)
#x2 = Lambda(lambda x: K.squeeze(x, -2))(x2)
x = Concatenate(axis=-1)([x1, target])
x = Dense(2048,
kernel_initializer='glorot_normal',
bias_initializer='zeros')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Dense(np.prod(input_shape),
activation='sigmoid',
bias_initializer='zeros')(x)
x = Reshape(input_shape)(x)
cnn = Model(inputs=[image, target], outputs=x)
elif architecture == 1:
image = Input(shape=input_shape)
target = Input(shape=(num_classes, ))
x1 = Flatten()(image)
x = Concatenate(axis=-1)([x1, target])
x = Dense(1024,
kernel_initializer='glorot_normal',
bias_initializer='zeros')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Dense(1024,
kernel_initializer='glorot_normal',
bias_initializer='zeros')(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Dense(np.prod(input_shape),
activation='sigmoid',
bias_initializer='zeros')(x)
x = Reshape(input_shape)(x)
cnn = Model(inputs=[image, target], outputs=x)
elif architecture == -1:
cnn = Sequential()
cnn.add(Flatten(input_shape=input_shape))
cnn.add(
Dense(2048,
activation='relu',
kernel_initializer='glorot_normal',
bias_initializer='zeros'))
cnn.add(Dropout(0.25))
cnn.add(
Dense(np.prod(input_shape),
activation='sigmoid',
kernel_initializer='glorot_normal',
bias_initializer='zeros'))
cnn.add(Reshape(input_shape))
elif architecture == -2:
cnn = Sequential()
cnn.add(
Conv2D(
64,
kernel_size=(3, 3),
activation='relu',
kernel_initializer='glorot_normal',
bias_initializer='zeros', #Constant(-0.5),
kernel_regularizer=l2(0.005),
input_shape=input_shape,
padding='same'))
cnn.add(
Conv2D(128,
kernel_size=(3, 3),
activation='relu',
kernel_initializer='glorot_normal',
bias_initializer='zeros',
kernel_regularizer=l2(0.005),
padding='same'))
cnn.add(MaxPooling2D(pool_size=(2, 2)))
cnn.add(
Conv2D(256,
kernel_size=(3, 3),
activation='relu',
kernel_initializer='glorot_normal',
bias_initializer='zeros',
kernel_regularizer=l2(0.005),
padding='same'))
cnn.add(MaxPooling2D(pool_size=(2, 2)))
cnn.add(Flatten())
cnn.add(
Dense(2048,
activation='relu',
kernel_initializer='glorot_normal',
bias_initializer='zeros',
kernel_regularizer=l2(0.05)))
cnn.add(Dropout(0.25))
cnn.add(
Dense(np.prod(input_shape),
activation='sigmoid',
kernel_initializer='glorot_normal',
bias_initializer='zeros'))
cnn.add(Reshape(input_shape))
elif architecture == 2:
cnn = Sequential()
cnn.add(
Conv2D(256,
kernel_size=(3, 3),
activation='relu',
input_shape=input_shape,
padding='same',
use_bias=True,
kernel_initializer='glorot_normal',
bias_initializer='zeros'))
cnn.add(MaxPooling2D(pool_size=(2, 2)))
cnn.add(Dropout(0.5))
cnn.add(
Conv2D(512,
kernel_size=(3, 3),
activation='relu',
padding='same',
use_bias=True,
kernel_initializer='glorot_normal',
bias_initializer='zeros'))
#cnn.add(MaxPooling2D(pool_size=(2, 2)))
#cnn.add(Dropout(0.5))
#cnn.add(Conv2D(512, kernel_size=(3, 3),activation='relu',padding='same',
# use_bias=True, kernel_initializer='glorot_normal', bias_initializer='zeros'))
#cnn.add(UpSampling2D(data_format='channels_last'))
#cnn.add(Dropout(0.5))
#cnn.add(Conv2DTranspose(256, kernel_size=(3,3), padding='same', activation='relu',
# use_bias=True, kernel_initializer='glorot_normal', bias_initializer='zeros'))
cnn.add(UpSampling2D(data_format='channels_last'))
cnn.add(Dropout(0.5))
cnn.add(
Conv2DTranspose(256,
kernel_size=(3, 3),
padding='same',
activation='relu',
use_bias=True,
kernel_initializer='glorot_normal',
bias_initializer='zeros'))
cnn.add(Dropout(0.5))
cnn.add(
Conv2DTranspose(1,
kernel_size=(3, 3),
padding='same',
activation='sigmoid',
use_bias=True,
kernel_initializer='glorot_normal',
bias_initializer='zeros'))
return cnn