def build_discriminator(self):
d_in = Input(shape=IMG_SIZE)
d = Conv2D(filters=16,
kernel_size=9,
padding='same')(d_in)
d = Dropout(rate=DROPOUT_RATE)(d)
d = LeakyReLU(alpha=LRELU_FACTOR)(d)
d = MaxPooling2D(pool_size=4)(d)
d = Conv2D(filters=32,
kernel_size=5,
padding='same')(d)
d = Dropout(rate=DROPOUT_RATE)(d)
d = LeakyReLU(alpha=LRELU_FACTOR)(d)
d = MaxPooling2D(pool_size=2)(d)
d = Conv2D(filters=32,
kernel_size=5,
padding='same')(d)
d = Dropout(rate=DROPOUT_RATE)(d)
d = LeakyReLU(alpha=LRELU_FACTOR)(d)
d = MaxPooling2D(pool_size=2)(d)
d = Conv2D(filters=64,
kernel_size=5,
padding='same')(d)
d = Dropout(rate=DROPOUT_RATE)(d)
d = LeakyReLU(alpha=LRELU_FACTOR)(d)
d = Flatten()(d)
d = Dense(units=128)(d)
d = Dropout(rate=DROPOUT_RATE)(d)
d = LeakyReLU(alpha=LRELU_FACTOR)(d)
d = Dense(units=1, activation='sigmoid')(d)
d = Model(inputs=d_in, outputs=d)
print("SUMMARY DISCRIMINATOR: ")
d.summary()
return d
def main(data_path, output_path):
X_trainS1, Y_train, X_valS1, Y_val = load_data(data_path)
epochs = 10
batch_size = 256
dropout_rate = 0.15
n_classes = 6
# 三个子模型的输入数据
main_input1 = Input(shape=(128, 3), name='main_input1')
def lstm_cell(main_input):
"""
基于DeepConvLSTM算法, 创建子模型
:param main_input: 输入数据
:return: 子模型
"""
sub_model = TimeDistributed(Dense(384),
input_shape=(128, 3))(main_input)
# sub_model = Flatten()(main_input)
print(sub_model)
sub_model = LSTM(256, return_sequences=True)(sub_model)
sub_model = LSTM(128, return_sequences=True)(sub_model)
sub_model = LSTM(128)(sub_model)
main_output = Dropout(dropout_rate)(sub_model)
return main_output
model = lstm_cell(main_input1)
model = Dropout(0.4)(model)
model = Dense(n_classes)(model)
model = BatchNormalization()(model)
output = Activation('softmax', name="softmax")(model)
model = Model([main_input1], output)
model.compile(loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['accuracy'])
print(model.summary())
# graph_path = os.path.join(output_path, "merged_model.png")
# plot_model(model, to_file=graph_path, show_shapes=True) # 绘制模型图
metrics = Metrics() # 度量FPR
history = model.fit(X_trainS1,
Y_train,
batch_size=batch_size,
validation_data=(X_valS1, Y_val),
epochs=epochs,
callbacks=[metrics]) # 增加FPR输出
model_path = os.path.join(output_path, "merged_dcl.h5")
model.save(model_path) # 存储模型
print(history.history)
def genre_model(num_genres, input_shape):
input = Input(shape=input_shape)
model = Conv1D(filters=128,
kernel_size=9,
activation='relu',
dilation_rate=1)(input)
model = MaxPooling1D(pool_size=3)(model)
model = Dropout(0.25)(model)
model = Conv1D(filters=128,
kernel_size=9,
activation='relu',
dilation_rate=2)(model)
model = MaxPooling1D(pool_size=3)(model)
model = Dropout(0.25)(model)
model = Conv1D(filters=64,
kernel_size=9,
activation='relu',
dilation_rate=4)(model)
model = MaxPooling1D(pool_size=3)(model)
model = Dropout(0.25)(model)
model = Conv1D(filters=64,
kernel_size=9,
activation='relu',
dilation_rate=8)(model)
model = MaxPooling1D(pool_size=3)(model)
model = Dropout(0.25)(model)
model = Conv1D(filters=32,
kernel_size=9,
activation='relu',
dilation_rate=16)(model)
model = MaxPooling1D(pool_size=3)(model)
model = Dropout(0.25)(model)
model = Conv1D(filters=32,
kernel_size=7,
activation='relu',
dilation_rate=32)(model)
model = MaxPooling1D(pool_size=3)(model)
model = Flatten()(model)
model = Dropout(0.25)(model)
output = Dense(num_genres, activation='softmax')(model)
model = Model(input, output)
model.summary()
return (model)
def get_model():
inp = Input(shape=(MAX_TEXT_LENGTH, 1))
#model = Embedding(MAX_FEATURES, EMBED_SIZE)(inp)
model = Dropout(0.5)(inp)
model = Conv1D(filters=32,
kernel_size=2,
padding='same',
activation='relu')(model)
model = MaxPooling1D(pool_size=2)(model)
model = Flatten()(model)
model = Dense(1024, activation='relu')(model)
model = Dropout(0.5)(model)
model = Dense(NB_CLASS, activation="softmax")(model)
model = Model(inputs=inp, outputs=model)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
model.summary()
return model
tangential_strain_layer = Activation('relu')(tangential_strain_layer)
tangential_strain_layer = (Conv1D(filters=filters, kernel_size=kernel_size))(tangential_strain_layer)
tangential_strain_layer = Activation('relu')(tangential_strain_layer)
tangential_strain_layer = (Dense(32, activation='relu'))(tangential_strain_layer)
tangential_strain_layer = Dropout(0.5)(tangential_strain_layer)
tangential_strain_layer = Flatten()(tangential_strain_layer)
merge = concatenate([energy_layer, maximum_amplitude_layer, radial_strain_layer, tangential_strain_layer])
model = (Dense(32, activation='relu'))(merge)
model = Dropout(0.5)(model)
model = (Dense(1, activation='sigmoid'))(model)
model = Model(inputs=[energy_input, maximum_amplitude_input, radial_strain_input, tangential_strain_input],
outputs=model)
model.compile(loss='binary_crossentropy', optimizer="rmsprop", metrics=['accuracy'])
print model.summary()
print "Training....................................................."
model.fit([energies, maximum_amplitudes, radial_strains, tangential_strains], train_labels, epochs=epochs, verbose=1,
batch_size=batch_size, validation_split=0.2)
print("--- train %s seconds ---" % (time.time() - start_time))
original_classified = model.predict([energies, maximum_amplitudes, radial_strains, tangential_strains],
batch_size=batch_size, verbose=0)
best_threshold = find_best_threshold(train_labels, original_classified)
if run_test:
print 'test'
for test in test_sets:
print test
cnn = Conv2D(256, cnn_kernel, activation=cnn_act, padding='same'
kernel_initializer=kernel_init, name='conv4')(cnn)
cnn = MaxPooling2D(pool_size=(1,2), strides=2)(cnn)
cnn = Dropout(0.25)(cnn)
cnn = Conv2D(512, cnn_kernel, activation=cnn_act, padding='same'
kernel_initializer=kernel_init, name='conv5')(cnn)
cnn = BatchNormalization(axis=1)(cnn)
cnn = Conv2D(512, cnn_kernel, activation=cnn_act, padding='same'
kernel_initializer=kernel_init, name='conv6')(cnn)
cnn = BatchNormalization(axis=1)(cnn)
cnn = MaxPooling2D(pool_size=(1,2), strides=2)(cnn)
cnn = Conv2D(512, cnn_kernel, activation=cnn_act, padding='same'
kernel_initializer=kernel_init, name='conv6')(cnn)
#cnn.add(Conv2D(512,(1,2), activation=cnn_act))
cnn = Dropout(0.25)(cnn)
cnn.summary()
conv_to_rnn_dims = (img_width//(1**2), (img_height // (2**2))*512)
rnn = Reshape(target_shape=conv_to_rnn_dims, name='reshape')(cnn)
rnn = Dense(32, activation=cnn_act, name='dense1')(rnn)
rnn = Bidirectional(LSTM(rnn_size, return_sequences=True,
kernel_initializer=kernel_init, name='lstm1'))(rnn)
rnn = Dropout(0.25)(rnn)
rnn = Bidirectional(LSTM(rnn_size, return_sequences=True,
kernel_initializer=kernel_init, name='lstm2'))(rnn)
rnn = Dense(list_chars_len+1, kernel_initializer=kernel_init
,name='dense2')(rnn)
y_pred = Activation('softmax', name='softmax')
Model(inputs=Input_data, outputs=y_pred).summary()
model = Flatten()(model)
dense_layer = Dense(nclass, activation=activations.softmax)(model)
model = models.Model(inputs=input_layer, outputs=dense_layer)
# Compile model
epochs = 10
opt = optimizers.Adam(
) # Default: lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False,
model.compile(optimizer=opt,
loss=losses.binary_crossentropy,
metrics=['accuracy'])
print(get_available_gpus())
# Model Summary
model.summary()
# Start timer
start_time = time.time()
# Fit model
model.fit(x_train,
y_train,
batch_size=16,
validation_data=(x_valid, y_valid),
epochs=epochs,
shuffle=True,
verbose=2,
callbacks=[metrics])
# Get time
total_train_time = time.time() - start_time
print('Total training time in seconds: ')
print(total_train_time)
def BiLSTM_model(filename,
train,
output,
X_train,
X_test,
word2ind,
maxWords,
y_train,
y_test,
ind2label,
validation=False,
X_valid=None,
y_valid=None,
word_embeddings=True,
pretrained_embedding="",
word_embedding_size=100,
maxChar=0,
char_embedding_type="",
char2ind="",
char_embedding_size=50,
lstm_hidden=32,
nbr_epochs=5,
batch_size=32,
dropout=0,
optimizer='rmsprop',
early_stopping_patience=-1,
folder_path="model_results",
gen_confusion_matrix=False):
"""
Build, train and test a BiLSTM Keras model. Works for multi-tasking learning.
The model architecture looks like:
- Words representations:
- Word embeddings
- Character-level representation [Optional]
- Dropout
- Bidirectional LSTM
- Dropout
- Softmax/CRF for predictions
:param filename: File to redirect the printing
:param train: Boolean if the model must be trained or not. If False, the model's wieght are expected to be stored in "folder_path/filename/filename.h5"
:param otput: "crf" or "softmax". Type of prediction layer to use
:param X_train: Data to train the model
:param X_test: Data to test the model
:param word2ind: Dictionary containing all words in the training data and a unique integer per word
:param maxWords: Maximum number of words in a sequence
:param y_train: Labels to train the model for the prediction task
:param y_test: Labels to test the model for the prediction task
:param ind2label: Dictionary where all labels for task 1 are mapped into a unique integer
:param validation: Boolean. If true, the validation score will be computed from 'X_valid' and 'y_valid'
:param X_valid: Optional. Validation dataset
:param y_valid: Optional. Validation dataset labels
:param word_embeddings: Boolean value. Add word embeddings into the model.
:param pretrained_embedding: Use the pretrained word embeddings.
Three values:
- "": Do not use pre-trained word embeddings (Default)
- False: Use the pre-trained embedding vectors as the weights in the Embedding layer
- True: Use the pre-trained embedding vectors as weight initialiers. The Embedding layer will still be trained.
:param word_embedding_size: Size of the pre-trained word embedding to use (100 or 300)
:param maxChar: The maximum numbers of characters in a word. If set to 0, the model will not use character-level representations of the words
:param char_embedding_type: Type of model to use in order to compute the character-level representation of words: Two values: "CNN" or "BILSTM"
:param char2ind: A dictionary where each character is maped into a unique integer
:param char_embedding_size: size of the character-level word representations
:param lstm_hidden: Dimentionality of the LSTM output space
:param nbr_epochs: Number of epochs to train the model
:param batch_size: Size of batches while training the model
:param dropout: Rate to apply for each Dropout layer in the model
:param optimizer: Optimizer to use while compiling the model
:param early_stopping_patience: Number of continuous tolerated epochs without improvement during training.
:param folder_path: Path to the directory storing all to-be-generated files
:param gen_confusion_matrix: Boolean value. Generated confusion matrices or not.
:return: The classification scores for both tasks.
"""
print("====== {0} start ======".format(filename))
end_string = "====== {0} end ======".format(filename)
# Create directory to store results
os.makedirs(folder_path + "/" + filename)
filepath = folder_path + "/" + filename + "/" + filename
# Set print outputs file
file, stdout_original = setPrintToFile("{0}.txt".format(filepath))
# Model params
nbr_words = len(word2ind) + 1
out_size = len(ind2label) + 1
best_results = ""
embeddings_list = []
inputs = []
# Input - Word Embeddings
if word_embeddings:
word_input = Input((maxWords, ))
inputs.append(word_input)
if pretrained_embedding == "":
word_embedding = Embedding(nbr_words,
word_embedding_size)(word_input)
else:
# Retrieve embeddings
embedding_matrix = word2VecEmbeddings(word2ind,
word_embedding_size)
word_embedding = Embedding(nbr_words,
word_embedding_size,
weights=[embedding_matrix],
trainable=pretrained_embedding,
mask_zero=False)(word_input)
embeddings_list.append(word_embedding)
# Input - Characters Embeddings
if maxChar != 0:
character_input = Input((
maxWords,
maxChar,
))
char_embedding = character_embedding_layer(char_embedding_type,
character_input, maxChar,
len(char2ind) + 1,
char_embedding_size)
embeddings_list.append(char_embedding)
inputs.append(character_input)
# Model - Inner Layers - BiLSTM with Dropout
embeddings = concatenate(embeddings_list) if len(
embeddings_list) == 2 else embeddings_list[0]
model = Dropout(dropout)(embeddings)
model = Bidirectional(
LSTM(lstm_hidden, return_sequences=True, dropout=dropout))(model)
model = Dropout(dropout)(model)
if output == "crf":
# Output - CRF
crfs = [[CRF(out_size), out_size]
for out_size in [len(x) + 1 for x in ind2label]]
outputs = [x[0](Dense(x[1])(model)) for x in crfs]
model_loss = [x[0].loss_function for x in crfs]
model_metrics = [x[0].viterbi_acc for x in crfs]
if output == "softmax":
outputs = [
Dense(out_size, activation='softmax')(model)
for out_size in [len(x) + 1 for x in ind2label]
]
model_loss = ['categorical_crossentropy' for x in outputs]
model_metrics = None
# Model
model = Model(inputs=inputs, outputs=outputs)
model.compile(loss=model_loss,
metrics=model_metrics,
optimizer=get_optimizer(optimizer))
print(model.summary(line_length=150), "\n\n\n\n")
# Training Callbacks:
callbacks = []
value_to_monitor = 'val_f1'
best_model_weights_path = "{0}.h5".format(filepath)
# 1) Classifition scores
classification_scores = Classification_Scores([X_train, y_train],
ind2label,
best_model_weights_path)
callbacks.append(classification_scores)
# 2) EarlyStopping
if early_stopping_patience != -1:
early_stopping = EarlyStopping(monitor=value_to_monitor,
patience=early_stopping_patience,
mode='max')
callbacks.append(early_stopping)
# Train
if train:
# Train the model. Keras's method argument 'validation_data' is referred as 'testing data' in this code.
hist = model.fit(X_train,
y_train,
validation_data=[X_test, y_test],
epochs=nbr_epochs,
batch_size=batch_size,
callbacks=callbacks,
verbose=2)
print()
print('-------------------------------------------')
print(
"Best F1 score:", early_stopping.best,
" (epoch number {0})".format(
1 + np.argmax(hist.history[value_to_monitor])))
# Save Training scores
save_model_training_scores("{0}".format(filepath), hist,
classification_scores)
# Print best testing classification report
best_epoch = np.argmax(hist.history[value_to_monitor])
print(classification_scores.test_report[best_epoch])
# Best epoch results
best_results = model_best_scores(classification_scores, best_epoch)
# Load weigths from best training epoch into model
save_load_utils.load_all_weights(model, best_model_weights_path)
# Create confusion matrices
if gen_confusion_matrix:
for i, y_target in enumerate(y_test):
# Compute predictions, flatten
predictions, target = compute_predictions(model, X_test, y_target,
ind2label[i])
# Generate confusion matrices
save_confusion_matrix(
target, predictions, list(ind2label[i].values()),
"{0}_task_{1}_confusion_matrix_test".format(
filepath, str(i + 1)))
# Validation dataset
if validation:
print()
print("Validation dataset")
print("======================")
# Compute classification report
for i, y_target in enumerate(y_valid):
# Compute predictions, flatten
predictions, target = compute_predictions(model,
X_valid,
y_target,
ind2label[i],
nbrTask=i)
# Only for multi-task
if len(y_train) > 1:
print("For task " + str(i + 1) + "\n")
print(
"===================================================================================="
)
print("")
print("With padding into account")
print(
metrics.flat_classification_report([target], [predictions],
digits=4))
print("")
print('----------------------------------------------')
print("")
print("Without the padding:")
print(
metrics.flat_classification_report([target], [predictions],
digits=4,
labels=list(
ind2label[i].values())))
# Generate confusion matrices
save_confusion_matrix(
target, predictions, list(ind2label[i].values()),
"{0}_task_{1}_confusion_matrix_validation".format(
filepath, str(i + 1)))
# Close file
closePrintToFile(file, stdout_original)
print(end_string)
return best_results
def __init__(self, model_name="Xception", train_class_name=None, training_batch_size=100, existing_weight=None, test_percentage=0.02, learning_rate=0.000, validation_every_X_batch=5):
if train_class_name == None:
print("You must specify train_class_name")
return
self.validation_every_X_batch = validation_every_X_batch
self.Y = []
self.model_file = model_name + "-{date:%Y-%m-%d-%H-%M-%S}".format( date=datetime.datetime.now())
print("model_folder: ", self.model_file)
self.train_class_name = train_class_name
if not os.path.exists(os.path.join("models", train_class_name)):
os.makedirs(os.path.join("models", train_class_name))
self.training_batch_size = training_batch_size
# We know that MNIST images are 28 pixels in each dimension.
img_size = 512
self.img_size_flat = img_size * img_size * 3
self.img_shape_full = (img_size, img_size, 3)
self.test = {}
with open('base/Annotations/label.csv', 'r') as csvfile:
reader = csv.reader(csvfile)
all_class_samples = []
for row in reader:
if row[1] == self.train_class_name:
self.num_classes = len(row[2])
break
# Start construction of the Keras Sequential model.
input_tensor = Input(shape=self.img_shape_full)
if model_name == "Xception":
base_model = Xception(input_tensor=input_tensor, weights='imagenet', include_top=False, classes=self.num_classes)
elif model_name == "MobileNet":
base_model = MobileNet(input_tensor=input_tensor, weights='imagenet', include_top=False, classes=self.num_classes)
elif model_name == "DenseNet121":
base_model = DenseNet121(input_tensor=input_tensor, weights='imagenet', include_top=False, classes=self.num_classes)
x = base_model.output
x = GlobalAveragePooling2D()(x)
model = Dropout(0.2)(x)
predictions = Dense(self.num_classes, activation='softmax')(x)
# this is the model we will train
model = Model(inputs=base_model.input, outputs=predictions)
self.model = model
print(model.summary())
self.optimizer = optimizers.Adam(lr=learning_rate)
model.compile(optimizer=self.optimizer, loss='categorical_crossentropy', metrics=['accuracy'])
with open('base/Annotations/label.csv', 'r') as csvfile:
reader = csv.reader(csvfile)
all_class_samples = []
for row in reader:
if row[1] != self.train_class_name:
continue
all_class_samples.append(row)
self.X = np.zeros((len(all_class_samples), img_size, img_size, 3))
# self.X = np.zeros((10, img_size, img_size, 3))
test_count = int(test_percentage * len(all_class_samples))
index = 0
print("Training " + train_class_name + " with: " + str(int((1 - test_percentage) * len(all_class_samples))) + ", Testing with: " + str(test_count), str(self.num_classes), "Classes")
print("Loading images...")
for row in all_class_samples:
image = Image.open("base/" + row[0])
img_array = np.asarray(image)
if img_array.shape != self.img_shape_full:
image = image.resize((img_size, img_size), Image.ANTIALIAS)
img_array = np.asarray(image)
self.X[index] = img_array
self.Y.append(row[2].index("y"))
if index % 500 == 0:
print(index)
index += 1
self.Y = to_categorical(self.Y, num_classes=self.num_classes)
self.X_train, self.X_test, self.y_train, self.y_test = train_test_split(self.X, self.Y, test_size = test_percentage, random_state=42)
del self.X
class_weight = {}
class_count = np.sum(self.y_train, axis=0)
print("Training Sample for each Class", class_count)
for class_index in range(self.num_classes):
class_weight[class_index] = 1 /(class_count[class_index] / np.sum(class_count)) / self.num_classes
self.class_weight = class_weight
print("Class weights: ", self.class_weight)
os.makedirs(os.path.join("models", train_class_name, self.model_file))
model.save(os.path.join("models", train_class_name, self.model_file, train_class_name + "_" + "model.h5"))
vgg = cnn_algo()
# vgg.summary()
for layers in vgg.layers: # we are keeping the weights constant, it is already trained by the thousands of weights that is in imagenet
layers.trainable = False # we are not training all the layers
x = Flatten()(
vgg.output
) # need to add last layer, after flattning we can add the last layer
folders = glob.glob(
cwd + '/train/*') # check number of folders inside training folder
print(folders)
prediction = Dense(len(folders), activation='softmax')(
x) # its the sigmoid activation function
model = Dropout(0.2)
model = input_model()
model.summary() # view the structure of model
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
training_set = aggumentation_training()
test_set = aggumentation_training()
r = model.fit_generator(training_set,
epochs=3,
steps_per_epoch=len(training_set),
validation_steps=len(test_set))
model.save('My_face_features_model.h5')
def main(data_path, output_path):
X_trainS1, Y_train, X_valS1, Y_val = load_data(data_path)
epochs = 20
batch_size = 256
kernel_size = 3
pool_size = 2
dropout_rate = 0.15
n_classes = 6
f_act = 'relu'
# 三个子模型的输入数据
main_input1 = Input(shape=(128, 3), name='main_input1')
def cnn_cell(main_input):
"""
基于CNN-Model算法, 创建子模型
:param main_input: 输入数据
:return: 子模型
"""
sub_model = Conv1D(512,
kernel_size,
input_shape=(128, 3),
activation=f_act,
padding='same')(main_input)
sub_model = BatchNormalization()(sub_model)
sub_model = MaxPooling1D(pool_size=pool_size)(sub_model)
sub_model = Dropout(dropout_rate)(sub_model)
sub_model = Conv1D(64, kernel_size, activation=f_act,
padding='same')(sub_model)
sub_model = BatchNormalization()(sub_model)
sub_model = MaxPooling1D(pool_size=pool_size)(sub_model)
sub_model = Dropout(dropout_rate)(sub_model)
print('sub_model322:', sub_model)
sub_model = Conv1D(32, kernel_size, activation=f_act,
padding='same')(sub_model)
print('sub_model322:', sub_model)
sub_model = BatchNormalization()(sub_model)
sub_model = Flatten()(sub_model)
print('sub_modelFlatten:', sub_model)
return sub_model
model = cnn_cell(main_input1)
model = Dropout(0.4)(model)
model = Dense(n_classes)(model)
model = BatchNormalization()(model)
output = Activation('softmax', name="softmax")(model)
model = Model([main_input1], output)
model.compile(loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['accuracy'])
# graph_path = os.path.join(output_path, "merged_model.png")
# plot_model(model, to_file=graph_path, show_shapes=True) # 绘制模型图
print(model.summary())
metrics = Metrics() # 度量FPR
history = model.fit(X_trainS1,
Y_train,
batch_size=batch_size,
validation_data=(X_valS1, Y_val),
epochs=epochs,
callbacks=[metrics]) # 增加FPR输出
model_path = os.path.join(output_path, "merged_dcl.h5")
model.save(model_path) # 存储模型
print(history.history)
