class ModelA(ModelBase):
def __init__(self, wv, maxlen=50, max_num_deptag=50):
super().__init__(wv, maxlen, max_num_deptag)
embedding_layer = self.wv.model.wv.get_embedding_layer()
sequence_input = Input(shape=(self.maxlen,), dtype='int32')
mask = Masking(mask_value=-1)(sequence_input)
embedded_sequences = embedding_layer(mask)
x = (Conv1D(filters=32, kernel_size=3, activation='relu'))(embedded_sequences)
x = (MaxPooling1D(pool_size=2))(x)
x = (Bidirectional(GRU(100, dropout=0.15)))(x)
x = (Dense(16))(x)
x = (Dropout(0.2))(x)
preds = (Dense(1, activation='tanh'))(x)
self.model = Model(sequence_input, preds, name='ConvBiGRUModelA')
self.compile_model()
def compile_model(self):
self.model.compile(loss='mean_squared_error', optimizer='sgd', metrics=['mae'])
def fit(self, x, y, batch_size, epochs, verbose, callbacks):
x_preprocessed = np.array([[t[0] for t in row] for row in x])
x_preprocessed = sequence.pad_sequences(x_preprocessed, maxlen=self.maxlen, value=-1)
y_preprocessed = np.array(y)
self.model.fit(x_preprocessed, y_preprocessed, batch_size=batch_size, epochs=epochs, verbose=verbose,
callbacks=callbacks)
def predict(self, x):
x_preprocessed = np.array([[t[0] for t in row] for row in x])
x_preprocessed = sequence.pad_sequences(x_preprocessed, maxlen=self.maxlen, value=-1)
return self.model.predict(x_preprocessed)
def fit_model(X_train, y_train):
#define model
model = Model(inputs=[admiss_data], outputs=main_output)
#
print(model.summary())
#
adam = optimizers.Adam(lr=0.0001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0)
#
model.compile(optimizer=adam,
loss='binary_crossentropy',
metrics=['accuracy'])
#
class_weight = {
0: 1.,
1: cw, # 1: 20.
}
histories = my_callbacks.Histories()
#model fit
model.fit([X_train],
y_train,
epochs=n_epochs,
batch_size=n_batch_size,
validation_data=([[X_val], y_val]),
class_weight=class_weight,
callbacks=[histories])
model.save('base_nn.h5')
return model
def train(self, classdict, nb_topics, *args, **kwargs):
""" Train the autoencoder.
:param classdict: training data
:param nb_topics: number of topics, i.e., the number of encoding dimensions
:param args: arguments to be passed to keras model fitting
:param kwargs: arguments to be passed to keras model fitting
:return: None
:type classdict: dict
:type nb_topics: int
"""
self.nb_topics = nb_topics
self.generate_corpus(classdict)
vecsize = len(self.dictionary)
# define all the layers of the autoencoder
input_vec = Input(shape=(vecsize, ))
encoded = Dense(self.nb_topics, activation='relu')(input_vec)
decoded = Dense(vecsize, activation='sigmoid')(encoded)
# define the autoencoder model
autoencoder = Model(input=input_vec, output=decoded)
# define the encoder
encoder = Model(input=input_vec, output=encoded)
# define the decoder
encoded_input = Input(shape=(self.nb_topics, ))
decoder_layer = autoencoder.layers[-1]
decoder = Model(input=encoded_input,
output=decoder_layer(encoded_input))
# compile the autoencoder
autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')
# process training data
embedvecs = np.array(
reduce(add, [
map(
lambda shorttext: self.retrieve_bow_vector(
shorttext, normalize=True), classdict[classtype])
for classtype in classdict
]))
# fit the model
autoencoder.fit(embedvecs, embedvecs, *args, **kwargs)
# store the autoencoder models
self.autoencoder = autoencoder
self.encoder = encoder
self.decoder = decoder
# flag setting
self.trained = True
# classes topic vector precomputation
self.classtopicvecs = {}
for label in classdict:
self.classtopicvecs[label] = self.precalculate_liststr_topicvec(
classdict[label])
def test_works():
x = Input(shape=(30, 1), name="input")
e = GRU(128, return_sequences=True)(x)
s = Slice("[-1,:]")(e)
# s = Slice('[-1,:]')(e)
# s = theano.printing.Print("s")(s)
r = RepeatVector(30)(s)
m = Merge(mode='concat', concat_axis=2)([r, x])
d = GRU(128, return_sequences=True)(m)
p = Ptr_Layer(30)([x, e, d])
model = Model(input=x, output=p, name='test')
# print(Sort(nb_out=5).get_output_shape_for((1,2,3)))
inp = np.random.randint(size=(10000, 30, 1), low=0, high=100)
indicies = np.argsort(inp[:, :, 0])
# print(indicies)
target = np.array(
[np.take(inp[i], indicies[i], axis=-2) for i in range(inp.shape[0])])
# print("Input")
# print(inp)
# print("Target")
# print(target)
model.compile(optimizer=optimizers.Adam(), loss='mse')
model.fit(inp, target, nb_epoch=500, batch_size=100)
def train():
# extracting file saved by data_prep.py
data = np.load('face_data.npz')
x , y = data['x'], data['y']
#categorical conversion of data label
y = keras.utils.to_categorical(y, 6)
# using transfer learning to reduce the time required to train the algo
resnet = VGGFace(model='resnet50',input_shape=(224, 224, 3))
layer_name = resnet.layers[-2].name
#adding our own custom layers to make the model work on our datatset
out = resnet.get_layer(layer_name).output
out = Dense(6,activation='softmax')(out)
resnet_4 = Model(resnet.input, out)
# removing last layer of the model and adding my own layer to it
for layer in resnet_4.layers[:-1]:
layer.trainable = False
resnet_4.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
#checking the final created dataset
print (resnet_4.summary())
# training the model we have created with our own dataset
resnet_4.fit(x, y,batch_size=10,epochs=10,shuffle=True)
#saving the trained model so that it can be used afterwards
resnet_4.save("C:\\Users\\hseth\\Desktop\\face recogination\\model_save_face.h5")
# checking the accuracy of the model on training data only as i used a very small dataset
scores = resnet_4.evaluate(x, y, verbose=1)
print('Test accuracy:', scores[1])
class ModelB(ModelBase):
def __init__(self, wv, maxlen=50, max_num_deptag=50):
super().__init__(wv, maxlen, max_num_deptag)
embedding_layer = self.wv.model.wv.get_embedding_layer()
sequence_input = Input(shape=(self.maxlen,), dtype='int32')
mask = Masking(mask_value=-1)(sequence_input)
embedded_sequences = embedding_layer(mask)
x = Dense(50, activation='tanh')(embedded_sequences)
x = (Dropout(0.2))(x)
x = Flatten()(x)
preds = Dense(1)(x)
self.model = Model(sequence_input, preds, name='MLPModelB')
self.compile_model()
def compile_model(self):
self.model.compile(loss='mean_squared_error', optimizer='sgd', metrics=['mae'])
def fit(self, x, y, batch_size, epochs, verbose, callbacks):
x_preprocessed = np.array([[t[0] for t in row] for row in x])
x_preprocessed = sequence.pad_sequences(x_preprocessed, maxlen=self.maxlen, value=-1)
y_preprocessed = np.array(y)
self.model.fit(x_preprocessed, y_preprocessed, batch_size=batch_size, epochs=epochs, verbose=verbose,
callbacks=callbacks)
def predict(self, x):
x_preprocessed = np.array([[t[0] for t in row] for row in x])
x_preprocessed = sequence.pad_sequences(x_preprocessed, maxlen=self.maxlen, value=-1)
return self.model.predict(x_preprocessed)
def my_layer():
"""test one specify layer"""
a = Input(shape=(3, 3, 2))
b = WeightedAdd()(a)
model = Model(inputs=a, outputs=b)
data = np.ones((1, 3, 3, 2))
print(model.predict_on_batch(data))
model.compile(optimizer='Adam', loss=mean_squared_error)
model.fit(data, data, epochs=1000)
print(model.predict_on_batch(data))
class ModelE(ModelBase):
def __init__(self, wv, maxlen=50, max_num_deptag=50):
super().__init__(wv, maxlen, max_num_deptag)
wv_layer = self.wv.model.wv.get_embedding_layer()
wv_input = Input(shape=(self.maxlen,), dtype='int32')
wv_mask = Masking(mask_value=-1)(wv_input)
wv_sequences = wv_layer(wv_mask)
deptag_input = Input(shape=(self.maxlen,), dtype='int32')
deptag_mask = Masking(mask_value=0)(deptag_input)
deptag_sequences = Embedding(input_dim=max_num_deptag, output_dim=10, input_length=50)(deptag_mask)
x = concatenate([wv_sequences, deptag_sequences])
x = (Conv1D(filters=128, kernel_size=3, activation='relu'))(x)
# x = (AveragePooling1D(pool_size=3))(x)
x = (Dropout(0.2))(x)
x = (Conv1D(filters=32, kernel_size=17, activation='relu'))(x)
x = (Dropout(0.2))(x)
x = (GlobalMaxPooling1D())(x)
x = (Dense(16))(x)
x = (Dropout(0.2))(x)
preds = (Dense(1, activation='tanh'))(x)
self.model = Model(inputs=[wv_input, deptag_input], outputs=preds, name='OnlyConvolutions')
self.compile_model()
def compile_model(self):
self.model.compile(loss='mean_squared_error', optimizer='sgd', metrics=['mae'])
def fit(self, x, y, batch_size, epochs, verbose, callbacks):
x1_preprocessed = np.array([[t[0] for t in row] for row in x])
x1_preprocessed = sequence.pad_sequences(x1_preprocessed, maxlen=self.maxlen, value=-1)
x2_preprocessed = np.array([[t[1] for t in row] for row in x])
x2_preprocessed = sequence.pad_sequences(x2_preprocessed, maxlen=self.maxlen, value=-1)
y_preprocessed = np.array(y)
self.model.fit([x1_preprocessed, x2_preprocessed], y_preprocessed, batch_size=batch_size, epochs=epochs,
verbose=verbose,
callbacks=callbacks)
def predict(self, x):
x1_preprocessed = np.array([[t[0] for t in row] for row in x])
x1_preprocessed = sequence.pad_sequences(x1_preprocessed, maxlen=self.maxlen, value=-1)
x2_preprocessed = np.array([[t[1] + 1 for t in row] for row in x])
x2_preprocessed = sequence.pad_sequences(x2_preprocessed, maxlen=self.maxlen, value=0)
return self.model.predict([x1_preprocessed, x2_preprocessed])
class CNNSigmoid(ModelBase):
def __init__(self, wv, maxlen=50, max_num_deptag=50):
super().__init__(wv, maxlen, max_num_deptag)
embedding_layer = self.wv.model.wv.get_embedding_layer()
sequence_input = Input(shape=(self.maxlen, ), dtype='int32')
mask = Masking(mask_value=-1)(sequence_input)
embedded_sequences = embedding_layer(mask)
x = (Conv1D(filters=128, kernel_size=3,
activation='relu'))(embedded_sequences)
# x = (AveragePooling1D(pool_size=3))(x)
x = (Dropout(0.2))(x)
x = (Conv1D(filters=32, kernel_size=17, activation='relu'))(x)
x = (Dropout(0.2))(x)
x = (GlobalMaxPooling1D())(x)
x = (Dense(16))(x)
x = (Dropout(0.2))(x)
preds = Dense(1, activation='sigmoid')(x)
self.model = Model(sequence_input, preds, name='CNNSigmoid')
self.compile_model()
def compile_model(self):
self.model.compile(loss='mean_squared_error',
optimizer='sgd',
metrics=['mae'])
def fit(self, x, y, batch_size, epochs, verbose, callbacks):
x_preprocessed = np.array(x)
x_preprocessed = sequence.pad_sequences(x_preprocessed,
maxlen=self.maxlen,
value=-1)
y_preprocessed = np.array(y)
self.model.fit(x_preprocessed,
y_preprocessed,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=callbacks)
def predict(self, x):
x_preprocessed = np.array(x)
x_preprocessed = sequence.pad_sequences(x_preprocessed,
maxlen=self.maxlen,
value=-1)
return self.model.predict(x_preprocessed)
def embedding_binary_classification():
docs = [
'Well done!', 'Good work', 'Great effort', 'nice work', 'Excellent!',
'Weak', 'Poor effort!', 'not good', 'poor work',
'Could have done better.'
]
# define class labels
labels = [1, 1, 1, 1, 1, 0, 0, 0, 0, 0]
vocab_size = 50
encoded_docs = [one_hot(d, vocab_size)
for d in docs] # one_hot编码到[1,n],不包括0
print(encoded_docs)
max_length = 4
padded_docs = pad_sequences(encoded_docs,
maxlen=max_length,
padding='post')
print(padded_docs)
input = Input(shape=(4, ))
x = Embedding(vocab_size, 8,
input_length=max_length)(input) # 这一步对应的参数量为50*8
x = Flatten()(x)
x = Dense(1, activation='sigmoid')(x)
model = Model(inputs=input, outputs=x)
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['acc'])
model.summary()
model.fit(padded_docs, labels, epochs=100, verbose=0)
loss, accuracy = model.evaluate(padded_docs, labels, verbose=0)
print('loss: {0},accuracy:{1}'.format(loss, accuracy))
loss_test, accuracy_test = model.evaluate(padded_docs, labels, verbose=0)
print('loss_test: {0},accuracy_test:{1}'.format(loss_test, accuracy_test))
test = one_hot('Weak', 50)
padded_test = pad_sequences([test], maxlen=max_length, padding='post')
print(model.predict(padded_test))
def train(model: Model,
x_train,
y_train,
x_test,
y_test,
batch_size=128,
epochs=20):
time = int(datetime.now().timestamp())
name = "{}_{}".format("conv", time)
chkp_path = "./checkpoints/{}".format(name)
os.makedirs(chkp_path, exist_ok=True)
model.compile(loss=K.categorical_crossentropy,
optimizer=Adam(),
metrics=[metrics.categorical_accuracy])
model.fit(
x_train,
y_train,
batch_size=batch_size,
epochs=epochs,
validation_data=(x_test, y_test),
callbacks=[
TensorBoard(log_dir='/tmp/tensorflow/{}'.format(name)),
ModelCheckpoint(os.path.join(
chkp_path,
"weights-improvement-{epoch:02d}-{val_categorical_accuracy:.2f}.hdf5"
),
monitor='val_categorical_accuracy',
verbose=1,
save_best_only=True,
mode='auto')
])
export_model(tf.train.Saver(), ['conv2d_1_input'], 'dense_2/Softmax', name)
score = model.evaluate(x_test, y_test)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
def train_auto_encoder(self,
train_x,
test_x,
input_dim,
out_model_file='encoder_cnn.h5',
monitor='val_loss',
patience=4):
'''
'''
early_stop = EarlyStopping(monitor=monitor, patience=patience)
checkpoint = ModelCheckpoint(out_model_file, monitor='val_loss', verbose=1,
save_best_only=True, mode='min')
input_data = Input(shape=(input_dim,), name='Input')
encoder = Dense(512, activation='relu', name='Encoder1')(input_data)
encoder = Dense(256, activation='relu', name='Encoder2')(encoder)
encoder = Dense(128, activation='relu', name='Encoder3')(encoder)
decoder = Dense(256, activation='relu', name='Decoder1')(encoder)
decoder = Dense(512, activation='relu', name='Decoder2')(decoder)
decoder = Dense(input_dim, activation='linear', name='Output')(decoder)
autoencoder = Model(input_data, decoder)
autoencoder.compile(optimizer='adam', loss='mse')
autoencoder.fit(train_x,
train_x,
epochs=200,
batch_size=512,
callbacks=[checkpoint, early_stop],
shuffle=True,
validation_data=(test_x, test_x),
verbose=1)
autoencoder.save(out_model_file)
return encoder
def sda_training(features, labels):
encoder_dims = [1600, 1024, 768]
stacked_encoder = []
int_labels = label_to_category(labels=labels, type='training')
X_train, X_test, y_train, y_test = train_test_split(features, labels, test_size=0.2, random_state=10)
y_train = to_categorical([int_labels[i] for i in y_train])
y_test = to_categorical([int_labels[i] for i in y_test])
for encoder_dim in encoder_dims:
input_dim = X_train.shape[1]
input_img = Input(shape=(input_dim,))
n_layer = noise.GaussianNoise(0.3)(input_img)
encode = Dense(encoder_dim, activation='sigmoid')(n_layer)
decode = Dense(input_dim, activation='sigmoid')(encode)
ae = Model(input_img, decode)
ae.compile(optimizer='adam', loss='mape')
ae.fit(X_train, X_train, epochs=10, batch_size=32, shuffle=True, validation_data=(X_test, X_test))
encoder = Model(input_img, encode)
X_train = encoder.predict(X_train)
X_test = encoder.predict(X_test)
stacked_encoder.append(encoder.layers[-1])
def test_lkrelu(self):
batch_size = 32
num_classes = 10
(x_train, y_train), (x_test, y_test) = load_cifar10()
inputs = Input(shape=x_train.shape[1:])
x = Conv2D(self.width, (3, 3))(inputs)
x = LKReLU()(x)
x = Flatten()(x)
x = Dense(num_classes, activation='softmax', name='fc1000')(x)
model = Model(inputs=inputs, outputs=x)
print(model.summary())
opt = keras.optimizers.sgd()
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=self.metrics)
log_dir = 'summaries/width-lkrelu-{}-cifar10-{}-{}'.format(
self.width, self.seed, datetime.datetime.now())
model.fit(
x_train,
y_train,
batch_size=batch_size,
epochs=self.epochs,
validation_data=(x_test, y_test),
shuffle=False,
callbacks=[
TensorBoard(log_dir=log_dir),
# ModelCheckpoint(
# 'checkpoints/width-lkrelu-cifar10-{epoch:02d}-{val_loss:.2f}.hdf5')
])
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
def vgg16_model(img_width, img_height, nb_epoch, nb_classes):
base_model = VGG16(weights='imagenet',
include_top=False,
input_shape=(img_width, img_height, 3))
# load dataset
(X_train, y_train), (X_test, y_test) = cifar10.load_data()
y_train = np_utils.to_categorical(y_train, nb_classes)
y_test = np_utils.to_categorical(y_test, nb_classes)
# Extract the last layer from third block of vgg16 model
last = base_model.get_layer('block5_pool').output
# Add classification layers on top of it
x = Flatten()(last)
x = BatchNormalization()(x)
x = Dense(256, activation='relu')(x)
x = Dropout(0.5)(x)
output = Dense(10, activation='softmax')(x)
model = Model(base_model.input, output)
# model.summary()
# model compile & fit
model.compile(loss='binary_crossentropy',
optimizer=optimizers.SGD(lr=1e-3, momentum=0.9),
metrics=['accuracy'])
model.fit(X_train,
y_train,
validation_data=(X_test, y_test),
epochs=nb_epoch,
batch_size=100,
verbose=1)
return model
class CTCModel:
def __init__(self,
inputs,
outputs,
greedy=True,
beam_width=100,
top_paths=1,
padding=-1,
charset=None):
"""
A override ou réécrire. C'est dans cette fonction qu'il faudra
affecter self.inputs et self.outputs avec les listes des layers
d'entrée et de sortie du réseau
"""
self.model_train = None
self.model_pred = None
self.model_eval = None
self.inputs = inputs
self.outputs = outputs
self.greedy = greedy
self.beam_width = beam_width
self.top_paths = top_paths
self.padding = padding
self.charset = charset
def compile(self, optimizer):
"""
A appeler une fois le modèle créé. Compile le modèle en ajoutant
la loss CTC
:param optimizer: L'optimizer a utiliser pendant l'apprentissage
"""
# Calcul du CTC
labels = Input(name='labels', shape=[None])
input_length = Input(name='input_length', shape=[1])
label_length = Input(name='label_length', shape=[1])
# Lambda layer for computing the loss function
loss_out = Lambda(self.ctc_loss_lambda_func,
output_shape=(1, ),
name='CTCloss')(self.outputs +
[labels, input_length, label_length])
# Lambda layer for the decoding function
out_decoded_dense = Lambda(self.ctc_complete_decoding_lambda_func,
output_shape=(None, None),
name='CTCdecode',
arguments={
'greedy': self.greedy,
'beam_width': self.beam_width,
'top_paths': self.top_paths
},
dtype="float32")(self.outputs +
[input_length])
#Lambda layer for computing the label error rate
out_analysis = Lambda(
self.ctc_complete_analysis_lambda_func,
output_shape=(None, ),
name='CTCanalysis',
arguments={
'greedy': self.greedy,
'beam_width': self.beam_width,
'top_paths': self.top_paths
},
dtype="float32")(self.outputs +
[labels, input_length, label_length])
# create Keras models
self.model_init = Model(inputs=self.inputs, outputs=self.outputs)
self.model_train = Model(inputs=self.inputs +
[labels, input_length, label_length],
outputs=loss_out)
self.model_pred = Model(inputs=self.inputs + [input_length],
outputs=out_decoded_dense)
self.model_eval = Model(inputs=self.inputs +
[labels, input_length, label_length],
outputs=out_analysis)
# Compile models
self.model_train.compile(loss={
'CTCloss': lambda yt, yp: yp
},
optimizer=optimizer)
self.model_pred.compile(loss={
'CTCdecode': lambda yt, yp: yp
},
optimizer=optimizer)
self.model_eval.compile(loss={
'CTCanalysis': lambda yt, yp: yp
},
optimizer=optimizer)
def get_model_train(self):
"""
:return: Modèle utilisé en interne pour l'entraînement
"""
return self.model_train
def get_model_pred(self):
"""
:return: Modèle utilisé en interne pour la prédiction
"""
return self.model_pred
def get_model_eval(self):
"""
:return: Model used to evaluate a data set
"""
return self.model_eval
def fit(self,
x=None,
y=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.0,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None):
"""
Permet de lire les données sur un device (CPU) et en
parallèle de s'entraîner sur un autre device (GPU)
Les données d'entrée doivent être de la forme :
[input_sequences, label_sequences, inputs_lengths, labels_length]
:param: Paramètres identiques à ceux de keras.engine.Model.fit()
:return: L'objet History correspondant à l'entrainement
"""
out = self.model_train.fit(x=x,
y=y,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
validation_split=validation_split,
validation_data=validation_data,
shuffle=shuffle,
class_weight=class_weight,
sample_weight=sample_weight,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps)
self.model_pred.set_weights(self.model_train.get_weights())
self.model_eval.set_weights(self.model_train.get_weights())
return out
def predict(self, x, batch_size=None, verbose=0):
""" CTC prediction
Inputs:
x = Input data as a 3D Tensor (batch_size, max_input_len, dim_features)
x_len = 1D array with the length of each data in batch_size
y = Input data as a 2D Tensor (batch_size, max_label_len)
y_len = 1D array with the length of each labeling
label_array = list of labels
pred = return predictions from the ctc (from model_pred)
eval = return an analysis of ctc prediction (from model_eval)
Outputs: a list containing:
out_pred = output of model_pred
out_eval = output of model_eval
"""
#model_out = self.model_pred.evaluate(x=x, y=np.zeros(x[0].shape[0]), batch_size=batch_size, verbose=verbose)
model_out = self.model_pred.predict(x,
batch_size=batch_size,
verbose=verbose)
return model_out
def predict2(self, x, batch_size=None, verbose=0, steps=None):
"""
The same function as in the Keras Model but with a different function predict_loop for dealing with variable length predictions
Generates output predictions for the input samples.
Computation is done in batches.
# Arguments
x: The input data, as a Numpy array
(or list of Numpy arrays if the model has multiple outputs).
batch_size: Integer. If unspecified, it will default to 32.
verbose: Verbosity mode, 0 or 1.
steps: Total number of steps (batches of samples)
before declaring the prediction round finished.
Ignored with the default value of `None`.
# Returns
Numpy array(s) of predictions.
# Raises
ValueError: In case of mismatch between the provided
input data and the model's expectations,
or in case a stateful model receives a number of samples
that is not a multiple of the batch size.
"""
#[x, x_len] = x
# Backwards compatibility.
if batch_size is None and steps is None:
batch_size = 32
if x is None and steps is None:
raise ValueError('If predicting from data tensors, '
'you should specify the `steps` '
'argument.')
# Validate user data.
x = _standardize_input_data(x,
self.model_pred._feed_input_names,
self.model_pred._feed_input_shapes,
check_batch_axis=False)
if self.model_pred.stateful:
if x[0].shape[0] > batch_size and x[0].shape[0] % batch_size != 0:
raise ValueError('In a stateful network, '
'you should only pass inputs with '
'a number of samples that can be '
'divided by the batch size. Found: ' +
str(x[0].shape[0]) + ' samples. '
'Batch size: ' + str(batch_size) + '.')
# Prepare inputs, delegate logic to `_predict_loop`.
if self.model_pred.uses_learning_phase and not isinstance(
K.learning_phase(), int):
#ins = [x, x_len] + [0.]
ins = x + [0.]
else:
#ins = [x, x_len]
ins = x
self.model_pred._make_predict_function()
f = self.model_pred.predict_function
out_pred = self._predict_loop(f,
ins,
batch_size=batch_size,
verbose=verbose,
steps=steps)
list_pred = []
for elmt in out_pred:
pred = []
for val in elmt:
if val != -1:
pred.append(val)
list_pred.append(pred)
return list_pred
@staticmethod
def ctc_loss_lambda_func(args):
"""
Function for computing the ctc loss (can be put in a Lambda layer)
:param args:
y_pred, labels, input_length, label_length
:return: CTC loss
"""
y_pred, labels, input_length, label_length = args
return K.ctc_batch_cost(labels, y_pred, input_length, label_length)
@staticmethod
def ctc_complete_decoding_lambda_func(args, **arguments):
"""
Complete CTC decoding using Keras (function K.ctc_decode)
:param args:
y_pred, input_length
:param arguments:
greedy, beam_width, top_paths
:return:
K.ctc_decode with dtype='float32'
"""
y_pred, input_length = args
my_params = arguments
assert (K.backend() == 'tensorflow')
return K.cast(K.ctc_decode(y_pred,
tf.squeeze(input_length),
greedy=my_params['greedy'],
beam_width=my_params['beam_width'],
top_paths=my_params['top_paths'])[0][0],
dtype='float32')
@staticmethod
def ctc_complete_analysis_lambda_func(args, **arguments):
"""
Complete CTC analysis using Keras and tensorflow
WARNING : tf is required
:param args:
y_pred, labels, input_length, label_len
:param arguments:
greedy, beam_width, top_paths
:return:
ler = label error rate
"""
y_pred, labels, input_length, label_len = args
my_params = arguments
assert (K.backend() == 'tensorflow')
batch = tf.log(tf.transpose(y_pred, perm=[1, 0, 2]) + 1e-8)
input_length = tf.to_int32(tf.squeeze(input_length))
greedy = my_params['greedy']
beam_width = my_params['beam_width']
top_paths = my_params['top_paths']
if greedy:
(decoded,
log_prob) = ctc.ctc_greedy_decoder(inputs=batch,
sequence_length=input_length)
else:
(decoded, log_prob) = ctc.ctc_beam_search_decoder(
inputs=batch,
sequence_length=input_length,
beam_width=beam_width,
top_paths=top_paths)
cast_decoded = tf.cast(decoded[0], tf.float32)
sparse_y = K.ctc_label_dense_to_sparse(
labels, tf.cast(tf.squeeze(label_len), tf.int32))
ed_tensor = tf_edit_distance(cast_decoded, sparse_y, norm=True)
ler_per_seq = Kreshape_To1D(ed_tensor)
return K.cast(ler_per_seq, dtype='float32')
def _predict_loop(self,
f,
ins,
max_len=20,
max_value=-1,
batch_size=32,
verbose=0,
steps=None):
"""Abstract method to loop over some data in batches.
# Arguments
f: Keras function returning a list of tensors.
ins: list of tensors to be fed to `f`.
batch_size: integer batch size.
verbose: verbosity mode.
steps: Total number of steps (batches of samples)
before declaring `_predict_loop` finished.
Ignored with the default value of `None`.
# Returns
Array of predictions (if the model has a single output)
or list of arrays of predictions
(if the model has multiple outputs).
"""
num_samples = self.model_pred._check_num_samples(
ins, batch_size, steps, 'steps')
if steps is not None:
# Step-based predictions.
# Since we do not know how many samples
# we will see, we cannot pre-allocate
# the returned Numpy arrays.
# Instead, we store one array per batch seen
# and concatenate them upon returning.
unconcatenated_outs = []
for step in range(steps):
batch_outs = f(ins)
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
if step == 0:
for batch_out in batch_outs:
unconcatenated_outs.append([])
for i, batch_out in enumerate(batch_outs):
unconcatenated_outs[i].append(batch_out)
if len(unconcatenated_outs) == 1:
return np.concatenate(unconcatenated_outs[0], axis=0)
return [
np.concatenate(unconcatenated_outs[i], axis=0)
for i in range(len(unconcatenated_outs))
]
else:
# Sample-based predictions.
outs = []
batches = _make_batches(num_samples, batch_size)
index_array = np.arange(num_samples)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_ids = index_array[batch_start:batch_end]
if ins and isinstance(ins[-1], float):
# Do not slice the training phase flag.
ins_batch = _slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
else:
ins_batch = _slice_arrays(ins, batch_ids)
batch_outs = f(ins_batch)
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
if batch_index == 0:
# Pre-allocate the results arrays.
for batch_out in batch_outs:
#shape = (num_samples, ) + batch_out.shape[1:] # WARNING 10)
shape = (num_samples, max_len)
outs.append(np.zeros(shape, dtype=batch_out.dtype))
#outs.append(np.zeros(shape, dtype=batch_out.dtype))#batch_out.dtype))# WARNING CHANGE FROM THE MAIN CODE
for i, batch_out in enumerate(batch_outs):
#outs[i][batch_start:batch_end] = batch_out # WARNING
outs[i][batch_start:batch_end] = sequence.pad_sequences(
batch_out,
value=float(max_value),
maxlen=max_len,
dtype=batch_out.dtype,
padding="post")
if len(outs) == 1:
return outs[0]
return outs
def save_model(self, path_dir, charset=None):
""" Save a model in path_dir
save model_train, model_pred and model_eval in json
save inputs and outputs in json
save model CTC parameters in a pickle """
model_json = self.model_train.to_json()
with open(path_dir + "/model_train.json", "w") as json_file:
json_file.write(model_json)
model_json = self.model_pred.to_json()
with open(path_dir + "/model_pred.json", "w") as json_file:
json_file.write(model_json)
model_json = self.model_eval.to_json()
with open(path_dir + "/model_eval.json", "w") as json_file:
json_file.write(model_json)
model_json = self.model_init.to_json()
with open(path_dir + "/model_init.json", "w") as json_file:
json_file.write(model_json)
param = {
'greedy': self.greedy,
'beam_width': self.beam_width,
'top_paths': self.top_paths,
'charset': self.charset
}
output = open(path_dir + "/model_param.pkl", 'wb')
p = pickle.Pickler(output)
p.dump(param)
output.close()
def load_model(self, path_dir, optimizer, initial_epoch=None):
""" Load a model in path_dir
load model_train, model_pred and model_eval from json
load inputs and outputs from json
load model CTC parameters from a pickle """
json_file = open(path_dir + '/model_train.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
self.model_train = model_from_json(loaded_model_json)
json_file = open(path_dir + '/model_pred.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
self.model_pred = model_from_json(loaded_model_json,
custom_objects={"tf": tf})
json_file = open(path_dir + '/model_eval.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
self.model_eval = model_from_json(loaded_model_json,
custom_objects={
"tf": tf,
"ctc": ctc,
"tf_edit_distance":
tf_edit_distance,
"Kreshape_To1D": Kreshape_To1D
})
json_file = open(path_dir + '/model_init.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
self.model_init = model_from_json(loaded_model_json,
custom_objects={"tf": tf})
self.inputs = self.model_init.inputs
self.outputs = self.model_init.outputs
input = open(path_dir + "/model_param.pkl", 'rb')
p = pickle.Unpickler(input)
param = p.load()
input.close()
self.greedy = param['greedy'] if 'greedy' in param.keys(
) else self.greedy
self.beam_width = param['beam_width'] if 'beam_width' in param.keys(
) else self.beam_width
self.top_paths = param['top_paths'] if 'top_paths' in param.keys(
) else self.top_paths
self.charset = param['charset'] if 'charset' in param.keys(
) else self.charset
self.compile(optimizer)
if initial_epoch:
file_weight = path_dir + 'weights.' + "%02d" % (
initial_epoch) + '.hdf5'
print(file_weight)
if os.path.exists(file_weight):
self.model_train.load_weights(file_weight)
self.model_pred.set_weights(self.model_train.get_weights())
self.model_eval.set_weights(self.model_train.get_weights())
else:
print("Weights for epoch ", initial_epoch,
" can not be loaded.")
else:
print("Training will be start at the beginning.")
class AttentionSumReader(object):
def __init__(self,
word_dict,
embedding_matrix,
d_len,
q_len,
embedding_dim,
hidden_size,
num_layers,
weight_path,
use_lstm=False):
"""
初始化模型
b ... batch_size
t ... d_len
f ... hidden_size*2
i ... candidate_len
"""
self.weight_path = weight_path
self.word_dict = word_dict
self.vocab_size = len(embedding_matrix)
self.d_len = d_len
self.q_len = q_len
self.A_len = 10
logging.info("Embedding matrix shape:%d x %d" %
(len(embedding_matrix), embedding_dim))
self.rnn_cell = LSTM if use_lstm else GRU
self.cell_name = "LSTM" if use_lstm else "GRU"
# 模型的输入
q_input = Input(batch_shape=(None, self.q_len),
dtype="int32",
name="q_input")
d_input = Input(batch_shape=(
None,
self.d_len,
),
dtype="int32",
name="d_input")
context_mask = Input(batch_shape=(None, self.d_len),
dtype="float32",
name="context_mask")
candidates_bi = Input(batch_shape=(None, self.A_len),
dtype="int32",
name="candidates_bi")
# 问题的编码模型
# output shape: (None, max_q_length, embedding_dim)
q_encode = Embedding(
input_dim=self.vocab_size,
output_dim=embedding_dim,
weights=[embedding_matrix],
mask_zero=True,
)(q_input)
for i in range(1, num_layers):
q_encode = Bidirectional(self.rnn_cell(
units=hidden_size,
name="{}-{}-{}".format("q-encoder", self.cell_name, i),
kernel_initializer="glorot_uniform",
recurrent_initializer="glorot_uniform",
bias_initializer='zeros',
return_sequences=True),
merge_mode="concat",
dtype="float32")(q_encode)
# q_encoder output shape: (None, hidden_size * 2)
# TODO: 用最后一步的隐层状态表示q
q_encode = Bidirectional(self.rnn_cell(
units=hidden_size,
name="{}-{}-{}".format("q-encoder", self.cell_name, num_layers),
kernel_initializer="glorot_uniform",
recurrent_initializer="glorot_uniform",
bias_initializer='zeros',
return_sequences=False),
merge_mode="concat",
dtype="float32")(q_encode)
# 上下文文档的编码模型
# output shape: (None, max_d_length, embedding_dim)
d_encode = Embedding(input_dim=self.vocab_size,
output_dim=embedding_dim,
weights=[embedding_matrix],
mask_zero=True,
input_length=self.d_len)(d_input)
# d_encoder output shape: (None, max_d_length, hidden_size * 2)
for i in range(1, num_layers + 1):
d_encode = Bidirectional(self.rnn_cell(
units=hidden_size,
name="{}-{}-{}".format("d-encoder", self.cell_name, i),
kernel_initializer="glorot_uniform",
recurrent_initializer="glorot_uniform",
bias_initializer='zeros',
return_sequences=True),
merge_mode="concat",
dtype="float32")(d_encode)
# noinspection PyUnusedLocal
def my_dot(x):
"""注意力点乘函数，原始版本"""
c = [
tf.reduce_sum(tf.multiply(x[0][:, inx, :], x[1]),
-1,
keep_dims=True) for inx in range(self.d_len)
]
return tf.concat(c, -1)
def my_dot_v2(x):
"""注意力点乘函数，快速版本"""
d_btf, q_bf = x
res = K.batch_dot(tf.expand_dims(q_bf, -1), d_btf, (1, 2))
return K.reshape(res, [-1, self.d_len])
mem_attention_pre_soft_bt = Lambda(
my_dot_v2, name="attention")([d_encode, q_encode])
mem_attention_pre_soft_masked_bt = Multiply(name="mask")(
[mem_attention_pre_soft_bt, context_mask])
mem_attention_bt = Activation(
activation="softmax",
name="softmax")(mem_attention_pre_soft_masked_bt)
# 注意力求和，attention-sum过程
# TODO: Get rid of sentence-by-sentence processing?
# TODO: Rewrite into matrix notation instead of scans?
def sum_prob_of_word(word_ix, sentence_ixs, sentence_attention_probs):
word_ixs_in_sentence = tf.where(tf.equal(sentence_ixs, word_ix))
return tf.reduce_sum(
tf.gather(sentence_attention_probs, word_ixs_in_sentence))
# noinspection PyUnusedLocal
def sum_probs_single_sentence(prev, cur):
candidate_indices_i, sentence_ixs_t, sentence_attention_probs_t = cur
result = tf.scan(fn=lambda previous, x: sum_prob_of_word(
x, sentence_ixs_t, sentence_attention_probs_t),
elems=[candidate_indices_i],
initializer=K.constant(0., dtype="float32"),
swap_memory=True)
return result
def sum_probs_batch(candidate_indices_bi, sentence_ixs_bt,
sentence_attention_probs_bt):
result = tf.scan(fn=sum_probs_single_sentence,
elems=[
candidate_indices_bi, sentence_ixs_bt,
sentence_attention_probs_bt
],
initializer=K.variable([0] * self.A_len,
dtype="float32"),
swap_memory=True)
return result
# output shape: (None, i) i = max_candidate_length = 10
y_hat = Lambda(lambda x: sum_probs_batch(x[0], x[1], x[2]),
name="attention_sum")(
[candidates_bi, d_input, mem_attention_bt])
self.model = Model(
inputs=[q_input, d_input, context_mask, candidates_bi],
outputs=y_hat)
plot_model(self.model,
to_file=__file__ + ".png",
show_shapes=True,
show_layer_names=True)
self.model.summary()
# noinspection PyUnusedLocal
def train(self, train_data, valid_data, batch_size, epochs, opt_name, lr,
grad_clip):
"""
模型训练。
"""
def save_weight_on_epoch_end(epoch, e_logs):
filename = "{}weight-epoch{}-{}-{}.h5".format(
self.weight_path,
time.strftime("%Y-%m-%d-(%H-%M)", time.localtime()), epoch,
e_logs['val_acc'])
self.model.save_weights(filepath=filename)
checkpointer = LambdaCallback(on_epoch_end=save_weight_on_epoch_end)
# tensorboard = TensorBoard(log_dir="./logs", histogram_freq=1, write_images=True)
earlystopping = EarlyStopping(monitor="val_loss",
patience=3,
verbose=1)
# 对输入进行预处理
questions_ok, documents_ok, context_mask, candidates_ok, y_true = self.preprocess_input_sequences(
train_data)
v_questions, v_documents, v_context_mask, v_candidates, v_y_true = self.preprocess_input_sequences(
valid_data)
if opt_name == "SGD":
optimizer = SGD(lr=lr, decay=1e-6, clipvalue=grad_clip)
elif opt_name == "ADAM":
optimizer = Adam(lr=lr, clipvalue=grad_clip)
else:
raise NotImplementedError("Other Optimizer Not Implemented.-_-||")
self.model.compile(optimizer=optimizer,
loss="categorical_crossentropy",
metrics=["accuracy"])
# 载入之前训练的权重
self.load_weight()
data = {
"q_input": questions_ok,
"d_input": documents_ok,
"context_mask": context_mask,
"candidates_bi": candidates_ok
}
v_data = {
"q_input": v_questions,
"d_input": v_documents,
"context_mask": v_context_mask,
"candidates_bi": v_candidates
}
logs = self.model.fit(x=data,
y=y_true,
batch_size=batch_size,
epochs=epochs,
validation_data=(v_data, v_y_true),
callbacks=[checkpointer, earlystopping])
def test(self, test_data, batch_size):
# 对输入进行预处理
questions_ok, documents_ok, context_mask, candidates_ok, y_true = self.preprocess_input_sequences(
test_data)
data = {
"q_input": questions_ok,
"d_input": documents_ok,
"context_mask": context_mask,
"candidates_bi": candidates_ok
}
y_pred = self.model.predict(x=data, batch_size=batch_size)
acc_num = np.count_nonzero(
np.equal(np.argmax(y_pred, axis=-1), np.zeros(len(y_pred))))
test_acc = acc_num / len(y_pred)
logging.info("Test accuracy is {}".format(test_acc))
return acc_num, test_acc
def load_weight(self, weight_path=None):
weight_file = self.weight_path if not weight_path else weight_path
if os.path.exists(weight_file + "weight.h5"):
logging.info("Load pre-trained weights:{}".format(weight_file +
"weight.h5"))
self.model.load_weights(filepath=weight_file + "weight.h5",
by_name=True)
@staticmethod
def union_shuffle(data):
d, q, a, A = data
c = list(zip(d, q, a, A))
random.shuffle(c)
return zip(*c)
def preprocess_input_sequences(self, data, shuffle=True):
"""
预处理输入：
shuffle
PAD/TRUNC到固定长度的序列
y_true是长度为self.A_len的向量，index=0为正确答案，one-hot编码
"""
documents, questions, answer, candidates = self.union_shuffle(
data) if shuffle else data
d_lens = [len(i) for i in documents]
questions_ok = pad_sequences(questions,
maxlen=self.q_len,
dtype="int32",
padding="post",
truncating="post")
documents_ok = pad_sequences(documents,
maxlen=self.d_len,
dtype="int32",
padding="post",
truncating="post")
context_mask = K.eval(
tf.sequence_mask(d_lens, self.d_len, dtype=tf.float32))
candidates_ok = pad_sequences(candidates,
maxlen=self.A_len,
dtype="int32",
padding="post",
truncating="post")
y_true = np.zeros_like(candidates_ok)
y_true[:, 0] = 1
return questions_ok, documents_ok, context_mask, candidates_ok, y_true
def fit_cnn():
X_train = train_embed
y_train = np.array(train_labels)
X_test = test_embed
y_test = test_labels
X_train = X_train.reshape(
(X_train.shape[0], X_train.shape[1], X_train.shape[2], 1))
X_test = X_test.reshape(
(X_test.shape[0], X_test.shape[1], X_test.shape[2], 1))
y_train, y_test = convert_one_hot(y_train, y_test)
sequence_length = train_embed.shape[1] # 60
embedding_dim = train_embed.shape[2]
filter_sizes = [2, 3, 4, 5, 6, 7, 8]
num_filters = 64
drop = 0.6
input_shape = train_embed[0].shape
epochs = 300
batch_size = 32
inputs = Input(shape=(sequence_length, embedding_dim, 1), dtype='float32')
#batch_norm = BatchNormalization(input_shape = input_shape)(inputs)
conv_0 = Conv2D(num_filters,
kernel_size=(filter_sizes[0], embedding_dim),
padding='valid',
kernel_initializer='normal',
activation='relu',
input_shape=input_shape)(inputs)
conv_1 = Conv2D(num_filters,
kernel_size=(filter_sizes[1], embedding_dim),
padding='valid',
kernel_initializer='normal',
activation='relu',
input_shape=input_shape)(inputs)
conv_2 = Conv2D(num_filters,
kernel_size=(filter_sizes[2], embedding_dim),
padding='valid',
kernel_initializer='normal',
activation='relu',
input_shape=input_shape)(inputs)
conv_3 = Conv2D(num_filters,
kernel_size=(filter_sizes[3], embedding_dim),
padding='valid',
kernel_initializer='normal',
activation='relu',
input_shape=input_shape)(inputs)
conv_4 = Conv2D(num_filters,
kernel_size=(filter_sizes[4], embedding_dim),
padding='valid',
kernel_initializer='normal',
activation='relu',
input_shape=input_shape)(inputs)
conv_5 = Conv2D(num_filters,
kernel_size=(filter_sizes[5], embedding_dim),
padding='valid',
kernel_initializer='normal',
activation='relu',
input_shape=input_shape)(inputs)
conv_6 = Conv2D(num_filters,
kernel_size=(filter_sizes[6], embedding_dim),
padding='valid',
kernel_initializer='normal',
activation='relu',
input_shape=input_shape)(inputs)
maxpool_0 = MaxPool2D(pool_size=(sequence_length - filter_sizes[0] + 1, 1),
strides=(1, 1),
padding='valid')(conv_0)
maxpool_1 = MaxPool2D(pool_size=(sequence_length - filter_sizes[1] + 1, 1),
strides=(1, 1),
padding='valid')(conv_1)
maxpool_2 = MaxPool2D(pool_size=(sequence_length - filter_sizes[2] + 1, 1),
strides=(1, 1),
padding='valid')(conv_2)
maxpool_3 = MaxPool2D(pool_size=(sequence_length - filter_sizes[3] + 1, 1),
strides=(1, 1),
padding='valid')(conv_3)
maxpool_4 = MaxPool2D(pool_size=(sequence_length - filter_sizes[4] + 1, 1),
strides=(1, 1),
padding='valid')(conv_4)
maxpool_5 = MaxPool2D(pool_size=(sequence_length - filter_sizes[5] + 1, 1),
strides=(1, 1),
padding='valid')(conv_5)
maxpool_6 = MaxPool2D(pool_size=(sequence_length - filter_sizes[6] + 1, 1),
strides=(1, 1),
padding='valid')(conv_6)
concatenated_tensor = Concatenate(axis=1)(
[maxpool_0, maxpool_1, maxpool_2, maxpool_3])
flatten = Flatten()(concatenated_tensor)
dropout = Dropout(drop)(flatten)
output = Dense(units=3, activation='softmax')(dropout)
# this creates a model that includes
model = Model(inputs=inputs, outputs=output)
adam = Adam(lr=1e-4, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
model.compile(optimizer=adam,
loss='categorical_crossentropy',
metrics=['accuracy'])
print("Traning Model...")
model.fit(X_train,
y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(X_test, y_test))
model.save("model/cnn.h5")
return model
def start_training(self):
# Vectorize the data.
input_texts = []
target_texts = []
input_characters = set()
target_characters = set()
lines = open(self.data_path, encoding='UTF-8').read().split('\n')
for line in lines[:min(self.num_samples, len(lines) - 1)]:
input_text, target_text = line.split('\t')
# We use "tab" as the "start sequence" character
# for the targets, and "\n" as "end sequence" character.
target_text = '\t' + target_text + '\n'
input_texts.append(input_text)
target_texts.append(target_text)
for char in input_text:
if char not in input_characters:
input_characters.add(char)
for char in target_text:
if char not in target_characters:
target_characters.add(char)
input_characters = sorted(list(input_characters))
target_characters = sorted(list(target_characters))
num_encoder_tokens = len(input_characters)
num_decoder_tokens = len(target_characters)
max_encoder_seq_length = max([len(txt) for txt in input_texts])
max_decoder_seq_length = max([len(txt) for txt in target_texts])
print('Number of samples:', len(input_texts))
print('Number of unique input tokens:', num_encoder_tokens)
print('Number of unique output tokens:', num_decoder_tokens)
print('Max sequence length for inputs:', max_encoder_seq_length)
print('Max sequence length for outputs:', max_decoder_seq_length)
input_token_index = dict([(char, i)
for i, char in enumerate(input_characters)])
target_token_index = dict([
(char, i) for i, char in enumerate(target_characters)
])
encoder_input_data = np.zeros(
(len(input_texts), max_encoder_seq_length, num_encoder_tokens),
dtype='float32')
decoder_input_data = np.zeros(
(len(input_texts), max_decoder_seq_length, num_decoder_tokens),
dtype='float32')
decoder_target_data = np.zeros(
(len(input_texts), max_decoder_seq_length, num_decoder_tokens),
dtype='float32')
for i, (input_text,
target_text) in enumerate(zip(input_texts, target_texts)):
for t, char in enumerate(input_text):
encoder_input_data[i, t, input_token_index[char]] = 1.
for t, char in enumerate(target_text):
# decoder_target_data is ahead of decoder_target_data by one timestep
decoder_input_data[i, t, target_token_index[char]] = 1.
if t > 0:
# decoder_target_data will be ahead by one timestep
# and will not include the start character.
decoder_target_data[i, t - 1,
target_token_index[char]] = 1.
# Define an input sequence and process it.
encoder_inputs = Input(shape=(None, num_encoder_tokens))
encoder = LSTM(self.latent_dim, return_state=True)
encoder_outputs, state_h, state_c = encoder(encoder_inputs)
# We discard `encoder_outputs` and only keep the states.
encoder_states = [state_h, state_c]
# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None, num_decoder_tokens))
# We set up our decoder to return full output sequences,
# and to return internal states as well. We don't use the
# return states in the training model, but we will use them in inference.
decoder_lstm = LSTM(self.latent_dim,
return_sequences=True,
return_state=True)
decoder_outputs, _, _ = decoder_lstm(decoder_inputs,
initial_state=encoder_states)
decoder_dense = Dense(num_decoder_tokens, activation='softmax')
decoder_outputs = decoder_dense(decoder_outputs)
# Define the model that will turn
# `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
#
# Run training
model.compile(optimizer='rmsprop', loss='categorical_crossentropy')
model.summary()
tbCallBack = callbacks.TensorBoard(log_dir='./data/Graph',
histogram_freq=0,
write_graph=True,
write_images=True)
model.fit([encoder_input_data, decoder_input_data],
decoder_target_data,
batch_size=self.batch_size,
epochs=self.epochs,
validation_split=0.2,
verbose=1,
callbacks=[tbCallBack])
#
# # Save model
model.save('./data/s2s2.h5')
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.20, random_state=0)
y_train = keras.utils.to_categorical(y_train, 4)
y_test = keras.utils.to_categorical(y_test, 4)
resnet = VGGFace(model='resnet50',input_shape=(224, 224, 3))
layer_name = resnet.layers[-2].name
out = resnet.get_layer(layer_name).output
out = Dense(4,activation='softmax')(out)
resnet_4 = Model(resnet.input, out)
for layer in resnet_4.layers[:-1]:
layer.trainable = False
resnet_4.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
print (resnet_4.summary())
resnet_4.fit(x_train, y_train,batch_size=16,epochs=5,validation_data=(x_test, y_test),shuffle=True)
scores = resnet_4.evaluate(x_test, y_test, verbose=1)
print('Test accuracy:', scores[1])
out1, out2, out3 = c3d(inputs)
model_c3d = Model(inputs=inputs, outputs=[out1, out2, out3])
# model_c3d.summary()
model_c3d.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
# %% Network training
print('Not using data augmentation.')
logging.debug("Running training...")
hist = model_c3d.fit(X_train_c3d, [y_train_frame, y_train_frame, y_train_frame],
batch_size=32, epochs=100, verbose=1,
shuffle=True, callbacks=[mc])
# %% Prediction
logging.debug("Running test...")
model_c3d.load_weights('/home/kdh/Desktop/c3da/checkpoint/c3da_weights.h5')
proba_c3d = model_c3d.predict(X_test_c3d, batch_size=4, verbose=1)
y_pred = [np.argmax(prob) for prob in proba_c3d[2]]
y_true = [np.argmax(true) for true in y_test_frame]
count = 0
for i in range(len(y_pred)):
if y_test[i] == y_pred[i]:
count += 1
for layer in range(num_layers - 1):
final_model.layers[layer].trainable = False
final_model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# get training data
rootdir_train = '/training'
filenames = load_filenames('all_train', rootdir_train)
#print(filenames)
X, Y = compile_images(filenames, (250, 250, 3))
# fit updated model on senators
final_model.fit(X, Y, batch_size=8, epochs=10, verbose=1, validation_split=0.2)
# test
rootdir_test = '/test'
filenames_test = load_filenames('all_test', rootdir_test)
X_test, Y_test = compile_images(filenames_test, (250, 250, 3))
print("Test results: ", final_model.test_on_batch(X_test, Y_test))
predictions = final_model.predict(X_test)
for i, f in enumerate(filnames_test):
print(f, " with truth value: ", Y_test[i], predictions[i])
final_model.save("/output/transfer_modle120.h5")
print('model compile time: {}'.format(time.time() - start_time))
print('')
#############################################################################
# TRAINING
batch_size = 256
nb_epoch = 50
# Model saving callback
checkpointer = ModelCheckpoint(filepath=WEIGHTS_PRESENCE_FILEPATH,
verbose=1,
save_best_only=True)
# Early stopping
early_stopping = EarlyStopping(monitor='val_loss', patience=5)
history = model_presence_a.fit(data_images_train,
data_presence_train,
batch_size=batch_size,
nb_epoch=nb_epoch,
verbose=2,
validation_data=(data_images_val,
data_presence_val),
shuffle=True,
callbacks=[checkpointer, early_stopping])
with open(HISTORY_PRESENCE_FILEPATH, 'w') as f_out:
json.dump(history.history, f_out)
print('model compile time: {}'.format(time.time() - start_time))
print('')
#############################################################################
# TRAINING
batch_size = 48
nb_epoch = 100
# Model saving callback
checkpointer = ModelCheckpoint(filepath=WEIGHTS_SEGMENT_FILEPATH,
verbose=1,
save_best_only=True)
# Early stopping
early_stopping = EarlyStopping(monitor='val_loss', patience=5)
history = model_segment.fit(data_images_train,
data_masks_train,
batch_size=batch_size,
nb_epoch=nb_epoch,
verbose=2,
shuffle=True,
validation_data=(data_images_val,
data_masks_val),
callbacks=[checkpointer, early_stopping])
with open(HISTORY_SEGMENT_FILEPATH, 'w') as f_out:
json.dump(history.history, f_out)
def main(i, RNN_TYPE):
# TODO: change to GRU, Recurrent, and SimpleRNN
# print(i, RNN_TYPE)
RNN = recurrent.LSTM
if RNN_TYPE == "gru":
# print("Starting a GRU: ")
RNN = recurrent.GRU
file_name = "history_gru_" + str(i) + ".csv"
elif RNN_TYPE == "recurrent":
# print("Starting a Recurrent Unit: ")
RNN = recurrent.Recurrent
file_name = "history_recurrent_" + str(i) + ".csv"
elif RNN_TYPE == "simplernn":
# print("Starting a SimpleRNN: ")
RNN = recurrent.SimpleRNN
file_name = "history_simple_" + str(i) + ".csv"
else:
# print("Starting an LSTM")
file_name = "history_lstm_" + str(i) + ".csv"
EMBED_HIDDEN_SIZE = 100
SENT_HIDDEN_SIZE = 100
QUERY_HIDDEN_SIZE = 100
BATCH_SIZE = 50
EPOCHS = 100
# print('RNN / Embed / Sent / Query = {}, {}, {}, {}'.format(RNN,
# EMBED_HIDDEN_SIZE,
# SENT_HIDDEN_SIZE,
# QUERY_HIDDEN_SIZE))
# print(os.getcwd())
file_list = (os.getcwd())
base_file = os.getcwd() + "/tasks_1-20_v1-2/en/"
file_list = (os.listdir(base_file))
# print(file_list)
test_file = ""
train_file = ""
for file in file_list:
if file.startswith("qa" + str(i) + "_"):
if file.endswith("_test.txt"):
test_file = file
# print(test_file)
elif file.endswith("_train.txt"):
train_file = file
# print(train_file)
print(train_file)
print(test_file)
f_train = open(base_file + train_file)
f_test = open(base_file + test_file)
# try:
# path = get_file('babi-tasks-v1-2.tar.gz',
# origin='https://s3.amazonaws.com/text-datasets/babi_tasks_1-20_v1-2.tar.gz')
# except:
# print('Error downloading dataset, please download it manually:\n'
# '$ wget http://www.thespermwhale.com/jaseweston/babi/tasks_1-20_v1-2.tar.gz\n'
# '$ mv tasks_1-20_v1-2.tar.gz ~/.keras/datasets/babi-tasks-v1-2.tar.gz')
# raise
# tar = tarfile.open(path)
# # Default QA1 with 1000 samples
# # challenge = 'tasks_1-20_v1-2/en/qa1_single-supporting-fact_{}.txt'
# # QA1 with 10,000 samples
# # challenge = 'tasks_1-20_v1-2/en-10k/qa1_single-supporting-fact_{}.txt'
# # QA2 with 1000 samples
# challenge = 'tasks_1-20_v1-2/en/qa2_two-supporting-facts_{}.txt'
# # QA2 with 10,000 samples
# # challenge = 'tasks_1-20_v1-2/en-10k/qa2_two-supporting-facts_{}.txt'
# train = get_stories(tar.extractfile(challenge.format('train')))
# test = get_stories(tar.extractfile(challenge.format('test')))
# print("training stories:")
train = get_stories(f_train)
# print(len(train))
# print("testing stories:")
test = get_stories(f_test)
# print(len(test))
vocab = set()
for story, q, answer in train + test:
vocab |= set(story + q + [answer])
vocab = sorted(vocab)
# check_existence(vocab)
# get_word_vectors_from_pretr_embeddings(train, test, vocab)
# Reserve 0 for masking via pad_sequences
vocab_size = len(vocab) + 1
print("Vocabulary size: ", vocab_size)
word_idx = dict((c, i + 1) for i, c in enumerate(vocab))
story_maxlen = max(map(len, (x for x, _, _ in train + test)))
query_maxlen = max(map(len, (x for _, x, _ in train + test)))
x, xq, y = vectorize_stories(train, word_idx, story_maxlen, query_maxlen)
tx, txq, ty = vectorize_stories(test, word_idx,story_maxlen, query_maxlen)
print('vocab = {}'.format(vocab))
print('x.shape = {}'.format(x.shape))
print('xq.shape = {}'.format(xq.shape))
print('y.shape = {}'.format(y.shape))
print('story_maxlen, query_maxlen = {}, {}'.format(story_maxlen, query_maxlen))
print('Build model...')
pre_trained_emb_weights = get_pre_trained_emb(vocab)
sentence = layers.Input(shape=(story_maxlen,), dtype='float32')
encoded_sentence = layers.Embedding(vocab_size, EMBED_HIDDEN_SIZE)(sentence)
encoded_sentence = layers.Dropout(0.3)(encoded_sentence)
question = layers.Input(shape=(query_maxlen,), dtype='float32')
encoded_question = layers.Embedding(vocab_size, EMBED_HIDDEN_SIZE)(question)
encoded_question = layers.Dropout(0.3)(encoded_question)
encoded_question = RNN(EMBED_HIDDEN_SIZE)(encoded_question)
encoded_question = layers.RepeatVector(story_maxlen)(encoded_question)
merged = layers.add([encoded_sentence, encoded_question])
merged = RNN(EMBED_HIDDEN_SIZE)(merged)
merged = layers.Dropout(0.3)(merged)
preds = layers.Dense(vocab_size, activation='softmax')(merged)
model = Model([sentence, question], preds)
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
print('Training')
history = model.fit([x, xq], y, batch_size=BATCH_SIZE, epochs=EPOCHS, validation_split=0.05)
pandas.DataFrame(history.history).to_csv("__pre_"+file_name)
loss, acc = model.evaluate([tx, txq], ty,
batch_size=BATCH_SIZE)
pandas.DataFrame([str(loss)+"_"+ str(acc)]).to_csv("__test_"+RNN_TYPE+"_"+str(i)+".csv")
print('Test loss / test accuracy = {:.4f} / {:.4f}'.format(loss, acc))
def main(_):
# The data, shuffled and split between train and test sets:
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
# Convert class vectors to binary class matrices.
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
inputs = Input(shape=(32, 32, 3))
out1, out2 = simple_cnn(inputs)
model = Model(inputs=inputs, outputs=[out1, out2])
# model.summary()
opt = keras.optimizers.rmsprop(lr=0.0001, decay=1e-6)
model.compile(
optimizer=opt,
loss=['categorical_crossentropy', 'categorical_crossentropy'],
metrics=['accuracy'])
if not data_augmentation:
print('Not using data augmentation.')
model.fit(x_train, [y_train, y_train],
batch_size=batch_size,
epochs=epochs,
validation_data=(x_test, [y_test, y_test]),
shuffle=True)
else:
print('Using real-time data augmentation.')
def train_generator(x, y, batch_size):
train_datagen = ImageDataGenerator(
width_shift_range=
0.1, # randomly shift images horizontally (fraction of total width)
height_shift_range=
0.1, # randomly shift images vertically (fraction of total height)
horizontal_flip=True) # randomly flip images
generator = train_datagen.flow(x, y, batch_size=batch_size)
while 1:
x_batch, y_batch = generator.next()
yield (x_batch, [y_batch, y_batch])
# Fit the model on the batches generated by datagen.flow().
model.fit_generator(generator=train_generator(x_train, y_train,
batch_size),
steps_per_epoch=int(y_train.shape[0] / batch_size),
epochs=epochs,
validation_data=(x_test, [y_test, y_test]),
callbacks=[])
# Save model and weights
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
model_path = os.path.join(save_dir, model_name)
model.save(model_path)
print('Saved trained model at %s ' % model_path)
# Score trained model.
scores = model.evaluate(x_test, [y_test, y_test], verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])
# When we see that the validation loss stopped improving, so that we can start
# learning faster and then get more precision
rp_callback = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss',
factor=0.5,
patience=2,
min_lr=0.000001,
verbose=1)
callbacks.append(rp_callback)
# In[ ]:
model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
model.fit(x=train_dataset, epochs=50, steps_per_epoch=len(train_gen), validation_data=valid_dataset, validation_steps=len(valid_gen), callbacks=callbacks)
# In[ ]:
# Save the model
now = datetime.now().strftime('%b%d_%H-%M-%S')
print(str(now))
model_name = os.path.join(models_dir, str(now))
classification_name = str(prediction_dir) + '/' +str(now)
model.save(model_name)
def train_label_none_label_classification(label_folder, non_label_folder, model_file=None):
c = Config()
# Build or load model
if model_file is None:
# create model
img_input = Input(shape=(28, 28, 3))
# prediction = model_cnn_2_layer.nn_classify_label_non_label(img_input)
# prediction = model_cnn_3_layer.nn_classify_label_non_label(img_input)
prediction = nn_cnn_3_layer.nn_classify_label_non_label(img_input)
model = Model(inputs=img_input, outputs=prediction)
model.compile(loss='categorical_crossentropy', optimizer=RMSprop(), metrics=['accuracy'])
else:
model = load_model(model_file)
model.summary()
# Load and normalize data
x_train, y_train, x_test, y_test = load_train_validation_data(label_folder, non_label_folder)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train[:, :, :, 0] -= c.img_channel_mean[0]
x_train[:, :, :, 1] -= c.img_channel_mean[1]
x_train[:, :, :, 2] -= c.img_channel_mean[2]
x_test[:, :, :, 0] -= c.img_channel_mean[0]
x_test[:, :, :, 1] -= c.img_channel_mean[1]
x_test[:, :, :, 2] -= c.img_channel_mean[2]
x_train /= 255
x_test /= 255
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
# x_train.reshape(x_train.shape[0], 28, 28, 3)
# x_test.reshape(x_test.shape[0], 28, 28, 3)
# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, 2)
y_test = keras.utils.to_categorical(y_test, 2)
# Checkpointing is to save the network weights only when there is an improvement in classification accuracy
# on the validation dataset (monitor=’val_acc’ and mode=’max’).
file_path = "weights-improvement-{epoch:04d}-{val_acc:.4f}.hdf5"
checkpoint = ModelCheckpoint(file_path, monitor='val_acc', verbose=1, save_best_only=True, mode='max')
callbacks_list = [checkpoint]
model.fit(x_train, y_train,
batch_size=128,
epochs=100,
verbose=1,
callbacks=callbacks_list,
validation_data=(x_test, y_test)
)
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
model.save('final_model.h5')
embedded_sequences = embedding_layer(sequence_input)
cnns = []
for filter_length in filter_lengths:
x = Conv1D(nb_filter=nb_filter,
filter_length=filter_length,
border_mode='valid',
activation='relu',
W_constraint=maxnorm(3),
W_regularizer=l2(0.0001),
subsample_length=1)(embedded_sequences)
x = MaxPooling1D(pool_length=MAX_SEQUENCE_LENGTH - filter_length + 1)(x)
x = Flatten()(x)
cnns.append(x)
x = merge(cnns, mode='concat')
x = Dropout(0.2)(x)
x = Dense(128, activation='relu')(x)
preds = Dense(len(labels_index), activation='softmax')(x)
model = Model(sequence_input, preds)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# happy learning!
model.fit(x_train, y_train, validation_data=(x_val, y_val),
nb_epoch=5, batch_size=128)
def ent_rel_pred_nn(cv_dat, drop, max_len, embedding_length, rdf2vec_model,
rev_char_dict):
print "We are into the training block"
train_x, train_y, test_x, test_y = cv_dat
print train_x.shape, test_x.shape
print "getting validation data"
train_x, val_x, train_y, val_y = train_test_split(train_x,
train_y,
test_size=0.1,
random_state=666,
stratify=train_y)
print "Validation shapes:" + str(val_x.shape)
print "Data distributions:"
print Counter([np.argmax(a) for a in train_y])
print Counter([np.argmax(a) for a in test_y])
print Counter([np.argmax(a) for a in val_y])
#get the rdf2vec_vectors for all
train_rdf2vec = get_rdf_vecs_phrases(train_x, rdf2vec_model, rev_char_dict)
val_rdf2vec = get_rdf_vecs_phrases(val_x, rdf2vec_model, rev_char_dict)
train_x = np.array(sequence.pad_sequences(train_x, maxlen=max_len),
dtype=np.int)
test_x = np.array(sequence.pad_sequences(test_x, maxlen=max_len),
dtype=np.int)
val_x = np.array(sequence.pad_sequences(val_x, maxlen=max_len),
dtype=np.int)
print "Training the neural net now"
print train_x[0]
#define the neural net model
input_vec = Input(shape=(max_len, ), dtype='int32', name="inp_vec")
embedded_l = Embedding(embedding_length,
128,
mask_zero=False,
input_length=max_len,
trainable=True)(input_vec)
lstm_1 = LSTM(128, return_sequences=False, dropout_W=0.3,
dropout_U=0.3)(embedded_l)
dense_1 = Dense(512,
activation='tanh',
kernel_initializer="glorot_uniform")(lstm_1)
drop_l = Dropout(drop)(dense_1)
dense_2 = Dense(256,
activation='tanh',
kernel_initializer="glorot_uniform")(drop_l)
drop_l_2 = Dropout(drop)(dense_2)
#batch_norm = BatchNormalization()(drop_l)
output_l = Dense(2, activation='softmax', name="output_layer")(drop_l_2)
output_rdf2vec = Dense(500, activation='softmax',
name="output_layer_2")(drop_l_2)
model = Model([input_vec], output=[output_l, output_rdf2vec])
adam = Adam(lr=0.0001)
#weighted categorical crooss entropy
w_array = np.ones((2, 2))
w_array[0, 1] = 1.2
ncce = functools.partial(w_categorical_crossentropy, weights=w_array)
ncce.__name__ = 'w_categorical_crossentropy'
model.compile(loss='categorical_crossentropy',
optimizer=adam,
metrics=['accuracy'],
loss_weights={
'output_layer': 1.,
'output_layer_2': 0.4
})
earlystop = EarlyStopping(monitor='val_output_layer_loss',
min_delta=0.0001,
patience=3,
verbose=1,
mode='auto')
callbacks_list = [earlystop]
print model.summary()
model.fit([train_x], [train_y, train_rdf2vec],
batch_size=128,
epochs=30,
verbose=1,
shuffle=True,
callbacks=callbacks_list,
validation_data=[[val_x], [val_y, val_rdf2vec]])
model_json = model.to_json()
with open(model_j, "w") as json_file:
json_file.write(model_json)
model.save_weights(model_wgt)
print "Saved the model to disk."
test_predictions = model.predict([test_x], verbose=False)
test_pred = [np.argmax(pred) for pred in test_predictions[0]]
test_y = [np.argmax(label) for label in test_y]
f1_value = f1_score(test_y, test_pred, average="macro")
print f1_value
return f1_value
model = Model(inputs=base_model.input, outputs=preds)
for layer in model.layers[:18]:
layer.trainable = False
for layer in model.layers[19:]:
layer.trainable = True
tensorboard = TensorBoard(log_dir="logs/{}".format(NAME))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
model.fit(X_train,
y_train,
batch_size=32,
epochs=3,
validation_split=0.3,
callbacks=[tensorboard])
model.save('emotion_cnn_transfer-vgg.model')
# get the predictions for the test data
predicted_classes = model.predict(X_test)
predicted_classes = predicted_classes.argmax(axis=-1)
predicted_classes = keras.utils.to_categorical(predicted_classes, num_classes)
# get the indices to be plotted
correct = np.where(predicted_classes == y_test)[0]
incorrect = np.where(predicted_classes != y_test)[0]
target_names = ["Class {}".format(i) for i in range(num_classes)]
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
batch_size = 128
nb_epoch = 10
data_augmentation = True
# Model saving callback
#checkpointer = ModelCheckpoint(filepath='stochastic_depth_cifar10.hdf5', verbose=1, save_best_only=True)
if not data_augmentation:
print('Not using data augmentation.')
history = model.fit(x_train,
y_train,
batch_size=batch_size,
nb_epoch=nb_epoch,
verbose=1,
validation_data=(x_test, y_test),
shuffle=True,
callbacks=[])
else:
print('Using real-time data augmentation.')
# realtime data augmentation
datagen_train = ImageDataGenerator(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.125,
height_shift_range=0.125,