def test1():
seq_size = 10
batch_size = 10
rnn_size = 1
xin = Input(batch_shape=(batch_size, seq_size,1))
xtop = Input(batch_shape=(batch_size, seq_size))
xbranch, xsummary = RTTN(rnn_size, return_sequences=True)([xin, xtop])
model = Model(input=[xin, xtop], output=[xbranch, xsummary])
model.compile(loss='MSE', optimizer='SGD')
data_gen = generate_data_batch(batch_size, seq_size)
model.fit_generator(generator=data_gen, samples_per_epoch=1000, nb_epoch=100)
predictions = Dense(12, activation='softmax')(x)
model = Model(input=base_model.input, output=predictions)
# first: train only the top layers (which were randomly initialized)
# i.e. freeze all convolutional InceptionV3 layers
for layer in base_model.layers:
layer.trainable = False
# compile the model (should be done *after* setting layers to non-trainable)
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
model.fit_generator(train_generator,
samples_per_epoch=20000,
nb_epoch=10,
callbacks=[TensorBoard(log_dir='/tmp/attempt1')])
# Save the top-only trained model.
model.save('object_classifierfit1.h5')
for i, layer in enumerate(base_model.layers):
print(i, layer.name, "Input: ", layer.input_shape, "Output: ",
layer.output_shape)
# Unlcok the "bottom" of the model and lock the "top"
for layer in model.layers[:172]:
layer.trainable = False
for layer in model.layers[172:]:
layer.trainable = True
height_shift_range=0.1,
shear_range=0.,
zoom_range=0.,
channel_shift_range=0.,
fill_mode='constant',
cval=0.,
horizontal_flip=False,
vertical_flip=False,
dim_ordering='tf')
datagen.fit(data_images_train)
# Model saving callback
checkpointer = ModelCheckpoint(filepath=WEIGHTS_SEGMENT_FILEPATH,
verbose=1,
save_best_only=True)
# Early stopping
early_stopping = EarlyStopping(monitor='val_loss', patience=20)
history = model_segment.fit_generator(
datagen.flow(data_images_train, data_masks_train, batch_size=batch_size),
samples_per_epoch=data_images_train.shape[0],
nb_epoch=nb_epoch,
verbose=2,
validation_data=(data_images_val, data_masks_val[:, :, :, 0]),
callbacks=[checkpointer, early_stopping])
with open(HISTORY_SEGMENT_FILEPATH, 'w') as f_out:
json.dump(history.history, f_out)
class FergusNModel(object):
def __init__(self, igor):
now = datetime.now()
self.run_name = "fergusn_{}mo_{}day_{}hr_{}min".format(
now.month, now.day, now.hour, now.minute)
log_location = join(igor.log_dir, self.run_name + ".log")
self.logger = igor.logger = make_logger(igor, log_location)
igor.verify_directories()
self.igor = igor
@classmethod
def from_yaml(cls, yamlfile, kwargs=None):
igor = Igor.from_file(yamlfile)
igor.prep()
model = cls(igor)
model.make(kwargs)
return model
@classmethod
def from_config(cls, config, kwargs=None):
igor = Igor(config)
model = cls(igor)
igor.prep()
model.make(kwargs)
return model
def load_checkpoint_weights(self):
weight_file = join(self.igor.model_location, self.igor.saving_prefix,
self.igor.checkpoint_weights)
if exists(weight_file):
self.logger.info("+ Loading checkpoint weights")
self.model.load_weights(weight_file, by_name=True)
else:
self.logger.warning(
"- Checkpoint weights do not exist; {}".format(weight_file))
def plot(self):
filename = join(self.igor.model_location, self.igor.saving_prefix,
'model_visualization.png')
kplot(self.model, to_file=filename)
self.logger.debug("+ Model visualized at {}".format(filename))
def make(self, theano_kwargs=None):
'''Make the model and compile it.
Igor's config options control everything.
Arg:
theano_kwargs as dict for debugging theano or submitting something custom
'''
if self.igor.embedding_type == "convolutional":
make_convolutional_embedding(self.igor)
elif self.igor.embedding_type == "token":
make_token_embedding(self.igor)
elif self.igor.embedding_type == "shallowconv":
make_shallow_convolutional_embedding(self.igor)
elif self.igor.embedding_type == "minimaltoken":
make_minimal_token_embedding(self.igor)
else:
raise Exception("Incorrect embedding type")
B = self.igor.batch_size
spine_input_shape = (B, self.igor.max_num_supertags)
child_input_shape = (B, 1)
parent_input_shape = (B, 1)
E, V = self.igor.word_embedding_size, self.igor.word_vocab_size # for word embeddings
repeat_N = self.igor.max_num_supertags # for lex
mlp_size = self.igor.mlp_size
## dropout parameters
p_emb = self.igor.p_emb_dropout
p_W = self.igor.p_W_dropout
p_U = self.igor.p_U_dropout
w_decay = self.igor.weight_decay
p_mlp = self.igor.p_mlp_dropout
def predict_params():
return {
'output_dim': 1,
'W_regularizer': l2(w_decay),
'activation': 'relu',
'b_regularizer': l2(w_decay)
}
dspineset_in = Input(batch_shape=spine_input_shape,
name='daughter_spineset_in',
dtype='int32')
pspineset_in = Input(batch_shape=spine_input_shape,
name='parent_spineset_in',
dtype='int32')
dhead_in = Input(batch_shape=child_input_shape,
name='daughter_head_input',
dtype='int32')
phead_in = Input(batch_shape=parent_input_shape,
name='parent_head_input',
dtype='int32')
dspine_in = Input(batch_shape=child_input_shape,
name='daughter_spine_input',
dtype='int32')
inputs = [dspineset_in, pspineset_in, dhead_in, phead_in, dspine_in]
### Layer functions
############# Convert the word indices to vectors
F_embedword = Embedding(input_dim=V,
output_dim=E,
mask_zero=True,
W_regularizer=l2(w_decay),
dropout=p_emb)
if self.igor.saved_embeddings is not None:
self.logger.info("+ Cached embeddings loaded")
F_embedword.initial_weights = [self.igor.saved_embeddings]
###### Prediction Functions
## these functions learn a vector which turns a tensor into a matrix of probabilities
### P(Parent supertag | Child, Context)
F_parent_predict = ProbabilityTensor(
name='parent_predictions',
dense_function=Dense(**predict_params()))
### P(Leaf supertag)
F_leaf_predict = ProbabilityTensor(
name='leaf_predictions', dense_function=Dense(**predict_params()))
###### Network functions.
##### Input word, correct its dimensions (basically squash in a certain way)
F_singleword = compose(Fix(), F_embedword)
##### Input spine, correct diemnsions, broadcast across 1st dimension
F_singlespine = compose(RepeatVector(repeat_N), Fix(),
self.igor.F_embedspine)
##### Concatenate and map to a single space
F_alignlex = compose(
RepeatVector(repeat_N), Dropout(p_mlp),
Dense(mlp_size, activation='relu', name='dense_align_lex'), concat)
F_alignall = compose(
Distribute(Dropout(p_mlp), name='distribute_align_all_dropout'),
Distribute(Dense(mlp_size,
activation='relu',
name='align_all_dense'),
name='distribute_align_all_dense'), concat)
F_alignleaf = compose(
Distribute(
Dropout(p_mlp * 0.66), name='distribute_leaf_dropout'
), ### need a separate oen because the 'concat' is different for the two situations
Distribute(Dense(mlp_size, activation='relu', name='leaf_dense'),
name='distribute_leaf_dense'),
concat)
### embed and form all of the inputs into their components
### note: spines == supertags. early word choice, haven't refactored.
leaf_spines = self.igor.F_embedspine(dspineset_in)
pspine_context = self.igor.F_embedspine(pspineset_in)
dspine_single = F_singlespine(dspine_in)
dhead = F_singleword(dhead_in)
phead = F_singleword(phead_in)
### combine the lexical material
lexical_context = F_alignlex([dhead, phead])
#### P(Parent Supertag | Daughter Supertag, Lexical Context)
### we know the daughter spine, want to know the parent spine
### size is (batch, num_supertags)
parent_problem = F_alignall(
[lexical_context, dspine_single, pspine_context])
### we don't have the parent, we just have a leaf
leaf_problem = F_alignleaf([lexical_context, leaf_spines])
parent_predictions = F_parent_predict(parent_problem)
leaf_predictions = F_leaf_predict(leaf_problem)
predictions = [parent_predictions, leaf_predictions]
theano_kwargs = theano_kwargs or {}
## make it quick so i can load in the weights.
self.model = Model(input=inputs,
output=predictions,
preloaded_data=self.igor.preloaded_data,
**theano_kwargs)
#mask_cache = traverse_nodes(parent_prediction)
#desired_masks = ['merge_3.in.mask.0']
#self.p_tensor = K.function(inputs+[K.learning_phase()], [parent_predictions, F_parent_predict.inbound_nodes[0].input_masks[0]])
if self.igor.from_checkpoint:
self.load_checkpoint_weights()
elif not self.igor.in_training:
raise Exception("No point in running this without trained weights")
if not self.igor.in_training:
expanded_children = RepeatVector(repeat_N, axis=2)(leaf_spines)
expanded_parent = RepeatVector(repeat_N, axis=1)(pspine_context)
expanded_lex = RepeatVector(repeat_N, axis=1)(
lexical_context
) # axis here is arbitary; its repeating on 1 and 2, but already repeated once
huge_tensor = concat(
[expanded_lex, expanded_children, expanded_parent])
densely_aligned = LastDimDistribute(
F_alignall.get(1).layer)(huge_tensor)
output_predictions = Distribute(
F_parent_predict, force_reshape=True)(densely_aligned)
primary_inputs = [phead_in, dhead_in, pspineset_in, dspineset_in]
leaf_inputs = [phead_in, dhead_in, dspineset_in]
self.logger.info("+ Compiling prediction functions")
self.inner_func = K.Function(primary_inputs + [K.learning_phase()],
output_predictions)
self.leaf_func = K.Function(leaf_inputs + [K.learning_phase()],
leaf_predictions)
try:
self.get_ptensor = K.function(
primary_inputs + [K.learning_phase()], [
output_predictions,
])
except:
import pdb
pdb.set_trace()
else:
optimizer = Adam(self.igor.LR,
clipnorm=self.igor.max_grad_norm,
clipvalue=self.igor.grad_clip_threshold)
theano_kwargs = theano_kwargs or {}
self.model.compile(loss="categorical_crossentropy",
optimizer=optimizer,
metrics=['accuracy'],
**theano_kwargs)
#self.model.save("here.h5")
def likelihood_function(self, inputs):
if self.igor.in_training:
raise Exception("Not in testing mode; please fix the config file")
return self.inner_func(tuple(inputs) + (0., ))
def leaf_function(self, inputs):
if self.igor.in_training:
raise Exception("Not in testing mode; please fix the config file")
return self.leaf_func(tuple(inputs) + (0., ))
def train(self):
replacers = {
"daughter_predictions": "child",
"parent_predictions": "parent",
"leaf_predictions": "leaf"
}
train_data = self.igor.train_gen(forever=True)
dev_data = self.igor.dev_gen(forever=True)
N = self.igor.num_train_samples
E = self.igor.num_epochs
# generator, samplers per epoch, number epochs
callbacks = [ProgbarV2(3, 10, replacers=replacers)]
checkpoint_fp = join(self.igor.model_location, self.igor.saving_prefix,
self.igor.checkpoint_weights)
self.logger.info("+ Model Checkpoint: {}".format(checkpoint_fp))
callbacks += [
ModelCheckpoint(filepath=checkpoint_fp,
verbose=1,
save_best_only=True)
]
callbacks += [LearningRateScheduler(lambda epoch: self.igor.LR * 0.9)]
csv_location = join(self.igor.log_dir, self.run_name + ".csv")
callbacks += [CSVLogger(csv_location)]
self.model.fit_generator(generator=train_data,
samples_per_epoch=N,
nb_epoch=E,
callbacks=callbacks,
verbose=1,
validation_data=dev_data,
nb_val_samples=self.igor.num_dev_samples)
def debug(self):
dev_data = self.igor.dev_gen(forever=False)
X, Y = next(dev_data)
self.model.predict_on_batch(X)
#self.model.evaluate_generator(dev_data, self.igor.num_dev_samples)
def profile(self, num_iterations=1):
train_data = self.igor.train_gen(forever=True)
dev_data = self.igor.dev_gen(forever=True)
# generator, samplers per epoch, number epochs
callbacks = [ProgbarV2(1, 10)]
self.logger.debug("+ Beginning the generator")
self.model.fit_generator(generator=train_data,
samples_per_epoch=self.igor.batch_size * 10,
nb_epoch=num_iterations,
callbacks=callbacks,
verbose=1,
validation_data=dev_data,
nb_val_samples=self.igor.batch_size)
self.logger.debug(
"+ Calling theano's pydot print.. this might take a while")
theano.printing.pydotprint(self.model.train_function.function,
outfile='theano_graph.png',
var_with_name_simple=True,
with_ids=True)
self.logger.debug("+ Calling keras' print.. this might take a while")
self.plot("keras_graph.png")
inputs = pretrained.input
outputs = pretrained.output
print(inputs)
print(outputs)
hidden = Dense(128, activation='relu')(outputs)
dropout = Dropout(.3)(hidden)
preds = Dense(2, activation='softmax')(dropout)
model = Model(inputs, preds)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=1e-4),
metrics=['acc'])
model.fit_generator(train_batches,
epochs=50,
validation_data=val_batches,
steps_per_epoch=1105 // 4,
validation_steps=255 // 4)
# serialize model to JSON
model_json = model.to_json()
with open("model.json", "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("Dogs_vs_cats_model.h5")
print("Saved model to disk")
model.summary()
class Cnn_model:
"""
Klasa umożliwiająca detekcję twarzy z użyciem konwolucyjnych sieci neuronowych. Zawiera wszystkie
dane wejściowe, obliczeniowe i model sieci służący do obliczeń. Dodatkowo zawiera funkcje
umożliwiające dodanie nowej osoby do bazy oraz identyfikację osób.
"""
def __init__(self):
# Prametry początkowe dla modelu
self.label_id = []
self.img_data = []
self.index = 0
# Ilość epok równa 50, ilość klas do określenia przy dodawaniu osób (poprzez kolejne dodane osoby)
self.nb_class = 0
self.epochs_num = 50
self.fotosperclass_number = 60
def initialize_networkmodel(self):
"""
Funkcja inicjalizuje model sieci. Model użyty to VGGFace bez trzech ostatnich warstw, które
zostały zamienione na warstwy o 64, 32 neuronach, z ostatnią warstwą o ilości neuronów
równej ilości klas. Następuje inicjalizacja parametrów sieci takich jak momentum czy wskaźnik
uczenia się. Określane są metody ewaluacji precyzji i straty.
"""
###### Struktura modelu ######
vgg_model = VGGFace(include_top=False,
weights='vggface',
input_shape=(180, 180, 3))
last_layer = vgg_model.get_layer('pool5').output
# Add layers
x = Flatten(name='flatten')(last_layer)
x = Dense(64, activation='relu', name='fc6')(x)
x = Dense(32, activation='relu', name='fc7')(x)
out = Dense(self.nb_class, activation='softmax', name='fc8')(x)
self.custom_vgg_model = Model(vgg_model.input, out)
# Zamrożenie uczenia wszystkich warstw poza trzema ostatnimi, które zostały dodane
layer_count = 0
for layer in self.custom_vgg_model.layers:
layer_count = layer_count + 1
for l in range(layer_count - 3):
self.custom_vgg_model.layers[l].trainable = False
###### Kompilacja modelu i parametrów uczenia (learning rate, momentum) ######
sgd = optimizers.SGD(lr=5e-3, decay=1e-6, momentum=0.9, nesterov=True)
self.custom_vgg_model.compile(optimizer=sgd,
loss='categorical_crossentropy',
metrics=['accuracy'])
def add_person(self, **kwargs):
"""
Funkcja umożliwia dodanie osoby do bazy do treningu sieci neuronowej. Jeśli parametr *gui*
jest przekazany wizualizacja danych z kamery nastapi w aplikacji GUI. W innym wypadku
wyświetlone zostanie okno z podglądem kamery.
Args:
**kwargs:
gui: Obiekt głównej aplikacji GUI, dzięki któremu możliwa jest wizualizacja danych
z kamery w aplikacji.
"""
# Sprawdzenie czy parametr został przekazany
gui = kwargs.get('gui', False)
# Przekazanie parametru gui do kolejnej funkcji, gdzie jeden z elementów gui zostanie wykorzystany
data = face_recording(gui=gui,
im_count=self.fotosperclass_number,
cnn=True)
# Jeśli dodawana jest pierwsza twarz
if not self.label_id:
curr_id = 0
# Jeśli twarz nie jest pierwsza
else:
curr_id = max(self.label_id) + 1
self.label_id.extend([curr_id] * len(data))
self.img_data.extend(data)
self.nb_class += 1
def data_processing(self):
"""
Funkcja generuje batche zdjęć i dzieli je na zestawy do uczenia i do walidacji. Zestaw danych
do uczenia jest powiększony przez generowane dodatkowo zdjęcia o różnych przekształceniach
(rotacja, przesunięcia, odwrócenie zdjęcia).
"""
# Hot encoding - wymagana forma etykiet dla uczenia sieci
self.hot_label_id = to_categorical(self.label_id)
# Rozbicie danych na treningowe i do walidacji
X_train, X_test, y_train, y_test = train_test_split(np.array(
self.img_data),
self.hot_label_id,
test_size=0.2)
train_datagen = ImageDataGenerator(rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
validation_datagen = ImageDataGenerator()
# Ustawienie parametru batchsize równego długości danych treningowych podzielonych przez 10 % długości
self.train_batchsize = int(round(len(X_train)) / 10)
self.train_generator = train_datagen.flow(
X_train, y_train, shuffle=True, batch_size=self.train_batchsize)
self.validation_generator = validation_datagen.flow(
X_test, y_test, batch_size=self.train_batchsize)
def train_cnn(self):
"""
Funkcja odpowiada za trening konwolucyjnej sieci neuronowej. Do modelu przekazywane są
zestawy zdjęć do uczenia i walidacji, a także następuje przekierowanie informacji nt. statusu
uczenia sieci do oddzielnego strumienia. Umożliwia on wyświetlanie takiej informacji w
aplikacji GUI. Dodatkowo dodany został Callback EarlyStopping, który w razie zmiany celności
nie większej niż 1% - po 5 epokach bez zmiany skończy proces uczenia.
"""
# Stworzenie listy Callbacków - jeden odpowiada za wyświetlanie danych w GUI - drugi za zatrzymanie uczenia
# w przypadku gdy przez 5 epok accuracy nie urośnie o więcej niż 1%.
self.stream_output = OutputStream()
early_stopping = EarlyStopping(monitor='acc',
min_delta=0.01,
patience=5,
verbose=0,
mode='max')
callback_list = [early_stopping, self.stream_output]
# Historia rejestrująca zmiany w procesie uczenia
self.history = self.custom_vgg_model.fit_generator(
self.train_generator,
steps_per_epoch=self.train_generator.n /
self.train_generator.batch_size,
epochs=self.epochs_num,
validation_data=self.validation_generator,
validation_steps=self.validation_generator.n /
self.validation_generator.batch_size,
verbose=1,
callbacks=callback_list)
# # Zapisz model do pliku
# self.custom_vgg_model.save('trained_model_{}'.format(
# datetime.datetime.fromtimestamp(time.time()).strftime('%Y%m%d_%H%M%S')))
def recognize_face(self, gui=False):
"""
Identyfikacja osoby z użyciem wcześniej wytrenowanej sieci neuronowej. Jeśli parametr *gui*
jest dodany nastąpi wyświetlenie obrazu z kamery w aplikacji GUI, jeśli nie - pojawi się
dodatkowe okno.
Args:
gui: Obiekt głównej aplikacji GUI, dzięki któremu możliwa jest wizualizacja danych
z kamery w aplikacji.
Returns:
image: poszukiwana twarz (zdjęcie).
model.predict: wartości prawdopodobieństwa przynależności do danej klasy.
"""
try:
assert self.custom_vgg_model
except:
raise AttributeError()
image = take_image(gui=gui, cnn=True)
return image, self.custom_vgg_model.predict(
np.expand_dims(image, axis=0))[0]
def accuracy_statistics(self):
"""
Wyświetlenie danych dotyczących precyzji i straty w trakcie uczenia sieci neuronowej. Dane
są podzielone na precyzję i stratę dla zestawu uczącego i zestawu walidacyjnego.
"""
##### Pokaż wyniki uczenia #####
acc = self.history.history['acc']
val_acc = self.history.history['val_acc']
loss = self.history.history['loss']
val_loss = self.history.history['val_loss']
epochs = range(len(acc))
# Precyzja uczelnia
plt.subplot(121)
plt.plot(epochs, acc)
plt.plot(epochs, val_acc)
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
# "Strata" uczenia
plt.subplot(122)
plt.plot(epochs, loss)
plt.plot(epochs, val_loss)
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()
'batch_size': 1,
'n_classes': 3,
'n_channels': 4,
'shuffle': True}
# Generators
print(len(train_val_test_dict['train']))
training_generator = SingleDataGenerator(train_val_test_dict['train'], **params)
validation_generator = SingleDataGenerator(train_val_test_dict['val'], **params)
cb_1=keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=2, verbose=0, mode='auto')
cb_2=keras.callbacks.ModelCheckpoint(filepath="1Bpred_weights.{epoch:02d}-{val_loss:.2f}.hdf5", monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', period=1)
results = model.fit_generator(generator=training_generator,
validation_data=validation_generator,
epochs=100,
callbacks=[cb_1,cb_2])
model.save_weights("model_1_weights.h5")
history_1_pred = results.history
pickle.dump( history_1_pred, open( "history_1_pred.pkl", "wb" ) )
params = {'dim': (160,32,160),
'batch_size': 1,
'n_classes': 3,
'n_channels': 4,
'shuffle': False}
validation_generator = SingleDataGenerator(train_val_test_dict['test'], **params)
class SiameseNetwork:
def __init__(self, shape, modelName, learningRate=1.0):
self.learningRate = learningRate
self.modelName = modelName
left_input = Input(shape)
right_input = Input(shape)
# Define Siamese Network using shared weights
L1_layer = Lambda(lambda tensors:K.abs(tensors[0] - tensors[1]))
L1_distance = L1_layer([left_input, right_input])
hidden = Dense(512, activation='relu')(L1_distance)
hidden2 = Dense(64, activation='relu')(hidden)
prediction = Dense(1, activation='sigmoid')(hidden2)
self.siamese_net = Model(inputs=[left_input, right_input], outputs=prediction)
# Compile and prepare network
self.siamese_net.compile(loss="binary_crossentropy", optimizer=Adadelta(self.learningRate), metrics=['accuracy'])
def trainModel(self, trainDatagen, valGen, epochs, batch_size, verbose=1):
early_stop = EarlyStopping(monitor='val_loss', min_delta=0.1, patience=5, verbose=verbose)
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=5, min_lr=0.01, verbose=verbose)
self.siamese_net.fit_generator(trainDatagen
,steps_per_epoch=320000 / batch_size, epochs=epochs
#,validation_data=valGen, validation_steps = 80000 / batch_size
,callbacks=[early_stop, reduce_lr])
def finetune(self, X, Y, epochs, batch_size, verbose=1):
early_stop = EarlyStopping(monitor='val_loss', min_delta=0.1, patience=5, verbose=1)
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=5, min_lr=0.01, verbose=verbose)
self.siamese_net.fit(self.preprocess(X), Y, batch_size=batch_size, epochs=epochs, validation_split=0.2, verbose=verbose, callbacks=[early_stop, reduce_lr])
def testAccuracy(self, X, Y, batch_size=512):
n_correct, total = 0, 0
X_left, X_right, Y_send = [], [], []
for i, x in enumerate(X):
for j, y in enumerate(X):
X_left.append(x)
X_right.append(y)
Y_send.append(1*(Y[i] == Y[j]))
if len(X_left) == batch_size:
Y_send = np.stack(Y_send)
predictions = np.argmax(self.siamese_net.predict([np.stack(X_left), np.stack(X_right)]), axis=1)
n_correct += np.sum(predictions == Y_send)
total += len(X_left)
X_left, X_right, Y_send = [], [], []
if len(X_left) > 0:
Y_send = np.stack(Y_send)
predictions = np.argmax(self.siamese_net.predict([np.stack(X_left), np.stack(X_right)]), axis=1)
n_correct += np.sum(predictions == Y_send)
total += len(X_left)
return n_correct / float(total)
def customTrainModel(self, dataGen, epochs, batch_size, valRatio=0.2, n_steps=320000):
steps_per_epoch = int(n_steps / batch_size)
for _ in range(epochs):
train_loss, val_loss = 0, 0
train_acc, val_acc = 0, 0
for i in range(steps_per_epoch):
x, y = next(dataGen)
# Split into train and val
indices = np.random.permutation(len(y))
splitPoint = int(len(y) * valRatio)
x_train, y_train = [ pp[indices[splitPoint:]] for pp in x], y[indices[splitPoint:]]
x_test, y_test = [ pp[indices[:splitPoint]] for pp in x], y[indices[:splitPoint]]
train_metrics = self.siamese_net.train_on_batch(x_train, y_train)
train_loss += train_metrics[0]
train_acc += train_metrics[1]
val_metrics = self.siamese_net.test_on_batch(x_test, y_test)
val_loss += val_metrics[0]
val_acc += val_metrics[1]
sys.stdout.write("%d / %d : Tr loss: %f, Tr acc: %f, Vl loss: %f, Vl acc: %f \r" % (i+1, steps_per_epoch, train_loss/(i+1), train_acc/(i+1), val_loss/(i+1), val_acc/(i+1)))
sys.stdout.flush()
print("\n")
def maybeLoadFromMemory(self):
try:
self.siamese_net.load_weights(self.modelName + ".h5")
return True
except:
return False
def save(self, customName=None):
if not customName:
self.siamese_net.save_weights(self.modelName + ".h5")
else:
self.siamese_net.save_weights(customName + ".h5")
def preprocess(self, X):
return X
def predict(self, X):
return self.siamese_net.predict(self.preprocess(X), batch_size=1024)
def train():
"""Construct model and train it"""
inputs = Input(shape=data.INPUT_SHAPE)
# load pre-trained model
base_model = ResNet50(input_shape=data.INPUT_SHAPE,
weights='imagenet',
include_top=False)
# freeze pre-trained weights
for layer in base_model.layers:
layer.trainable = False
x = base_model(inputs)
# flatten multi-dimensional structure into one dimension
x = Flatten()(x)
class_output = Dense(2, activation='softmax')(x)
# no idea what activation would be good here
box_coordinates = Dense(4, activation='relu')(x)
model = Model(inputs=inputs, outputs=[class_output, box_coordinates])
# compile the model
# (this should be done after setting layers to non-trainable)
model.compile(optimizer='rmsprop',
loss=['binary_crossentropy', 'mean_squared_error'],
loss_weights=[1 / 0.6, 1 / 10000])
# callbacks
# saves network architecture and weights after each epoch
cp_dir = CHECKPOINT_ROOT_DIR + os.sep + time.strftime("%Y-%m-%d_%H-%M-%S")
os.makedirs(cp_dir)
filepath = cp_dir + os.sep + \
"model_epoch{epoch:02d}_val_loss{val_loss:.2f}.hdf5"
model_checkpoint = callbacks.ModelCheckpoint(filepath,
monitor='val_loss',
verbose=0,
save_best_only=False,
save_weights_only=False,
mode='auto',
period=1)
# writes information about training process to the specified folder in
# TesorBoard format. TensorBoard can then visualize these logs in a
# browser
batch_size = 10
tensor_board = callbacks.TensorBoard(log_dir=TENSORBOARD_LOG_PATH,
histogram_freq=0,
batch_size=batch_size,
write_graph=True,
write_grads=False,
write_images=False,
embeddings_freq=0,
embeddings_layer_names=None,
embeddings_metadata=None)
# `steps_per_epoch` is the number of batches to draw from the generator
# at each epoch.
model.fit_generator(generator=train_generator(data.train_data, batch_size,
data.INPUT_SHAPE),
validation_data=validation_generator(
data.validation_data, batch_size,
data.INPUT_SHAPE),
validation_steps=3,
steps_per_epoch=500,
epochs=10,
workers=1,
callbacks=[model_checkpoint, tensor_board])
return model
)
train_batches = train_generator.flow_from_directory('flowers/train', target_size=(224, 224), batch_size=4)
val_generator = ImageDataGenerator(preprocessing_function=preprocess_input)
val_batches = val_generator.flow_from_directory('flowers/val', target_size=(224, 224), batch_size=4)
indices = train_batches.class_indices
labels = [None] * 17
for key in indices:
labels[indices[key]] = key
pretrained = VGG19(include_top=False, input_shape=(224, 224, 3), weights='imagenet', pooling='max')
for layer in pretrained.layers:
layer.trainable = False
inputs = pretrained.input
outputs = pretrained.output
print(inputs)
print(outputs)
hidden = Dense(128, activation='relu')(outputs)
dropout = Dropout(.3)(hidden)
preds = Dense(17, activation='softmax')(dropout)
model = Model(inputs, preds)
model.compile(loss='categorical_crossentropy', optimizer=Adam(lr=1e-4), metrics=['acc'])
model.fit_generator(train_batches, epochs=100, validation_data=val_batches)
self.best_acc = []
self.dataset = dataset
def on_epoch_end(self, epoch, logs={}):
score, acc = self.model.evaluate_generator(get_batch(self.dataset),
steps=self.dataset.epoch)
self.best_acc.append(acc)
print(("best test -los:{} -acc:{}".format(score, acc)))
def on_train_end(self, logs={}):
print(("test acc list: " + str(self.best_acc)))
print(("BestTest acc:{}".format(max(self.best_acc))))
print(model.summary())
# plot_model(model, to_file='../figure/hierarchicalCNN_No_Shape.png', show_shapes=False)
# plot_model(sent_model, to_file='../figure/sentModel_No_Shape.png', show_shapes=False)
checkpointer = ModelCheckpoint(filepath="../save/weights.hdf5",
verbose=1,
monitor='val_acc',
save_best_only=True,
save_weights_only=False)
testCallback = TestHistory(dataset=testset)
model.fit_generator(get_batch(dataset=trainset),
steps_per_epoch=trainset.epoch,
epochs=15,
validation_data=get_batch(dataset=devset),
validation_steps=devset.epoch,
callbacks=[testCallback])
def fit(self, sp, X_tr, X_tr_aux, y_tr, X_val, X_val_aux, y_val):
keras.backend.clear_session()
print(X_tr.shape)
input_img = Input(shape=(X_tr.shape[1:]))
# We don't use it in this CNN.
input_aux = Input(shape=[X_tr_aux.shape[1]])
x = Conv2D(32, kernel_size=(3, 3))(input_img)
x = LeakyReLU()(x)
x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(x)
x = Dropout(0.1)(x)
x = Conv2D(64, kernel_size=(3, 3))(x)
x = LeakyReLU()(x)
x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(x)
x = Dropout(0.4)(x)
x = Flatten()(x)
x = Dense(128)(x)
x = LeakyReLU()(x)
x = Dropout(sp['dropout1'])(x)
x = Dense(10, activation='softmax')(x)
model = Model([input_img, input_aux], x)
model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=Adam(lr=sp['lr']),
metrics=['accuracy'])
batch_size = sp['batch_size']
generator = ImageDataGenerator(
horizontal_flip=True,
vertical_flip=True,
rotation_range=20,
zoom_range=0.2,
shear_range=0.2,
)
iter_ = generator.flow(
X_tr,
range(len(X_tr)),
batch_size=batch_size,
seed=123,
)
def datagen():
while True:
X_batch, idx = iter_.next()
yield [X_batch, X_tr_aux[idx]], y_tr[idx]
best_weights = '.best_weights.h5'
model.fit_generator(generator=datagen(),
steps_per_epoch=int(X_tr.shape[0] / batch_size),
epochs=100,
verbose=2,
validation_data=([X_val, X_val_aux], y_val),
callbacks=[
EarlyStopping(monitor='val_loss',
patience=10,
verbose=1),
ModelCheckpoint(best_weights,
save_best_only=True,
monitor='val_loss'),
])
model.load_weights(filepath=best_weights)
return model
model_dict = get_model_dict_from_db(model_str, opts)
input = model_dict['in']
output = model_dict['out']
for i in np.arange(opt_utils.get_epoch(opts) / reset_step):
model = Model(model_dict['in'], model_dict['out'])
# for layer in model.layers:
# k = np.random.binomial(1,.5)
# layer.trainable = bool(k)
model.summary()
model.compile(loss=opt_utils.get_loss(opts),
optimizer=optimizer,
metrics=opt_utils.get_metrics(opts))
# TRAIN
samples_per_epoch = data_train.shape[0] if opts[
'training_opts']['samples_per_epoch'] == -1 else opts[
'training_opts']['samples_per_epoch']
model.fit_generator(data_gen.flow(
data_train,
label_train,
batch_size=opts['training_opts']['batch_size'],
shuffle=True,
seed=opts['seed']),
samples_per_epoch=samples_per_epoch,
nb_epoch=reset_step,
callbacks=callback_list,
validation_data=(data_test,
label_test))
except:
print(model_str, dataset_str)
traceback.print_exc()
indices = train_batches.class_indices
labels = [None] * 17
for key in indices:
labels[indices[key]] = key
pretrained = VGG19(include_top=False,
input_shape=(224, 224, 3),
weights='imagenet',
pooling='max')
for layer in pretrained.layers:
layer.trainable = False
inputs = pretrained.input
outputs = pretrained.output
hidden = Dense(128, activation='relu')(outputs)
dropout = Dropout(.3)(hidden)
preds = Dense(17, activation='softmax')(dropout)
model = Model(inputs, preds)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=1e-4),
metrics=['acc'])
model.fit_generator(train_batches,
epochs=100,
validation_data=val_batches,
steps_per_epoch=len(train_batches),
validation_steps=len(val_batches))
# custom_vgg_model = vgg_model
# fine-tuning by GAP
print(custom_vgg_model.summary())
train_datagen = ImageDataGenerator(rescale=1. / 255, shear_range=0.2, zoom_range=0.2, horizontal_flip=True,
validation_split=0)
train_generator = train_datagen.flow_from_directory(
'../../data/train',
target_size=(200, 200),
batch_size=32,
class_mode='categorical',
subset="training")
validation_generator = train_datagen.flow_from_directory(
'../../data/train',
target_size=(200, 200),
batch_size=32,
class_mode='categorical',
subset="validation")
custom_vgg_model.compile(optimizer=keras.optimizers.Adam(lr=1e-7), loss=classification_loss_stage2,
metrics=["accuracy", square_loss_tec, cross_loss_tec])
# custom_vgg_model.compile(optimizer=keras.optimizers.Adam(lr=1e-7), loss=keras.losses.categorical_crossentropy, metrics=["accuracy", square_loss_tec, cross_loss_tec])
model_save_callback = keras.callbacks.ModelCheckpoint("./saved_models/model%s.h5" % time.strftime('%Y-%m-%d-%H-%M-%S'),
monitor="val_loss", verbose=1, save_best_only=False, mode="auto")
custom_vgg_model.fit_generator(train_generator, steps_per_epoch=5000, epochs=2, validation_data=validation_generator,
validation_steps=200, callbacks=[model_save_callback])
# custom_vgg_model.fit(x_train, y_train, batch_size=32, epochs=300, verbose=1, validation_data=(x_test, y_test), shuffle=True, callbacks=[model_save_callback])
# keras.models.save_model(custom_vgg_model, "./saved_models/final_model_" + time.strftime('%Y-%m-%d-%H-%M-%S') + ".h5")
monitor='val_acc',
verbose=1,
save_best_only=True,
mode='max')
print(" --nb_train_samples: " + str(nb_train_samples) + "\n" +
" --nb_validation_samples: " + str(nb_validation_samples) + "\n" +
" --nb_epoch: " + str(nb_epoch) + "\n" + " --nb_classes: " +
str(nb_classes) + "\n" + " --learning rate: " + str(learningrate) +
"\n" + " --momentum: " + str(momentum) + "\n" + " --batchtrainsize: " +
str(batch_trainsize) + "\n" + " --batchvalidationsize: " +
str(batch_testsize) + "\n" + " --optimizer: SGD\n" +
" --metrics: accuracy\n" + " --model: VGG16 (until block3_pool layer)\n")
# fine-tune the model
model.fit_generator(train_generator,
steps_per_epoch=nb_train_samples,
epochs=nb_epoch,
validation_data=validation_generator,
validation_steps=nb_validation_samples,
callbacks=[tb, checkpoint])
# serialize model to JSON
model_json = model.to_json()
with open("./outputs/model.json", "w") as json_file:
json_file.write(model_json)
# save the model !!!Be careful, if you want to load it again, delete "_v2" tag in name
model.save('./trained_models\cifar10-vgg16_model_alllayers_v2.h5')
del model #prevent memory leak
pred = Dense(10, activation='sigmoid')(x)
model = Model(base_model.input, pred)
for layer in base_model.layers:
layer.trainable = False
model.compile(loss='binary_crossentropy',
optimizer=optimizers.SGD(lr=1e-5, momentum=0.9),
metrics=['accuracy'])
model.summary()
train_datagen = ImageDataGenerator(
rescale=1. / 255,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
train_datagen.fit(X_train)
train_generator = train_datagen.flow(X_train, Y_train, batch_size=32)
test_datagen = ImageDataGenerator(rescale=1. / 255)
validation_generator = test_datagen.flow(X_test, Y_test, batch_size=32)
model.fit_generator(
train_generator,
samples_per_epoch=nb_train_samples,
nb_epoch=nb_epoch,
validation_data=validation_generator,
nb_val_samples=nb_validation_samples)
top_model.add(Flatten(input_shape=base_model.output_shape[1:]))
top_model.add(Dense(256, activation='relu'))
top_model.add(Dropout(.5))
top_model.add(Dense(1, activation='sigmoid'))
model = Model(input=base_model.input, output=top_model(base_model.output))
for layer in model.layers[:19]:
layer.trainable = False
model.compile(loss='binary_crossentropy',
optimizer=Adam(),
metrics=['accuracy'])
tensorboard = TensorBoard(log_dir=os.path.join(SAVE_PATH, 'logs'),
histogram_freq=1,
write_graph=True,
write_images=False)
model.fit_generator(train_generator,
steps_per_epoch=nb_train_samples // batch_size,
epochs=2000,
validation_data=validation_generator,
validation_steps=nb_validation_samples // batch_size,
callbacks=[
ModelCheckpoint(filepath=os.path.join(
SAVE_PATH,
'checkpoint-{epoch:02d}-{val_loss:.2f}.hdf5'),
save_best_only=True), tensorboard
])
x = BatchNormalization()(x)
x = Dropout(DROPOUT)(x)
out = Dense(NUM_CLASSES, activation="softmax")(x)
model = Model(inputs=densenet_model.input, outputs=out)
#################################################################################################################################################################################################################
# COMPILE, FINE-TUNE/TRAIN AND EVALUATE THE MODEL
#################################################################################################################################################################################################################
model.compile(loss=CLASS_MODE + "_crossentropy",
optimizer="adam",
metrics=["accuracy"])
K.set_value(model.optimizer.lr, LEARNING_RATE)
model.fit_generator(train_it,
shuffle=True,
epochs=NUM_EPOCHS,
steps_per_epoch=int(TOT_TRAINING_IMAGES / BATCH_SIZE),
validation_data=val_it,
validation_steps=int(TOT_VALIDATION_IMAGES /
BATCH_SIZE))
scores = model.evaluate_generator(test_it, steps=24)
print("Accuracy: %.2f%%" % (scores[1] * 100))
################################################################################################################################################################
# SAVE THE CURRENT SESSION AND THE FINAL MODEL
################################################################################################################################################################
saver = tf.train.Saver()
sess = K.get_session()
saver.save(sess, SAVE_DIR + timestamp + "/session.ckpt")
model.compile(optimizer=opt, loss=custom_mse, metrics=[root_mean_squared_error, mean_squared_error])
model.summary()
start_from = 0
save_every_n_epoch = 1
n_epochs = 10000
# model.load_weights("../weights/implementation9-bn-24.h5")
# start image downloader
ip = None
g = h5_small_vgg_generator(b_size, "../data/h5_small_train", ip)
gval = h5_small_vgg_generator(b_size, "../data/h5_small_validation", None)
for i in range(start_from // save_every_n_epoch, n_epochs // save_every_n_epoch):
print("START", i * save_every_n_epoch, "/", n_epochs)
history = model.fit_generator(g, steps_per_epoch=100000//b_size, epochs=save_every_n_epoch,
validation_data=gval, validation_steps=(10000//b_size))
model.save_weights("../weights/implementation9-bn-" + str(i * save_every_n_epoch) + ".h5")
# save sample images
whole_image_check_overlapping(model, 80, "imp9-bn-" + str(i * save_every_n_epoch) + "-")
# save history
output = open('../history/imp9-bn-{:0=4d}.pkl'.format(i * save_every_n_epoch), 'wb')
pickle.dump(history.history, output)
output.close()
def fit(self, sp, X_tr, X_tr_aux, y_tr, X_val, X_val_aux, y_val):
keras.backend.clear_session()
print(sp)
## Init model
input_bands = Input(shape=X_tr.shape[1:])
base_model = ResNet50(input_shape=input_bands.shape[1:],
input_tensor=input_bands,
include_top=False,
classes=1,
pooling='avg')
x = base_model.output
x = Dropout(sp['dropout1'])(x)
if sp['batchnorm'] == 1:
x = BatchNormalization()(x)
x = Dense(sp['n_filter'], activation='relu')(x)
x = LeakyReLU()(x)
x = Dropout(sp['dropout2'])(x)
x = Dense(1,
activation='sigmoid',
kernel_initializer='he_normal',
name='pred')(x)
model = Model(input_bands, x)
model.compile(loss='binary_crossentropy',
optimizer=Adam(lr=sp['lr']),
metrics=['accuracy'])
batch_size = sp['batch_size']
now = datetime.datetime.now()
## Train
### Warm up (train only classify layer )
for layer in base_model.layers:
layer.trainable = False
model.compile(loss='binary_crossentropy',
optimizer=Adam(lr=sp['lr']),
metrics=['accuracy'])
generator = ImageDataGenerator(horizontal_flip=True,
vertical_flip=True,
rotation_range=25,
fill_mode='wrap')
iter_ = generator.flow(X_tr,
range(len(X_tr)),
batch_size=batch_size,
seed=123)
def datagen():
while True:
X_batch, idx = iter_.next()
yield X_batch, y_tr[idx]
best_weights = 'weights/weights.hdf5'
model.fit_generator(
generator=datagen(),
steps_per_epoch=int(len(X_tr) / batch_size),
epochs=10,
validation_data=(X_val, y_val),
callbacks=[
EarlyStopping(monitor='val_loss', patience=25, verbose=1),
ModelCheckpoint(best_weights,
save_best_only=True,
monitor='val_loss'),
keras.callbacks.TensorBoard(
log_dir="logs/chipping_{:%Y%m%dT%H%M}".format(now),
histogram_freq=0,
write_graph=True,
write_images=False),
])
### Train all model
for layer in base_model.layers:
layer.trainable = True
model.compile(loss='binary_crossentropy',
optimizer=Adam(lr=sp['lr']),
metrics=['accuracy'])
model.fit_generator(generator=datagen(),
steps_per_epoch=int(len(X_tr) / batch_size),
epochs=150,
validation_data=(X_val, y_val),
callbacks=[
EarlyStopping(monitor='val_loss',
patience=25,
verbose=1),
ModelCheckpoint(best_weights,
save_best_only=True,
monitor='val_loss'),
])
model.load_weights(filepath=best_weights)
return model
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.')
n_process = 4
pool = multiprocessing.Pool(processes=n_process)
def train_generator(x, y, batch_size):
train_datagen = T.ImageDataGenerator(
pool=pool,
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=[])
pool.terminate()
# 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])
train_datagen.fit(X_train)
train_generator = train_datagen.flow(X_train, Y_train, batch_size=BATCH_SIZE)
#%%
val_datagen = ImageDataGenerator(rescale=1. / 255, horizontal_flip=False)
val_datagen.fit(X_val)
val_generator = val_datagen.flow(X_val, Y_val, batch_size=BATCH_SIZE)
#%% md
## Train the Model
#%%
train_steps_per_epoch = X_train.shape[0] // BATCH_SIZE
val_steps_per_epoch = X_val.shape[0] // BATCH_SIZE
history = model.fit_generator(train_generator,
steps_per_epoch=train_steps_per_epoch,
validation_data=val_generator,
validation_steps=val_steps_per_epoch,
epochs=EPOCHS,
verbose=1)
#%% md
## Analyze Model Performance
#%%
f, (ax1, ax2) = plt.subplots(1, 2, figsize=(15, 5))
t = f.suptitle('Deep Neural Net Performance', fontsize=12)
f.subplots_adjust(top=0.85, wspace=0.3)
epochs = list(range(1, EPOCHS + 1))
ax1.plot(epochs, history.history['acc'], label='Train Accuracy')
ax1.plot(epochs, history.history['val_acc'], label='Validation Accuracy')
ax1.set_xticks(epochs)
ax1.set_ylabel('Accuracy Value')
ax1.set_xlabel('Epoch')
model.compile(optimizer=opt, loss=custom_mse)
model.summary()
start_from = 160
save_every_n_epoch = 10
n_epochs = 300
model.load_weights("../weights/implementation7l-150.h5")
g = image_processing.image_generator_parts(list_dir,
b_size,
im_size=(224, 224))
for i in range(start_from // save_every_n_epoch,
n_epochs // save_every_n_epoch):
print("START", i * save_every_n_epoch, "/", n_epochs)
history = model.fit_generator(g,
steps_per_epoch=len(list_dir) // b_size,
epochs=save_every_n_epoch)
model.save_weights("../weights/implementation7l-" +
str(i * save_every_n_epoch) + ".h5")
# save sample images
whole_image_check(model, 20, "imp7l-" + str(i) + "-")
# save history
output = open('../history/imp7l-{:0=4d}.pkl'.format(i), 'wb')
pickle.dump(history.history, output)
output.close()
class NonConvexEstimator(BaseEstimator):
def __init__(self,
alpha=1.,
n_components=25,
step_size=1e-3,
latent_dropout_rate=0.,
input_dropout_rate=0.,
coef_init=None,
optimizer='adam',
intercept_init=None,
batch_size=256,
latent_sparsity=None,
use_generator=False,
random_state=None,
n_jobs=1,
max_iter=1000):
self.alpha = alpha
self.n_components = n_components
self.max_iter = max_iter
self.step_size = step_size
self.latent_dropout_rate = latent_dropout_rate
self.input_dropout_rate = input_dropout_rate
self.coef_init = coef_init
self.intercept_init = intercept_init
self.batch_size = batch_size
self.optimizer = optimizer
self.latent_sparsity = latent_sparsity
self.random_state = random_state
self.n_jobs = n_jobs
self.use_generator = use_generator
def fit(self, Xs, ys, dataset_weights=None):
self.random_state = check_random_state(self.random_state)
n_datasets = len(Xs)
n_samples = sum(X.shape[0] for X in Xs)
n_features = Xs[0].shape[1]
# Data
sizes = np.array([y.shape[1] for y in ys])
limits = [0] + np.cumsum(sizes).tolist()
n_targets = limits[-1]
if self.n_components == 'auto':
n_components = n_targets
else:
n_components = self.n_components
self.slices_ = []
for iter in range(n_datasets):
self.slices_.append(np.array([limits[iter], limits[iter + 1]]))
self.slices_ = tuple(self.slices_)
if dataset_weights is None:
dataset_weights = [1.] * n_datasets
dataset_weights /= np.sum(dataset_weights) / n_datasets
# dataset_weights = np.array(dataset_weights) * np.sqrt([X.shape[0] for X in Xs])
padded_ys = []
masks = []
Xs = [X.copy() for X in Xs]
for y, this_slice in zip(ys, self.slices_):
padded_y = np.zeros((y.shape[0], n_targets))
mask = np.zeros((y.shape[0], n_targets), dtype='bool')
mask[:, this_slice[0]:this_slice[1]] = 1
padded_y[:, this_slice[0]:this_slice[1]] = y
padded_ys.append(padded_y)
masks.append(mask)
if self.use_generator:
our_generator = generator(Xs,
padded_ys,
masks,
batch_size=self.batch_size,
dataset_weights=dataset_weights,
random_state=self.random_state)
if self.batch_size is not None:
steps_per_epoch = n_samples / self.batch_size
else:
steps_per_epoch = n_datasets
else:
Xs_cat = np.concatenate(Xs, axis=0)
padded_ys_cat = np.concatenate(padded_ys, axis=0)
masks_cat = np.concatenate(masks, axis=0)
sample_weight = np.concatenate(
[[dataset_weight * n_samples / X.shape[0] / n_datasets] *
X.shape[0] for dataset_weight, X in zip(dataset_weights, Xs)])
if self.batch_size is None:
batch_size = n_samples
else:
batch_size = self.batch_size
# Model
if self.intercept_init is not None:
supervised_bias_initializer = \
lambda shape: K.constant(self.intercept_init)
else:
supervised_bias_initializer = 'zeros'
if self.coef_init is not None:
U, S, VT = svd(self.coef_init, full_matrices=False)
latent_init = U[:, :n_components]
latent_init *= S[:n_components]
supervised_init = VT[:n_components]
if n_components > latent_init.shape[1]:
latent_init = np.concatenate([
latent_init,
np.zeros((latent_init.shape[0],
n_components - latent_init.shape[1]))
],
axis=1)
supervised_init = np.concatenate([
supervised_init,
np.zeros((n_components - supervised_init.shape[0],
supervised_init.shape[1]))
],
axis=0)
supervised_kernel_initializer = \
lambda shape: K.constant(supervised_init)
latent_kernel_initializer = \
lambda shape: K.constant(latent_init)
else:
supervised_kernel_initializer = 'glorot_uniform'
latent_kernel_initializer = 'glorot_uniform'
data = Input(shape=(n_features, ), name='data', dtype='float32')
mask = Input(shape=(n_targets, ), name='mask', dtype='bool')
dropout_data = Dropout(rate=self.input_dropout_rate)(data)
if n_components is not None:
latent = Dense(n_components,
name='latent',
use_bias=False,
kernel_initializer=latent_kernel_initializer,
kernel_regularizer=l2(self.alpha))(dropout_data)
latent = Dropout(rate=self.latent_dropout_rate)(latent)
else:
latent = dropout_data
logits = Dense(n_targets,
use_bias=True,
name='supervised',
kernel_initializer=supervised_kernel_initializer,
bias_initializer=supervised_bias_initializer,
kernel_regularizer=l2(self.alpha))(latent)
prediction = PartialSoftmax()([logits, mask])
self.model_ = Model(inputs=[data, mask], outputs=prediction)
if self.optimizer == 'adam':
optimizer = Adam
elif self.optimizer == 'sgd':
optimizer = SGD
elif self.optimizer == 'rmsprop':
optimizer = RMSprop
else:
raise ValueError('Wrong optimizer')
self.model_.compile(optimizer(self.step_size),
loss='categorical_crossentropy',
metrics=['accuracy'])
schedule = lambda i: 1e-3 if i < self.max_iter // 2 else 1e-4
scheduler = LearningRateScheduler(schedule)
callbacks = [scheduler]
if self.latent_sparsity is not None:
callbacks.append(L1ProjCallback(radius=self.latent_sparsity))
sess = tf.Session(
config=tf.ConfigProto(device_count={'CPU': self.n_jobs},
inter_op_parallelism_threads=self.n_jobs,
intra_op_parallelism_threads=self.n_jobs,
use_per_session_threads=True))
K.set_session(sess)
if self.use_generator:
self.model_.fit_generator(our_generator,
steps_per_epoch=steps_per_epoch,
verbose=2,
epochs=self.max_iter,
callbacks=callbacks)
else:
self.model_.fit([Xs_cat, masks_cat],
padded_ys_cat,
sample_weight=sample_weight,
batch_size=batch_size,
verbose=2,
epochs=self.max_iter,
callbacks=callbacks)
supervised, self.intercept_ = self.model_.get_layer(
'supervised').get_weights()
if self.n_components is not None:
latent = self.model_.get_layer('latent').get_weights()[0]
self.coef_ = latent.dot(supervised)
else:
self.coef_ = supervised
def predict(self, Xs):
n_datasets = len(Xs)
X_cat = np.concatenate(Xs, axis=0)
n_targets = self.model_.output_shape[1]
sample_sizes = np.array([X.shape[0] for X in Xs])
limits = [0] + np.cumsum(sample_sizes).tolist()
sample_slices = []
for i in range(n_datasets):
sample_slices.append(np.array([limits[i], limits[i + 1]]))
sample_slices = tuple(sample_slices)
masks = []
for X, this_slice in zip(Xs, self.slices_):
mask = np.zeros((X.shape[0], n_targets), dtype='bool')
mask[:, this_slice[0]:this_slice[1]] = 1
masks.append(mask)
mask_cat = np.concatenate(masks, axis=0)
y_cat = self.model_.predict([X_cat, mask_cat])
ys = []
for X, sample_slice, target_slice in zip(Xs, sample_slices,
self.slices_):
y = y_cat[sample_slice[0]:sample_slice[1],
target_slice[0]:target_slice[1]]
ys.append(y)
ys = tuple(ys)
return ys
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,
horizontal_flip=True,
vertical_flip=False)
datagen_train.fit(x_train)
# fit the model on the batches generated by datagen.flow()
history = model.fit_generator(datagen_train.flow(x_train, y_train, batch_size=batch_size, shuffle=True),
samples_per_epoch=x_train.shape[0],
nb_epoch=nb_epoch, verbose=1,
validation_data=(x_test, y_test),
callbacks=[])
model = Model(inputs=input_tensor, outputs=predictions)
for layer in model.layers:
layer.W_regularizer = l2(1e-2)
layer.trainable = True
model.compile(loss='categorical_crossentropy', optimizer=optimizers.Adam(
lr=config.lr), metrics=['accuracy'])
# print(model.summary())
# ----- TRAINING -----
model.fit_generator(train_generator,
steps_per_epoch=config.TRAIN_NUM // 12,
validation_data=validate_generator,
validation_steps=config.VAL_NUM // 12,
callbacks=callbacks,
epochs=config.EPOCHS,
verbose=1)
# ------ TESTING ------
label_map = validate_generator.class_indices
print(label_map)
# No augmentation on test dataset
test_datagen = ImageDataGenerator(rescale=1. / 255)
generator = test_datagen.flow_from_directory(
config.TEST_DATA,
target_size=(config.IMAGE_SIZE, config.IMAGE_SIZE),
batch_size=12,
class_mode=None, # No labels for test dataset
EMBEDDING_DIMENSION,
input_length=TARGET_LENGTH,
weights=emb_weights)(target_string)
ts = Conv1D(1000, 3, activation='relu')(ts)
ts = GlobalMaxPooling1D()(ts)
ts = Dropout(0.5)(ts)
inp = concatenate([cwp, cws, esp, ess, l2v, ts])
inp = Dense(units=len(ENCODING_MAP_2x2), activation=u'softmax')(inp)
model = Model(inputs=[
context_words_pair, context_words_single, entities_strings_pair,
entities_strings_single, mapvec, target_string
],
outputs=[inp])
model.compile(loss=u'categorical_crossentropy',
optimizer=u'rmsprop',
metrics=[u'accuracy'])
print(u'Finished building model...')
# --------------------------------------------------------------------------------------------------------------------
checkpoint = ModelCheckpoint(
filepath=u"../data/weights.{epoch:02d}-{acc:.2f}.hdf5", verbose=0)
early_stop = EarlyStopping(monitor=u'acc', patience=5)
file_name = u"../data/train_wiki_uniform.txt"
model.fit_generator(
generate_arrays_from_file(file_name, word_to_index),
steps_per_epoch=int(check_output(["wc", file_name]).split()[0]) /
BATCH_SIZE,
epochs=250,
callbacks=[checkpoint, early_stop])
#--FC Layer hidden1 = Dense(512,activation='relu',name='hidden1')(last_layer) hidden1 = Dropout(0.5)(hidden1) #--FC Layer hidden2 = Dense(512,activation='relu',name='hidden2')(hidden1) hidden2 = Dropout(0.5)(hidden2) #--Regression Layer out = Dense(1,activation='linear',name='classifier')(hidden2) custom_vgg_model = Model(rgb_model.input,out) custom_vgg_model.compile(loss='mse',optimizer=keras.optimizers.Adam(lr=0.0001,decay=0.0005)) custom_vgg_model.fit_generator(generator(),samples_per_epoch=1000,validation_data=val_generator(),validation_steps=10,epochs=2) #--Here you read the label Y=readCSV(dirLabelsTest+'_Depression.csv') buf=0 numberOfFrames = NUM_FRAMES # while (buf < numberOfFrames): #Insert here the directory of the images of test set imagem = image.load_img(dirTesting,target_size=(FRAME_HEIGHT,FRAME_WIDTH)) imagem = image.img_to_array(imagem) #Subtract the mean of VGGFace2 dataset #---put your function here imga = utils.preprocess_input(imagem,version=2) #here it is the mean value of VGGFace dataset X.append(imga) X = np.array(X) X = np.expand_dims(X,axis=0)