def main(path):
train_images, train_labels = utils.load_training_data(path)
test_images, test_labels = utils.load_test_data(path)
n_classes = 10
model = KMeans(n_clusters=n_classes, n_init=1, max_iter=1)
model.fit(train_images)
# which images are assigned to each cluster:
# 1. check all data points assigned to each cluster
# 2. check actual labels of the data points assigned to each cluster
# 3. assign the mode of actual labels to be the label for that cluster
cluster_label_dict = {}
for cluster_num in range(n_classes):
idx = utils.cluster_indices(cluster_num, model.labels_)
original_labels = np.take(train_labels, idx)
mode = stats.mode(original_labels)[0][0]
cluster_label_dict.update({cluster_num: mode})
# prediction
predicted_cluster = model.predict(test_images)
predicted_labels = np.vectorize(cluster_label_dict.get)(predicted_cluster)
accuracy = utils.classification_accuracy(predicted_labels, test_labels)
print(" K means clustering accuracy for cifar 10 = {}".format(accuracy))
# visualise clusters
cluster_centroids = model.cluster_centers_
utils.visualize(cluster_centroids)
def __init__(self, hp):
super(DataGenerator, self).__init__()
self.hp = hp
self.train_data_path = hp['train_data_path']
self.train_images, self.train_labels = utils.load_training_data(
self.train_data_path)
self.total_train_images = self.train_images.shape[0]
def test_removeOutliersFromDataset(self):
data = load_training_data()
data2 = remove_outliers_from_dataset(data)
self.assertNotEqual(data, data2)
self.assertEqual(len(data['t']), 956)
self.assertEqual(len(data2['t']), 901)
def fetch_data():
"""
Saves and returns image training data
"""
# Load training data from face_profiles/
face_profile_data, face_profile_name_index, face_profile_names = utils.load_training_data(
)
# Build the classifier
face_profile = build_svc(face_profile_data, face_profile_name_index,
face_profile_names)
data_dir = os.path.join(os.path.dirname(__file__), "../temp")
data_path = os.path.join(data_dir, "SVM.pkl")
if not os.path.exists(data_dir):
os.mkdir(data_dir)
# Save the classifier
with open(data_path, 'wb') as file:
dump(face_profile, file)
print("\nTraining data is successfully saved\n")
return face_profile
def test_loadData(self):
test_data = load_test_data()
training_data = load_training_data()
self.assertEqual(len(training_data), 7, msg='Column number incorrect')
self.assertEqual(len(test_data), 7, msg='Column number incorrect')
self.assertEqual(len(training_data['t']),
956,
msg='Row number incorrect')
self.assertEqual(len(test_data['p']), 506, msg='Row number incorrect')
def main(data_folder_path, model_folder_path):
if not isinstance(data_folder_path, Path):
data_folder_path = Path(data_folder_path)
if not isinstance(model_folder_path, Path):
model_folder_path = Path(model_folder_path)
training_label_path = data_folder_path.joinpath('training_label.txt')
training_nolabel_path = data_folder_path.joinpath('training_nolabel.txt')
testing_path = data_folder_path.joinpath('testing_data.txt')
print("loading training data ...")
train_x, y = load_training_data(str(training_label_path))
train_x_no_label = load_training_data(str(training_nolabel_path))
print("loading testing data ...")
test_x = load_testing_data(str(testing_path))
# model = train_word2vec(train_x[:100])
model = train_word2vec(train_x + train_x_no_label + test_x)
print("saving model ...")
model.save(str(model_folder_path.joinpath('w2v_all.model')))
def train_non_private_model(args):
"""Train a non-private baseline model on the given dataset for comparison."""
# Dataloaders
trainloader = utils.load_training_data(args)
evalloader = utils.load_evaluation_data(args)
# Logs
file = open(os.path.join(args.non_private_model_path,
'logs-(initial-lr:{:.2f})-(num-epochs:{:d}).txt'.format(
args.lr, args.num_epochs)), 'w')
utils.augmented_print("##########################################", file)
utils.augmented_print(
"Training a non-private model on '{}' dataset!".format(args.dataset),
file)
utils.augmented_print("Initial learning rate: {:.2f}".format(args.lr), file)
utils.augmented_print(
"Number of training epochs: {:d}".format(args.num_epochs), file)
# Non-private model
model = models.Private_Model('model(non-private)', args)
if args.cuda:
model.cuda()
# Optimizer
optimizer = optim.SGD(model.parameters(), lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
decay_steps = [int(args.num_epochs * 0.5), int(args.num_epochs * 0.75),
int(args.num_epochs * 0.9)]
# Training steps
for epoch in range(args.num_epochs):
if epoch in decay_steps:
for param_group in optimizer.param_groups:
param_group['lr'] *= 0.1
utils.train(model, trainloader, optimizer, args)
eval_loss, acc = utils.evaluate(model, evalloader, args)
acc_detailed = utils.evaluate_detailed(model, evalloader, args)
utils.augmented_print(
"Epoch {:d} | Evaluation Loss: {:.4f} | Accuracy: {:.2f}% | Detailed Accuracy(%): {}".format(
epoch + 1, eval_loss, acc,
np.array2string(acc_detailed, precision=2, separator=', ')),
file, flush=True)
# Checkpoints
state = {'epoch': args.num_epochs, 'accuracy': acc, 'eval_loss': eval_loss,
'state_dict': model.state_dict()}
filename = "checkpoint-{}.pth.tar".format(model.name)
filepath = os.path.join(args.non_private_model_path, filename)
torch.save(state, filepath)
utils.augmented_print("##########################################", file)
file.close()
train_with_label = sys.argv[1]
train_no_label = sys.argv[2]
#testing_data = os.path.join(path_prefix, 'testing_data.txt')
model_dir = 'model/'
w2v_path = os.path.join(model_dir, 'w2v_all.model') # 處理word to vec model的路徑
# 定義句子長度、要不要固定embedding、batch大小、要訓練幾個epoch、learning rate的值、model的資料夾路徑
sen_len = 20
fix_embedding = True # fix embedding during training
batch_size = 32
epoch = 5
lr = 1e-3
print("loading data ...") # 把'training_label.txt'跟'training_nolabel.txt'讀進來
train_x, y = load_training_data(train_with_label)
#train_x_no_label = load_training_data(train_no_label)
#train_x_no_label = train_x_no_label[:160000]
# 對input跟labels做預處理
#preprocess = Preprocess(train_x, sen_len, train_x_no_label, w2v_path=w2v_path)
preprocess = Preprocess(train_x, sen_len, w2v_path=w2v_path)
embedding = preprocess.make_embedding(load=True)
#print(preprocess.embedding.most_similar("love"))
train_x = preprocess.sentence_word2idx()
#train_x, train_x_no_label = preprocess.sentence_word2idx()
y = preprocess.labels_to_tensor(y)
# 製作一個model的對象3model = LSTM_Net(embedding, embedding_dim=250, hidden_dim=150, num_layers=1, dropout=0.1, fix_embedding=fix_embedding)
model = LSTM_Net(embedding,
embedding_dim=256,
hidden_dim=128,
num_layers=3,
import utils as ut
#Support Vector Machine
import svm
import logging
import warnings
print(__doc__)
###############################################################################
# Building SVC from database
FACE_DIM = (50, 50) # h = 50, w = 50
# Load training data from face_profiles/
face_profile_data, face_profile_name_index, face_profile_names = ut.load_training_data(
"./Paragon/Drivers/Backend/m_lowNeurals/vInter/face_profiles/")
print("\n", face_profile_name_index.shape[0], " samples from ",
len(face_profile_names), " people are loaded")
# Build the classifier
clf, pca = svm.build_SVC(face_profile_data, face_profile_name_index, FACE_DIM)
###############################################################################
# Facial Recognition In Live Tracking
DISPLAY_FACE_DIM = (600, 600) # the displayed video stream screen dimension
SKIP_FRAME = 2 # the fixed skip frame
frame_skip_rate = 0 # skip SKIP_FRAME frames every other frame
SCALE_FACTOR = 1 # used to resize the captured frame for face detection for faster processing speed
face_cascade = cv2.CascadeClassifier(
def train_word2vec(x):
# 訓練 word to vector 的 word embedding
model = word2vec.Word2Vec(x,
size=250,
window=5,
min_count=5,
workers=12,
iter=10,
sg=1)
return model
if __name__ == "__main__":
path_prefix = sys.argv[1]
print("loading training data ...")
train_x, y = load_training_data(
os.path.join(path_prefix, 'training_label.txt'))
train_x_no_label = load_training_data(
os.path.join(path_prefix, 'training_nolabel.txt'))
print("loading testing data ...")
test_x = load_testing_data(os.path.join(path_prefix, 'testing_data.txt'))
#model = train_word2vec(train_x + train_x_no_label + test_x)
model = train_word2vec(train_x + test_x)
print("saving model ...")
#model.save(os.path.join(path_prefix, 'model/w2v_all.model'))
model.save(os.path.join(path_prefix, 'w2v_all.model'))
def main():
original_train = pd.read_csv(args.train_dataset)
test = pd.read_csv(args.test_dataset)
print("Original Training Dataset ", original_train.shape)
print("Test Dataset ", test.shape)
augment_train_images(args.train_images_folder)
augmented_train = pd.read_csv('train_augmented.csv')
print(augmented_train.shape)
print(augmented_train.head())
print("Original Dataset ")
print(original_train['target'].value_counts())
print("Augmented Dataset ")
print(augmented_train['target'].value_counts())
train_fnames = os.listdir(args.train_images_folder)
test_fnames = os.listdir(args.test_images_folder)
train_fnames = [x for x in train_fnames if x.endswith('.jpg')]
test_fnames = [x for x in test_fnames if x.endswith('.jpg')]
print("Training Images Count: ", len(train_fnames))
print("Test Images Count: ", len(test_fnames))
train_data, train_labels = utils.load_training_data(
train_fnames, augmented_train, args.train_images_folder)
train_data = np.array(train_data)
train_labels = np.array(train_labels)
print("Shape of training data: ", train_data.shape)
print("Shape of training labels: ", train_labels.shape)
le = LabelEncoder()
train_labels = le.fit_transform(train_labels)
np.save('classes.npy', le.classes_)
X_train, X_val, y_train, y_val = train_test_split(
train_data,
train_labels,
test_size=args.test_split_size,
random_state=42)
y_train = to_categorical(y_train, num_classes=8)
y_val = to_categorical(y_val, num_classes=8)
print("Shape of test_x: ", X_train.shape)
print("Shape of train_y: ", y_train.shape)
print("Shape of test_x: ", X_val.shape)
print("Shape of test_y: ", y_val.shape)
train_datagenerator = ImageDataGenerator(rescale=1. / 255)
val_datagenerator = ImageDataGenerator(rescale=1. / 255)
train_datagenerator.fit(X_train)
val_datagenerator.fit(X_val)
num_classes = 8
model = models.vgg_model(num_classes)
print(model.summary())
epochs = 30
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
history = model.fit(X_train,
y_train,
batch_size=30,
epochs=epochs,
validation_data=(X_val, y_val))
acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(epochs)
plt.figure(1)
plt.plot(epochs, acc)
plt.plot(epochs, val_acc)
plt.title('Training and validation accuracy')
plt.savefig('statics/accuracy.png')
plt.show()
plt.figure(2)
plt.plot(epochs, loss)
plt.plot(epochs, val_loss)
plt.title('Training and validation loss')
plt.savefig('statics/validation.png')
plt.show()
model.save('dance-form.h5')
def main(args):
# initial parameters
embedding_size = args.embedding_size
mention_context_size = args.mention_context_size
type_context_size = args.type_context_size
embedding_file = args.embedding_path
hidden_units = args.hidden_units
learning_rate = args.learning_rate
margin = args.margin
batch_size = args.batch_size
n_epochs = args.num_epochs
relation_size = 82
nkerns = [500]
filter_size = [1, 1]
pool = [1, 1]
l1 = 0.000001
l2 = 0.000002
newbob = False
network_file = args.model_path
test_file = args.test
test_result_file = args.test_result
label_file = args.ontology_path
label_file_norm = args.norm_ontology_path
relation_file = args.relation_path
train_type_flag = args.seen_types
tup_representation_size = embedding_size * 2
# load word vectors
word_vectors, vector_size = load_word_vec(embedding_file)
# read train and dev file
print("start loading train and dev file ... ")
doc_id_list_test, type_list_test, trigger_list_test, left_word_list_test, relation_list_test, \
right_word_list_test = load_training_data(test_file)
print("start loading arg and relation files ... ")
all_type_list, all_type_structures = load_types_1(label_file_norm)
rel_index, index_rel = read_relation_index(relation_file)
type_size = len(all_type_list)
# using a matrix to represent each relation
relation_matrix = random_init_rel_vec_factor(
relation_file, tup_representation_size * tup_representation_size)
train_types = get_types_for_train(train_type_flag, label_file)
# prepare data structure
print("start preparing data structures ... ")
curSeed = 23455
rng = numpy.random.RandomState(curSeed)
seed = rng.get_state()[1][0]
print("seed: ", seed)
result_index_test_matrix, result_vector_test_matrix, input_context_test_matrix, input_trigger_test_matrix, \
relation_binary_test_matrix, pos_neg_test_matrix = input_matrix_1_test(
type_list_test, trigger_list_test, left_word_list_test, relation_list_test, right_word_list_test,
embedding_size, mention_context_size, relation_size, label_file, word_vectors, rel_index, train_type_flag)
input_type_matrix, input_type_structure_matrix = type_matrix(
all_type_list, all_type_structures, embedding_file, type_context_size)
time1 = time.time()
dt = theano.config.floatX
test_set_content = theano.shared(
numpy.matrix(input_context_test_matrix, dtype=dt))
test_set_trigger = theano.shared(
numpy.matrix(input_trigger_test_matrix, dtype=dt))
test_set_relation_binary = theano.shared(
numpy.matrix(relation_binary_test_matrix, dtype=dt))
test_set_posneg = theano.shared(numpy.matrix(pos_neg_test_matrix,
dtype=dt))
test_set_y = theano.shared(
numpy.array(result_index_test_matrix, dtype=numpy.dtype(numpy.int32)))
test_set_y_vector = theano.shared(
numpy.matrix(result_vector_test_matrix, dtype=dt))
train_set_type = theano.shared(numpy.matrix(input_type_matrix, dtype=dt))
train_set_type_structure = theano.shared(
numpy.matrix(input_type_structure_matrix, dtype=dt))
train_types = theano.shared(numpy.matrix(train_types, dtype=dt))
# compute number of minibatches for training, validation and testing
n_test_batches = input_trigger_test_matrix.shape[0]
n_test_batches /= batch_size
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x_content = T.matrix(
'x_content') # the data is presented as rasterized images
x_trigger = T.matrix(
'x_trigger') # the data is presented as rasterized images
x_relation_binary = T.matrix('x_relation_binary')
x_pos_neg_flag = T.matrix('x_pos_neg_flag')
x_type = T.matrix('x_type')
x_type_structure = T.matrix('x_type_structure')
y = T.ivector('y') # the labels are presented as 1D vector of
y_vector = T.matrix('y_vector') # the labels are presented as 1D vector of
x_train_types = T.matrix('x_train_types')
# [int] labels
i_shape = [tup_representation_size,
mention_context_size] # this is the size of context matrizes
time2 = time.time()
print("time for preparing data structures: ", time2 - time1)
# build actual model
print('start building the model ... ')
time1 = time.time()
rel_w = theano.shared(value=relation_matrix, borrow=True) ## 26*400
# Construct the mention structure input Layer
layer0_input = x_content.reshape((batch_size, 1, i_shape[0], i_shape[1]))
layer0_input_binary_relation = x_relation_binary.reshape(
(batch_size, 1, relation_size, i_shape[1])) ## 100*1*26*5
# compose amr relation matrix to each tuple
compose_layer = ComposeLayerMatrix(
input=layer0_input,
input_binary_relation=layer0_input_binary_relation,
rel_w=rel_w,
rel_vec_size=tup_representation_size)
layer1_input = compose_layer.output
# initialize the convolution weight matrix
filter_shape = (nkerns[0], 1, tup_representation_size, filter_size[1])
pool_size = (pool[0], pool[1])
fan_in = numpy.prod(filter_shape[1:])
fan_out = (filter_shape[0] * numpy.prod(filter_shape[2:]) /
numpy.prod(pool_size))
w_bound = numpy.sqrt(6. / (fan_in + fan_out))
conv_w = theano.shared(numpy.asarray(rng.uniform(low=-w_bound,
high=w_bound,
size=filter_shape),
dtype=theano.config.floatX),
borrow=True)
b_values = numpy.zeros((filter_shape[0], ), dtype=theano.config.floatX)
conv_b = theano.shared(value=b_values, borrow=True)
# conv with pool layer
layer1_conv = LeNetConvPoolLayer(rng,
W=conv_w,
b=conv_b,
input=layer1_input,
image_shape=(batch_size, 1, i_shape[0],
i_shape[1]),
filter_shape=filter_shape,
poolsize=pool_size)
layer1_output = layer1_conv.output
layer1_flattened = layer1_output.flatten(2)
trigger_features_shaped = x_trigger.reshape((batch_size, embedding_size))
layer2_input = T.concatenate([layer1_flattened, trigger_features_shaped],
axis=1)
# Construct the type structure input Layer
layer_type_input = x_type_structure.reshape(
(type_size, 1, tup_representation_size, type_context_size))
filter_shape_type = (nkerns[0], 1, tup_representation_size, filter_size[1])
pool_size_type = (pool[0], pool[1])
# initialize the implicit relation tensor
type_tensor_shape = (tup_representation_size, tup_representation_size,
tup_representation_size)
type_tensor_w = theano.shared(numpy.asarray(rng.uniform(
low=-w_bound, high=w_bound, size=type_tensor_shape),
dtype=theano.config.floatX),
borrow=True)
# compose relation tensor to each tuple
compose_type_layer = ComposeLayerTensor(input=layer_type_input,
tensor=type_tensor_w)
layer_type_input1 = compose_type_layer.output
# conv with pool layer
layer1_conv_type = LeNetConvPoolLayer(rng,
W=conv_w,
b=conv_b,
input=layer_type_input1,
image_shape=(type_size, 1,
tup_representation_size,
type_context_size),
filter_shape=filter_shape_type,
poolsize=pool_size_type)
layer1_type_output = layer1_conv_type.output
layer1_type_flattened = layer1_type_output.flatten(2)
types_shaped = x_type.reshape((type_size, embedding_size))
layer2_type_input = T.concatenate([layer1_type_flattened, types_shaped],
axis=1)
layer2_type_input_size = nkerns[0]**pool[1] + embedding_size
# ranking based max margin loss layer
train_types_signal = x_train_types.reshape((type_size, 1))
pos_neg_flag = x_pos_neg_flag.reshape((batch_size, 1))
layer3 = MaxRankingMarginCosine1(rng=rng,
input=layer2_input,
input_label=layer2_type_input,
true_label=y_vector,
n_in=layer2_type_input_size,
margin=margin,
batch_size=batch_size,
type_size=type_size,
train_type_signal=train_types_signal,
pos_neg_flag=pos_neg_flag)
cost = layer3.loss
# create a list of all model parameters to be fit by gradient descent
param_list = [
compose_layer.params, layer1_conv.params, compose_type_layer.params
]
params = []
for p in param_list:
params += p
# the cost we minimize during training is the NLL of the model
lambd1 = T.scalar('lambda1', dt)
lambd2 = T.scalar('lambda2', dt)
# L1 and L2 regularization possible
reg2 = 0
reg1 = 0
for p in param_list:
reg2 += T.sum(p[0]**2)
reg1 += T.sum(abs(p[0]))
cost += lambd2 * reg2
cost += lambd1 * reg1
lr = T.scalar('lr', dt)
start = index * batch_size
end = (index + 1) * batch_size
testVariables = {}
testVariables[x_content] = test_set_content[start:end]
testVariables[x_trigger] = test_set_trigger[start:end]
testVariables[x_relation_binary] = test_set_relation_binary[start:end]
testVariables[x_type] = train_set_type
testVariables[x_type_structure] = train_set_type_structure
testVariables[y] = test_set_y[start:end]
testVariables[y_vector] = test_set_y_vector[start:end]
testVariables[x_train_types] = train_types
testVariables[x_pos_neg_flag] = test_set_posneg[start:end]
print("length of train variables ", len(testVariables))
# create a list of gradients for all model parameters
grads = T.grad(cost, params)
# train_model is a function that updates the model parameters by SGD Since this model has many parameters,
# it would be tedious to manually create an update rule for each model parameter. We thus create the updates
# list by automatically looping over all (params[i],grads[i]) pairs.
updates = []
for param_i, grad_i in zip(params, grads):
updates.append((param_i, param_i - lr * grad_i))
test_model_confidence = theano.function([index],
layer3.results(y),
on_unused_input='ignore',
givens=testVariables)
time2 = time.time()
print("time for building the model: ", time2 - time1)
print("loading saved network")
netfile = open(network_file)
relW = cPickle.load(netfile)
compose_layer.params[0].set_value(relW, borrow=True)
convolW = cPickle.load(netfile)
convolB = cPickle.load(netfile)
layer1_conv.params[0].set_value(convolW, borrow=True)
layer1_conv.params[1].set_value(convolB, borrow=True)
layer1_conv_type.params[0].set_value(convolW, borrow=True)
layer1_conv_type.params[1].set_value(convolB, borrow=True)
typeW = cPickle.load(netfile)
compose_type_layer.params[0].set_value(typeW, borrow=True)
netfile.close()
print("finish loading network")
test_batch_size = 100
all_batches = len(result_index_test_matrix) / test_batch_size
confidence_prob = []
confidence_value = []
confidence_list = []
confidence = [test_model_confidence(i) for i in xrange(all_batches)]
for r in range(0, len(confidence)):
for r1 in range(0, test_batch_size):
hypo_result = confidence[r][0].item(r1)
confidence_prob.append(confidence[r][2][r1])
confidence_value.append(confidence[r][1][r1])
confidence_list.append(hypo_result)
y_pred = confidence_list
f = open(test_result_file, "w")
for i in range(0, len(y_pred)):
f.write(str(y_pred[i]) + "\t" + str(confidence_value[i]) + "\t")
for j in range(0, type_size):
f.write(str(confidence_prob[i][j]) + " ")
f.write("\n")
f.close()
def train(epochs, batch_size, input_dir, model_save_dir):
# Make an instance of the VGG class
vgg_model = VGG_MODEL(image_shape)
# Get High-Resolution(HR) [148,148,3] in this case and corresponding Low-Resolution(LR) images
x_train_lr, x_train_hr = utils.load_training_data(input_dir, [148, 148, 3])
#Based on the the batch size, get the total number of batches
batch_count = int(x_train_hr.shape[0] / batch_size)
#Get the downscaled image shape based on the downscale factor
image_shape_downscaled = utils.get_downscaled_shape(
image_shape, downscale_factor)
# Initialize the generator network with the input image shape as the downscaled image shape (shape of LR images)
generator = networks.Generator(input_shape=image_shape_downscaled)
# Initialize the discriminator with the input image shape as the original image shape (HR image shape)
discriminator = networks.Discriminator(image_shape)
# Get the optimizer to tweak parameters based on loss
optimizer = vgg_model.get_optimizer()
# Compile the three models - generator, discriminator and gan(comb of both gen and disc - this network will train generator and will not tweak discriminator)
generator.compile(loss=vgg_model.vgg_loss, optimizer=optimizer)
discriminator.compile(loss="binary_crossentropy", optimizer=optimizer)
gan = networks.GAN_Network(generator, discriminator,
image_shape_downscaled, optimizer,
vgg_model.vgg_loss)
# Run training for the number of epochs defined
for e in range(1, epochs + 1):
print('-' * 15, 'Epoch %d' % e, '-' * 15)
for _ in tqdm(range(batch_count)):
# Get the next batch of LR and HR images
image_batch_lr, image_batch_hr = utils.get_random_batch(
x_train_lr, x_train_hr, x_train_hr.shape[0], batch_size)
generated_images_sr = generator.predict(image_batch_lr)
print(generated_images_sr.shape)
real_data_Y = np.ones(
batch_size) - np.random.random_sample(batch_size) * 0.2
fake_data_Y = np.random.random_sample(batch_size) * 0.2
discriminator.trainable = True
print(real_data_Y.shape)
d_loss_real = discriminator.train_on_batch(image_batch_hr,
real_data_Y)
d_loss_fake = discriminator.train_on_batch(generated_images_sr,
fake_data_Y)
discriminator_loss = 0.5 * np.add(d_loss_fake, d_loss_real)
rand_nums = np.random.randint(0,
x_train_hr.shape[0],
size=batch_size)
image_batch_hr = x_train_hr[rand_nums]
image_batch_lr = x_train_lr[rand_nums]
gan_Y = np.ones(
batch_size) - np.random.random_sample(batch_size) * 0.2
discriminator.trainable = False
gan_loss = gan.train_on_batch(image_batch_lr,
[image_batch_hr, gan_Y])
print("discriminator_loss : %f" % discriminator_loss)
print("gan_loss :", gan_loss)
gan_loss = str(gan_loss)
if e % 50 == 0:
generator.save_weights(model_save_dir + 'gen_model%d.h5' % e)
discriminator.save_weights(model_save_dir + 'dis_model%d.h5' % e)
networks.save_model(gan)
path_prefix = "/home/shannon/Downloads/dataset"
model_dir = "./model"
train_w_filename = os.path.join(path_prefix, 'training_label.txt')
train_wo_filename = os.path.join(path_prefix, 'training_label_semi.txt')
if len(sys.argv) > 2:
train_w_filename = sys.argv[1]
train_wo_filename = sys.argv[2]
w2v_model_filename = os.path.join(model_dir, 'w2v_labeled.model')
# checking device to cpu or gpu
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("device:", device)
# loading data
print("loading training data ...")
train_x, y = load_training_data(train_w_filename)
#train_x_no_label, y_no_label = load_training_data(train_wo_filename)
#train_x = train_x+train_x_no_label
#y = y+y_no_label
# parameters
sen_len = 32 #32
fix_embedding = True # fix embedding during training
batch_size = 16 #1024
epoch = 20
lr = 0.0002 #0.0002 0.00005
# preprocessing data
print("preprocessing training data ...")
preprocess = Preprocess(train_x, sen_len, w2v_path=w2v_model_filename)
embedding_matrix = preprocess.make_embedding(load=True)
import pandas as pd
import argparse
from gensim.models import word2vec
from utils import load_training_data, load_testing_data
def train_word2vec(x):
# 訓練 word to vector 的 word embedding
model = word2vec.Word2Vec(x,
size=250,
window=5,
min_count=10,
workers=12,
iter=10,
sg=1)
return model
if __name__ == "__main__":
path_prefix = './'
print("loading training data ...")
train_x, y = load_training_data('./training_label.txt')
train_x_no_label = load_training_data('./training_nolabel.txt')
print("loading testing data ...")
test_x = load_testing_data('./testing_data.txt')
model = train_word2vec(train_x + train_x_no_label + test_x)
print("saving model ...")
model.save(os.path.join(path_prefix, 'w2v_all.model'))
def train_word2vec(x):
# 訓練word to vector 的 word embedding
model = word2vec.Word2Vec(x,
size=250,
window=5,
min_count=5,
workers=12,
iter=10,
sg=1)
return model
if __name__ == "__main__":
print("loading training data ...")
if len(sys.argv) == 3:
train_x, y = u.load_training_data(sys.argv[1])
train_x_no_label = u.load_training_data(sys.argv[2])
else:
train_x, y = u.load_training_data(
os.path.join(path_prefix, 'training_label.txt'))
train_x_no_label = u.load_training_data(
os.path.join(path_prefix, 'training_nolabel.txt'))
print("loading testing data ...")
test_x = u.load_testing_data(os.path.join(path_prefix, 'testing_data.txt'))
print('start training...')
#model = train_word2vec(train_x + train_x_no_label + test_x)
model = train_word2vec(train_x + test_x)
print("saving model ...")
import numpy as np
import utils
import matplotlib.pyplot as plt
import time
train_images, train_labels = utils.load_training_data(
"cifar-10-batches-py") # 50000,3072
test_images, test_labels = utils.load_test_data(
"cifar-10-batches-py") # 10000,3072
mean_channel_train_images = utils.cifar_10_color(train_images) # 50000,3
mean_channel_test_images = utils.cifar_10_color(test_images) # 10000,3
# task 1
tic = time.time()
mu, sigma, prior = utils.naive_bayes_learn(mean_channel_train_images,
train_labels) # 10x3 10x3 10x1
prediction = utils.cifar10_classifier_naivebayes(mean_channel_test_images, mu,
sigma, prior)
accuracy = utils.classification_accuracy(prediction, test_labels)
print("naive bayes accuracy = {}".format(accuracy))
toc = time.time()
print("Time taken for Naive Bayes = ", toc - tic)
print("----------------------------------")
# # task 2
tic = time.time()
mu, covariance, prior = utils.bayes_learn(mean_channel_train_images,
train_labels)
prediction = utils.cifar10_classifier_bayes(mean_channel_test_images, mu,
covariance, prior)
import utils as ut
import svm
import sys
import logging
import warnings
print(__doc__)
###############################################################################
# Building SVC from database
FACE_DIM = (50,50) # h = 50, w = 50
# Load training data from face_profiles/
face_profile_data, face_profile_name_index, face_profile_names = ut.load_training_data("../face_profiles/")
print "\n", face_profile_name_index.shape[0], " samples from ", len(face_profile_names), " people are loaded"
# Build the classifier
clf, pca = svm.build_SVC(face_profile_data, face_profile_name_index, FACE_DIM)
###############################################################################
# Facial Recognition In Live Tracking
DISPLAY_FACE_DIM = (200, 200) # the displayed video stream screen dimention
SKIP_FRAME = 2 # the fixed skip frame
frame_skip_rate = 0 # skip SKIP_FRAME frames every other frame
SCALE_FACTOR = 4 # used to resize the captured frame for face detection for faster processing speed
# ----------
# script to train a simple cnn (later maybe AlexNet) with the German Trafic Sign Dataset from https://sid.erda.dk/public/archives/daaeac0d7ce1152aea9b61d9f1e19370/published-archive.html
#
# run image_preprocessing.py once before running this script
# ----------
from utils import load_training_data, get_basic_model, get_complex_model, compute_confusion_matrix, plot_confusion_matrix, plot_history, show_images, classes
from matplotlib import pyplot as plt
# load the dataset
train_image_dir = "data/train"
X_data, y_data = load_training_data(train_image_dir)
# splitting
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X_data,
y_data,
test_size=0.2,
random_state=15)
X_train, X_val, y_train, y_val = train_test_split(X_train,
y_train,
test_size=0.2,
random_state=33)
print("Train set size: {0}, Val set size: {1}, Test set size: {2}".format(
len(X_train), len(X_val), len(X_test)))
# base model
basic_model = get_basic_model()
history = basic_model.fit(X_train,
y_train,
epochs=10,
node.op = 'Sub'
if 'use_locking' in node.attr:
del node.attr['use_locking']
with tf.Graph().as_default() as graph:
tf.import_graph_def(graph_def,
name=node_name_prefix,
input_map=None,
return_elements=None,
op_dict=None,
producer_op_list=None)
for op in graph.get_operations():
print(op.name, op.values())
x = graph.get_tensor_by_name(input_node_name)
keep_prob = graph.get_tensor_by_name(keep_prob_node_name)
y = graph.get_tensor_by_name(prediction_node_name)
dataset = utils.load_training_data(batch_size=1)
with tf.Session(graph=graph) as sess:
image_batch, label_batch, path_batch = dataset.next_batch(
augmented=False)
visualization(image_batch[0])
sess.run(tf.global_variables_initializer())
pred = sess.run(y, feed_dict={x: image_batch, keep_prob: .5})
print("path: {}, prediction: {}, label: {}".format(
path_batch[0], pred[0], label_batch[0]))
def main(args):
# initial parameters
embedding_size = args.embedding_size
mention_context_size = args.mention_context_size
type_context_size = args.type_context_size
embedding_file = args.embedding_path
hidden_units = args.hidden_units
learning_rate = args.learning_rate
margin = args.margin
batch_size = args.batch_size
n_epochs = args.num_epochs
relation_size = 82
nkerns = [500]
filter_size = [1, 1]
pool = [1, 1]
l1 = 0.000001
l2 = 0.000002
newbob = False
network_file = args.model_path
train_file = args.train
dev_file = args.dev
test_file = args.test
label_file = args.ontology_path
label_file_norm = args.norm_ontology_path
relation_file = args.relation_path
train_type_flag = args.seen_types
tup_representation_size = embedding_size * 2
# load word vectors
word_vectors, vector_size = load_word_vec(embedding_file)
# read train and dev file
print ("start loading train and dev file ... ")
doc_id_list_train, type_list_train, trigger_list_train, left_word_list_train, relation_list_train, \
right_word_list_train = load_training_data(train_file)
doc_id_list_dev, type_list_dev, trigger_list_dev, left_word_list_dev, relation_list_dev, \
right_word_list_dev = load_training_data(dev_file)
doc_id_list_test, type_list_test, trigger_list_test, left_word_list_test, relation_list_test, \
right_word_list_test = load_training_data(test_file)
print ("start loading arg and relation files ... ")
all_type_list, all_type_structures = load_types_1(label_file_norm)
rel_index, index_rel = read_relation_index(relation_file)
type_size = len(all_type_list)
# using a matrix to represent each relation
relation_matrix = random_init_rel_vec_factor(relation_file, tup_representation_size*tup_representation_size)
train_types = get_types_for_train(train_type_flag, label_file)
# prepare data structure
print ("start preparing data structures ... ")
curSeed = 23455
rng = numpy.random.RandomState(curSeed)
seed = rng.get_state()[1][0]
print ("seed: ", seed)
result_index_train_matrix, result_vector_train_matrix, input_context_train_matrix, input_trigger_train_matrix, \
relation_binary_train_matrix, pos_neg_train_matrix = input_matrix_1(
type_list_train, trigger_list_train, left_word_list_train, relation_list_train, right_word_list_train,
embedding_size, mention_context_size, relation_size, label_file, word_vectors, rel_index, train_type_flag)
result_index_dev_matrix, result_vector_dev_matrix, input_context_dev_matrix, input_trigger_dev_matrix, \
relation_binary_dev_matrix, pos_neg_dev_matrix = input_matrix_1_test(
type_list_dev, trigger_list_dev, left_word_list_dev, relation_list_dev, right_word_list_dev, embedding_size,
mention_context_size, relation_size, label_file, word_vectors, rel_index, train_type_flag)
result_index_test_matrix, result_vector_test_matrix, input_context_test_matrix, input_trigger_test_matrix, \
relation_binary_test_matrix, pos_neg_test_matrix = input_matrix_1_test(
type_list_test, trigger_list_test, left_word_list_test, relation_list_test, right_word_list_test,
embedding_size, mention_context_size, relation_size, label_file, word_vectors, rel_index, train_type_flag)
input_type_matrix, input_type_structure_matrix = type_matrix(
all_type_list, all_type_structures, embedding_file, type_context_size)
time1 = time.time()
dt = theano.config.floatX
train_set_content = theano.shared(numpy.matrix(input_context_train_matrix, dtype=dt))
valid_set_content = theano.shared(numpy.matrix(input_context_dev_matrix, dtype=dt))
test_set_content = theano.shared(numpy.matrix(input_context_test_matrix, dtype=dt))
train_set_trigger = theano.shared(numpy.matrix(input_trigger_train_matrix, dtype=dt))
valid_set_trigger = theano.shared(numpy.matrix(input_trigger_dev_matrix, dtype=dt))
test_set_trigger = theano.shared(numpy.matrix(input_trigger_test_matrix, dtype=dt))
train_set_relation_binary = theano.shared(numpy.matrix(relation_binary_train_matrix, dtype=dt))
valid_set_relation_binary = theano.shared(numpy.matrix(relation_binary_dev_matrix, dtype=dt))
test_set_relation_binary = theano.shared(numpy.matrix(relation_binary_test_matrix, dtype=dt))
train_set_posneg = theano.shared(numpy.matrix(pos_neg_train_matrix, dtype=dt))
valid_set_posneg = theano.shared(numpy.matrix(pos_neg_dev_matrix, dtype=dt))
test_set_posneg = theano.shared(numpy.matrix(pos_neg_test_matrix, dtype=dt))
train_set_y = theano.shared(numpy.array(result_index_train_matrix, dtype=numpy.dtype(numpy.int32)))
valid_set_y = theano.shared(numpy.array(result_index_dev_matrix, dtype=numpy.dtype(numpy.int32)))
test_set_y = theano.shared(numpy.array(result_index_test_matrix, dtype=numpy.dtype(numpy.int32)))
train_set_y_vector = theano.shared(numpy.matrix(result_vector_train_matrix, dtype=dt))
valid_set_y_vector = theano.shared(numpy.matrix(result_vector_dev_matrix, dtype=dt))
test_set_y_vector = theano.shared(numpy.matrix(result_vector_test_matrix, dtype=dt))
train_set_type = theano.shared(numpy.matrix(input_type_matrix, dtype=dt))
train_set_type_structure = theano.shared(numpy.matrix(input_type_structure_matrix, dtype=dt))
train_types = theano.shared(numpy.matrix(train_types, dtype=dt))
# compute number of minibatches for training, validation and testing
n_train_batches = input_trigger_train_matrix.shape[0]
n_valid_batches = input_trigger_dev_matrix.shape[0]
n_test_batches = input_trigger_test_matrix.shape[0]
n_train_batches /= batch_size
n_valid_batches /= batch_size
n_test_batches /= batch_size
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x_content = T.matrix('x_content') # the data is presented as rasterized images
x_trigger = T.matrix('x_trigger') # the data is presented as rasterized images
x_relation_binary = T.matrix('x_relation_binary')
x_pos_neg_flag = T.matrix('x_pos_neg_flag')
x_type = T.matrix('x_type')
x_type_structure = T.matrix('x_type_structure')
y = T.ivector('y') # the labels are presented as 1D vector of
y_vector = T.matrix('y_vector') # the labels are presented as 1D vector of
x_train_types = T.matrix('x_train_types')
# [int] labels
i_shape = [tup_representation_size, mention_context_size] # this is the size of context matrizes
time2 = time.time()
print ("time for preparing data structures: ", time2 - time1)
# build actual model
print ('start building the model ... ')
time1 = time.time()
rel_w = theano.shared(value=relation_matrix, borrow=True) ## 26*400
# Construct the mention structure input Layer
layer0_input = x_content.reshape((batch_size, 1, i_shape[0], i_shape[1]))
layer0_input_binary_relation = x_relation_binary.reshape((batch_size, 1, relation_size, i_shape[1])) ## 100*1*26*5
# compose amr relation matrix to each tuple
compose_layer = ComposeLayerMatrix(input=layer0_input, input_binary_relation=layer0_input_binary_relation,
rel_w=rel_w, rel_vec_size=tup_representation_size)
layer1_input = compose_layer.output
# initialize the convolution weight matrix
filter_shape = (nkerns[0], 1, tup_representation_size, filter_size[1])
pool_size = (pool[0], pool[1])
fan_in = numpy.prod(filter_shape[1:])
fan_out = (filter_shape[0] * numpy.prod(filter_shape[2:]) / numpy.prod(pool_size))
w_bound = numpy.sqrt(6. / (fan_in + fan_out))
conv_w = theano.shared(numpy.asarray(
rng.uniform(low=-w_bound, high=w_bound, size=filter_shape),
dtype=theano.config.floatX),
borrow=True)
b_values = numpy.zeros((filter_shape[0],), dtype=theano.config.floatX)
conv_b = theano.shared(value=b_values, borrow=True)
# conv with pool layer
layer1_conv = LeNetConvPoolLayer(rng, W=conv_w, b=conv_b, input=layer1_input,
image_shape=(batch_size, 1, i_shape[0], i_shape[1]),
filter_shape=filter_shape, poolsize=pool_size)
layer1_output = layer1_conv.output
layer1_flattened = layer1_output.flatten(2)
trigger_features_shaped = x_trigger.reshape((batch_size, embedding_size))
layer2_input = T.concatenate([layer1_flattened, trigger_features_shaped], axis=1)
# Construct the type structure input Layer
layer_type_input = x_type_structure.reshape((type_size, 1, tup_representation_size, type_context_size))
filter_shape_type = (nkerns[0], 1, tup_representation_size, filter_size[1])
pool_size_type = (pool[0], pool[1])
# initialize the implicit relation tensor
type_tensor_shape = (tup_representation_size, tup_representation_size, tup_representation_size)
type_tensor_w = theano.shared(numpy.asarray(rng.uniform(low=-w_bound, high=w_bound, size=type_tensor_shape),
dtype=theano.config.floatX), borrow=True)
# compose relation tensor to each tuple
compose_type_layer = ComposeLayerTensor(input=layer_type_input, tensor=type_tensor_w)
layer_type_input1 = compose_type_layer.output
# conv with pool layer
layer1_conv_type = LeNetConvPoolLayer(rng, W=conv_w, b=conv_b, input=layer_type_input1,
image_shape=(type_size, 1, tup_representation_size, type_context_size),
filter_shape=filter_shape_type, poolsize=pool_size_type)
layer1_type_output = layer1_conv_type.output
layer1_type_flattened = layer1_type_output.flatten(2)
types_shaped = x_type.reshape((type_size, embedding_size))
layer2_type_input = T.concatenate([layer1_type_flattened, types_shaped], axis=1)
layer2_type_input_size = nkerns[0] ** pool[1] + embedding_size
# ranking based max margin loss layer
train_types_signal = x_train_types.reshape((type_size, 1))
pos_neg_flag = x_pos_neg_flag.reshape((batch_size, 1))
layer3 = MaxRankingMarginCosine1(rng=rng, input=layer2_input, input_label=layer2_type_input, true_label=y_vector,
n_in=layer2_type_input_size, margin=margin, batch_size=batch_size,
type_size=type_size, train_type_signal=train_types_signal,
pos_neg_flag=pos_neg_flag)
cost = layer3.loss
# create a list of all model parameters to be fit by gradient descent
param_list = [compose_layer.params, layer1_conv.params, compose_type_layer.params]
params = []
for p in param_list:
params += p
# the cost we minimize during training is the NLL of the model
lambd1 = T.scalar('lambda1', dt)
lambd2 = T.scalar('lambda2', dt)
# L1 and L2 regularization possible
reg2 = 0
reg1 = 0
for p in param_list:
reg2 += T.sum(p[0] ** 2)
reg1 += T.sum(abs(p[0]))
cost += lambd2 * reg2
cost += lambd1 * reg1
lr = T.scalar('lr', dt)
start = index * batch_size
end = (index + 1) * batch_size
validVariables = {}
validVariables[x_content] = valid_set_content[start: end]
validVariables[x_trigger] = valid_set_trigger[start: end]
validVariables[x_relation_binary] = valid_set_relation_binary[start: end]
validVariables[x_type] = train_set_type
validVariables[x_type_structure] = train_set_type_structure
validVariables[y] = valid_set_y[start: end]
validVariables[y_vector] = valid_set_y_vector[start: end]
validVariables[x_train_types] = train_types
validVariables[x_pos_neg_flag] = valid_set_posneg[start: end]
testVariables = {}
testVariables[x_content] = test_set_content[start: end]
testVariables[x_trigger] = test_set_trigger[start: end]
testVariables[x_relation_binary] = test_set_relation_binary[start: end]
testVariables[x_type] = train_set_type
testVariables[x_type_structure] = train_set_type_structure
testVariables[y] = test_set_y[start: end]
testVariables[y_vector] = test_set_y_vector[start: end]
testVariables[x_train_types] = train_types
testVariables[x_pos_neg_flag] = test_set_posneg[start: end]
trainVariables = {}
trainVariables[x_content] = train_set_content[start: end]
trainVariables[x_trigger] = train_set_trigger[start: end]
trainVariables[x_relation_binary] = train_set_relation_binary[start: end]
trainVariables[x_type] = train_set_type
trainVariables[x_type_structure] = train_set_type_structure
trainVariables[y] = train_set_y[start: end]
trainVariables[y_vector] = train_set_y_vector[start: end]
trainVariables[x_train_types] = train_types
trainVariables[x_pos_neg_flag] = train_set_posneg[start: end]
print ("length of train variables ", len(trainVariables))
# create a list of gradients for all model parameters
grads = T.grad(cost, params)
# train_model is a function that updates the model parameters by SGD Since this model has many parameters,
# it would be tedious to manually create an update rule for each model parameter. We thus create the updates
# list by automatically looping over all (params[i],grads[i]) pairs.
updates = []
for param_i, grad_i in zip(params, grads):
updates.append((param_i, param_i - lr * grad_i))
test_model_confidence = theano.function([index], layer3.results(y), on_unused_input='ignore', givens=testVariables)
eval_model_confidence = theano.function([index], layer3.results(y), on_unused_input='ignore', givens=validVariables)
train_model = theano.function([index, lr, lambd1, lambd2], [cost, layer3.loss], updates=updates,
on_unused_input='ignore', givens=trainVariables)
time2 = time.time()
print ("time for building the model: ", time2 - time1)
# Train the MODEL
print ('start training ... ')
time1 = time.time()
num_examples_per_epoch = len(trigger_list_train)
validation_frequency = num_examples_per_epoch / batch_size # validate after each epoch
best_params = []
best_micro_fscore = -1
last_fscore = -1
best_macro_fscore = -1
best_iter = 0
best_micro_fscore_eval = -1
best_macro_fscore_eval = -1
best_iter_eval = 0
start_time = time.clock()
epoch = 0
done_looping = False
max_no_improvement = 5
no_improvement = 0
while (epoch < n_epochs) and (not done_looping):
print ('epoch = ', epoch)
epoch += 1
this_n_train_batches = num_examples_per_epoch / batch_size
for minibatch_index in xrange(this_n_train_batches):
iter = (epoch - 1) * this_n_train_batches + minibatch_index
if iter % 100 == 0:
print ('training @ iter = ', iter)
cost_ij, loss = train_model(minibatch_index, learning_rate, l1, l2)
print ("cost: " + str(cost_ij))
print ("loss: " + str(loss))
if (iter + 1) % validation_frequency == 0:
# test
confidence_eval = [test_model_confidence(i) for i in xrange(n_test_batches)]
confidence_list_eval = []
for r in range(0, len(confidence_eval)):
for r1 in range(0, batch_size):
hypo_result_eval = confidence_eval[r][0].item(r1)
confidence_list_eval.append(hypo_result_eval)
y_pred_eval = confidence_list_eval
y_true_eval = result_index_test_matrix[:n_test_batches * batch_size]
y_true_eval_2 = []
for i in range(len(y_true_eval)):
y_true_eval_2.append(int(y_true_eval[i]))
labels1 = []
for l in range(1, 380):
labels1.append(l)
this_micro_fscore_eval = f1_score(y_true_eval_2, y_pred_eval, labels=labels1, average='micro')
this_macro_fscore_eval = f1_score(y_true_eval_2, y_pred_eval, labels=labels1, average='macro')
print('EVAL: *** epoch %i, best_validation %f, best_validation_m1 %f, learning_rate %f, '
'minibatch %i/%i, validation fscore %f %%' % (epoch, best_micro_fscore_eval * 100.,
best_macro_fscore_eval * 100, learning_rate,
minibatch_index + 1, this_n_train_batches,
this_micro_fscore_eval * 100.))
if this_micro_fscore_eval > best_micro_fscore_eval:
best_micro_fscore_eval = this_micro_fscore_eval
best_macro_fscore_eval = this_macro_fscore_eval
best_iter_eval = iter
# validate
confidence = [eval_model_confidence(i) for i in xrange(n_valid_batches)]
confidence_list = []
for r in range(0, len(confidence)):
for r1 in range(0, batch_size):
hypo_result = confidence[r][0].item(r1)
confidence_list.append(hypo_result)
y_pred = confidence_list
y_true = result_index_dev_matrix[:n_valid_batches * batch_size]
y_true_2 = []
for i in range(len(y_true)):
y_true_2.append(int(y_true[i]))
labels = []
for l in range(1, 380):
labels.append(l)
this_micro_fscore = f1_score(y_true_2, y_pred, labels=labels, average='micro')
this_macro_fscore = f1_score(y_true_2, y_pred, labels=labels, average='macro')
print('epoch %i, best_validation %f, best_validation_m1 %f, learning_rate %f, minibatch %i/%i, '
'validation fscore %f %%' % (epoch, best_micro_fscore * 100., best_macro_fscore * 100,
learning_rate, minibatch_index + 1, this_n_train_batches,
this_micro_fscore * 100.))
# if we got the best validation score until now
if this_micro_fscore > best_micro_fscore:
best_micro_fscore = this_micro_fscore
best_macro_fscore = this_macro_fscore
best_iter = iter
best_params = []
for p in param_list:
p_param = []
for part in p:
p_param.append(part.get_value(borrow=False))
best_params.append(p_param)
no_improvement = 0
else:
if this_micro_fscore > last_fscore:
no_improvement -= 1
no_improvement = max(no_improvement, 0)
else:
no_improvement += 1
updatestep = minibatch_index + this_n_train_batches * (epoch - 1)
if newbob: # learning rate schedule depending on dev result
learning_rate /= 1.2
print ("reducing learning rate to ", learning_rate)
last_fscore = this_micro_fscore
if newbob: # learning rate schedule depending on dev result
if no_improvement > max_no_improvement or learning_rate < 0.0000001:
done_looping = True
break
if not newbob:
if epoch + 1 > 10:
learning_rate /= 1.2
print ("reducing learning rate to ", learning_rate)
if epoch + 1 > 50:
done_looping = True
break
end_time = time.clock()
print('Optimization complete.')
print('Best validation score of %f %% obtained for c=%i, nk=%i, f=%i, h=%i at iteration %i,' %
(best_micro_fscore * 100., mention_context_size, nkerns[0], filter_size[1], hidden_units, best_iter + 1))
time2 = time.time()
print ("time for training: ", time2 - time1)
print('Saving net.')
save_file = open(network_file, 'wb')
for p in best_params:
for p_part in p:
cPickle.dump(p_part, save_file, -1)
save_file.close()
# 3份数据的路径
train_with_label = './data/training_label.txt'
train_no_label = './data/training_nolabel.txt'
testing_data = './data/testing_data.txt'
# word2vec模型的路径
w2v_path = os.path.join('./data/w2v_all.model')
sen_len = 20 # 统一句子长度
fix_embedding = True # 训练时固定embedding模型
batch_size = 128
epoch = 5
lr = 0.001
print("loading data ...") # 读取'training_label.txt'和'training_nolabel.txt'
train_x, y = utils.load_training_data(train_with_label)
train_x_no_label = utils.load_training_data(train_no_label)
# 对input和labels做预处理
preprocess = Preprocess(train_x, sen_len, w2v_path=w2v_path)
embedding = preprocess.make_embedding(load=True)
train_x = preprocess.sentence_word2idx()
y = preprocess.labels_to_tensor(y)
# 定义model
model = LSTM_Net(embedding,
embedding_dim=250,
hidden_dim=150,
num_layers=1,
dropout=0.5,
fix_embedding=fix_embedding)
def main():
try:
import sklearn
if sklearn.__version__ < "0.20":
gs.fatal("Package python3-scikit-learn 0.20 or newer is not installed")
except ImportError:
gs.fatal("Package python3-scikit-learn 0.20 or newer is not installed")
try:
import pandas as pd
except ImportError:
gs.fatal("Package python3-pandas 0.25 or newer is not installed")
# parser options ---------------------------------------------------------------------------------------------------
group = options["group"]
training_map = options["training_map"]
training_points = options["training_points"]
field = options["field"]
model_save = options["save_model"]
model_name = options["model_name"]
hyperparams = {
"penalty": options["penalty"],
"alpha": options["alpha"],
"l1_ratio": options["l1_ratio"],
"C": options["c"],
"epsilon": options["epsilon"],
"min_samples_leaf": options["min_samples_leaf"],
"n_estimators": options["n_estimators"],
"learning_rate": options["learning_rate"],
"subsample": options["subsample"],
"max_depth": options["max_depth"],
"max_features": options["max_features"],
"n_neighbors": options["n_neighbors"],
"weights": options["weights"],
"hidden_layer_sizes": options["hidden_units"],
}
cv = int(options["cv"])
group_raster = options["group_raster"]
importances = flags["f"]
preds_file = options["preds_file"]
classif_file = options["classif_file"]
fimp_file = options["fimp_file"]
param_file = options["param_file"]
norm_data = flags["s"]
random_state = int(options["random_state"])
load_training = options["load_training"]
save_training = options["save_training"]
n_jobs = int(options["n_jobs"])
balance = flags["b"]
category_maps = option_to_list(options["category_maps"])
# define estimator -------------------------------------------------------------------------------------------------
hyperparams, param_grid = process_param_grid(hyperparams)
estimator, mode = predefined_estimators(
model_name, random_state, n_jobs, hyperparams
)
# remove dict keys that are incompatible for the selected estimator
estimator_params = estimator.get_params()
param_grid = {
key: value for key, value in param_grid.items() if key in estimator_params
}
scoring, search_scorer = scoring_metrics(mode)
# checks of input options ------------------------------------------------------------------------------------------
if (
mode == "classification"
and balance is True
and model_name not in check_class_weights()
):
gs.warning(model_name + " does not support class weights")
balance = False
if mode == "regression" and balance is True:
gs.warning("Balancing of class weights is only possible for classification")
balance = False
if classif_file:
if cv <= 1:
gs.fatal(
"Output of cross-validation global accuracy requires cross-validation cv > 1"
)
if not os.path.exists(os.path.dirname(classif_file)):
gs.fatal("Directory for output file {} does not exist".format(classif_file))
# feature importance file selected but no cross-validation scheme used
if importances:
if sklearn.__version__ < "0.22":
gs.fatal("Feature importances calculation requires scikit-learn version >= 0.22")
if fimp_file:
if importances is False:
gs.fatal('Output of feature importance requires the "f" flag to be set')
if not os.path.exists(os.path.dirname(fimp_file)):
gs.fatal("Directory for output file {} does not exist".format(fimp_file))
# predictions file selected but no cross-validation scheme used
if preds_file:
if cv <= 1:
gs.fatal(
"Output of cross-validation predictions requires cross-validation cv > 1"
)
if not os.path.exists(os.path.dirname(preds_file)):
gs.fatal("Directory for output file {} does not exist".format(preds_file))
# define RasterStack -----------------------------------------------------------------------------------------------
stack = RasterStack(group=group)
if category_maps is not None:
stack.categorical = category_maps
# extract training data --------------------------------------------------------------------------------------------
if load_training != "":
X, y, cat, class_labels, group_id = load_training_data(load_training)
if class_labels is not None:
a = pd.DataFrame({"response": y, "labels": class_labels})
a = a.drop_duplicates().values
class_labels = {k: v for (k, v) in a}
else:
gs.message("Extracting training data")
if group_raster != "":
stack.append(group_raster)
if training_map != "":
X, y, cat = stack.extract_pixels(training_map)
y = y.flatten()
with RasterRow(training_map) as src:
class_labels = {v: k for (k, v, m) in src.cats}
if "" in class_labels.values():
class_labels = None
elif training_points != "":
X, y, cat = stack.extract_points(training_points, field)
y = y.flatten()
if y.dtype in (np.object_, np.object):
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
y = le.fit_transform(y)
class_labels = {k: v for (k, v) in enumerate(le.classes_)}
else:
class_labels = None
# take group id from last column and remove from predictors
if group_raster != "":
group_id = X[:, -1]
X = np.delete(X, -1, axis=1)
stack.drop(group_raster)
else:
group_id = None
# check for labelled pixels and training data
if y.shape[0] == 0 or X.shape[0] == 0:
gs.fatal(
"No training pixels or pixels in imagery group "
"...check computational region"
)
from sklearn.utils import shuffle
if group_id is None:
X, y, cat = shuffle(X, y, cat, random_state=random_state)
else:
X, y, cat, group_id = shuffle(
X, y, cat, group_id, random_state=random_state
)
if save_training != "":
save_training_data(
save_training, X, y, cat, class_labels, group_id, stack.names
)
# cross validation settings ----------------------------------------------------------------------------------------
# inner resampling method (cv=2)
from sklearn.model_selection import GridSearchCV, StratifiedKFold, GroupKFold, KFold
if any(param_grid) is True:
if group_id is None and mode == "classification":
inner = StratifiedKFold(n_splits=2, random_state=random_state)
elif group_id is None and mode == "regression":
inner = KFold(n_splits=2, random_state=random_state)
else:
inner = GroupKFold(n_splits=2)
else:
inner = None
# outer resampling method (cv=cv)
if cv > 1:
if group_id is None and mode == "classification":
outer = StratifiedKFold(n_splits=cv, random_state=random_state)
elif group_id is None and mode == "regression":
outer = KFold(n_splits=cv, random_state=random_state)
else:
outer = GroupKFold(n_splits=cv)
# modify estimators that take sample_weights -----------------------------------------------------------------------
if balance is True:
from sklearn.utils import compute_class_weight
class_weights = compute_class_weight(class_weight="balanced", classes=(y), y=y)
fit_params = {"sample_weight": class_weights}
else:
class_weights = None
fit_params = {}
# preprocessing ----------------------------------------------------------------------------------------------------
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
# standardization
if norm_data is True and category_maps is None:
scaler = StandardScaler()
trans = ColumnTransformer(
remainder="passthrough",
transformers=[("scaling", scaler, np.arange(0, stack.count))],
)
# one-hot encoding
elif norm_data is False and category_maps is not None:
enc = OneHotEncoder(handle_unknown="ignore", sparse=False)
trans = ColumnTransformer(
remainder="passthrough", transformers=[("onehot", enc, stack.categorical)]
)
# standardization and one-hot encoding
elif norm_data is True and category_maps is not None:
scaler = StandardScaler()
enc = OneHotEncoder(handle_unknown="ignore", sparse=False)
trans = ColumnTransformer(
remainder="passthrough",
transformers=[
("onehot", enc, stack.categorical),
("scaling", scaler, np.setxor1d(
range(stack.count), stack.categorical).astype('int')),
],
)
# combine transformers
if norm_data is True or category_maps is not None:
estimator = Pipeline([("preprocessing", trans), ("estimator", estimator)])
param_grid = wrap_named_step(param_grid)
fit_params = wrap_named_step(fit_params)
if any(param_grid) is True:
estimator = GridSearchCV(
estimator=estimator,
param_grid=param_grid,
scoring=search_scorer,
n_jobs=n_jobs,
cv=inner,
)
# estimator training -----------------------------------------------------------------------------------------------
gs.message(os.linesep)
gs.message(("Fitting model using " + model_name))
if balance is True and group_id is not None:
estimator.fit(X, y, groups=group_id, **fit_params)
elif balance is True and group_id is None:
estimator.fit(X, y, **fit_params)
else:
estimator.fit(X, y)
# message best hyperparameter setup and optionally save using pandas
if any(param_grid) is True:
gs.message(os.linesep)
gs.message("Best parameters:")
optimal_pars = [
(k.replace("estimator__", "").replace("selection__", "") + " = " + str(v))
for (k, v) in estimator.best_params_.items()
]
for i in optimal_pars:
gs.message(i)
if param_file != "":
param_df = pd.DataFrame(estimator.cv_results_)
param_df.to_csv(param_file)
# cross-validation -------------------------------------------------------------------------------------------------
if cv > 1:
from sklearn.metrics import classification_report
from sklearn import metrics
if (
mode == "classification"
and cv > np.histogram(y, bins=np.unique(y))[0].min()
):
gs.message(os.linesep)
gs.fatal(
"Number of cv folds is greater than number of "
"samples in some classes"
)
gs.message(os.linesep)
gs.message("Cross validation global performance measures......:")
if (
mode == "classification"
and len(np.unique(y)) == 2
and all([0, 1] == np.unique(y))
):
scoring["roc_auc"] = metrics.roc_auc_score
from sklearn.model_selection import cross_val_predict
preds = cross_val_predict(
estimator, X, y, group_id, cv=outer, n_jobs=n_jobs, fit_params=fit_params
)
test_idx = [test for train, test in outer.split(X, y)]
n_fold = np.zeros((0,))
for fold in range(outer.get_n_splits()):
n_fold = np.hstack((n_fold, np.repeat(fold, test_idx[fold].shape[0])))
preds = {"y_pred": preds, "y_true": y, "cat": cat, "fold": n_fold}
preds = pd.DataFrame(data=preds, columns=["y_pred", "y_true", "cat", "fold"])
gs.message(os.linesep)
gs.message("Global cross validation scores...")
gs.message(os.linesep)
gs.message("Metric \t Mean \t Error")
for name, func in scoring.items():
score_mean = (
preds.groupby("fold")
.apply(lambda x: func(x["y_true"], x["y_pred"]))
.mean()
)
score_std = (
preds.groupby("fold")
.apply(lambda x: func(x["y_true"], x["y_pred"]))
.std()
)
gs.message(
name + "\t" + str(score_mean.round(3)) + "\t" + str(score_std.round(3))
)
if mode == "classification":
gs.message(os.linesep)
gs.message("Cross validation class performance measures......:")
report_str = classification_report(
y_true=preds["y_true"],
y_pred=preds["y_pred"],
sample_weight=class_weights,
output_dict=False,
)
report = classification_report(
y_true=preds["y_true"],
y_pred=preds["y_pred"],
sample_weight=class_weights,
output_dict=True,
)
report = pd.DataFrame(report)
gs.message(report_str)
if classif_file != "":
report.to_csv(classif_file, mode="w", index=True)
# write cross-validation predictions to csv file
if preds_file != "":
preds.to_csv(preds_file, mode="w", index=False)
text_file = open(preds_file + "t", "w")
text_file.write('"Real", "Real", "integer", "integer"')
text_file.close()
# feature importances ----------------------------------------------------------------------------------------------
if importances is True:
from sklearn.inspection import permutation_importance
fimp = permutation_importance(
estimator,
X,
y,
scoring=search_scorer,
n_repeats=5,
n_jobs=n_jobs,
random_state=random_state,
)
feature_names = deepcopy(stack.names)
feature_names = [i.split("@")[0] for i in feature_names]
fimp = pd.DataFrame(
{
"feature": feature_names,
"importance": fimp["importances_mean"],
"std": fimp["importances_std"],
}
)
gs.message(os.linesep)
gs.message("Feature importances")
gs.message("Feature" + "\t" + "Score")
for index, row in fimp.iterrows():
gs.message(
row["feature"] + "\t" + str(row["importance"]) + "\t" + str(row["std"])
)
if fimp_file != "":
fimp.to_csv(fimp_file, index=False)
# save the fitted model
import joblib
joblib.dump((estimator, y, class_labels), model_save)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
# 通過 torch.cuda.is_available() 的回傳值進行判斷是否有使用 GPU 的環境，如果有的話 device 就設為 "cuda"，沒有的話就設為 "cpu"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_prefix = sys.argv[1]
no_prefix = sys.argv[2]
train_with_label = os.path.join(train_prefix, 'training_label.txt')
train_no_label = os.path.join(no_prefix, 'training_nolabel.txt')
w2v_path = './w2v_all.model'
print("loading data ...")
train_x, y = load_training_data(train_with_label)
train_x_no_label = load_training_data(train_no_label)
# Variables
batch_size = 128
fix_embedding = True
epoch = 6
lr = 0.001
sen_len = 30
# 對 input 跟 labels 做預處理
preprocess = Preprocess(train_x, sen_len, w2v_path=w2v_path)
embedding = preprocess.make_embedding(load=True)
train_x = preprocess.sentence_word2idx()
y = preprocess.labels_to_tensor(y)
train_wo_filename = os.path.join(path_prefix, 'training_nolabel.txt')
testing_filename = os.path.join(path_prefix, 'testing_data.txt')
def train_word2vec(x):
# training word to vector by word embedding
model = word2vec.Word2Vec(x,
size=200,
window=5,
min_count=5,
workers=12,
iter=8,
sg=1)
return model
if __name__ == "__main__":
print("loading training data ...")
train_x, y = load_training_data(train_w_filename)
train_x_no_label = load_training_data(train_wo_filename)
print("loading testing data ...")
test_x = load_testing_data(testing_filename)
#embedding_x = train_x + train_x_no_label + test_x #1578614
embedding_x = train_x + test_x #400000
print("embedding 2d list ... lenght=", len(embedding_x))
model = train_word2vec(embedding_x)
print("saving model ...")
model.save(os.path.join('model/w2v_200.model'))
import matplotlib.pyplot as plt
import utils as ut
import svm
import sys
import logging
import warnings
print(__doc__)
###############################################################################
# Building SVC from database
FACE_DIM = (50, 50) # h = 50, w = 50
# Load training data from face_profiles/
face_profile_data, face_profile_name_index, face_profile_names = ut.load_training_data(
"../face_profiles/")
print "\n", face_profile_name_index.shape[0], " samples from ", len(
face_profile_names), " people are loaded"
# Build the classifier
clf, pca = svm.build_SVC(face_profile_data, face_profile_name_index, FACE_DIM)
###############################################################################
# Facial Recognition In Live Tracking
DISPLAY_FACE_DIM = (200, 200) # the displayed video stream screen dimention
SKIP_FRAME = 2 # the fixed skip frame
frame_skip_rate = 0 # skip SKIP_FRAME frames every other frame
SCALE_FACTOR = 4 # used to resize the captured frame for face detection for faster processing speed
face_cascade = cv2.CascadeClassifier(
def build(self):
# load training df
self.tdf = utils.load_training_data(self.dataloc, stage=2)
# drop missing image
drop_idx = [
i for i, row in self.tdf['filename'].iteritems()
if fnmatch.fnmatch(row, '*ID_33fcf4d99*')
]
self.tdf = self.tdf.drop(drop_idx)
# set up training fraction
## train and validate dataframes
shuff = self.tdf.sample(frac=self.training_fraction,
random_state=self.random_state)
self.train_df = shuff.iloc[:int(0.90 * len(shuff))]
self.validate_df = shuff.iloc[int(0.90 * len(shuff)):]
len(shuff), len(self.train_df), len(self.validate_df)
# set up generators
self.categories = utils.define_categories(self.train_df)
self.train_generator = utils.Three_Channel_Generator(
self.train_df.reset_index(),
ycols=self.categories,
desired_size=self.img_size,
batch_size=self.batch_size,
random_transform=self.random_transform,
rgb=True)
self.validate_generator = utils.Three_Channel_Generator(
self.validate_df.reset_index(),
ycols=self.categories,
desired_size=self.img_size,
batch_size=self.batch_size,
random_transform=False,
rgb=True)
# load model
self.model = models(self.model_name,
input_image_size=self.img_size,
number_of_output_categories=len(self.categories))
if self.weights_path is not None:
self.model.load_weights(self.weights_path)
# setup callbacks
earlystop = EarlyStopping(patience=10)
learning_rate_reduction = ReduceLROnPlateau(
monitor='categorical_accuracy',
patience=2,
verbose=1,
factor=0.5,
min_lr=0.00001)
checkpoint_name = "model_weights_vgg19_{}.h5".format(self.datestamp)
checkpoint = ModelCheckpoint(checkpoint_name,
monitor='val_acc',
verbose=0,
save_best_only=True,
save_weights_only=True,
mode='auto')
self.callbacks = [earlystop, checkpoint]
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
import numpy as np
import os
import sys
import music21
from utils import create_RNN_model, load_dictionaries, load_training_data, music_to_network_input
seq_length = 32
notes, durations, offsets = load_training_data()
network_input, network_output = music_to_network_input(notes, durations,
offsets, seq_length)
notes_to_int, int_to_notes, duration_to_int, int_to_duration, offset_to_int, int_to_offset = load_dictionaries(
)
notes_num = len(notes_to_int)
dur_num = len(duration_to_int)
off_num = len(offset_to_int)
model = create_RNN_model(notes_num,
dur_num,
off_num,
embedding_size=300,
rnn_units=256,
add_attention=False)
model.summary()
weights_folder = os.path.join(os.getcwd(), 'weights')
import torch.nn as nn
import utils
import cv2
import os
import time
from PIL import Image
from torch.autograd import Variable
from torch.utils.data import DataLoader, TensorDataset
import torch.optim as optim
from torchsummary import summary
cuda = True if torch.cuda.is_available() else False
device = torch.device("cuda" if cuda else "cpu")
Data_Train = utils.load_training_data("./faces")
Data_Train = [np.swapaxes(image, 0, 2) for image in Data_Train]
Data_Train = [np.swapaxes(image, 1, 2) for image in Data_Train]
Data_Train = torch.stack([torch.Tensor(i) for i in Data_Train])
Data_Train = TensorDataset(Data_Train)
class Generator(nn.Module):
def __init__(self, noise_dim=100, input_size=65):
super().__init__()
self.noise_dim = noise_dim
self.input_size = input_size
# simple SGD (later there will be step decay added to the train function)
sgd = optimizers.SGD(lr=args.base_lr, decay=0, momentum=0.9, nesterov=True)
# compile
model.compile(loss=args.loss, optimizer=sgd, metrics=[args.loss])
print model.summary() # print a summary
if args.cnn_weights is not None: # load model weights if specified
model.load_weights(args.cnn_weights, by_name=True)
###############################################################################
# load / prepare data
data = load_training_data(args.survey,
args.targets,
args.colors,
args.split,
args.test_size,
args.train_size,
args.aux_inputs,
seed=args.seed,
drop_lowq=args.drop_lowq)
X_train, X_test, y_train, y_test = data
if args.scale_targets: # scale targets to std == 1
scales = y_train.std(axis=0)
y_train /= scales
y_test /= scales
# create data generators for on the fly augmentations
dg_train = DataGenerator(X_train,
y_train,
batch_size=args.batch_size,
'variation3': variations.variation3()}
model = versions[args.version]
if not args.iscontinue:
try:
os.system("rm -r {}".format(working_dir))
except:
pass
if not os.path.isdir(model.working_dir):
os.mkdir(model.working_dir)
model.create_train_data()
model.create_test_data()
train_data_path = os.path.join(model.working_dir, 'train_data.txt')
training_data = utils.load_training_data(model.working_dir,
train_data_path)
# create lexicon
utils.create_lexicon_file(model.working_dir, train_data_path)
# create wfst
utils.create_wfst(training_data, model.working_dir)
# run the evaluation with different configurations
utils.find_best_config(training_data,
model.working_dir,
ngrams=ngrams,
smooth_methods=smooth_methods)