def __init__(self, trainpath, testpath, detector='sift'):
self.trainstream = data_loader(filepath=trainpath, channels=CHANNELS,
preprocess=PREPROCESS_QUEUE, randomize=RANDOM)
self.teststream = data_loader(filepath=testpath, channels=CHANNELS,
preprocess=PREPROCESS_QUEUE, randomize=RANDOM)
# self.SIFT = cv2.xfeatures.SIFT_create()
# Model
if detector.lower() == 'orb':
self.KPD = cv2.ORB_create(nfeatures=FEATURES, scoreType=cv2.ORB_FAST_SCORE)
self.bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
elif detector.lower() == 'surf':
self.KPD = cv2.xfeatures2d.SURF_create(hessianThreshold=1000)
index_params = dict(algorithm=FLANN_INDEX_KDTREE, trees=5)
search_params = dict(checks=50) # or pass empty dictionary
self.bf = cv2.FlannBasedMatcher(index_params, search_params)
elif detector.lower() == 'sift':
self.KPD = cv2.xfeatures2d.SIFT_create(nfeatures=FEATURES)
self.bf = cv2.BFMatcher(crossCheck=True)
else:
print('Unknown Keypoint Detector')
exit(1)
# Cluster method
self.kMeans = MiniBatchKMeans(VOCAB_SIZE, batch_size=64)
def __init__(self,
trainpath,
testpath,
n=3,
detector='orb',
use_n_matches=30):
self.trainstream = data_loader(filepath=trainpath,
channels=CHANNELS,
preprocess=PREPROCESS_QUEUE,
randomize=RANDOM)
self.teststream = data_loader(filepath=testpath,
channels=CHANNELS,
preprocess=PREPROCESS_QUEUE,
randomize=RANDOM)
self.num_matches = use_n_matches
if detector.lower() == 'orb':
self.KPD = cv2.ORB_create(nfeatures=FEATURES,
scoreType=cv2.ORB_FAST_SCORE)
self.bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
elif detector.lower() == 'surf':
self.KPD = cv2.xfeatures2d.SURF_create(hessianThreshold=1000)
index_params = dict(algorithm=FLANN_INDEX_KDTREE, trees=5)
search_params = dict(checks=50) # or pass empty dictionary
self.bf = cv2.FlannBasedMatcher(index_params, search_params)
elif detector.lower() == 'sift':
self.KPD = cv2.xfeatures2d.SIFT_create(nfeatures=FEATURES)
self.bf = cv2.BFMatcher(crossCheck=True)
else:
print('Unknown Keypoint Detector')
exit(1)
# Nearest neighbors
self.n = n
# Descriptor dictionary
self.desc = dict()
def main():
cuda = torch.cuda.is_available() and True
embedding_size = 200
convolution_size = 3
LR= 0.01
CNN_size = 667
batch_size = 5
num_epoch = 1 # 10
print("-- cnn model details --")
print("embedding Size:", CNN_size, "convolution size:", convolution_size, "learning rate:", LR, "batch size:", batch_size, "num epochs:", num_epoch)
padding = "<padding>"
train_file = "../data/ask_ubuntu/train_random.txt"
dev_file = "../data/ask_ubuntu/dev.txt"
test_file = "../data/ask_ubuntu/test.txt"
corpus_file = "../data/ask_ubuntu/text_tokenized.txt"
embedding_path = "../data/ask_ubuntu/vectors_pruned.200.txt"
data_loader = util.data_loader(corpus_file, cut_off = 2, padding=padding)
encoder = util.Encoder(data_loader.vocab_map[padding], data_loader, embedding_path, cuda)
print("loaded encoder")
CNN = util.CNN(embedding_size, CNN_size, convolution_size)
if cuda:
encoder = encoder.cuda()
CNN = CNN.cuda()
print("loading annotations...")
dev = data_loader.read_annotations(dev_file, 20, 10)
dev_data = data_loader.create_eval_batches(dev)
test = data_loader.read_annotations(test_file, 20, 10)
test_data = data_loader.create_eval_batches(test)
train_data = data_loader.read_annotations(train_file)
print("loaded annotations")
train_losses, dev_metrics, test_metrics = \
util.train(encoder, CNN, num_epoch, data_loader, train_data, dev_data, test_data, batch_size, util.CNN_forward, True, cuda, LR=LR)
CNN = CNN.cpu()
torch.save(CNN, "cnn.model")
return train_losses, dev_metrics, test_metrics
def main():
cuda = torch.cuda.is_available() and True
num_epoch = 10
batch_size = 2
input_size = 200
output_size = 120
LR = 0.001
dev_file = "../data/ask_ubuntu/dev.txt"
test_file = "../data/ask_ubuntu/test.txt"
train_file = "../data/ask_ubuntu/train_random.txt"
corpus_file = "../data/ask_ubuntu/text_tokenized.txt"
padding = "<padding>"
embedding_path = "../data/ask_ubuntu/vectors_pruned.200.txt"
print("-- lstm model --")
print("embedding size:", output_size, "learning rate:", LR, "batch size:",
batch_size, "num epoch:", num_epoch)
# Represent each question as a word sequence (and not as a bag of words)
data_loader = util.data_loader(corpus_file, cut_off=1, padding=padding)
dev = data_loader.read_annotations(dev_file, 20, 10)
dev_data = data_loader.create_eval_batches(dev)
test = data_loader.read_annotations(test_file, 20, 10)
test_data = data_loader.create_eval_batches(test)
train_data = data_loader.read_annotations(train_file, 10, 2)
# Utilize an exisiting vector representation of the words
encoder = util.Encoder(data_loader.vocab_map[padding], data_loader,
embedding_path, cuda)
print("embeddings done")
model = util.LSTM(input_size, output_size)
if cuda:
model = model.cuda()
encoder = encoder.cuda()
train_losses, dev_metrics, test_metrics \
= util.train(encoder, model, num_epoch, data_loader, train_data, dev_data, test_data, batch_size, util.LSTM_forward, True, cuda, LR)
model = model.cpu()
torch.save(model, "lstm.model")
return train_losses, dev_metrics, test_metrics
def __init__(self, trainpath, testpath, detectors=('surf', 'sift', 'orb'), weights=None, params=None):
if weights:
assert len(detectors) == len(weights)
self.weights = weights
else:
self.weights = [1] * len(detectors)
self.trainpath = trainpath
self.testpath = testpath
self.detectors = detectors
self.members = []
for detector in detectors:
self.members.append(BF(trainpath,
testpath,
n=1,
detector=detector,
use_n_matches=NUM_MATCHES))
self.labels = [l[1] for l in data_loader(filepath=testpath)]
def main():
######################### Hyperparameter#########################
learning_rate = 3e-3
decay_rate = 0.9 # decay per epoch
model_name = 'deep_CNN_modified'
num_epochs = 12
#################################################################
train_dir = 'saved_model/{}/'.format(model_name)
if not os.path.exists(train_dir):
os.makedirs(train_dir)
result_dir = 'result_plot/{}/'.format(model_name)
if not os.path.exists(result_dir):
os.makedirs(result_dir)
# full dataset
dataset = data_loader('data/kitti_full.npy',
'data/kitti_full_metadata.json')
# # mini dataset
# dataset = data_loader('data/kitti_mini.npy', 'data/kitti_mini_metadata.json')
model = deep_CNN_model(learning_rate, model_name, num_epochs, dataset,
train_dir, result_dir)
sess = tf.Session()
model = initialize_model(sess, model, train_dir)
model.train(sess)
return img
def chop_lower(img):
return img[:2 * int(img.shape[0] / 3), :]
def zoomy(img, shave=SHAVE):
y, x, _ = img.shape
chunk = int(shave * y)
return img[chunk:y - chunk, :, :]
def zoomx(img, shave=SHAVE):
y, x, _ = img.shape
chunk = int(shave * y)
return img[chunk:y - chunk, :, :]
if __name__ == '__main__':
loader = data_loader(
filepath='/home/benedict/classes/cv/project/lists/splits/0_train.txt',
channels=cv2.IMREAD_GRAYSCALE,
randomize=False)
for img, target in loader:
img = resize(img)
img = binarize(img)
img = inverse(img)
show(img)
args = get_args() if args.version == 'v1': model = modelV1(input_shape=3*32*32, num_classes=10).to(dev) elif args.version == 'v2': model = modelV2(num_classes=10).to(dev) elif args.version == 'v3': model = modelV3(num_classes=10).to(dev) elif args.version == 'v4': model = modelV4(num_classes=10).to(dev) else: model = standardArchs(name=args.version, num_output=10).to(dev) print(model) train_dataloader = data_loader(args.root_dir + '/train', batch_size=32, train=True) test_dataloader = data_loader(args.root_dir + '/test', batch_size=8, train=False) if args.optim == 'sgd': optimizer = optim.SGD(model.parameters(), lr=args.base_lr, momentum=0.9, weight_decay=1e-4) if args.optim == 'adam': optimizer = optim.Adam(model.parameters(), lr=args.base_lr) if args.optim == 'adagrad': optimizer = optim.Adagrad(model.parameters(), lr=args.base_lr) #We can use more if need to for epoch in range(args.num_epochs): train_epoch(epoch, model, train_dataloader, optimizer, args.base_lr, args.version) test_epoch(epoch, model, test_dataloader, args.version)
def main():
use_lstm = True
cuda = torch.cuda.is_available() and True
embedding_size = 300
convolution_size = 3
output_size = 200
cnn_output_size = 600
batch_size = 2
num_epoch = 10
classifier_hidden_size = 20
lr = 0.01
padding = "<padding>"
embedding_path = "../data/glove.txt"
corpus_file = "../data/ask_ubuntu/text_tokenized.txt"
corpus_2 = "../data/stackexchange_android/corpus.txt"
train_file = "../data/ask_ubuntu/train_random.txt"
pos_dev = "../data/stackexchange_android/dev.pos.txt"
neg_dev = "../data/stackexchange_android/dev.neg.txt"
pos_test = "../data/stackexchange_android/test.pos.txt"
neg_test = "../data/stackexchange_android/test.neg.txt"
if use_lstm:
print("-- lstm model --")
print("embedding size:", embedding_size, "learning rate:", lr,
"batch size:", batch_size, "num epoch:", num_epoch)
else:
print("-- cnn model --")
print("embedding size:", embedding_size, "learning rate:", lr, "batch size:", batch_size, "num epoch:", num_epoch,\
"convolution size:", convolution_size, "cnn output size:", cnn_output_size)
data_loader = util.data_loader(corpus_file, cut_off=1, padding=padding)
data_loader.read_new_corpus(corpus_2)
dev_annotations = util.read_annotations_2(pos_dev, neg_dev, -1, -1)
test_annotations = util.read_annotations_2(pos_test, neg_test, -1, -1)
dev_data = data_loader.create_eval_batches(dev_annotations,
first_corpus=False)
test_data = data_loader.create_eval_batches(test_annotations,
first_corpus=False)
train_data = data_loader.read_annotations(train_file)
if use_lstm:
print("-- lstm model --")
print("embedding size:", embedding_size, "learning rate:", lr,
"batch size:", batch_size, "num epoch:", num_epoch)
model = util.LSTM(embedding_size, output_size)
forward = util.LSTM_forward
classifier = util.Classifier(output_size, classifier_hidden_size)
else:
print("-- cnn model --")
print("embedding size:", embedding_size, "learning rate:", lr, "batch size:", batch_size, "num epoch:", num_epoch,\
"convolution size:", convolution_size, "cnn output size:", cnn_output_size)
model = util.CNN(embedding_size, cnn_output_size, convolution_size)
forward = util.CNN_forward
classifier = util.Classifier(cnn_output_size, classifier_hidden_size)
encoder = util.Encoder(data_loader.vocab_map[padding], data_loader,
embedding_path, cuda, embedding_size)
print("encoder loaded")
if cuda:
model = model.cuda()
encoder = encoder.cuda()
classifier = classifier.cuda()
return util.advisarial_trainer(encoder,
model,
classifier,
num_epoch,
data_loader,
train_data,
dev_data,
test_data,
batch_size,
forward,
cuda,
LR=0.0001,
L=False)
def main():
cuda = torch.cuda.is_available() and True
num_epoch = 1 # 10
batch_size = 2
embedding_size = 300
output_size = 120
convolution_size = 3
corpus_file = "../data/ask_ubuntu/text_tokenized.txt"
embedding_path = "../data/glove.txt"
train_file = "../data/ask_ubuntu/train_random.txt"
corpus_2 = "../data/stackexchange_android/corpus.txt"
pos_dev = "../data/stackexchange_android/dev.pos.txt"
neg_dev = "../data/stackexchange_android/dev.neg.txt"
pos_test = "../data/stackexchange_android/test.pos.txt"
neg_test = "../data/stackexchange_android/test.neg.txt"
use_lstm = False
padding = "<padding>"
if use_lstm:
lr = 0.0001
embedding_size = 300
print("-- lstm model --")
print("embedding size:", embedding_size, "learning rate:", lr,
"batch size:", batch_size, "num epoch:", num_epoch)
else:
lr = 0.001
cnn_output_size = 500
print("-- cnn model --")
print("embedding size:", embedding_size, "learning rate:", lr,
"batch size:", batch_size, "num epoch:", num_epoch,
"convolution size:", convolution_size, "cnn output size:",
cnn_output_size)
data_loader = util.data_loader(corpus_file, cut_off=0, padding=padding)
data_loader.read_new_corpus(corpus_2)
print("encoder loading...")
encoder = util.Encoder(data_loader.vocab_map[padding], data_loader,
embedding_path, cuda, embedding_size)
print("encoder loaded")
dev_annotations = util.read_annotations_2(pos_dev, neg_dev, -1, -1)
test_annotations = util.read_annotations_2(pos_test, neg_test, -1, -1)
dev_data = data_loader.create_eval_batches(dev_annotations,
first_corpus=False)
test_data = data_loader.create_eval_batches(test_annotations,
first_corpus=False)
train_data = data_loader.read_annotations(train_file, 10, 3)
print("run model")
if use_lstm:
model = util.LSTM(embedding_size, output_size)
forward = util.LSTM_forward
else:
model = util.CNN(embedding_size, cnn_output_size, convolution_size)
forward = util.CNN_forward
if cuda:
model = model.cuda()
encoder = encoder.cuda()
return util.train_cross(encoder,
model,
num_epoch,
data_loader,
train_data,
dev_data,
test_data,
batch_size,
forward,
lr,
pre_trained_encoder=True,
cuda=cuda)
def train(args):
torch.manual_seed(args.seed)
# Get data loader
train_data, dev_data, word2id, id2word, char2id, new_args = data_loader(
args)
model = QAxl(new_args)
if args.use_cuda:
model = model.cuda()
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
# Get optimizer and scheduler
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adamax(parameters, lr=args.lrate)
lrate = args.lrate
if args.eval:
model.load_state_dict(torch.load(args.model_dir))
model.eval()
model.SelfEvaluate(dev_batches,
args.data_dir + 'dev_eval.json',
answer_file=args.answer_file,
drop_file=args.data_dir + 'drop.json',
dev=args.data_dir + 'dev.json')
exit()
if args.load_model:
model.load_state_dict(torch.load(args.model_dir))
best_score = 0.0
## Training
for epoch in range(1, args.epochs + 1):
train_batches = get_batches(train_data, args.batch_size)
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
model.train()
for i, train_batch in enumerate(train_batches):
loss = model(train_batch)
model.zero_grad()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(parameters,
new_args['grad_clipping'])
optimizer.step()
model.reset_parameters()
if i % 100 == 0:
print(
'epoch = %d, loss = %.5f, step = %d, lrate = %.5f best_score = %.3f'
% (epoch, model.train_loss.value, i, lrate, best_score))
sys.stdout.flush()
model.eval()
exact_match_score, F1 = model.SelfEvaluate(
dev_batches,
args.data_dir + 'dev_eval.json',
answer_file=args.answer_file,
drop_file=args.data_dir + 'drop.json',
dev=args.data_dir + 'dev-v2.0.json')
if best_score < F1:
best_score = F1
print('saving %s ...' % args.model_dir)
torch.save(model.state_dict(), args.model_dir)
if epoch > 0 and epoch % args.decay_period == 0:
lrate *= args.decay
for param_group in optimizer.param_groups:
param_group['lr'] = lrate
def main():
parser = argparse.ArgumentParser(description='reads input file and trains\
a model for toxicity classification')
# Required arguments
parser.add_argument("-i",
"--input",
required=True,
help="Path to training set")
parser.add_argument("-t", "--test", required=True, help="Path to test set")
parser.add_argument("-o",
"--output",
required=True,
help="Path to submission file")
parser.add_argument("-e",
"--embedding",
required=True,
help="Path to embedding file")
# Optional arguments
parser.add_argument("-c",
"--text_column",
type=str,
default="commnet_text",
help="name of the column that contains the text")
parser.add_argument("-l",
"--label_column",
type=str,
default="target",
help="name of the column that contains the labels")
parser.add_argument("-p",
"--preprocess",
type=bool,
default=False,
help="if set the text will be processed")
parser.add_argument("-m",
"--model",
type=str,
choices=['LSTM', 'CNN'],
default="LSTM",
help="the deep neural network")
args = vars(parser.parse_args())
# Loading the data
X_train, y_train, X_test, tokenizer = data_loader(args['input'],
args['test'],
args['text_column'],
args['label_column'],
args['preprocess'])
# Loading the embeddings
word_index = tokenizer.word_index
embedding_matrix = build_embeddings(args['embedding'], word_index)
# building the model
model_selection = args['model']
Model = Model()
if model_selection == 'LSTM':
Model.buil_lstm()
elif model_selection == 'CNN':
Model.build_cnn()
else:
raise ValueError("the model must be CNN or LSTM")
# Training the model
Model.train(X_train, y_train)
# Predicting the labels for test set
test_df = pd.read_csv(args['test'])
Model.predict(X_test, test_df, args['output'])
Logging.add_log(f'Read Parameter: model: {model}, word: {word}', 'debug')
config = load.Config()
text_model = load.Model(config)
# 加载数据
if model == 'FastText':
train_data, train_label = FastText_dataloader('train', word, config)
test_data, test_label = FastText_dataloader('test', word, config)
eva_data, eva_label = FastText_dataloader('dev', word, config)
else:
_, train_data, train_label = load_data('train', word, config)
_, test_data, test_label = load_data('test', word, config)
_, eva_data, eva_label = load_data('dev', word, config)
Logging.add_log(f'Completed loading data', 'debug')
dataloader = data_loader(data=train_data, label=train_label, config=config)
test_dataloader = data_loader(data=test_data,
label=test_label,
config=config)
eva_dataloader = data_loader(data=eva_data, label=eva_label, config=config)
Logging.add_log(f'Start Training', 'debug')
train(config=config,
model=text_model,
data=dataloader,
name=model,
test_dataloader=test_dataloader,
eva_dataloader=eva_dataloader,
logger=Logging)
default='results',
help='results save path')
parser.add_argument('--inverse_order',
type=bool,
default=True,
help='True: [input, target], False: [target, input]')
opt = parser.parse_args()
print(opt)
# results save path
if not os.path.isdir(opt.dataset + '_results/test_results'):
os.mkdir(opt.dataset + '_results/test_results')
# data_loader
test_loader = util.data_loader('data/' + opt.dataset + '/' +
opt.test_subfolder,
1,
shuffle=False)
img_size = test_loader.shape[1]
# variables
x = tf.placeholder(tf.float32,
shape=(None, opt.input_size, opt.input_size,
test_loader.shape[3]))
# network
G = network.generator(x, opt.ngf)
# open session and initialize all variables
saver = tf.train.Saver()
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
default=True,
help='0: [input, target], 1 - [target, input]')
opt = parser.parse_args()
print(opt)
# results save path
root = opt.dataset + '_' + opt.save_root + '/'
model = opt.dataset + '_'
if not os.path.isdir(root):
os.mkdir(root)
if not os.path.isdir(root + 'Fixed_results'):
os.mkdir(root + 'Fixed_results')
# data_loader
train_loader = util.data_loader('data/' + opt.dataset + '/' +
opt.train_subfolder,
opt.batch_size,
shuffle=True)
test_loader = util.data_loader('data/' + opt.dataset + '/' +
opt.test_subfolder,
opt.test_batch_size,
shuffle=True)
img_size = train_loader.shape[1]
test_img = test_loader.next_batch()
if opt.inverse_order:
fixed_x_ = test_img[:, :, img_size:, :]
fixed_y_ = test_img[:, :, 0:img_size, :]
else:
fixed_x_ = test_img[:, :, 0:img_size, :]
fixed_y_ = test_img[:, :, img_size:, :]
if img_size != opt.input_size:
