def task3():
sgd_params = {
'GMM': {
'alpha': 1.0,
'mb_num': 200
},
'Peaks': {
'alpha': 0.1,
'mb_num': 250
},
'SwissRoll': {
'alpha': 0.08,
'mb_num': 250
}
}
for data_set, params in sgd_params.items():
X_tr, y_tr, X_te, y_te = get_data(data_set)
_ = sgd(X_tr,
y_tr,
X_te,
y_te,
alpha=params['alpha'],
mb_num=params['mb_num'],
max_epochs=200,
data_set=data_set)
def feature_to_vector(self):
"""
创建并初始化ont-hot向量：从csv中读取药物数据，然后将功效特征转换为one-hot向量
:return:
"""
series = get_data(medicine_path)
# print("series", series)
root_2_word = root_to_word(thesaurus_path) # 获取同义词根到词的映射字典
# print("root_2_word", len(root_2_word), root_2_word)
word_2_root = word_to_root(thesaurus_path) # 获取词到同义词根的映射字典
# print("word_2_root", len(word_2_root), word_2_root)
# 创建并初始化一个DataFrame存储one-hot向量，第一行的列索引为词根
self.df = pd.DataFrame(np.zeros((len(series), len(root_2_word))),
columns=root_2_word.keys())
for indexs in series.index: # series去掉了nan值，index是不连贯的,所以用这种方法遍历
item_str = series[indexs]
if item_str == '':
continue
# item_list = item_str.strip().split() # 针对以空格作为分隔符的症状数据
item_list = re.split("、", item_str) # 针对以“、”作为分隔符的功效数据
for item in item_list:
if item in word_2_root:
# 找到每个功效特征词的词根，然后在one-hot向量的相应索引处进行激活
self.df[word_2_root[item]].loc[indexs] = 1
else:
print(item) # 输出没有匹配的词，进行人工处理
# 删除没有任何匹配的词根
max_value = self.df.max() # 返回df中每一列的最大值
# print("max_value:", max_value)
drop_list = list(
max_value[max_value == 0].index) # 找到最大值为0的列的索引（即没有出现过的词根）
# print("drop_list:", len(drop_list))
self.df = self.df.drop(drop_list, axis=1) # 删除未出现过的词根
def evaluate(model, batch):
"""
Evaluate the training and test set accuracy
:return:
"""
# Here goes one batch of training
q_seq, q_mask, d_seq, d_mask, target_span = get_data(
batch,
config.mode.lower() == 'train')
with torch.no_grad():
# The loss is individual loss for each pair of question, context and answer
loss, start_pos_pred, end_pos_pred = model(q_seq, q_mask, d_seq,
d_mask, target_span)
start_pos_pred = start_pos_pred.tolist()
end_pos_pred = end_pos_pred.tolist()
f1 = 0
for i, (pred_ans_start, pred_ans_end, true_ans_tokens) in enumerate(
zip(start_pos_pred, end_pos_pred, batch.ans_tokens)):
pred_ans_tokens = batch.context_tokens[i][
pred_ans_start:pred_ans_end + 1]
prediction = " ".join(pred_ans_tokens)
ground_truth = " ".join(true_ans_tokens)
f1 += f1_score(prediction, ground_truth)
f1 = f1 / (i + 1)
return f1
def load_data(train_size, test_size):
keep_channels = ['C3']
trial_len = 1.5
# X, y = get_data("../data/CLASubjectA1601083StLRHand.mat", trial_len, keep_channels)
X, y = get_data("../data/CLASubjectB1512153StLRHand.mat", trial_len,
keep_channels)
X = X[y != 3]
y = y[y != 3]
# 0 is left hand
y[y == 1] = 0
# 1 is right hand
y[y == 2] = 1
interval_len = .45
X = trim_intervals(X, .15, interval_len)
num_channels = len(keep_channels)
d2 = np.ceil(num_channels * interval_len / 0.005).astype(int)
X = X.reshape(642, d2)
X_train, X_test, y_train, y_test = train_test_split(X,
y,
train_size=train_size,
test_size=test_size)
return X_train, X_test, y_train, y_test
def load_data(self):
keep_channels = ['C3']
trial_len = 1.5
# X, y = get_data("../data/CLASubjectA1601083StLRHand.mat", trial_len, keep_channels)
# "../data/CLASubjectB1512153StLRHand.mat"
X, y = get_data(self.filename, trial_len, keep_channels)
X = X[y != 3]
y = y[y != 3]
# 0 is left hand
y[y == 1] = 0
# 1 is right hand
y[y == 2] = 1
interval_len = .45
X = trim_intervals(X, .15, interval_len)
num_channels = len(keep_channels)
d2 = np.ceil(num_channels * interval_len / 0.005).astype(int)
X = X.reshape(642, d2)
if 'pca' is self.features:
X = self.PCA(X).numpy()
if 'nn' is self.features:
self.get_nn_features()
self.X = X
self.y = y
def task7():
nn_sgd_params = {
'GMM': {
'layers_nums': (1, 3, 5, 10, 15),
'alpha': 1.0,
'mb_num': 200
},
'Peaks': {
'layers_nums': (1, 3, 5, 10, 15),
'alpha': 1.0,
'mb_num': 200
},
'SwissRoll': {
'layers_nums': (1, 3, 5, 10),
'alpha': 1.0,
'mb_num': 200
}
}
for data_set, params in nn_sgd_params.items():
X_tr, y_tr, X_te, y_te = get_data(data_set)
n = X_tr.shape[0]
layers_nums = params['layers_nums']
for layers_num in layers_nums:
layers = [n + 5 * i for i in range(layers_num)]
_ = nn_sgd(X_tr,
y_tr,
X_te,
y_te,
layers=layers,
alpha=params['alpha'],
mb_num=params['mb_num'],
max_epochs=200,
data_set=data_set)
def main():
# Prepare dataset.
raw_file = './data/y_n_all' # raw data file.
prepared_dir = './data/' # dir of prepared data(sentences cut down and labels are number)
# cut_mode = 'character'
cut_mode = 'jieba'
# cut_mode = '2-gram'
vocab_dir = './data/Bayes_vocabulary'
prepared_data,prepared_label = data_utils.prepare_data(raw_file,prepared_dir,cut_method = cut_mode, vocab_dir = vocab_dir)
print('Get prepared dataset.')
# pdb.set_trace()
# print(prepared_data)
# Get training and test dataset.
traning_dir = './data/train/'
test_dir = './data/test/'
ratio = 0
train,test = data_utils.get_data(list(zip(prepared_data,prepared_label)),traning_dir,test_dir,ratio = ratio, cut_method = cut_mode)
print('Get training and test dataset.')
# pdb.set_trace()
# print(train.data)
# train
model_name = 'Bayes'
config_dir = './model/'
train_model = init_model(model_name,config_dir)
print('Initialize the model.')
training(train_model,train)
print('Training finished.')
# store the variable.
v = {'model':train_model}
model_file = train_model.model_path + 'model.pickle'
with open(model_file,'wb') as f:
pickle.dump(v, f)
def test(self, num_test=5):
x_train, x_test, y_train, y_test = get_data()
rand = np.random.randint(low=0, high=x_test.shape[0], size=num_test)
x_test = x_test[rand]
y_test = y_test[rand]
predicts = self.model.predict(x_test)
for i in range(num_test):
compare([x_test[i], predicts[i], y_test[i]])
def train(self):
x_train, x_test, y_train, y_test = get_data()
self.model.fit(x_train,
y_train,
epochs=36,
batch_size=32,
shuffle=True,
validation_data=(x_test, y_test))
self.model.save('pretrained/model.h5')
def main():
training_images, training_labels, testing_images, testing_labels = get_data(CIFAR10_FOLDER)
(training_set, training_labels,
testing_set, testing_labels) = subsample_data(training_images, training_labels,
testing_images, testing_labels,
training_num=5000, testing_num=500)
normalize_data(training_set, testing_set)
cross_validate_all_classifiers(training_set, training_labels,
testing_set, testing_labels)
def evaluate_model(weight_name):
batch_size = 10
x_train, y_train = get_data(aug=True, name='train')
x_test, y_test = get_data(aug=False, name='test')
num_data = len(x_test)
[x_test] = img_standardization(x_train, x_test)
x_test = _parse_function(x_test, im_size=224)
dataset_test = get_dataset(x_test, y_test, batch_size, resize=False)
model = tf.keras.models.load_model('./weight/' + weight_name, compile=True)
# because evaluate() will calculate loss and metrics['accuracy'], so recompiling the loaded model is necessary
[loss, acc] = model.evaluate(dataset_test,
steps=math.ceil(num_data / batch_size))
print('TEST loss: ', loss)
print('TEST acc: ', acc)
return
def main(argv):
conf = configparser.ConfigParser()
print(conf.read(argv))
max_smi_len = conf.getint('model', 'max_smi_len')
max_seq_len = conf.getint('model', 'max_seq_len')
data_path = conf.get('model', 'data_path')
ligands = pd.read_csv(data_path + 'ligands.csv',
header=None,
names=['id', 'smi'])
proteins = pd.read_csv(data_path + 'proteins.csv',
header=None,
names=['id', 'seq'])
pairs = pd.read_csv(data_path + 'pairs.csv', header=None)
print(ligands.shape, proteins.shape, pairs.shape)
char_smi_set = json.load(open(conf.get('model', 'char_smi')))
char_seq_set = json.load(open(conf.get('model', 'char_seq')))
smi_feature, seq_feature = get_data(ligands, proteins, max_smi_len,
max_seq_len, char_smi_set,
char_seq_set)
print(smi_feature.shape, seq_feature.shape)
sess = tf.InteractiveSession(config=tf.ConfigProto(
allow_soft_placement=True))
model = CNN(filter_num=conf.getint('model', 'filter_num'),
smi_window_len=conf.getint('model', 'smi_window_len'),
seq_window_len=conf.getint('model', 'seq_window_len'),
max_smi_len=max_smi_len,
max_seq_len=max_seq_len,
char_smi_set_size=len(char_smi_set),
char_seq_set_size=len(char_seq_set),
embed_dim=conf.getint('model', 'embed_dim'))
trainy = np.asarray(pairs.iloc[:, 2]).reshape([-1, 1])
trainX = []
for idx, row in pairs.iterrows():
ligand_index = ligands[ligands.id == row[0]].index.values[0]
protein_index = proteins[proteins.id == row[1]].index.values[0]
trainX.append([smi_feature[ligand_index], seq_feature[protein_index]])
trainX = np.asarray(trainX)
model_path = os.path.join(conf.get('model', 'path', fallback='tmp'),
'all.model')
model.train(sess,
trainX,
trainy,
nb_epoch=conf.getint('model', 'num_epoch'),
batch_size=conf.getint('model', 'batch_size'),
model_path=model_path)
def main(argv):
conf = configparser.ConfigParser()
print(conf.read(argv))
max_smi_len = conf.getint('model', 'max_smi_len')
max_seq_len = conf.getint('model', 'max_seq_len')
char_smi_set = json.load(open(conf.get('model', 'char_smi')))
char_seq_set = json.load(open(conf.get('model', 'char_seq')))
data_path = conf.get('data', 'path')
data_predicted = conf.get('data', 'prediction').split(',')
sess = tf.InteractiveSession(
config=tf.ConfigProto(allow_soft_placement=True))
''' SMILES + seq '''
model = CNN(filter_num=conf.getint('model', 'filter_num'),
smi_window_len=conf.getint('model', 'smi_window_len'),
seq_window_len=conf.getint('model', 'seq_window_len'),
max_smi_len=max_smi_len,
max_seq_len=max_seq_len,
char_smi_set_size=len(char_smi_set),
char_seq_set_size=len(char_seq_set),
embed_dim=conf.getint('model', 'embed_dim'))
''' ECFP + seq '''
# model = ECFPCNN(filter_num=conf.getint('model', 'filter_num'),
# seq_window_len=conf.getint('model', 'seq_window_len'),
# char_seq_set_size=len(char_seq_set),
# embed_dim=conf.getint('model', 'embed_dim'),
# max_smi_len=max_smi_len,
# max_seq_len=max_seq_len)
model_path = os.path.join(
conf.get('model', 'path', fallback='tmp'), 'all.model')
for data_name in data_predicted:
path = data_path + data_name + '/'
ligands = pd.read_csv(path + 'ligands.csv', header=None)
proteins = pd.read_csv(path + 'proteins.csv', header=None)
smi_feature, seq_feature = get_data(
ligands, proteins, max_smi_len, max_seq_len, char_smi_set, char_seq_set)
inputs = []
for smif in smi_feature:
inputs.append([smif, seq_feature[0]])
res = model.predict(sess, np.asarray(inputs), batch_size=conf.getint(
'model', 'batch_size'), model_path=model_path)
names = [x.split('.')[0] for x in list(ligands.iloc[:, 0])]
final_data = pd.DataFrame(np.asarray(list(zip(names, res))))
final_data.to_csv(path + 'res.csv', index=None, header=None)
def main(input_file_path, output_dir_path, main_task, protect_att):
""" Runs data processing scripts to turn raw data from (../raw) into
cleaned data ready to be analyzed (saved in ../processed).
"""
logger = logging.getLogger(__name__)
logger.info('making final data set from raw data')
df = get_data(input_file_path)
logger.info('read all twits and removed duplicates')
if main_task == 'sentiment':
if protect_att == 'race':
logger.info('making sentiment-race')
pos_pos = get_attr_sentiments(df, happy, sad, 'aa',
MIN_SENTENCE_LEN)
pos_neg = get_attr_sentiments(df, happy, sad, 'wh',
MIN_SENTENCE_LEN)
neg_pos = get_attr_sentiments(df, sad, happy, 'aa',
MIN_SENTENCE_LEN)
neg_neg = get_attr_sentiments(df, sad, happy, 'wh',
MIN_SENTENCE_LEN)
else:
logger.error('not supporting this task...')
exit(-1)
elif main_task == 'mention':
if protect_att == 'race':
logger.info('making mention-race')
wh, aa = get_race(df, MIN_SENTENCE_LEN)
pos_pos, neg_pos = mention_split(aa, MIN_SENTENCE_LEN)
pos_neg, neg_neg = mention_split(wh, MIN_SENTENCE_LEN)
else:
logger.error('not supporting this task...')
exit(-1)
else:
logger.error('not supporting this task...')
exit(-1)
logger.info('done collecting data')
size = 100000
sentences = pos_pos[:size] + pos_neg[:size] + neg_pos[:size] + neg_neg[:
size]
vocab = list(set([item for sublist in sentences for item in sublist]))
id2voc = dict(enumerate(vocab))
voc2id = {v: k for k, v in id2voc.iteritems()}
to_file(output_dir_path, voc2id, vocab, pos_pos[:size], pos_neg[:size],
neg_pos[:size], neg_neg[:size])
logger.info('written to file. exiting.')
def step(model, optimizer, batch):
"""
One batch of training
:return: loss
"""
# Here goes one batch of training
q_seq, q_mask, d_seq, d_mask, target_span = get_data(
batch,
config.mode.lower() == 'train')
model.zero_grad()
# The loss is individual loss for each pair of question, context and answer
loss, _, _ = model(q_seq, q_mask, d_seq, d_mask, target_span)
loss.backward(retain_graph=True)
optimizer.step()
return loss
def step(model, optimizer, batch, params):
"""
One batch of training
:return: loss
"""
# Here goes one batch of training
q_seq, q_mask, d_seq, d_mask, target_span = get_data(
batch,
config.mode.lower() == 'train')
model.zero_grad()
# The loss is individual loss for each pair of question, context and answer
loss, start_pos, end_pos = model(q_seq, q_mask, d_seq, d_mask, target_span)
loss = torch.sum(loss)
loss.backward(retain_graph=True
) # TODO : Is this causing memory consumption to increase
clip_grad_norm_(params, config.max_grad_norm)
optimizer.step()
del start_pos, end_pos
del q_mask, q_seq, d_mask, d_seq, target_span
return loss
def objective(params):
train_data, val_data, test_data = get_data(params)
# define a keras model only based on DNA
try:
K.clear_session()
model = Janggu.create(get_model, params, train_data[0], train_data[1], name=params['name'])
model.compile(optimizer=get_opt(params['opt']), loss='binary_crossentropy',
metrics=['acc'])
hist = model.fit(train_data[0], train_data[1], epochs=params['epochs'], batch_size=64,
validation_data=val_data,
callbacks=[EarlyStopping(patience=5, restore_best_weights=True)])
except ValueError:
traceback.print_stack()
exc_type, exc_value, exc_traceback = sys.exc_info()
print(repr(traceback.extract_tb(exc_traceback)))
return {'status': 'fail'}
print('#' * 40)
for key in hist.history:
print('{}: {}'.format(key, hist.history[key][-1]))
print('#' * 40)
pred_test = model.predict(test_data[0])
pred_val = model.predict(val_data[0])
model.evaluate(val_data[0], val_data[1], callbacks=['auprc', 'auroc'], datatags=['val'])
model.evaluate(test_data[0], test_data[1], callbacks=['auprc', 'auroc'], datatags=['test'])
auprc_val = average_precision_score(val_data[1][:], pred_val)
auprc_test = average_precision_score(test_data[1][:], pred_test)
model.summary()
print('auprc_val: {:.2%}'.format(auprc_val))
print('auprc_test: {:.2%}'.format(auprc_test))
return {'loss': hist.history['val_loss'][-1], 'status': 'ok', 'all_losses': hist.history,
'auprc_val': auprc_val,
'auprc_test': auprc_test,
'model_config': model.kerasmodel.to_json(),
'model_weights': model.kerasmodel.get_weights(),
'concrete_params': params,
'modelname': model.name}
def infer(data_filepath='data/flowers.hdf5',
z_dim=128,
out_dir='gan',
n_steps=10):
G = load_model(out_dir)
val_data = get_data(data_filepath, 'train')
val_data = next(iterate_minibatches(val_data, 1))
emb_fixed, txt_fixed = val_data[1], val_data[2]
z_start = np.random.uniform(-1, 1, size=(1, z_dim))
z_end = np.random.uniform(-1, 1, size=(1, z_dim))
G.trainable = False
for i in range(n_steps + 1):
p = i / float(n_steps)
z = z_start * (1 - p) + z_end * p
fake_image = G.predict([z, emb_fixed])[0]
img = ((fake_image + 1) * 0.5)
plt.imsave("{}/fake_z_interpolation_i{}".format(out_dir, i), img)
print(i,
str(txt_fixed[0]).strip(),
file=open("{}/fake_z_interpolation.txt".format(out_dir), "a"))
def train(data_filepath='data/flowers.hdf5',
ndf=64,
ngf=128,
z_dim=128,
emb_dim=128,
lr_d=5e-5,
lr_g=5e-5,
n_iterations=int(1e6),
batch_size=64,
iters_per_checkpoint=100,
n_checkpoint_samples=16,
out_dir='wgan_gp_lr5e-5'):
global BATCH_SIZE
BATCH_SIZE = batch_size
logger = SummaryWriter(out_dir)
logger.add_scalar('d_lr', lr_d, 0)
logger.add_scalar('g_lr', lr_g, 0)
train_data = get_data(data_filepath, 'train')
val_data = get_data(data_filepath, 'valid')
data_iterator = iterate_minibatches(train_data, batch_size)
val_data_iterator = iterate_minibatches(val_data, n_checkpoint_samples)
val_data = next(val_data_iterator)
img_fixed = images_from_bytes(val_data[0])
emb_fixed = val_data[1]
txt_fixed = val_data[2]
img_shape = img_fixed[0].shape
emb_shape = emb_fixed[0].shape
print("emb shape {}".format(img_shape))
print("img shape {}".format(emb_shape))
z_shape = (z_dim, )
# plot real text for reference
log_images(img_fixed, 'real', '0', logger)
log_text(txt_fixed, 'real', '0', logger)
# build models
D = build_discriminator(img_shape, emb_shape, emb_dim, ndf)
G = build_generator(z_shape, emb_shape, emb_dim, ngf)
# build model outputs
real_inputs = Input(shape=img_shape)
txt_inputs = Input(shape=emb_shape)
z_inputs = Input(shape=(z_dim, ))
fake_samples = G([z_inputs, txt_inputs])
averaged_samples = RandomWeightedAverage()([real_inputs, fake_samples])
D_real = D([real_inputs, txt_inputs])
D_fake = D([fake_samples, txt_inputs])
D_averaged = D([averaged_samples, txt_inputs])
# The gradient penalty loss function requires the input averaged samples to
# get gradients. However, Keras loss functions can only have two arguments,
# y_true and y_pred. We get around this by making a partial() of the
# function with the averaged samples here.
loss_gp = partial(loss_gradient_penalty,
averaged_samples=averaged_samples,
gradient_penalty_weight=GRADIENT_PENALTY_WEIGHT)
# Functions need names or Keras will throw an error
loss_gp.__name__ = 'loss_gradient_penalty'
# define D graph and optimizer
G.trainable = False
D.trainable = True
D_model = Model(inputs=[real_inputs, txt_inputs, z_inputs],
outputs=[D_real, D_fake, D_averaged])
D_model.compile(optimizer=Adam(lr_d, beta_1=0.5, beta_2=0.9),
loss=[loss_wasserstein, loss_wasserstein, loss_gp])
# define D(G(z)) graph and optimizer
G.trainable = True
D.trainable = False
G_model = Model(inputs=[z_inputs, txt_inputs], outputs=D_fake)
G_model.compile(Adam(lr=lr_g, beta_1=0.5, beta_2=0.9),
loss=loss_wasserstein)
ones = np.ones((batch_size, 1), dtype=np.float32)
minus_ones = -ones
dummy = np.zeros((batch_size, 1), dtype=np.float32)
# fix a z vector for training evaluation
z_fixed = np.random.uniform(-1, 1, size=(n_checkpoint_samples, z_dim))
for i in range(n_iterations):
D.trainable = True
G.trainable = False
for j in range(N_CRITIC_ITERS):
z = np.random.normal(0, 1, size=(batch_size, z_dim))
real_batch = next(data_iterator)
losses_d = D_model.train_on_batch(
[images_from_bytes(real_batch[0]), real_batch[1], z],
[ones, minus_ones, dummy])
D.trainable = False
G.trainable = True
z = np.random.normal(0, 1, size=(batch_size, z_dim))
real_batch = next(data_iterator)
loss_g = G_model.train_on_batch([z, real_batch[1]], ones)
print("iter", i)
if (i % iters_per_checkpoint) == 0:
G.trainable = False
fake_image = G.predict([z_fixed, emb_fixed])
log_images(fake_image, 'val_fake', i, logger)
log_images(img_fixed, 'val_real', i, logger)
log_text(txt_fixed, 'val_fake', i, logger)
log_losses(losses_d, loss_g, i, logger)
# --- get settings --- #
# parse command line arguments, or use defaults
parser = utils.rgan_options_parser()
settings = vars(parser.parse_args())
# if a settings file is specified, it overrides command line arguments/defaults
if settings['settings_file']:
settings = utils.load_settings_from_file(settings)
# --- get data, split --- #
data_path = './experiments/data/' + settings['data_load_from'] + '.data.npy'
print('Loading data from', data_path)
settings["eval_single"] = False
settings["eval_an"] = False
samples, labels, index = data_utils.get_data(
settings["data"], settings["seq_length"], settings["seq_step"],
settings["num_signals"], settings["sub_id"], settings["eval_single"],
settings["eval_an"], data_path)
# --- save settings, data --- #
# no need
print('Ready to run with settings:')
for (k, v) in settings.items():
print(v, '\t', k)
# add the settings to local environment
# WARNING: at this point a lot of variables appear
locals().update(settings)
json.dump(settings,
open('./experiments/settings/' + identifier + '.txt', 'w'),
indent=0)
class myADclass():
import os
import model
import tensorflow as tf
import data_utils
import configuration
data, labels, w2idx = data_utils.get_data(configuration.config['dataset'])
configuration.config['n_words'] = len(w2idx) + 1
with tf.Session() as sess:
net = model.CNN(configuration.config, sess, w2idx)
net.train(data, labels)
import sys
sys.path.append("F:\学习\python项目")
from matplotlib import pyplot
import data_utils
import numpy
def draw_pic(data):
x = [i for i in range(len(data))]
y = data
pyplot.scatter(x, y, 1, "red")
pyplot.show()
if __name__ == "__main__":
data = data_utils.get_data("F:/工作/正常.txt")
data = data_utils.change_bad_data(data)
draw_pic(data)
def create_model(session, forward_only, beam_search):
dtype = tf.float16 if FLAGS.use_fp16 else tf.float32
#加载预训练好的词向量文件
vec_post, vec_response = data_utils.get_data(FLAGS.data_dir, FLAGS.post_vocab_size, FLAGS.response_vocab_size)
print('============-===============', vec_post)
print(len(vec_post[1]))
model = seq2seq_model.Seq2SeqModel(
FLAGS.post_vocab_size,
FLAGS.response_vocab_size,
_buckets,
FLAGS.size,
FLAGS.num_layers,
FLAGS.max_gradient_norm,
FLAGS.batch_size,
FLAGS.learning_rate,
FLAGS.learning_rate_decay_factor,
wordEmbedding=vec_post,
embedding_size=FLAGS.embedding_size,
forward_only=forward_only,
beam_search=beam_search,
beam_size=FLAGS.beam_size,
category=FLAGS.category,
use_emb=FLAGS.use_emb,
use_imemory=FLAGS.use_imemory,
use_ememory=FLAGS.use_ememory,
emotion_size=FLAGS.emotion_size,
imemory_size=FLAGS.imemory_size,
dtype=dtype)
see_variable = True
if see_variable == True:
for i in tf.all_variables():
print(i.name, i.get_shape())
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
pre_ckpt = tf.train.get_checkpoint_state(FLAGS.pretrain_dir)
#判断是否已经存在模型文件
if ckpt: #and tf.gfile.Exists(ckpt.model_checkpoint_path+".index"):
if FLAGS.load_model == 0:
print("Reading model parameters from %s" % ckpt.model_checkpoint_path)
model.saver.restore(session, ckpt.model_checkpoint_path)
else:
path = ckpt.model_checkpoint_path[:ckpt.model_checkpoint_path.find('-')+1]+str(FLAGS.load_model)
print("Reading model parameters from %s" % path)
model.saver.restore(session, path)
else:
#初始化，从新训练
if pre_ckpt:
session.run(tf.initialize_variables(model.initial_var))
if FLAGS.pretrain > -1:
path = pre_ckpt.model_checkpoint_path[:pre_ckpt.model_checkpoint_path.find('-')+1]+str(FLAGS.pretrain)
print("Reading pretrain model parameters from %s" % path)
model.pretrain_saver.restore(session, path)
else:
print("Reading pretrain model parameters from %s" % pre_ckpt.model_checkpoint_path)
model.pretrain_saver.restore(session, pre_ckpt.model_checkpoint_path)
else:
print("Created model with fresh parameters.")
session.run(tf.global_variables_initializer())
# vec_post, vec_response = data_utils.get_data(FLAGS.data_dir, FLAGS.post_vocab_size, FLAGS.response_vocab_size)
# print('vec_post:', vec_post.shape)
# print('vec_res:', vec_response)
# initvec_post = tf.constant(vec_post, dtype=dtype, name='init_wordvector_post')
#定位decoder词向量初始化，用预训练的词向量替换
initvec_response = tf.constant(vec_response, dtype=dtype, name='init_wordvector_response')
# embedding_post = [x for x in tf.trainable_variables() if x.name == 'embedding_attention_seq2seq/rnn/embedding_wrapper/embedding:0'][0]
embedding_response = [x for x in tf.trainable_variables() if x.name == 'embedding_attention_seq2seq/embedding_attention_decoder/embedding:0'][0]
print(type(embedding_response))
print(embedding_response)
# session.run(tf.assign(embedding_post, initvec_post))
# session.run(tf.assign(embedding_response, initvec_response))
# session.run(embedding_post.assign(initvec_post))
session.run(embedding_response.assign(initvec_response))
#
return model
import data_utils
import numpy as np
from sklearn.model_selection import train_test_split
import GAN
keep_channels = ['C3']
trial_len = 1.5
X, y = data_utils.get_data("../data/CLASubjectB1512153StLRHand.mat", trial_len,
keep_channels)
X = X[y != 3]
y = y[y != 3]
# 0 is left hand
y[y == 1] = 0
# 1 is right hand
y[y == 2] = 1
interval_len = .45
X = data_utils.trim_intervals(X, .15, interval_len)
num_channels = len(keep_channels)
d2 = np.ceil(num_channels * interval_len / 0.005).astype(int)
X = X.reshape(642, d2)
X_train, X_test, y_train, y_test = train_test_split(X,
y,
train_size=0.8,
test_size=0.2)
gan = GAN.GAN((X, y),
g_in=X.shape[1],
import time
import pandas as pd
import numpy as np
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.preprocessing import StandardScaler
from sklearn.impute import SimpleImputer
from sklearn.model_selection import RandomizedSearchCV
from sklearn.svm import SVR
from sklearn.metrics import mean_squared_error
from scipy.stats import uniform
from data_utils import get_data, split_data
from utils import DataFrameSelector, CombinedAttributesAdder, CustomLabelBinarizer
# laoding the dataset
housing = get_data()
# split into train_set and test_set
train_set, test_set = split_data(housing)
housing_train = train_set.drop(
"median_house_value",
axis=1) # median_house_value column contains the target values
housing_test = test_set.drop("median_house_value", axis=1)
housing_train_labels = train_set["median_house_value"].copy()
housing_test_labels = test_set["median_house_value"].copy()
# data preparation and prediction going to be done in a pipeline
# pipeline to preprocess numerical features
numerical_attributes = train_set.drop(
def main(unused_argv): train_data, valid_data = data_utils.get_data() trainer = Trainer(train_data, valid_data, data_utils.IMAGE_NEW_SIZE ** 2) trainer.run()
arg_parser = ArgumentParser(description='train network')
arg_parser.add_argument('--train', required=True,
help='HDF5 training data')
arg_parser.add_argument('--test', required=True,
help='HDF5 testing data')
arg_parser.add_argument('--epochs', type=int, default=100,
help='epochs for the training process')
arg_parser.add_argument('--batch', type=int, default=64,
help='training batch size')
arg_parser.add_argument('--seed', type=int, default=1234,
help='seed for the RNG')
arg_parser.add_argument('file', help='HDF5 to store network')
options = arg_parser.parse_args()
np.random.seed(options.seed)
(x_train, x_val, x_test,
y_train, y_val, y_test) = get_data(options.train, options.test)
input_shape = x_train.shape[1:]
output_shape = y_train.shape[1:]
model = config_model(input_shape, output_shape)
history = model.fit(x_train, y_train, epochs=options.epochs,
batch_size=options.batch, verbose=0,
validation_data=(x_val, y_val))
model.save(options.file)
hist_filename = change_path_suffix(options.file, '_hist.h5')
store_history(hist_filename, history)
loss, accuracy = model.evaluate(x_train, y_train, verbose=0)
print(f'training: loss = {loss:.3f}, accuracy = {accuracy:.3f}')
loss, accuracy = model.evaluate(x_val, y_val, verbose=0)
print(f'validation: loss = {loss:.3f}, accuracy = {accuracy:.3f}')
loss, accuracy = model.evaluate(x_test, y_test, verbose=0)
print(f'test: loss = {loss:.3f}, accuracy = {accuracy:.3f}')
from tensorflow.contrib import layers
from tqdm import *
ENCODER_UNITS = 10
DECODER_UNITS = 10
EMBEDDING_SIZE = 10
tf.app.flags.DEFINE_boolean('train_mode', True, 'Run in a training mode')
tf.app.flags.DEFINE_integer('num_epochs', 5000, 'Number of epochs')
FLAGS = tf.app.flags.FLAGS
sess = tf.Session()
questions, q_seq_length, answers_inputs, answers_targets, a_seq_length = data_utils.get_data('questions_tokenized.txt',
'answers_tokenized.txt')
inputs = tf.placeholder(tf.int32, [None, None], name='encoder_inputs')
decoder_inputs = tf.placeholder(tf.int32, [None, None], name='decoder_inputs')
decoder_targets = tf.placeholder(tf.int32, [None, None], name='decoder_targets')
vocabulary, rev_vocabulary, vocabulary_size = data_utils.initialize_vocabulary('vocabulary.txt')
embeddings = tf.Variable(tf.random_uniform([vocabulary_size, EMBEDDING_SIZE], -1.0, 1.0), dtype=tf.float32,
name='embeddings')
inputs_embed = tf.nn.embedding_lookup(embeddings, inputs)
decoder_inputs_embed = tf.nn.embedding_lookup(embeddings, decoder_inputs)
questions_seq_length_pc = tf.placeholder(tf.int32, [None], name='questions_sequence_length')
answers_seq_length_pc = tf.placeholder(tf.int32, [None], name='answers_sequence_length')
with tf.variable_scope('encoder'):
import json
dataset_used = 'MNIST'
device = 'cuda'
log_interval = 100
epochs = 5
batch_size = 32
model = Classifier(n_class=10).to(device)
model_name = 'classifier' + '_' + dataset_used + '_' + str(session_num)
if session_num > 1:
last_model_name = 'classifier' + '_' + dataset_used + '_' + str(
session_num - 1)
model.load_state_dict(torch.load('models/' + last_model_name + '.pth'))
train_data_loader, val_data_loader, train_size = get_data(dataset_used,
batch_size,
get_mean_std=False)
optimizer = optim.Adam([{
'params': model.encoder.parameters(),
'lr': 1e-5
}, {
'params': model.fc1.parameters()
}, {
'params': model.fc2.parameters()
}],
lr=1e-3)
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.95)
losses = []
for epoch in range(epochs):
epoch_loss = 0.0
def main(unused_argv):
train_data, valid_data = data_utils.get_data()
trainer = Trainer(train_data, valid_data, data_utils.IMAGE_NEW_SIZE**2)
trainer.run()
* initialize weights for generator and discriminator
* learning rate scheduler
'''
# LOAD DATA
data_type = 'sine'
options = {
'seq_length': 20,
'num_samples': 16000,
'num_signals': 1,
'freq_low': 1,
'freq_high': 5,
'amplitude_low': 0.3,
'amplitude_high': 0.9
}
data, _, _ = get_data(data_type, options)
print(data.shape)
batch_size = 8
shuffle = True
dataloader = DataLoader(data, batch_size=batch_size, shuffle=shuffle)
# INITIALIZE GEN AND DISC
z_dim = 10
device = 'cpu' # GPU
hidden_dim_g = 200
hidden_dim_d = 100
in_dim = options['num_signals']
num_layers = 2
class_names = ['no break', 'break'] # 0 is no break and 1 is break
frame_range = list(range(0, 56, 4))
num_classes = len(class_names)
part_frame_range = list(chunks(frame_range, 2))
for i, sub_range in enumerate(part_frame_range):
print()
print('PULLING PARTITION %d OF %d' % (i, len(part_frame_range) - 1))
num_train = 34 * len(sub_range) # 3400
num_val = 2 * len(sub_range) # 200
num_test = 10 * len(sub_range) # 188
(X_train, y_train,
X_val, y_val, X_test, y_test) = \
data_utils.get_data(frame_range=sub_range,
num_train=num_train,
num_validation=num_val,
num_test=num_test,
feature_list=None,
reshape_frames=False,
crop_at_constr=True,
blur_im=True)
# create tesor objects, normalize and zero center and pass into data
#loaders
# hardcoded means and standard deviation of pixel values
#data
X_train = torch.from_numpy(X_train)
y_train = torch.from_numpy(y_train)
X_val = torch.from_numpy(X_val)
y_val = torch.from_numpy(y_val)
X_test = torch.from_numpy(X_test)
y_test = torch.from_numpy(y_test)
def test_mlp(
initial_learning_rate,
learning_rate_decay,
squared_filter_length_limit,
n_epochs,
batch_size,
dropout,
results_file_name,
dataset,
use_bias,
get_mlp=get_mlp_default):
"""
The dataset is the one from the mlp demo on deeplearning.net. This training
function is lifted from there almost exactly.
:type dataset: string
:param dataset: the path of the MNIST dataset file from
http://www.iro.umontreal.ca/~lisa/deep/data/mnist/mnist.pkl.gz
"""
datasets = get_data(True)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
data_y = datasets[3]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
epoch = T.scalar()
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
learning_rate = theano.shared(np.asarray(initial_learning_rate,
dtype=theano.config.floatX))
rng = np.random.RandomState()
# construct the MLP class
classifier = get_mlp(rng, use_bias, x)
# Build the expresson for the cost function.
cost = classifier.negative_log_likelihood(y)
dropout_cost = classifier.dropout_negative_log_likelihood(y)
# Compile theano function for testing.
test_model = theano.function(inputs=[index],
outputs=classifier.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]})
# theano.printing.pydotprint(test_model, outfile="test_file.png")
test_auc = theano.function([index],
(classifier.layers[-1]).p_y_given_x,
givens={
x: test_set_x[index * batch_size: (index + 1) * batch_size],
# y: valid_set_y[index * batch_size: (index + 1) * batch_size]
})
# Compile theano function for validation.
validate_model = theano.function(inputs=[index],
outputs=classifier.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]})
# validate_auc = theano.function([index], classifier.layers[-1].p_y_given_x,
validate_auc = theano.function([index],
(classifier.layers[-1]).p_y_given_x,
givens={
x: valid_set_x[index * batch_size: (index + 1) * batch_size],
# y: valid_set_y[index * batch_size: (index + 1) * batch_size]
})
#theano.printing.pydotprint(validate_model, outfile="validate_file.png",
# var_with_name_simple=True)
# Compute gradients of the model wrt parameters
gparams = []
for param in classifier.params:
# Use the right cost function here to train with or without dropout.
gparam = T.grad(dropout_cost if dropout else cost, param)
gparams.append(gparam)
# ... and allocate mmeory for momentum'd versions of the gradient
gparams_mom = []
for param in classifier.params:
gparam_mom = theano.shared(np.zeros(param.get_value(borrow=True).shape,
dtype=theano.config.floatX))
gparams_mom.append(gparam_mom)
T_mom = 50.
# Compute momentum for the current epoch
mom = ifelse(epoch < T_mom,
# bug fix...
(epoch / T_mom) * 0.5 + (1 - epoch / T_mom) * 0.5,
0.99)
# Update the step direction using momentum
updates = {}
for gparam_mom, gparam in zip(gparams_mom, gparams):
updates[gparam_mom] = mom * gparam_mom + (1. - mom) * gparam
# ... and take a step along that direction
for param, gparam_mom in zip(classifier.params, gparams_mom):
stepped_param = param - learning_rate * updates[gparam_mom]
# This is a silly hack to constrain the norms of the rows of the weight
# matrices. This just checks if there are two dimensions to the
# parameter and constrains it if so... maybe this is a bit silly but it
# should work for now.
if param.get_value(borrow=True).ndim == 2:
squared_norms = T.sum(stepped_param ** 2, axis=1).reshape((stepped_param.shape[0], 1))
scale = T.clip(T.sqrt(squared_filter_length_limit / squared_norms), 0., 1.)
updates[param] = stepped_param * scale
else:
updates[param] = stepped_param
output = dropout_cost if dropout else cost
# grads = T.grad(output, classifier.params)
# updates = []
# for param_i, grad_i in zip(classifier.params, grads):
# updates.append((param_i, param_i - learning_rate * grad_i))
# Compile theano function for training. This returns the training cost and
# updates the model parameters.
train_model = theano.function(inputs=[epoch, index], outputs=output,
updates=updates,
on_unused_input='warn',
givens=OrderedDict({
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]
}))
# theano.printing.pydotprint(train_model, outfile="train_file.png")
# Theano function to decay the learning rate, this is separate from the
# training function because we only want to do this once each epoch instead
# of after each minibatch.
decay_learning_rate = theano.function(inputs=[], outputs=learning_rate,
updates={learning_rate: learning_rate * learning_rate_decay})
if True:
for loc, param in enumerate(classifier.params):
if param.get_value().shape == (32,5,4,5,5):
print 'saving images...'
ff = param.get_value()[:,0,0,:,:].reshape((32,25))
img = PIL.Image.fromarray(tile_raster_images(ff, (5, 5), (3, 5), tile_spacing=(1, 1)))
img.save("ff-before" + str(loc) + ".png")
ft = param.get_value()[0, :, 2, :, :].reshape((5, 25))
img = PIL.Image.fromarray(tile_raster_images(ft, (5, 5), (1, 5), tile_spacing = (1,1)))
img.save("ft-before" + str(loc) + ".png")
###############
# TRAIN MODEL #
###############
print '... training'
best_params = None
best_validation_errors = np.inf
best_iter = 0
test_score = 0.
epoch_counter = 0
start_time = time.clock()
patient = True
patience_limit = 15
# results_file = open(results_file_name, 'wb')
while epoch_counter < n_epochs and patient:
# Train this epoch
epoch_counter = epoch_counter + 1
for minibatch_index in xrange(n_train_batches):
minibatch_avg_cost = train_model(epoch_counter, minibatch_index)
# Compute loss on validation set
# validation_losses = [validate_model(i) for i in xrange(n_valid_batches)]
# this_validation_errors = np.sum(validation_losses)
probs = numpy.concatenate([validate_auc(i) for i
in xrange(n_valid_batches)])[:, 1]
fpr, tpr, _ = metrics.roc_curve(data_y[test_end:test_end + len(probs)], probs)
this_validation_errors = 1 - metrics.auc(fpr, tpr)
# only calc test error if needed
if this_validation_errors < best_validation_errors:
t_probs = numpy.concatenate([test_auc(i) for i
in xrange(n_test_batches)])[:, 1]
fpr, tpr, _ = metrics.roc_curve(data_y[train_end:train_end + len(t_probs)], t_probs)
test_score = 1 - metrics.auc(fpr, tpr)
# Report and save progress.
print "epoch {}, valid error {}, learning_rate={}{}".format(
epoch_counter, this_validation_errors * 100,
learning_rate.get_value(borrow=True),
" ** " + str(test_score) if this_validation_errors < best_validation_errors else "")
if this_validation_errors < best_validation_errors:
best_iter = epoch_counter
best_validation_errors = this_validation_errors
if False:
for loc, param in enumerate(classifier.params):
if param.get_value().shape == (32, 5, 4, 5, 5):
print 'saving images...'
ff = param.get_value()[:, 0, 0, :, :].reshape((32, 25))
img = PIL.Image.fromarray(tile_raster_images(ff, (5, 5), (3, 5), tile_spacing=(1, 1)))
img.save("ff-after" + str(epoch_counter) + ".png")
ft = param.get_value()[0, :, 2, :, :].reshape((5, 25))
img = PIL.Image.fromarray(tile_raster_images(ft, (5, 5), (1, 5), tile_spacing=(1, 1)))
img.save("ft-after" + str(epoch_counter) + ".png")
# results_file.write("{0}\n".format(this_validation_errors))
# results_file.flush()
decay_learning_rate()
if epoch_counter - best_iter > patience_limit:
patient = False
end_time = time.clock()
print(('Optimization complete. Best validation score of %f %% '
'obtained at iteration %i, with test performance %f %%') %
(best_validation_errors * 100., best_iter, test_score * 100.))
return test_score