def run():
"""
Runs the full data gathering, processing and moddeling pipeline
"""
# Get raw data
raw_data = data.get_data()
# Calculate new cases
new_cases = data.new_cases(raw_data)
# Get smoothed data
original, smoothed = data.smooth_cases(new_cases)
# Plot orginal vs smoothed
plots.original_smoothed(original, smoothed)
# Get posteriors from optimized Bayesian model
posterior = model.bayesian_model(smoothed,
p=0.95,
save_posteriors_df=True,
optimize=True)
posterior = pd.read_csv("data/most_likely_rt.csv",
parse_dates=["date"],
index_col=["date"])
# Plot Rt estimate
plots.plot_rt(posterior, annotate=True)
def main(args):
if args.verbose:
sprint(0, '[ args ]')
for k in sorted(args.__dict__.keys()):
sprint(2, "({}): {}".format(k, args.__dict__[k]))
# fin = os.path.join(args.sdir, args.data + '.pkl')
# if not args.save and os.path.exists(fin):
# sprint(1, '> loading %s' % fin)
# with open(fin, 'r') as f:
# masks = pkl.load(f)
# train, C = get_data(args.data, 'train')
# else:
# train, C = get_data(args.data, 'train')
# jdicts = [build(args, t, i) for i,t in enumerate(train)]
# masks = map(fmask, [build_mask(jdicts, c) for c in C])
# masks = [np.swapaxes(np.swapaxes(m, 0, 2), 1, 2) for m in masks]
# # masks = [np.swapaxes(np.swapaxes(build_mask(jdicts, c), 1, 3), 2, 3) for c in C]
# # # masks = map(fmask, [build_mask(jdicts, c) for c in C])
# # masks = map(fmask, masks)
# if args.save:
# if not os.path.exists(args.sdir):
# sprint(2, '! creating directory %s' % args.sdir)
# os.mkdir(args.sdir)
# fout = os.path.join(args.sdir, args.data + '.pkl')
# sprint(1, '+ writing to %s' % fout)
# with open(fout, 'w') as f:
# pkl.dump(masks, f)
train, C = get_data(args.data, 'train')
X, y = train[0]['X'] / train[0]['X'].max(), train[0]['y']
masks = map(sum,
[X[filter(lambda i: y[i] == c, range(len(y)))] for c in C])
sprint(1, '[ calculating stats')
stats = zip(*all_stats(train, masks).reshape(-1, 2))
# del train
return map(lambda x: x.reshape(-1, 28, 28), masks), stats
def test_new_cases():
data = get_data()
cases = new_cases(data)
assert isinstance(cases, pd.DataFrame)
assert len(cases.columns) == 1
assert cases.dtypes["cases"] == int
assert cases.index.nunique() == len(cases)
def main(
log_dir,
model_type,
vocab_size,
emb_size,
batch_size,
epochs,
maxlen,
min_acc,
num_samples,
oov_token,
):
_create_dir_if_not_exist(log_dir)
(str_X_train, y_train), (str_X_val, y_val), (str_X_test,
y_test) = get_data()
test_samples = np.random.choice(str_X_test, num_samples)
preprocessor = Preprocessor(maxlen=maxlen,
vocab_size=vocab_size,
oov_token=oov_token)
preprocessor.fit_on_texts(str_X_train + str_X_val + str_X_test)
X_train = preprocessor.transform(str_X_train)
X_val = preprocessor.transform(str_X_val)
X_test = preprocessor.transform(str_X_test)
# Define model
model = get_model(model_type, maxlen, vocab_size, emb_size)
model.compile(loss="binary_crossentropy",
optimizer="adam",
metrics=["accuracy"])
model.fit(
X_train,
y_train,
validation_data=(X_val, y_val),
batch_size=batch_size,
epochs=epochs,
)
score = model.evaluate(X_test, y_test, verbose=0)
print("Test accuracy:", score[1])
assert score[
1] > min_acc, f"score doesnt meet the minimum threshold {min_acc}"
test_requests = _generate_test_requests(model, preprocessor, test_samples)
_save_json(join(log_dir, "test_requests.json"), test_requests)
preprocessor.save(join(log_dir, "preprocessor.pkl"))
# HACK
# For some reason savedModel format is not working with `lstm` model
if model_type == "lstm":
model.save(join(log_dir, "model.h5"))
else:
model.save(join(log_dir, "saved_model/model"))
def main(args, data='train', dir='plot'):
if args.verbose:
sprint(0, '[ args ]')
for k in sorted(args.__dict__.keys()):
sprint(2, "({}): {}".format(k, args.__dict__[k]))
train, C = get_data(args.data, 'train')
jdicts = [build(args, t, i) for i, t in enumerate(train)]
masks = {c: build_mask(jdicts, c) for c in C}
plot(args, jdicts, masks)
return jdicts, masks, train
def test_smooth_cases():
data = get_data()
cases = new_cases(data)
processed = smooth_cases(cases)
assert isinstance(processed, tuple)
assert len(processed) == 2
original = processed[0]
smoothed = processed[1]
assert isinstance(original, pd.DataFrame)
assert isinstance(smoothed, pd.DataFrame)
assert original.shape == smoothed.shape
assert len(cases) != len(smoothed)
def main(args):
if not os.path.exists(args.params_dir):
os.makedirs(args.params_dir)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = args.log_level
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
if args.dataset == 'synthetic':
global N_VOCAB
train_data, val_data, test_data, N_VOCAB = get_data(args)
Ss, Ts, Xs, Ys = train_data['Is'], train_data['Ts'], train_data['Xs'], train_data['Ys']
dSs, dTs, dXs, dYs = val_data['Is'], val_data['Ts'], val_data['Xs'], val_data['Ys']
tSs, tTs, tXs, tYs = test_data['Is'], test_data['Ts'], test_data['Xs'], test_data['Ys']
else:
Ss, Ts, Is, As, Ys, Yis = make_train_dataset()
dSs, dTs, dIs, dAs, dYs, dYis = make_dev_dataset()
tSs, tTs, tIs, tAs, tYs, tYis = make_dev_dataset()
Xs = AtoI(As)
dXs = AtoI(dAs)
tXs = AtoI(tAs)
print('*' * 80)
print('Train samples: {} | Validation samples: {} | Test samples: {}'.format(len(Ss), len(dSs), len(tSs)))
print('Vocab size: {}'. format(N_VOCAB))
print('*' * 80)
model = build_2d_model(args)
param_fpath = '../params/params_unet' + str(args.run) + '.h5'
os.system('rm ' + param_fpath)
checkpointer = tfk.callbacks.ModelCheckpoint(param_fpath, monitor='val_acc', verbose=0,
save_weights_only=True, save_best_only=True, mode='max', period=1)
model.fit(x=[Ts, Xs], y=[Ys], validation_data=([dTs, dXs], [dYs]), batch_size=args.batch_size, epochs=args.epochs, callbacks=[checkpointer])
# Loads the weights
model.load_weights(param_fpath)
# Re-evaluate the model
print('*' * 80)
print('Validation set performance')
model.evaluate([dTs, dXs], [dYs], batch_size=args.batch_size)
print('*' * 80)
print('Test set performance')
print('-' * 80)
model.evaluate(x=[tTs, tXs], y=[tYs], batch_size=args.batch_size)
print('*' * 80)
parser.add_argument('-p', '--prepick', nargs='*')
args = parser.parse_args()
if args.sets:
import src.sets as s
if 'teams' in args.sets:
s.make_team_sets()
if 'player' in args.sets:
s.make_player_sets()
if 'picks' in args.sets:
s.make_picks_sets()
if args.data:
import src.data as d
if 'get' in args.data:
d.get_data()
if 'all' in args.data:
d.get_all_data()
if 'picks' in args.data:
d.make_pick_data()
if 'influ' in args.data:
d.plot_influential()
if args.freq:
import src.freq as f
if 'phrase' in args.freq:
f.find_freq_phrases()
if 'timeline' in args.freq:
f.plot_timelines()
if 'top' in args.freq:
f.plot_freq_bar()
def train():
data, word2index, index2word = get_data()
data = torch.from_numpy(data)
data_loader = torch.utils.data.DataLoader(data,
batch_size=Config['batch_size'],
shuffle=True)
model = PoetryModel(len(word2index), Config['embedding_dim'],
Config['hidden_dim'])
if os.path.exists(Config['model_path']):
model.load_state_dict(torch.load(Config['model_path']))
model.to(Config['device'])
optimizer = torch.optim.Adam(model.parameters(),
lr=Config['learning_rate'])
criterion = nn.CrossEntropyLoss()
print('training...')
for epoch in range(Config['epochs']):
log = [] # 日志类数据
for index, data in enumerate(data_loader):
# 转置成shape ( sequence_length, batch_size ) sequence_length 为max_len
# contiguous 为了改变底层的存储，contiguous重新按顺序存储，否则执行view会报错
data = data.long().transpose(1,
0).contiguous().to(Config['device'])
x, y = data[:-1, :], data[1:, :]
y_, _ = model(x)
optimizer.zero_grad()
loss = criterion(y_, y.view(-1))
loss.backward()
optimizer.step()
if (index + 1) % Config['print_every'] == 0:
print('epoch %d, iter %d, loss %.4f' %
(epoch, index + 1, loss.item()))
temp = {} # 用来保存过程数据
temp['epoch'] = epoch
temp['iter'] = index + 1
temp['loss'] = loss.item()
# 分别以不同的字开头续写诗
print('普通诗词')
temp['普通诗词'] = []
for w in list('山上一把火'):
poetry = generate(model, w, word2index, index2word)
print(poetry)
temp['普通诗词'].append(poetry)
# 生成藏头诗
print('藏头诗')
poerty_acrostic = generate_acrostic(model, '我想回校', word2index,
index2word)
print(poerty_acrostic)
temp['藏头诗'] = poerty_acrostic
log.append(temp)
with open(os.path.join(Config['log_json_path'],
'epoch_%d.json' % epoch),
'w',
encoding='utf-8',
errors='ignore') as f:
# 将log转字符串再转字典再保存
json.dump(json.loads(json.dumps(log)), f, ensure_ascii=False)
torch.save(model.state_dict(), Config['model_path'])
def one_class(c=0):
train, classes = get_data('mnist', 'train')
X, y = train[0]['X'], train[0]['y']
shape = train[0]['shape']
I = filter(lambda i: y[i] == c, range(len(y)))
return one_mask(X[I])
def main():
df = data.get_data()
chart = chart_all(df)
chart.save(CFG['OutputFile'], embed_options={'actions': False})
def normalize_adj(adj):
"""Symmetrically normalize adjacency matrix."""
adj = adj.coalesce()
sp_adj = sp.coo_matrix((adj.values(), adj.indices()),
shape=list(adj.size()))
rowsum = np.array(sp_adj.sum(axis=1))
d_inv_sqrt = np.power(rowsum, -0.5).flatten()
d_inv_sqrt[np.isinf(d_inv_sqrt)] = 0.
d_mat_inv_sqrt = sp.diags(d_inv_sqrt)
norm_sp_adj = sp_adj.dot(d_mat_inv_sqrt).transpose().dot(
d_mat_inv_sqrt).tocoo()
entries = norm_sp_adj.data
row = norm_sp_adj.row
col = norm_sp_adj.col
indices = torch.LongTensor([row, col])
entries = torch.FloatTensor(entries)
return sparse.FloatTensor(indices, entries, adj.size())
if __name__ == '__main__':
from src.data import Corpus, get_data, get_vocabulary, get_labels
vocab = get_vocabulary('data/20ng-vocabulary.txt')
labels = get_labels('data/20ng-labels.txt')
corpus = get_data('data/test.txt', labels)
clean_text(corpus, vocab)
adj = build_adj_matrix(corpus, vocab)
print(adj)
import os, sys
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from sklearn import metrics
import random
from src.data import get_data
from src.models import get_model
from src.train.trainer import train
from src.train.test import predict
from src.explainer import explain
#### Model Definition
data = get_data(load_existing=True)
(X_train, X_test, y_train, y_test, feature_names, label_encoder) = data
input_dim, output_dim = X_train.shape[1], y_train.shape[1]
print("input_dim, output_dim:", input_dim, output_dim)
model = get_model(input_dim, output_dim)
model = train(model, data, load_existing=True)
# predict(model, label_encoder, data)
explain(model, data, n_samples=1, submodular_pick=True)
model.load_state_dict(torch.load(save_path))
return train_losses, val_losses
if __name__ == '__main__':
from src.data import Corpus, get_data, get_vocabulary, get_labels
from src.preprocessing import clean_text, build_adj_matrix
from src.models.gcn import GCN
seed = 0
val_split = 0.1
vocab = get_vocabulary('data/20ng-vocabulary.txt')
labels = get_labels('data/20ng-labels.txt')
corpus = get_data('data/train-20news.txt', labels)
test_corpus = get_data('data/test-20news.txt', labels)
# Mask out unknown words
clean_text(corpus, vocab)
clean_text(test_corpus, vocab)
# Split validation set
corpus.shuffle(seed)
len_train = int(len(corpus) * (1 - val_split))
train_corpus = Corpus(corpus[:len_train])
val_corpus = Corpus(corpus[len_train:])
num_documents = len(train_corpus) + len(val_corpus) + len(test_corpus)
train_adj_matrix = build_adj_matrix(train_corpus,
vocab,
import os
import numpy as np
import tensorflow as tf
from tensorflow.keras.preprocessing import image
from tensorflow.keras.applications.mobilenet_v2 import preprocess_input
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from src.model import MobileNet
from src.data import get_data
from src.data import get_image_value
dim = (224,224)
model_type = 'Mobilenet'
bw = False
train_paths, train_labels = get_data('train')
train_images = np.array([get_image_value(i, dim, bw, model_type) for i in train_paths])
train_dict = dict(images = train_images, labels = train_labels)
valid_paths, valid_labels = get_data('valid')
valid_images = np.array([get_image_value(i, dim, bw, model_type) for i in valid_paths])
valid_dict = dict(images = valid_images, labels = valid_labels)
test_paths, test_labels = get_data('test')
test_images = np.array([get_image_value(i, dim, bw, model_type) for i in test_paths])
test_dict = dict(images = test_images, labels = test_labels)
data_dir = './data/images/syn/'
img_height = 224
img_width = 224
def test_get_data():
data = get_data()
assert isinstance(data, pd.DataFrame)
from src.train import train
from src.data import get_train_val_test, get_data
import sys
if __name__ == '__main__':
batch_size = 64
print(sys.argv)
model_name = sys.argv[1]
dataset = sys.argv[2]
epoch = int(sys.argv[3])
if dataset == 'fairface' or dataset == 'utkface':
train_set, val_set, test_set = get_train_val_test(
batch_size, 'rec/{}'.format(dataset))
else:
train_path = sys.argv[4]
val_path = sys.argv[5]
train_set = get_data(batch_size, train_path, True)
val_set = test_set = get_data(batch_size, val_path)
train(model_name,
dataset,
train_set,
val_set,
test_set,
batch_size,
epoch=epoch)
