def __init__(self, *args, **kw):
# ensure the parent"s __init__ is called
super(Window, self).__init__(*args, **kw)
self.conf = Configs()
self.api = YandexAPI()
self.builders = builders.Builders()
self.panel = main_panel.MainPanel(self)
self.login = login.Login(self)
self.SetTitle(self.conf.get_attr("APP_TITLE"))
self.main_pnl = self.panel.make_main_panel()
self.input = wx.TextCtrl()
self.gauge = self.panel.playback_slider
self.playlists_list = None
self.account_menu = None
self.playlist_selection = None
self.playlists = None
self.player = Player(parent=self.panel, slider=self.gauge)
self.Bind(events.FIRST_TRACK_APPEAR, self.on_first_track)
self.Bind(events.PLAYLIST_READY, self.on_playlist_ready)
self.gauge.Bind(wx.EVT_SLIDER, self.player.on_seek)
self.make_menu()
self.Center()
self.Show()
if not self.api.is_logged_in() and wx.FindWindowByName(
"login_popup") is None:
self.login.create_login_popup()
def init_config(self):
# self.config = Configs(model='default', dataset='default', epochs=10, batch_size=128)
# self.config = Configs(model='resnet', dataset='cifar10', epochs=10, batch_size=128)
# self.config = Configs(model='resnet152', dataset='mnist', epochs=10, batch_size=128)
self.config = Configs(model='default',
dataset='stock',
epochs=10000,
batch_size=128)
def __init__(self):
self.conf = Configs()
if 'RESOURCEPATH' in os.environ:
self.asset_dir = '{}/assets'.format(os.environ['RESOURCEPATH'])
else:
self.dir_name = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
self.asset_dir = os.path.join(self.dir_name, 'assets')
pass
def __init__(self, parent):
self.builders = Builders()
self.conf = Configs()
self.parent = parent
self.api = YandexAPI()
self.validation_error = wx.StaticText()
self.sizer = wx.BoxSizer()
self.dialog = wx.Dialog()
self.main_pnl = self.parent.panel
pass
def __init__(self):
self.conf = Configs()
self.client = self.login()
self.win = None
self.list_type = None
self.playlists_list = None
self.updating_thread = None
if 'RESOURCEPATH' in os.environ:
self.cache = '{}/cache'.format(os.environ['RESOURCEPATH'])
else:
self.dirName = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
self.cache = os.path.join(self.dirName, 'cache')
pass
class Builders(object):
def __init__(self):
self.conf = Configs()
if 'RESOURCEPATH' in os.environ:
self.asset_dir = '{}/assets'.format(os.environ['RESOURCEPATH'])
else:
self.dir_name = os.path.dirname(
os.path.dirname(os.path.abspath(__file__)))
self.asset_dir = os.path.join(self.dir_name, 'assets')
pass
def button_builder(self, parent, label, name):
button = wx.Button(parent, label=label, name=name)
button.SetForegroundColour(self.conf.get_attr("TEXT_COLOR"))
button.SetBackgroundColour(self.conf.get_attr("BACKGROUND_COLOR"))
return button
def input_builder(self, parent, name, size=(250, 20), value=""):
input_field = wx.TextCtrl(parent, size=size, name=name, value=value)
input_field.SetBackgroundColour(self.conf.get_attr("BACKGROUND_COLOR"))
input_field.SetForegroundColour(self.conf.get_attr("TEXT_COLOR"))
return input_field
def static_text_builder(self, parent, label):
text = wx.StaticText(parent, label=label)
text.SetForegroundColour(self.conf.get_attr("TEXT_COLOR"))
return text
def build_bitmap_button(self, btn_dict):
handler = btn_dict['handler']
name = btn_dict['name']
parent = btn_dict['parent']
size_h = btn_dict['size_h']
size_w = btn_dict['size_w']
img = wx.Image(os.path.join(self.asset_dir, btn_dict['bitmap']),
wx.BITMAP_TYPE_PNG)
img = img.Scale(size_w, size_h, wx.IMAGE_QUALITY_HIGH)
img = wx.Bitmap(img)
btn = buttons.GenBitmapButton(parent=parent, bitmap=img, name=name)
btn.Bind(wx.EVT_BUTTON, handler)
return btn
def build_playlist_cover(self, playlist_dict):
cover_name = playlist_dict['cover']
cover_parent = playlist_dict['parent']
cover = Animation('{}/{}.gif'.format(self.asset_dir, cover_name))
cover_ctrl = AnimationCtrl(cover_parent, -1, cover, name=cover_name)
return cover_ctrl
def __init__(self, parent):
self.conf = Configs()
self.parent = parent
self.api = YandexAPI()
self.builders = Builders()
if 'RESOURCEPATH' in os.environ:
self.asset_dir = '{}/assets'.format(os.environ['RESOURCEPATH'])
else:
self.dir_name = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
self.asset_dir = os.path.join(self.dir_name, 'assets')
pass
self.main_pnl = wx.Panel(parent)
self.song_band = None
self.song_name = None
self.playback_slider = None
self.play_pause_btn = None
self.next_track = None
self.prev_track = None
def main(configs: Configs = None, data_loader: DataLoader = None):
"""
main function for data processor from raw files SAP to tables in database
to be consumed by forecast model
usage example:
$ python spike-challenge/src/make_dataset.py
"""
if configs is None:
configs = Configs('default_config.yaml')
if data_loader is None:
data_loader = DataLoader()
data_loader.load_data()
class MainPanel(object):
def __init__(self, parent):
self.conf = Configs()
self.parent = parent
self.api = YandexAPI()
self.builders = Builders()
if 'RESOURCEPATH' in os.environ:
self.asset_dir = '{}/assets'.format(os.environ['RESOURCEPATH'])
else:
self.dir_name = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
self.asset_dir = os.path.join(self.dir_name, 'assets')
pass
self.main_pnl = wx.Panel(parent)
self.song_band = None
self.song_name = None
self.playback_slider = None
self.play_pause_btn = None
self.next_track = None
self.prev_track = None
def toggle_gauge_slider(self, slider_type='slider'):
if slider_type == 'gauge':
control = wx.Gauge(
self.main_pnl,
range=20,
size=(self.parent.GetSize()[0], 5),
style=wx.GA_HORIZONTAL | wx.GA_SMOOTH
).Pulse()
else:
control = wx.Slider(
self.main_pnl,
size=(self.parent.GetSize()[0], 5),
minValue=0,
maxValue=20
)
control.Disable()
return control
def make_main_panel(self):
self.main_pnl.SetBackgroundColour(self.conf.get_attr("BACKGROUND_COLOR"))
self.playback_slider = self.toggle_gauge_slider()
main_sizer = wx.BoxSizer(wx.VERTICAL)
toolbar = self.build_audio_bar()
main_sizer.Add(
self.playback_slider,
0,
wx.ALL | wx.EXPAND,
0
)
main_sizer.AddStretchSpacer(1)
main_sizer.Add(
self.playlist_list(),
0
)
main_sizer.AddStretchSpacer(1)
main_sizer.Add(
toolbar,
0
)
self.main_pnl.SetSizer(main_sizer)
return self.main_pnl
def build_audio_bar(self):
"""
Builds the audio bar controls
"""
audio_bar_sizer = wx.BoxSizer(wx.HORIZONTAL)
self.prev_track = self.builders.build_bitmap_button({
'bitmap': 'player_prev.png',
'handler': self.parent.on_prev,
'name': 'prev',
'parent': self.main_pnl,
'size_h': 30,
'size_w': 30
})
self.song_band = self.builders.static_text_builder(parent=self.main_pnl, label="")
self.song_name = self.builders.static_text_builder(parent=self.main_pnl, label="")
song_sizer = wx.BoxSizer(wx.VERTICAL)
song_sizer.Add(
self.song_band,
0
)
song_sizer.Add(
self.song_name,
0
)
audio_bar_sizer.Add(
self.prev_track,
0
)
img = wx.Image(os.path.join(self.asset_dir, "player_play.png"), wx.BITMAP_TYPE_ANY)
img = img.Scale(30, 30, wx.IMAGE_QUALITY_HIGH)
img = wx.Bitmap(img)
self.play_pause_btn = buttons.GenBitmapToggleButton(self.main_pnl, bitmap=img, name="play")
self.play_pause_btn.Enable(False)
img = wx.Image(os.path.join(self.asset_dir, "player_pause.png"), wx.BITMAP_TYPE_ANY)
img = img.Scale(30, 30, wx.IMAGE_QUALITY_HIGH)
img = wx.Bitmap(img)
self.play_pause_btn.SetBitmapSelected(img)
self.play_pause_btn.SetInitialSize()
self.play_pause_btn.Bind(wx.EVT_BUTTON, self.parent.on_play)
audio_bar_sizer.Add(
self.play_pause_btn,
0
)
self.next_track = self.builders.build_bitmap_button({
'bitmap': 'player_next.png',
'handler': self.parent.on_next,
'name': 'next',
'parent': self.main_pnl,
'size_h': 30,
'size_w': 30
})
audio_bar_sizer.Add(
self.next_track,
0
)
audio_bar_sizer.AddSpacer(5)
audio_bar_sizer.Add(
song_sizer,
0
)
self.next_track.Disable()
self.prev_track.Disable()
return audio_bar_sizer
def enable_play_button(self):
self.play_pause_btn.Enable(True)
def playlist_list(self):
playlists_sizer = wx.BoxSizer(wx.HORIZONTAL)
playlists_sizer.AddSpacer(5)
playlists = [
{
"cover": "playlistOfTheDay",
"parent": self.main_pnl
},
{
"cover": "neverHeard",
"parent": self.main_pnl
},
{
"cover": "missedLikes",
"parent": self.main_pnl
},
{
"cover": "recentTracks",
"parent": self.main_pnl
}
]
for playlist in playlists:
item = self.builders.build_playlist_cover(playlist)
playlists_sizer.Add(
item,
0
)
playlists_sizer.AddSpacer(5)
item.Bind(wx.EVT_ENTER_WINDOW, self.on_hover)
item.Bind(wx.EVT_LEAVE_WINDOW, self.on_unhover)
item.Bind(wx.EVT_LEFT_DOWN, self.on_click)
return playlists_sizer
def on_click(self, event):
playlist_type = event.EventObject.GetName()
self.api.preparation(playlist_type, self.parent)
def on_unhover(self, event):
event.EventObject.Stop()
def on_hover(self, event):
event.EventObject.Play()
def main(args):
# process config
c = Configs(args.config)
ROOT = os.environ['TENSOROFLOW']
model_path = '%s/examples/model/basic_nmt/model' % ROOT
PAD = c.const['PAD']
EOS = c.const['EOS']
train_step = c.option['train_step']
max_time = c.option['max_time']
batch_size = c.option['batch_size']
vocabulary_size = c.option['vocabulary_size']
input_embedding_size = c.option['embedding_size']
hidden_units = c.option['hidden_units']
source_train_data_path = c.data['source_train_data']
target_train_data_path = c.data['target_train_data']
source_valid_data_path = c.data['source_valid_data']
target_valid_data_path = c.data['target_valid_data']
source_test_data_path = c.data['source_test_data']
target_test_data_path = c.data['target_test_data']
# read data
source_dictionary, source_reverse_dictionary = build_dictionary(
read_words(source_train_data_path), vocabulary_size)
source_train_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_train_data_path)
]
target_dictionary, target_reverse_dictionary = build_dictionary(
read_words(target_train_data_path), vocabulary_size)
target_train_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_train_data_path)
]
source_valid_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_valid_data_path)
]
target_valid_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_valid_data_path)
]
source_test_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_test_data_path)
]
target_test_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_test_data_path)
]
# placeholder
encoder_inputs = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='encoder_inputs')
decoder_inputs = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='decoder_inputs')
decoder_labels = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='decoder_labels')
# embed
embeddings = tf.Variable(tf.random_uniform(
[vocabulary_size, input_embedding_size], -1.0, 1.0),
dtype=tf.float32,
name='embeddings')
encoder_inputs_embedded = tf.nn.embedding_lookup(embeddings,
encoder_inputs)
decoder_inputs_embedded = tf.nn.embedding_lookup(embeddings,
decoder_inputs)
# encoder
encoder_units = hidden_units
encoder_cell = tf.contrib.rnn.LSTMCell(encoder_units)
_, encoder_final_state = tf.nn.dynamic_rnn(encoder_cell,
encoder_inputs_embedded,
dtype=tf.float32,
time_major=True)
# decoder
decoder_units = encoder_units
decoder_cell = tf.contrib.rnn.LSTMCell(decoder_units)
decoder_output, decoder_final_state = tf.nn.dynamic_rnn(
decoder_cell,
decoder_inputs_embedded,
initial_state=encoder_final_state,
scope="plain_decoder",
dtype=tf.float32,
time_major=True)
decoder_logits = tf.contrib.layers.linear(decoder_output, vocabulary_size)
decoder_prediction = tf.argmax(
decoder_logits, 2) # max_time: axis=0, batch: axis=1, vocab: axis=2
# optimizer
stepwise_cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
labels=tf.one_hot(decoder_labels,
depth=vocabulary_size,
dtype=tf.float32),
logits=decoder_logits,
)
loss = tf.reduce_mean(stepwise_cross_entropy)
train_op = tf.train.AdamOptimizer().minimize(loss)
saver = tf.train.Saver()
batch_idx = {'train': 0, 'valid': 0, 'test': 0}
with tf.Session() as sess:
if args.mode == 'train':
# train
loss_freq = train_step // 100
loss_log = []
loss_suffix = ''
es = EarlyStopper(max_size=5, edge_threshold=0.1)
m = Monitor(train_step)
sess.run(tf.global_variables_initializer())
for i in range(train_step):
m.monitor(i, loss_suffix)
source_train_batch, _ = batchnize(source_train_datas,
batch_size,
batch_idx['train'])
target_train_batch, batch_idx['train'] = batchnize(
target_train_datas, batch_size, batch_idx['train'])
batch_data = seq2seq(source_train_batch, target_train_batch,
max_time, vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
sess.run(fetches=[train_op, loss], feed_dict=feed_dict)
if i % loss_freq == 0:
source_valid_batch, _ = batchnize(source_valid_datas,
batch_size,
batch_idx['valid'])
target_valid_batch, batch_idx['valid'] = batchnize(
target_valid_datas, batch_size, batch_idx['valid'])
batch_data = seq2seq(source_valid_batch,
target_valid_batch, max_time,
vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
loss_val = sess.run(fetches=loss, feed_dict=feed_dict)
loss_log.append(loss_val)
loss_suffix = 'loss: %f' % loss_val
es_status = es(loss_val)
if i > train_step // 2 and es_status:
print('early stopping at step: %d' % i)
break
saver.save(sess, model_path)
print('save at %s' % model_path)
plt.plot(np.arange(len(loss_log)) * loss_freq, loss_log)
plt.savefig('%s_loss.png' % model_path)
elif args.mode == 'eval':
saver.restore(sess, model_path)
print('load from %s' % model_path)
else:
raise
# evaluate
loss_val = []
input_vectors = None
predict_vectors = None
for i in range(len(source_test_datas) // batch_size + 1):
source_test_batch, _ = batchnize(source_test_datas, batch_size,
batch_idx['test'])
target_test_batch, batch_idx['test'] = batchnize(
target_test_datas, batch_size, batch_idx['test'])
batch_data = seq2seq(source_test_batch, target_test_batch,
max_time, vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
pred = sess.run(fetches=decoder_prediction, feed_dict=feed_dict)
if predict_vectors is None:
predict_vectors = pred.T
else:
predict_vectors = np.vstack((predict_vectors, pred.T))
input_ = batch_data['encoder_inputs']
if input_vectors is None:
input_vectors = input_.T
else:
input_vectors = np.vstack((input_vectors, input_.T))
loss_val.append(sess.run(fetches=loss, feed_dict=feed_dict))
input_sentences = ''
predict_sentences = ''
for i, (input_vector, predict_vector) in enumerate(
zip(input_vectors[:len(source_test_datas)],
predict_vectors[:len(target_test_datas)])):
input_sentences += ' '.join([
source_reverse_dictionary[vector] for vector in input_vector
if not vector == PAD
])
predict_sentences += ' '.join([
target_reverse_dictionary[vector] for vector in predict_vector
if not vector == PAD
])
if i < len(source_test_datas) - 1:
input_sentences += '\n'
predict_sentences += '\n'
evaluate_input_path = '%s.evaluate_input' % model_path
evaluate_predict_path = '%s.evaluate_predict' % model_path
with open(evaluate_input_path, 'w') as f1, \
open(evaluate_predict_path, 'w') as f2:
f1.write(input_sentences)
f2.write(predict_sentences)
print('input sequences at {}'.format(evaluate_input_path))
print('predict sequences at {}'.format(evaluate_predict_path))
print('mean of loss: %f' % np.mean(loss_val))
print('finish.')
def main():
parser = argparse.ArgumentParser(description="Shared Model")
parser.add_argument('--test_prompt_id', type=int, default=1, help='prompt id of test essay set')
parser.add_argument('--seed', type=int, default=12, help='set random seed')
args = parser.parse_args()
test_prompt_id = args.test_prompt_id
seed = args.seed
np.random.seed(seed)
tf.random.set_seed(seed)
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
print("Test prompt id is {} of type {}".format(test_prompt_id, type(test_prompt_id)))
print("Seed: {}".format(seed))
configs = Configs()
data_path = configs.DATA_PATH
train_path = data_path + '/train.tsv'
dev_path = data_path + '/dev.tsv'
pretrained_embedding = configs.PRETRAINED_EMBEDDING
embedding_path = configs.EMBEDDING_PATH
embedding_dim = configs.EMBEDDING_DIM
vocab_size = configs.VOCAB_SIZE
epochs = configs.EPOCHS
batch_size = configs.BATCH_SIZE
read_configs = {
'train_path': train_path,
'dev_path': dev_path,
'vocab_size': vocab_size
}
word_vocab = read_word_vocab(read_configs)
print('vocab complete')
train_data_src, train_data_tgt, dev_data_src, dev_data_tgt = \
read_essays_words(read_configs, word_vocab, test_prompt_id)
if pretrained_embedding:
embedd_dict, embedd_dim, _ = load_word_embedding_dict(embedding_path)
embedd_matrix = build_embedd_table(word_vocab, embedd_dict, embedd_dim, caseless=True)
embed_table = [embedd_matrix]
else:
embed_table = None
max_sentlen = max(train_data_src['max_sentlen'], train_data_tgt['max_sentlen'],
dev_data_src['max_sentlen'], dev_data_tgt['max_sentlen'])
max_sentnum = max(train_data_src['max_sentnum'], train_data_tgt['max_sentnum'],
dev_data_src['max_sentnum'], dev_data_tgt['max_sentnum'])
print('max sent length: {}'.format(max_sentlen))
print('max sent num: {}'.format(max_sentnum))
train_data_src['y_scaled'] = get_scaled_down_scores(train_data_src['data_y'], train_data_src['prompt_ids'])
train_data_tgt['y_scaled'] = get_scaled_down_scores(train_data_tgt['data_y'], train_data_tgt['prompt_ids'])
dev_data_src['y_scaled'] = get_scaled_down_scores(dev_data_src['data_y'], dev_data_src['prompt_ids'])
dev_data_tgt['y_scaled'] = get_scaled_down_scores(dev_data_tgt['data_y'], dev_data_tgt['prompt_ids'])
X_train_src = pad_hierarchical_text_sequences(train_data_src['words'], max_sentnum, max_sentlen)
X_train_tgt = pad_hierarchical_text_sequences(train_data_tgt['words'], max_sentnum, max_sentlen)
X_dev_src = pad_hierarchical_text_sequences(dev_data_src['words'], max_sentnum, max_sentlen)
X_dev_tgt = pad_hierarchical_text_sequences(dev_data_tgt['words'], max_sentnum, max_sentlen)
X_train_src = X_train_src.reshape((X_train_src.shape[0], X_train_src.shape[1] * X_train_src.shape[2]))
X_train_tgt = X_train_tgt.reshape((X_train_tgt.shape[0], X_train_tgt.shape[1] * X_train_tgt.shape[2]))
X_dev_src = X_dev_src.reshape((X_dev_src.shape[0], X_dev_src.shape[1] * X_dev_src.shape[2]))
X_dev_tgt = X_dev_tgt.reshape((X_dev_tgt.shape[0], X_dev_tgt.shape[1] * X_dev_tgt.shape[2]))
Y_train_src = np.array(train_data_src['y_scaled'])
Y_train_tgt = np.array(train_data_tgt['y_scaled'])
Y_dev_src = np.array(dev_data_src['y_scaled'])
Y_dev_tgt = np.array(dev_data_tgt['y_scaled'])
train_src_batches = batch_generator(
[X_train_src, Y_train_src], batch_size)
train_tgt_batches = batch_generator(
[X_train_tgt, Y_train_tgt], batch_size)
disc_loss_fn = tf.keras.losses.SparseCategoricalCrossentropy()
score_loss_fn = tf.keras.losses.MeanSquaredError()
optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
shared_model = SharedModelV2(len(word_vocab), max_sentnum, max_sentlen, embedding_dim, embed_table)
steps = (X_train_src.shape[0] // batch_size) * epochs
evaluator = SharedModelEvaluatorV2(test_prompt_id, X_dev_src, X_train_tgt, X_dev_tgt, dev_data_src['prompt_ids'],
train_data_tgt['prompt_ids'], dev_data_tgt['prompt_ids'], Y_dev_src, Y_train_tgt,
Y_dev_tgt)
evaluator.evaluate(shared_model, 0, print_info=True)
for step in range(steps):
src_label = tf.zeros((batch_size, 1))
tgt_label = tf.ones((batch_size, 1))
X_train_src_batch, Y_train_src_batch = next(train_src_batches)
X_train_tgt_batch, Y_train_tgt_batch = next(train_tgt_batches)
X_both = tf.concat([X_train_src_batch, X_train_tgt_batch], axis=0)
label_both = tf.concat([src_label, tgt_label], axis=0)
fe_loss, score_loss, disc_loss = full_train_step(X_train_src_batch, Y_train_src_batch, X_both, label_both,
shared_model, score_loss_fn, disc_loss_fn, optimizer)
current_step = step + 1
if current_step % (steps//epochs) == 0:
print(
"fe loss (for one batch) at step %d: %.4f"
% (current_step, float(fe_loss))
)
print(
"score loss (for one batch) at step %d: %.4f"
% (current_step, float(score_loss))
)
print(
"disc loss (for one batch) at step %d: %.4f"
% (current_step, float(disc_loss))
)
print('steps', steps)
print('step', current_step)
print('epochs', epochs)
print('batch_size', batch_size)
print('Evaluating epoch', current_step/(steps//epochs))
if step == 0:
evaluator.evaluate(shared_model, 0)
else:
evaluator.evaluate(shared_model, current_step/(steps//epochs))
evaluator.print_final_info()
class Window(wx.Frame):
def __init__(self, *args, **kw):
# ensure the parent"s __init__ is called
super(Window, self).__init__(*args, **kw)
self.conf = Configs()
self.api = YandexAPI()
self.builders = builders.Builders()
self.panel = main_panel.MainPanel(self)
self.login = login.Login(self)
self.SetTitle(self.conf.get_attr("APP_TITLE"))
self.main_pnl = self.panel.make_main_panel()
self.input = wx.TextCtrl()
self.gauge = self.panel.playback_slider
self.playlists_list = None
self.account_menu = None
self.playlist_selection = None
self.playlists = None
self.player = Player(parent=self.panel, slider=self.gauge)
self.Bind(events.FIRST_TRACK_APPEAR, self.on_first_track)
self.Bind(events.PLAYLIST_READY, self.on_playlist_ready)
self.gauge.Bind(wx.EVT_SLIDER, self.player.on_seek)
self.make_menu()
self.Center()
self.Show()
if not self.api.is_logged_in() and wx.FindWindowByName(
"login_popup") is None:
self.login.create_login_popup()
def make_menu(self):
self.account_menu = wx.Menu()
if self.api.is_logged_in():
logout = self.account_menu.Append(1, "&Logout\tCtrl-L",
"Logout from account")
self.Bind(wx.EVT_MENU, self.login.on_logout_menu, logout)
player_menu = wx.Menu()
self.playlists_list = wx.Menu(wx.ID_ANY)
self.playlist_selection = player_menu.Append(wx.ID_ANY, "Playlists",
self.playlists_list)
if self.api.is_logged_in():
self.playlists = self.api.get_play_lists_list()
for playlist in self.playlists:
self.Bind(
wx.EVT_MENU, self.on_list_select,
self.playlists_list.Append(wx.ID_ANY, playlist['name']))
pass
else:
self.playlist_selection.Enable(False)
pass
help_menu = wx.Menu()
about_item = help_menu.Append(wx.ID_ABOUT)
menu_bar = wx.MenuBar()
menu_bar.Append(self.account_menu, "Account")
menu_bar.Append(player_menu, "Player")
menu_bar.Append(help_menu, "Help")
self.SetMenuBar(menu_bar)
self.Bind(wx.EVT_MENU, self.on_about, about_item)
def on_first_track(self, event):
playlist_type = event.playlist_type
self.gauge.Enable()
self.panel.enable_play_button()
self.player.load_playlist(playlist_type)
pass
def on_playlist_ready(self, event):
playlist_name = event.playlist_name
playlist_type = event.playlist_type
self.player.load_playlist(playlist_type)
notify(playlist_name, "Playlist is ready", "")
pass
def on_prev(self, event):
self.player.on_prev(event)
pass
def on_next(self, event):
self.player.on_next(event)
pass
def on_play(self, event):
if not event.GetIsDown():
self.on_pause(event)
return
self.player.on_play(event)
pass
def on_pause(self, event):
self.player.on_pause(event)
pass
def on_list_select(self, event):
playlist_label = event.GetEventObject().GetLabelText(event.GetId())
playlist_type = None
for playlist in self.playlists:
if playlist['name'] == playlist_label:
playlist_type = playlist['type']
self.api.preparation(playlist_type, self)
def on_exit(self, event):
self.Close()
def on_about(self, event):
wx.MessageBox("This is a wxPython Hello World sample",
"About Hello World 2", wx.OK | wx.ICON_ASTERISK)
import os
from configs.consts import ProjectPath
from configs.configs import Configs
__all__ = ['ProjectPath', 'cfg']
cfg = Configs()
def _mkdir_dirs(*dir_paths):
for dir_path in dir_paths:
if not os.path.exists(dir_path):
os.makedirs(dir_path)
_mkdir_dirs(
ProjectPath.LOGS_DIR.value,
ProjectPath.CKPTS_DIR.value,
ProjectPath.VOC_CKPTS_DIR.value,
ProjectPath.CONSOLE_LOGS_DIR.value,
ProjectPath.TB_LOGS_DIR.value,
)
def main(args):
# process config
c = Configs(args.config)
ROOT = os.environ['TENSOROFLOW']
output = c.option.get('output', 'examples/model/buf')
model_directory = '%s/%s' % (ROOT, output)
model_path = '%s/model' % model_directory
dictionary_path = {
'source': '%s/source_dictionary.pickle' % model_directory,
'source_reverse':
'%s/source_reverse_dictionary.pickle' % model_directory,
'target': '%s/target_dictionary.pickle' % model_directory,
'target_reverse':
'%s/target_reverse_dictionary.pickle' % model_directory
}
PAD = c.const['PAD']
BOS = c.const['BOS']
EOS = c.const['EOS']
train_step = c.option['train_step']
max_time = c.option['max_time']
batch_size = c.option['batch_size']
vocabulary_size = c.option['vocabulary_size']
input_embedding_size = c.option['embedding_size']
hidden_units = c.option['hidden_units']
layers = c.option['layers']
source_train_data_path = c.data['source_train_data']
target_train_data_path = c.data['target_train_data']
source_valid_data_path = c.data['source_valid_data']
target_valid_data_path = c.data['target_valid_data']
source_test_data_path = c.data['source_test_data']
target_test_data_path = c.data['target_test_data']
# initialize output directory
if pathlib.Path(model_directory).exists():
print('Warning: model %s is exists.')
print('Old model will be overwritten.')
while True:
print('Do you wanna continue? [yes|no]')
command = input('> ')
if command == 'yes':
shutil.rmtree(model_directory)
break
elif command == 'no':
sys.exit()
else:
print('You can only input "yes" or "no".')
print('Make new model: %s' % model_directory)
pathlib.Path(model_directory).mkdir()
# read data
if args.mode == 'train':
source_dictionary, source_reverse_dictionary = build_dictionary(
read_words(source_train_data_path), vocabulary_size)
source_train_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_train_data_path)
]
target_dictionary, target_reverse_dictionary = build_dictionary(
read_words(target_train_data_path), vocabulary_size)
target_train_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_train_data_path)
]
source_valid_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_valid_data_path)
]
target_valid_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_valid_data_path)
]
if args.debug:
source_train_datas = source_train_datas[:1000]
target_train_datas = source_train_datas[:1000]
else:
with open(dictionary_path['source'], 'rb') as f1, \
open(dictionary_path['source_reverse'], 'rb') as f2, \
open(dictionary_path['target'], 'rb') as f3, \
open(dictionary_path['target_reverse'], 'rb') as f4:
source_dictionary = pickle.load(f1)
source_reverse_dictionary = pickle.load(f2)
target_dictionary = pickle.load(f3)
target_reverse_dictionary = pickle.load(f4)
source_test_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_test_data_path)
]
target_test_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_test_data_path)
]
# placeholder
encoder_inputs = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='encoder_inputs')
decoder_inputs = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='decoder_inputs')
decoder_labels = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='decoder_labels')
# embed
embeddings = tf.Variable(tf.random_uniform(
[vocabulary_size, input_embedding_size], -1.0, 1.0),
dtype=tf.float32,
name='embeddings')
encoder_inputs_embedded = tf.nn.embedding_lookup(embeddings,
encoder_inputs)
decoder_inputs_embedded = tf.nn.embedding_lookup(embeddings,
decoder_inputs)
# encoder with bidirection
encoder_units = hidden_units
encoder_layers_fw = [
tf.contrib.rnn.LSTMCell(size) for size in [encoder_units] * layers
]
encoder_cell_fw = tf.contrib.rnn.MultiRNNCell(encoder_layers_fw)
encoder_layers_bw = [
tf.contrib.rnn.LSTMCell(size) for size in [encoder_units] * layers
]
encoder_cell_bw = tf.contrib.rnn.MultiRNNCell(encoder_layers_bw)
(encoder_output_fw,
encoder_output_bw), encoder_state = tf.nn.bidirectional_dynamic_rnn(
encoder_cell_fw,
encoder_cell_bw,
encoder_inputs_embedded,
dtype=tf.float32,
time_major=True)
encoder_outputs = tf.concat((encoder_output_fw, encoder_output_bw), 2)
encoder_state = tuple(
tf.contrib.rnn.LSTMStateTuple(
tf.concat((encoder_state[0][layer].c,
encoder_state[1][layer].c), 1),
tf.concat((encoder_state[0][layer].h,
encoder_state[1][layer].h), 1))
for layer in range(layers))
# decoder with attention
decoder_units = encoder_units * 2
attention_units = decoder_units
decoder_layers = [
tf.contrib.rnn.LSTMCell(size) for size in [decoder_units] * layers
]
cell = tf.contrib.rnn.MultiRNNCell(decoder_layers)
sequence_length = tf.cast([max_time] * batch_size, dtype=tf.int32)
beam_width = 1
tiled_encoder_outputs = tf.contrib.seq2seq.tile_batch(
encoder_outputs, multiplier=beam_width)
tiled_encoder_final_state = tf.contrib.seq2seq.tile_batch(
encoder_state, multiplier=beam_width)
tiled_sequence_length = tf.contrib.seq2seq.tile_batch(
sequence_length, multiplier=beam_width)
attention_mechanism = tf.contrib.seq2seq.LuongAttention(
num_units=attention_units,
memory=tiled_encoder_outputs,
memory_sequence_length=tiled_sequence_length)
attention_cell = tf.contrib.seq2seq.AttentionWrapper(
cell, attention_mechanism, attention_layer_size=256)
decoder_initial_state = attention_cell.zero_state(dtype=tf.float32,
batch_size=batch_size *
beam_width)
decoder_initial_state = decoder_initial_state.clone(
cell_state=tiled_encoder_final_state)
if args.mode == 'train':
helper = tf.contrib.seq2seq.TrainingHelper(
inputs=decoder_inputs_embedded,
sequence_length=tf.cast([max_time] * batch_size, dtype=tf.int32),
time_major=True)
elif args.mode == 'eval':
"""
helper = tf.contrib.seq2seq.TrainingHelper(
inputs=decoder_inputs_embedded,
sequence_length=tf.cast([max_time] * batch_size, dtype=tf.int32),
time_major=True)
"""
helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(
embedding=embeddings,
start_tokens=tf.tile([BOS], [batch_size]),
end_token=EOS)
decoder = tf.contrib.seq2seq.BasicDecoder(
cell=attention_cell,
helper=helper,
initial_state=decoder_initial_state)
decoder_outputs = tf.contrib.seq2seq.dynamic_decode(
decoder=decoder,
output_time_major=True,
impute_finished=False,
maximum_iterations=max_time)
decoder_logits = tf.contrib.layers.linear(decoder_outputs[0][0],
vocabulary_size)
decoder_prediction = tf.argmax(
decoder_logits, 2) # max_time: axis=0, batch: axis=1, vocab: axis=2
# optimizer
stepwise_cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
labels=tf.one_hot(decoder_labels,
depth=vocabulary_size,
dtype=tf.float32),
logits=decoder_logits,
)
loss = tf.reduce_mean(stepwise_cross_entropy)
regularizer = 0.0 * tf.nn.l2_loss(decoder_outputs[0][0])
train_op = tf.train.AdamOptimizer().minimize(loss + regularizer)
saver = tf.train.Saver()
minibatch_idx = {'train': 0, 'valid': 0, 'test': 0}
with tf.Session() as sess:
if args.mode == 'train':
# train
global_max_step = train_step * (
len(source_train_datas) // batch_size + 1)
loss_freq = global_max_step // 100 if global_max_step > 100 else 1
loss_log = []
batch_loss_log = []
loss_suffix = ''
es = EarlyStopper(max_size=5, edge_threshold=0.1)
m = Monitor(global_max_step)
log = Logger('%s/log' % model_directory)
sess.run(tf.global_variables_initializer())
global_step = 0
stop_flag = False
for batch in range(train_step):
if stop_flag:
break
current_batch_loss_log = []
while True: # minibatch process
m.monitor(global_step, loss_suffix)
source_train_batch, _ = batchnize(source_train_datas,
batch_size,
minibatch_idx['train'])
target_train_batch, minibatch_idx['train'] = batchnize(
target_train_datas, batch_size, minibatch_idx['train'])
batch_data = seq2seq(source_train_batch,
target_train_batch,
max_time,
vocabulary_size,
reverse=True)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
sess.run(fetches=[train_op, loss], feed_dict=feed_dict)
if global_step % loss_freq == 0:
source_valid_batch, _ = batchnize(
source_valid_datas, batch_size,
minibatch_idx['valid'])
target_valid_batch, minibatch_idx['valid'] = batchnize(
target_valid_datas, batch_size,
minibatch_idx['valid'])
batch_data = seq2seq(source_valid_batch,
target_valid_batch,
max_time,
vocabulary_size,
reverse=True)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
loss_val = sess.run(fetches=loss, feed_dict=feed_dict)
loss_log.append(loss_val)
current_batch_loss_log.append(loss_val)
loss_suffix = 'loss: %f' % loss_val
global_step += 1
if minibatch_idx['train'] == 0:
batch_loss = np.mean(current_batch_loss_log)
batch_loss_log.append(batch_loss)
loss_msg = 'Batch: {}/{}, batch loss: {}'.format(
batch + 1, train_step, batch_loss)
print(loss_msg)
log(loss_msg)
es_status = es(batch_loss)
if batch > train_step // 2 and es_status:
print('early stopping at step: %d' % global_step)
stop_flag = True
break
# save tf.graph and variables
saver.save(sess, model_path)
print('save at %s' % model_path)
# save plot of loss
plt.plot(np.arange(len(loss_log)) * loss_freq, loss_log)
plt.savefig('%s_global_loss.png' % model_path)
plt.figure()
plt.plot(np.arange(len(batch_loss_log)), batch_loss_log)
plt.savefig('%s_batch_loss.png' % model_path)
# save dictionary
with open(dictionary_path['source'], 'wb') as f1, \
open(dictionary_path['source_reverse'], 'wb') as f2, \
open(dictionary_path['target'], 'wb') as f3, \
open(dictionary_path['target_reverse'], 'wb') as f4:
pickle.dump(source_dictionary, f1)
pickle.dump(source_reverse_dictionary, f2)
pickle.dump(target_dictionary, f3)
pickle.dump(target_reverse_dictionary, f4)
elif args.mode == 'eval':
saver.restore(sess, model_path)
print('load from %s' % model_path)
else:
raise # args.mode should be train or eval
# evaluate
loss_val = []
input_vectors = None
predict_vectors = None
for i in range(len(source_test_datas) // batch_size + 1):
source_test_batch, _ = batchnize(source_test_datas, batch_size,
minibatch_idx['test'])
target_test_batch, minibatch_idx['test'] = batchnize(
target_test_datas, batch_size, minibatch_idx['test'])
batch_data = seq2seq(source_test_batch,
target_test_batch,
max_time,
vocabulary_size,
reverse=True)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
pred = sess.run(fetches=decoder_prediction, feed_dict=feed_dict)
if predict_vectors is None:
predict_vectors = pred.T
else:
predict_vectors = np.vstack((predict_vectors, pred.T))
input_ = batch_data['encoder_inputs']
if input_vectors is None:
input_vectors = input_.T
else:
input_vectors = np.vstack((input_vectors, input_.T))
loss_val.append(sess.run(fetches=loss, feed_dict=feed_dict))
input_sentences = ''
predict_sentences = ''
ignore_token = EOS
for i, (input_vector, predict_vector) in enumerate(
zip(input_vectors[:len(source_test_datas)],
predict_vectors[:len(target_test_datas)])):
input_sentences += ' '.join([
source_reverse_dictionary[vector] for vector in input_vector
if not vector == ignore_token
])
predict_sentences += ' '.join([
target_reverse_dictionary[vector] for vector in predict_vector
if not vector == ignore_token
])
if i < len(source_test_datas) - 1:
input_sentences += '\n'
predict_sentences += '\n'
evaluate_input_path = '%s.evaluate_input' % model_path
evaluate_predict_path = '%s.evaluate_predict' % model_path
with open(evaluate_input_path, 'w') as f1, \
open(evaluate_predict_path, 'w') as f2:
f1.write(input_sentences)
f2.write(predict_sentences)
print('input sequences at {}'.format(evaluate_input_path))
print('predict sequences at {}'.format(evaluate_predict_path))
print('mean of loss: %f' % np.mean(loss_val))
print('finish.')
def main():
parser = argparse.ArgumentParser(description="PAES_attributes model")
parser.add_argument('--test_prompt_id',
type=int,
default=1,
help='prompt id of test essay set')
parser.add_argument('--seed', type=int, default=12, help='set random seed')
parser.add_argument('--attribute_name',
type=str,
help='name of the attribute to be trained on')
args = parser.parse_args()
test_prompt_id = args.test_prompt_id
attribute_name = args.attribute_name
seed = args.seed
np.random.seed(seed)
tf.random.set_seed(seed)
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
print("Test prompt id is {} of type {}".format(test_prompt_id,
type(test_prompt_id)))
print("Attribute: {}".format(attribute_name))
print("Seed: {}".format(seed))
configs = Configs()
data_path = configs.DATA_PATH
train_path = data_path + str(test_prompt_id) + '/train.pk'
dev_path = data_path + str(test_prompt_id) + '/dev.pk'
test_path = data_path + str(test_prompt_id) + '/test.pk'
features_path = configs.FEATURES_PATH
readability_path = configs.READABILITY_PATH
vocab_size = configs.VOCAB_SIZE
epochs = configs.EPOCHS
batch_size = configs.BATCH_SIZE
read_configs = {
'train_path': train_path,
'dev_path': dev_path,
'test_path': test_path,
'features_path': features_path,
'readability_path': readability_path,
'vocab_size': vocab_size
}
pos_vocab = read_pos_vocab(read_configs)
train_data, dev_data, test_data = read_essays_single_score(
read_configs, pos_vocab, attribute_name)
max_sentlen = max(train_data['max_sentlen'], dev_data['max_sentlen'],
test_data['max_sentlen'])
max_sentnum = max(train_data['max_sentnum'], dev_data['max_sentnum'],
test_data['max_sentnum'])
print('max sent length: {}'.format(max_sentlen))
print('max sent num: {}'.format(max_sentnum))
train_data['y_scaled'] = get_single_scaled_down_score(
train_data['data_y'], train_data['prompt_ids'], attribute_name)
dev_data['y_scaled'] = get_single_scaled_down_score(
dev_data['data_y'], dev_data['prompt_ids'], attribute_name)
test_data['y_scaled'] = get_single_scaled_down_score(
test_data['data_y'], test_data['prompt_ids'], attribute_name)
X_train_pos = pad_hierarchical_text_sequences(train_data['pos_x'],
max_sentnum, max_sentlen)
X_dev_pos = pad_hierarchical_text_sequences(dev_data['pos_x'], max_sentnum,
max_sentlen)
X_test_pos = pad_hierarchical_text_sequences(test_data['pos_x'],
max_sentnum, max_sentlen)
X_train_pos = X_train_pos.reshape(
(X_train_pos.shape[0], X_train_pos.shape[1] * X_train_pos.shape[2]))
X_dev_pos = X_dev_pos.reshape(
(X_dev_pos.shape[0], X_dev_pos.shape[1] * X_dev_pos.shape[2]))
X_test_pos = X_test_pos.reshape(
(X_test_pos.shape[0], X_test_pos.shape[1] * X_test_pos.shape[2]))
X_train_linguistic_features = np.array(train_data['features_x'])
X_dev_linguistic_features = np.array(dev_data['features_x'])
X_test_linguistic_features = np.array(test_data['features_x'])
X_train_readability = np.array(train_data['readability_x'])
X_dev_readability = np.array(dev_data['readability_x'])
X_test_readability = np.array(test_data['readability_x'])
Y_train = np.array(train_data['y_scaled'])
Y_dev = np.array(dev_data['y_scaled'])
Y_test = np.array(test_data['y_scaled'])
print('================================')
print('X_train_pos: ', X_train_pos.shape)
print('X_train_readability: ', X_train_readability.shape)
print('X_train_ling: ', X_train_linguistic_features.shape)
print('Y_train: ', Y_train.shape)
print('================================')
print('X_dev_pos: ', X_dev_pos.shape)
print('X_dev_readability: ', X_dev_readability.shape)
print('X_dev_ling: ', X_dev_linguistic_features.shape)
print('Y_dev: ', Y_dev.shape)
print('================================')
print('X_test_pos: ', X_test_pos.shape)
print('X_test_readability: ', X_test_readability.shape)
print('X_test_ling: ', X_test_linguistic_features.shape)
print('Y_test: ', Y_test.shape)
print('================================')
model = build_PAES(len(pos_vocab), max_sentnum, max_sentlen,
X_train_readability.shape[1],
X_train_linguistic_features.shape[1], configs)
dev_features_list = [
X_dev_pos, X_dev_linguistic_features, X_dev_readability
]
test_features_list = [
X_test_pos, X_test_linguistic_features, X_test_readability
]
evaluator = Evaluator(test_prompt_id, dev_data['prompt_ids'],
test_data['prompt_ids'], dev_features_list,
test_features_list, Y_dev, Y_test, attribute_name)
evaluator.evaluate(model, -1, print_info=True)
for ii in range(epochs):
print('Epoch %s/%s' % (str(ii + 1), epochs))
start_time = time.time()
model.fit(
[X_train_pos, X_train_linguistic_features, X_train_readability],
Y_train,
batch_size=batch_size,
epochs=1,
verbose=0,
shuffle=True)
tt_time = time.time() - start_time
print("Training one epoch in %.3f s" % tt_time)
evaluator.evaluate(model, ii + 1)
evaluator.print_final_info()
def main(args):
tf.reset_default_graph()
# process config
c = Configs(args.config)
ROOT = os.environ['TENSOROFLOW']
model_path = '%s/examples/model/multi_layer_seq2seq/model' % ROOT
PAD = c.const['PAD']
EOS = c.const['EOS']
train_step = c.option['train_step']
max_time = c.option['max_time']
batch_size = c.option['batch_size']
vocabulary_size = c.option['vocabulary_size']
input_embedding_size = c.option['embedding_size']
hidden_units = c.option['hidden_units']
layers = c.option['layers']
datas = []
# placeholder
encoder_inputs = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='encoder_inputs')
decoder_inputs = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='decoder_inputs')
decoder_labels = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='decoder_labels')
# embed
embeddings = tf.Variable(tf.random_uniform(
[vocabulary_size, input_embedding_size], -1.0, 1.0),
dtype=tf.float32,
name='embeddings')
encoder_inputs_embedded = tf.nn.embedding_lookup(embeddings,
encoder_inputs)
decoder_inputs_embedded = tf.nn.embedding_lookup(embeddings,
decoder_inputs)
# encoder
encoder_units = hidden_units
encoder_layers = [
tf.contrib.rnn.LSTMCell(size) for size in [encoder_units] * layers
]
encoder_cell = tf.contrib.rnn.MultiRNNCell(encoder_layers)
encoder_output, encoder_final_state = tf.nn.dynamic_rnn(
encoder_cell,
encoder_inputs_embedded,
dtype=tf.float32,
time_major=True)
del encoder_output
# decoder
decoder_units = encoder_units
decoder_layers = [
tf.contrib.rnn.LSTMCell(size) for size in [decoder_units] * layers
]
decoder_cell = tf.contrib.rnn.MultiRNNCell(decoder_layers)
decoder_output, decoder_final_state = tf.nn.dynamic_rnn(
decoder_cell,
decoder_inputs_embedded,
initial_state=encoder_final_state,
scope="plain_decoder",
dtype=tf.float32,
time_major=True)
decoder_logits = tf.contrib.layers.linear(decoder_output, vocabulary_size)
decoder_prediction = tf.argmax(
decoder_logits, 2) # max_time: axis=0, batch: axis=1, vocab: axis=2
# optimizer
stepwise_cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
labels=tf.one_hot(decoder_labels,
depth=vocabulary_size,
dtype=tf.float32),
logits=decoder_logits,
)
loss = tf.reduce_mean(stepwise_cross_entropy)
train_op = tf.train.AdamOptimizer().minimize(loss)
saver = tf.train.Saver()
with tf.Session() as sess:
if args.mode == 'train':
# train
loss_freq = train_step // 100
loss_log = []
loss_suffix = ''
es = EarlyStopper(max_size=5, edge_threshold=0.1)
m = Monitor(train_step)
sess.run(tf.global_variables_initializer())
for i in range(train_step):
m.monitor(i, loss_suffix)
batch_data = through(datas, max_time, batch_size,
vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
sess.run(fetches=[train_op, loss], feed_dict=feed_dict)
if i % loss_freq == 0:
batch_data = through(datas, max_time, batch_size,
vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
loss_val = sess.run(fetches=loss, feed_dict=feed_dict)
loss_log.append(loss_val)
loss_suffix = 'loss: %f' % loss_val
es_status = es(loss_val)
if i > train_step // 2 and es_status:
print('early stopping at step: %d' % i)
break
saver.save(sess, model_path)
print('save at %s' % model_path)
plt.plot(np.arange(len(loss_log)) * loss_freq, loss_log)
plt.savefig('%s_loss.png' % model_path)
elif args.mode == 'eval':
saver.restore(sess, model_path)
print('load from %s' % model_path)
else:
raise
# evaluate
batch_data = through(datas, max_time, batch_size, vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
pred = sess.run(fetches=decoder_prediction, feed_dict=feed_dict)
input_ = batch_data['encoder_inputs']
loss_val = sess.run(fetches=loss, feed_dict=feed_dict)
print('input sequences...\n{}'.format(input_))
print('predict sequences...\n{}'.format(pred))
print('loss: %f' % loss_val)
print('finish.')
class Login(object):
def __init__(self, parent):
self.builders = Builders()
self.conf = Configs()
self.parent = parent
self.api = YandexAPI()
self.validation_error = wx.StaticText()
self.sizer = wx.BoxSizer()
self.dialog = wx.Dialog()
self.main_pnl = self.parent.panel
pass
def create_login_popup(self):
self.dialog = dialog = wx.Dialog(self.parent,
wx.ID_ANY,
"",
style=wx.DEFAULT_DIALOG_STYLE
| wx.RESIZE_BORDER,
name="login_popup")
self.sizer = sizer = wx.BoxSizer(wx.VERTICAL)
dialog.BackgroundColour = self.conf.get_attr("BACKGROUND_COLOR")
text = self.builders.static_text_builder(dialog, label="Please login")
font = text.GetFont()
font.PointSize += 10
font = font.Bold()
text.SetFont(font)
login_label = self.builders.static_text_builder(dialog, label="Login:"******"login_input")
password_label = self.builders.static_text_builder(dialog,
label="Password:"******"password_input")
login_button = self.builders.button_builder(dialog, "Login",
"login_button")
login_button.Bind(wx.EVT_BUTTON, self.on_login)
self.validation_error = self.builders.static_text_builder(dialog,
label="")
self.validation_error.SetForegroundColour(wx.RED)
sizer.Add(text, 0,
wx.ALIGN_CENTER_HORIZONTAL | wx.LEFT | wx.RIGHT | wx.TOP, 20)
sizer.Add(login_label, 0,
wx.ALIGN_CENTER_HORIZONTAL | wx.LEFT | wx.RIGHT | wx.TOP, 5)
sizer.Add(login_input, 0,
wx.ALIGN_CENTER_HORIZONTAL | wx.LEFT | wx.RIGHT | wx.TOP, 10)
sizer.Add(password_label, 0,
wx.ALIGN_CENTER_HORIZONTAL | wx.LEFT | wx.RIGHT | wx.TOP, 5)
sizer.Add(password_input, 0,
wx.ALIGN_CENTER_HORIZONTAL | wx.LEFT | wx.RIGHT | wx.TOP, 10)
sizer.Add(login_button, 0,
wx.ALIGN_CENTER_HORIZONTAL | wx.LEFT | wx.RIGHT | wx.TOP, 10)
self.sizer.Add(self.validation_error, 1,
wx.ALIGN_CENTER_HORIZONTAL | wx.TOP, 5)
dialog.SetSizer(sizer)
dialog.Bind(wx.EVT_CLOSE, self.on_popup_close)
dialog.Center()
dialog.Show()
def on_popup_close(self, event):
if self.api.is_logged_in() is True:
event.Skip(True)
else:
self.parent.on_exit(event)
def on_login(self, event):
login_button = self.parent.FindWindowByName("login_button")
login_button.Disable()
login = self.parent.FindWindowByName("login_input").GetValue()
password = self.parent.FindWindowByName("password_input").GetValue()
try:
self.api.login(login=login, password=password)
popup = self.parent.FindWindowByName("login_popup")
popup.Destroy()
notify(subtitle="Hello " + self.api.get_display_name())
self.parent.playlist_selection.Enable(True)
self.parent.make_menu()
except BadRequest as e:
self.validation_error.SetLabel(str(e))
size = self.dialog.GetSize()
self.dialog.SetInitialSize()
self.dialog.SetSize(size)
login_button.Enable()
event.Skip()
def on_logout_menu(self, event):
self.api.logout()
self.create_login_popup()
self.parent.make_menu()
class XHNF(object):
def __init__(self):
self.config = None
self.data = None
self.model = None
self.start_date = '20140101'
self.end_date = '20160501'
def init_config(self):
# self.config = Configs(model='default', dataset='default', epochs=10, batch_size=128)
# self.config = Configs(model='resnet', dataset='cifar10', epochs=10, batch_size=128)
# self.config = Configs(model='resnet152', dataset='mnist', epochs=10, batch_size=128)
self.config = Configs(model='default',
dataset='stock',
epochs=10000,
batch_size=128)
def init_data(self):
self.data = get_data(self.config.dataset, self.start_date,
self.end_date)
def init_predict_data(self):
now = datetime.datetime.now()
end_date = datetime.datetime.strftime(now, "%Y%m%d")
start = now + datetime.timedelta(days=-7)
start_date = datetime.datetime.strftime(start, "%Y%m%d")
self.data = get_predict_data(self.config.dataset, start_date, end_date)
def init_test_data(self):
now = datetime.datetime.now()
end_date = datetime.datetime.strftime(now, "%Y%m%d")
start = now + datetime.timedelta(days=-10)
start_date = datetime.datetime.strftime(start, "%Y%m%d")
self.data = get_test_data(self.config.dataset, start_date, end_date)
def init_model(self):
self.model = get_model(self.config.model, self.data.input_shape,
self.data.nb_classes)
self.model.compile(loss=keras.losses.categorical_crossentropy,
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
def init(self):
self.init_config()
self.init_data()
self.init_model()
def init_predict(self):
self.init_config()
self.init_predict_data()
self.init_model()
def init_test(self):
self.init_config()
self.init_test_data()
self.init_model()
def getModelFileName(self):
records_dir = 'records' \
+ os.path.sep + self.config.get_model() \
+ os.path.sep + self.config.get_dataset()
weight_name = "Weight_C" + str(self.data.nb_classes)
weight_name = weight_name + "_I" + str(
self.data.input_shape[0]) + "_" + str(
self.data.input_shape[1]) + "_" + str(self.data.input_shape[2])
weight_name = weight_name + ".h5"
file_name = records_dir \
+ os.path.sep + weight_name
#print("input_shape",str(self.data.input_shape), "nb_classes", str(self.data.nb_classes))
if os.path.exists(records_dir) == False:
os.makedirs(records_dir)
return file_name
def train_network(self):
filepath = self.getModelFileName()
checkpoint = ModelCheckpoint(filepath, monitor='val_loss' \
, save_weights_only=True,verbose=1,save_best_only=True, period=1)
earlyStopping = EarlyStopping(monitor='val_loss',
patience=6,
verbose=0,
mode='auto')
if os.path.exists(filepath):
self.model.load_weights(filepath)
print("checkpoint_loaded")
print('start training, start:', self.start_date, '; end:',
self.end_date)
if self.data.x_train.shape[0] <= 0:
print("train shape:", self.data.x_train.shape)
print("do not read any train date, so just next loop.")
return False
self.model.fit(self.data.x_train,
self.data.y_train,
batch_size=self.config.batch_size,
epochs=self.config.epochs,
verbose=1,
validation_data=(self.data.x_test, self.data.y_test),
callbacks=[checkpoint, earlyStopping])
score = self.model.evaluate(self.data.x_test,
self.data.y_test,
verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
return True
def train(self):
print('do train.')
train_status = True
for i in range(1000):
print('train run', i)
if self.data == None:
end = datetime.datetime.strptime(self.end_date, '%Y%m%d')
now = datetime.datetime.now()
if end < now:
end = end + datetime.timedelta(days=20)
else:
print('all done, last train date ', self.end_date)
break
end_date_str = datetime.datetime.strftime(end, "%Y%m%d")
if train_status == True:
self.start_date = self.end_date
self.end_date = end_date_str
self.data = get_data(self.config.dataset, self.start_date,
self.end_date)
# del self.model
# self.init_model()
train_status = self.train_network()
del self.data
self.data = None
def save_predict(self, y):
now = datetime.datetime.now()
now_str = datetime.datetime.strftime(now, "%Y%m%d")
records_dir = 'money' \
+ os.path.sep + now_str
if os.path.exists(records_dir) == False:
os.makedirs(records_dir)
file_name = records_dir + os.path.sep + now_str + "_predict.csv"
np.savetxt(file_name,
y,
fmt='%s',
delimiter=',',
header='date,code,predict,real')
def display_predict(self, need_save=False):
y_predict = self.model.predict(self.data.x_train)
if self.data.nb_classes == 2:
y_res = y_predict[:, 1] - y_predict[:, 0]
y_res = y_res.reshape(y_res.shape[0], 1)
yt = np.delete(self.data.y_train, (0), axis=1)
y = np.concatenate((self.data.y_test, y_res), axis=1)
y = np.concatenate((y, yt), axis=1)
elif self.data.nb_classes == 4:
y_res = y_predict[:,
0] - y_predict[:,
1] - y_predict[:,
2] - y_predict[:,
3]
y_res = y_res.reshape(y_res.shape[0], 1)
yt = np.delete(self.data.y_train, (1, 2, 3), axis=1)
y = np.concatenate((self.data.y_test, y_res), axis=1)
y = np.concatenate((y, yt), axis=1)
else:
y = np.concatenate((self.data.y_test, y_predict), axis=1)
y = np.concatenate((y, self.data.y_train), axis=1)
print(y)
if need_save == True:
self.save_predict(y)
def test(self):
print('do test.')
filepath = self.getModelFileName()
if os.path.exists(filepath):
self.model.load_weights(filepath)
print("checkpoint_loaded")
if self.data.x_train.shape[0] <= 0:
print("train shape:", self.data.x_train.shape)
print("do not read any test date, so just exit.")
return
self.display_predict()
score = self.model.evaluate(self.data.x_train,
self.data.y_train,
verbose=1)
print('Test loss:', score[0])
print('Test accuracy:', score[1])
def predict(self):
print('do predict.')
filepath = self.getModelFileName()
if os.path.exists(filepath):
self.model.load_weights(filepath)
print("checkpoint_loaded")
if self.data.x_train.shape[0] <= 0:
print("train shape:", self.data.x_train.shape)
print("do not read any predict date, so just exit.")
return
self.display_predict(True)
def do_train(self):
self.init()
self.train()
def do_test(self):
self.init_test()
self.test()
def do_predict(self):
self.init_predict()
self.predict()
"""
if configs is None:
configs = Configs('default_config.yaml')
if data_loader is None:
data_loader = DataLoader()
data_loader.load_data()
# parser = Parser(
# data=data_loader
# )
if __name__ == '__main__':
START = time.time()
# find .env automatically by walking up directories until it's found, then
# load up the .env entries as environment variables
load_dotenv(find_dotenv())
configs = None
if len(sys.argv) > 1:
configs = Configs(sys.argv[1])
main(configs)
END = time.time()
Logger.info('Script completed in', '%i seconds' % int(END - START),
__file__)
def main(args):
# process config
c = Configs(args.config)
ROOT = os.environ['TENSOROFLOW']
model_directory = '%s/examples/model/multi_layer_nmt' % ROOT
model_path = '%s/model' % model_directory
dictionary_path = {
'source': '%s/source_dictionary.pickle' % model_directory,
'source_reverse':
'%s/source_reverse_dictionary.pickle' % model_directory,
'target': '%s/target_dictionary.pickle' % model_directory,
'target_reverse':
'%s/target_reverse_dictionary.pickle' % model_directory
}
PAD = c.const['PAD']
EOS = c.const['EOS']
train_step = c.option['train_step']
max_time = c.option['max_time']
batch_size = c.option['batch_size']
vocabulary_size = c.option['vocabulary_size']
input_embedding_size = c.option['embedding_size']
hidden_units = c.option['hidden_units']
layers = c.option['layers']
source_train_data_path = c.data['source_train_data']
target_train_data_path = c.data['target_train_data']
source_valid_data_path = c.data['source_valid_data']
target_valid_data_path = c.data['target_valid_data']
source_test_data_path = c.data['source_test_data']
target_test_data_path = c.data['target_test_data']
# read data
if args.mode == 'train':
source_dictionary, source_reverse_dictionary = build_dictionary(
read_words(source_train_data_path), vocabulary_size)
source_train_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_train_data_path)
]
target_dictionary, target_reverse_dictionary = build_dictionary(
read_words(target_train_data_path), vocabulary_size)
target_train_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_train_data_path)
]
source_valid_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_valid_data_path)
]
target_valid_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_valid_data_path)
]
if args.debug:
source_train_datas = source_train_datas[:1000]
target_train_datas = source_train_datas[:1000]
else:
with open(dictionary_path['source'], 'rb') as f1, \
open(dictionary_path['source_reverse'], 'rb') as f2, \
open(dictionary_path['target'], 'rb') as f3, \
open(dictionary_path['target_reverse'], 'rb') as f4:
source_dictionary = pickle.load(f1)
source_reverse_dictionary = pickle.load(f2)
target_dictionary = pickle.load(f3)
target_reverse_dictionary = pickle.load(f4)
source_test_datas = [
sentence_to_onehot(lines, source_dictionary)
for lines in read_data(source_test_data_path)
]
target_test_datas = [
sentence_to_onehot(lines, target_dictionary)
for lines in read_data(target_test_data_path)
]
# placeholder
encoder_inputs = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='encoder_inputs')
decoder_inputs = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='decoder_inputs')
decoder_labels = tf.placeholder(shape=(None, None),
dtype=tf.int32,
name='decoder_labels')
# embed
embeddings = tf.Variable(tf.random_uniform(
[vocabulary_size, input_embedding_size], -1.0, 1.0),
dtype=tf.float32,
name='embeddings')
encoder_inputs_embedded = tf.nn.embedding_lookup(embeddings,
encoder_inputs)
decoder_inputs_embedded = tf.nn.embedding_lookup(embeddings,
decoder_inputs)
# encoder
encoder_units = hidden_units
encoder_layers = [
tf.contrib.rnn.LSTMCell(size) for size in [encoder_units] * layers
]
encoder_cell = tf.contrib.rnn.MultiRNNCell(encoder_layers)
encoder_output, encoder_final_state = tf.nn.dynamic_rnn(
encoder_cell,
encoder_inputs_embedded,
dtype=tf.float32,
time_major=True)
del encoder_output
# decoder
decoder_units = encoder_units
decoder_layers = [
tf.contrib.rnn.LSTMCell(size) for size in [decoder_units] * layers
]
decoder_cell = tf.contrib.rnn.MultiRNNCell(decoder_layers)
decoder_output, decoder_final_state = tf.nn.dynamic_rnn(
decoder_cell,
decoder_inputs_embedded,
initial_state=encoder_final_state,
scope="plain_decoder",
dtype=tf.float32,
time_major=True)
decoder_logits = tf.contrib.layers.linear(decoder_output, vocabulary_size)
decoder_prediction = tf.argmax(
decoder_logits, 2) # max_time: axis=0, batch: axis=1, vocab: axis=2
# optimizer
stepwise_cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
labels=tf.one_hot(decoder_labels,
depth=vocabulary_size,
dtype=tf.float32),
logits=decoder_logits,
)
loss = tf.reduce_mean(stepwise_cross_entropy)
train_op = tf.train.AdamOptimizer().minimize(loss)
saver = tf.train.Saver()
minibatch_idx = {'train': 0, 'valid': 0, 'test': 0}
with tf.Session() as sess:
if args.mode == 'train':
# train
global_max_step = train_step * (
len(source_train_datas) // batch_size + 1)
loss_freq = global_max_step // 100 if global_max_step > 100 else 1
loss_log = []
batch_loss_log = []
loss_suffix = ''
es = EarlyStopper(max_size=5, edge_threshold=0.1)
m = Monitor(global_max_step)
sess.run(tf.global_variables_initializer())
global_step = 0
stop_flag = False
for batch in range(train_step):
if stop_flag:
break
current_batch_loss_log = []
while True: # minibatch process
m.monitor(global_step, loss_suffix)
source_train_batch, _ = batchnize(source_train_datas,
batch_size,
minibatch_idx['train'])
target_train_batch, minibatch_idx['train'] = batchnize(
target_train_datas, batch_size, minibatch_idx['train'])
batch_data = seq2seq(source_train_batch,
target_train_batch, max_time,
vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
sess.run(fetches=[train_op, loss], feed_dict=feed_dict)
if global_step % loss_freq == 0:
source_valid_batch, _ = batchnize(
source_valid_datas, batch_size,
minibatch_idx['valid'])
target_valid_batch, minibatch_idx['valid'] = batchnize(
target_valid_datas, batch_size,
minibatch_idx['valid'])
batch_data = seq2seq(source_valid_batch,
target_valid_batch, max_time,
vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
loss_val = sess.run(fetches=loss, feed_dict=feed_dict)
loss_log.append(loss_val)
current_batch_loss_log.append(loss_val)
loss_suffix = 'loss: %f' % loss_val
es_status = es(loss_val)
if batch > train_step // 2 and es_status:
print('early stopping at step: %d' % global_step)
stop_flag = True
break
global_step += 1
if minibatch_idx['train'] == 0:
batch_loss = np.mean(current_batch_loss_log)
batch_loss_log.append(batch_loss)
print('Batch: {}/{}, batch loss: {}'.format(
batch + 1, train_step, batch_loss))
break
# save tf.graph and variables
saver.save(sess, model_path)
print('save at %s' % model_path)
# save plot of loss
plt.plot(np.arange(len(loss_log)) * loss_freq, loss_log)
plt.savefig('%s_global_loss.png' % model_path)
plt.figure()
plt.plot(np.arange(len(batch_loss_log)), batch_loss_log)
plt.savefig('%s_batch_loss.png' % model_path)
# save dictionary
with open(dictionary_path['source'], 'wb') as f1, \
open(dictionary_path['source_reverse'], 'wb') as f2, \
open(dictionary_path['target'], 'wb') as f3, \
open(dictionary_path['target_reverse'], 'wb') as f4:
pickle.dump(source_dictionary, f1)
pickle.dump(source_reverse_dictionary, f2)
pickle.dump(target_dictionary, f3)
pickle.dump(target_reverse_dictionary, f4)
elif args.mode == 'eval':
saver.restore(sess, model_path)
print('load from %s' % model_path)
else:
raise # args.mode should be train or eval
# evaluate
loss_val = []
input_vectors = None
predict_vectors = None
for i in range(len(source_test_datas) // batch_size + 1):
source_test_batch, _ = batchnize(source_test_datas, batch_size,
minibatch_idx['test'])
target_test_batch, minibatch_idx['test'] = batchnize(
target_test_datas, batch_size, minibatch_idx['test'])
batch_data = seq2seq(source_test_batch, target_test_batch,
max_time, vocabulary_size)
feed_dict = {
encoder_inputs: batch_data['encoder_inputs'],
decoder_inputs: batch_data['decoder_inputs'],
decoder_labels: batch_data['decoder_labels']
}
pred = sess.run(fetches=decoder_prediction, feed_dict=feed_dict)
if predict_vectors is None:
predict_vectors = pred.T
else:
predict_vectors = np.vstack((predict_vectors, pred.T))
input_ = batch_data['encoder_inputs']
if input_vectors is None:
input_vectors = input_.T
else:
input_vectors = np.vstack((input_vectors, input_.T))
loss_val.append(sess.run(fetches=loss, feed_dict=feed_dict))
input_sentences = ''
predict_sentences = ''
for i, (input_vector, predict_vector) in enumerate(
zip(input_vectors[:len(source_test_datas)],
predict_vectors[:len(target_test_datas)])):
input_sentences += ' '.join([
source_reverse_dictionary[vector] for vector in input_vector
if not vector == PAD
])
predict_sentences += ' '.join([
target_reverse_dictionary[vector] for vector in predict_vector
if not vector == PAD
])
if i < len(source_test_datas) - 1:
input_sentences += '\n'
predict_sentences += '\n'
evaluate_input_path = '%s.evaluate_input' % model_path
evaluate_predict_path = '%s.evaluate_predict' % model_path
with open(evaluate_input_path, 'w') as f1, \
open(evaluate_predict_path, 'w') as f2:
f1.write(input_sentences)
f2.write(predict_sentences)
print('input sequences at {}'.format(evaluate_input_path))
print('predict sequences at {}'.format(evaluate_predict_path))
print('mean of loss: %f' % np.mean(loss_val))
print('finish.')
class YandexAPI(object):
def __init__(self):
self.conf = Configs()
self.client = self.login()
self.win = None
self.list_type = None
self.playlists_list = None
self.updating_thread = None
if 'RESOURCEPATH' in os.environ:
self.cache = '{}/cache'.format(os.environ['RESOURCEPATH'])
else:
self.dirName = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
self.cache = os.path.join(self.dirName, 'cache')
pass
def __new__(cls, *args, **kw):
if not hasattr(cls, '_instance'):
orig = super(YandexAPI, cls)
cls._instance = orig.__new__(cls)
return cls._instance
def login(self, login=None, password=None):
if self.conf.get_attr("token") is not False:
client = Client().from_token(self.conf.get_attr("token"))
elif login is not None and password is not None:
client = Client().from_credentials(login, password)
token = client.token
self.conf.set_attr("token", token)
else:
client = Client()
self.client = client
return client
def is_logged_in(self):
if self.client.account.display_name is None:
return False
else:
return True
def logout(self):
self.conf.remove_attr("token")
self.client = Client()
pass
def get_display_name(self):
return str(self.login().account.display_name)
def get_play_lists_list(self):
entities = self.client.landing(blocks="personalplaylists").blocks[0].entities
lists = []
for playlist in entities:
lists.append({
"name": playlist.data.data.title,
"type": playlist.data.type
})
self.playlists_list = lists
return lists
def preparation(self, list_type, win):
self.updating_thread = threading.Thread(target=self.update)
self.list_type = list_type
self.win = win
index = {
"date": date.today().__str__(),
"last_track_num": 1,
"tracks": []
}
if not os.path.exists('{}/{}/'.format(self.cache, list_type)):
os.mkdir('cache/{}'.format(list_type))
if not os.path.exists('{}/{}/index.json'.format(self.cache, list_type)):
with open('{}/{}/index.json'.format(self.cache, list_type), 'w+') as file:
json.dump(index, file, indent=4)
self.updating_thread.start()
else:
if self.is_need_update():
with open('{}/{}/index.json'.format(self.cache, list_type), 'w+') as file:
json.dump(index, file, indent=4)
self.updating_thread.start()
else:
wx.PostEvent(self.win, events.FirstTrackAppear(playlist_type=list_type))
playlist_title = ""
for playlist in self.playlists_list:
if playlist['type'] == list_type:
playlist_title = playlist['name']
wx.PostEvent(self.win, events.PlaylistReady(playlist_name=playlist_title, playlist_type=list_type))
return True
def is_need_update(self):
list_type = self.list_type
with open('{}/{}/index.json'.format(self.cache, list_type), 'r') as file:
index_date = datetime.strptime(json.load(file)['date'], '%Y-%m-%d').date()
if index_date == date.today():
return False
else:
return True
def update(self):
print("Starting update")
list_type = self.list_type
blocks = self.client.landing(blocks="personalplaylists").blocks[0].entities
playlist = ""
print("processing blocks")
for block in blocks:
if block.data.type == list_type:
playlist = block.data.data
tracks = self.client.users_playlists(playlist.kind, playlist.owner.uid)[0].tracks
index_file = json.load(open('{}/{}/index.json'.format(self.cache, list_type), 'r'))
index = 1
print("processing tracks")
for track in tracks:
if index == 2:
wx.PostEvent(self.win, events.FirstTrackAppear(playlist_name=playlist.title, playlist_type=list_type))
full_track_info = track.track
index_file['tracks'].append({
"id": full_track_info.id,
"title": full_track_info.title,
"artist": full_track_info.artists[0]['name'],
"duration": full_track_info.duration_ms,
"num": index
})
with open('{}/{}/index.json'.format(self.cache, list_type), 'w+') as file:
json.dump(index_file, file)
track.track.download_cover('{}/{}/{}.png'.format(self.cache, list_type, index))
track.track.download('{}/{}/{}.mp3'.format(self.cache, list_type, index), codec="mp3", bitrate_in_kbps=320)
if index == 3:
break
index = index + 1
print("finishing updating")
wx.PostEvent(self.win, events.PlaylistReady(playlist_name=playlist.title, playlist_type=list_type))
return True
import collections
import os
from typing import List, Sequence, TypeVar
import numpy as np
from configs.configs import Configs
ROOT = os.environ['TENSOROFLOW']
config_file = '%s/configs/const.ini' % ROOT
print('data.py: config file is %s' % config_file)
c = Configs(config_file)
PAD = c.const['PAD']
EOS = c.const['EOS']
BOS = c.const['BOS']
UNK = c.const['UNK']
END_TOKEN = c.const['END_TOKEN']
A = TypeVar('A')
def read_words(input_file: str) -> List[str]:
words = []
with open(input_file) as f:
words += f.read().split()
return words
def read_data(input_file: str) -> List[List[str]]:
with open(input_file) as f:
lines = f.readlines()
return lines
def main():
parser = argparse.ArgumentParser(description="AES_aug model")
parser.add_argument('--test_prompt_id',
type=int,
default=1,
help='prompt id of test essay set')
parser.add_argument('--seed', type=int, default=12, help='set random seed')
args = parser.parse_args()
test_prompt_id = args.test_prompt_id
seed = args.seed
np.random.seed(seed)
tf.random.set_seed(seed)
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
print("Test prompt id is {} of type {}".format(test_prompt_id,
type(test_prompt_id)))
print("Seed: {}".format(seed))
configs = Configs()
data_path = configs.DATA_PATH
train_path = data_path + str(test_prompt_id) + '/train.pk'
dev_path = data_path + str(test_prompt_id) + '/dev.pk'
test_path = data_path + str(test_prompt_id) + '/test.pk'
features_path = configs.FEATURES_PATH
readability_path = configs.READABILITY_PATH
vocab_size = configs.VOCAB_SIZE
epochs = configs.EPOCHS
batch_size = configs.BATCH_SIZE
pretrained_embedding = configs.PRETRAINED_EMBEDDING
embedding_path = configs.EMBEDDING_PATH
read_configs = {
'train_path': train_path,
'dev_path': dev_path,
'test_path': test_path,
'features_path': features_path,
'readability_path': readability_path,
'vocab_size': vocab_size
}
word_vocab = read_word_vocab(read_configs)
train_data, dev_data, test_data = read_essays_words_flat(
read_configs, word_vocab)
if pretrained_embedding:
embedd_dict, embedd_dim, _ = load_word_embedding_dict(embedding_path)
embedd_matrix = build_embedd_table(word_vocab,
embedd_dict,
embedd_dim,
caseless=True)
embed_table = [embedd_matrix]
else:
embed_table = None
max_essay_len = max(train_data['max_essay_len'], dev_data['max_essay_len'],
test_data['max_essay_len'])
print('max essay length: {}'.format(max_essay_len))
train_data['y_scaled'] = get_scaled_down_scores(train_data['data_y'],
train_data['prompt_ids'])
dev_data['y_scaled'] = get_scaled_down_scores(dev_data['data_y'],
dev_data['prompt_ids'])
test_data['y_scaled'] = get_scaled_down_scores(test_data['data_y'],
test_data['prompt_ids'])
X_train = pad_flat_text_sequences(train_data['words'], max_essay_len)
X_dev = pad_flat_text_sequences(dev_data['words'], max_essay_len)
X_test = pad_flat_text_sequences(test_data['words'], max_essay_len)
X_train_linguistic_features = np.array(train_data['features_x'])
X_dev_linguistic_features = np.array(dev_data['features_x'])
X_test_linguistic_features = np.array(test_data['features_x'])
X_train_readability = np.array(train_data['readability_x'])
X_dev_readability = np.array(dev_data['readability_x'])
X_test_readability = np.array(test_data['readability_x'])
Y_train = np.array(train_data['y_scaled'])
Y_dev = np.array(dev_data['y_scaled'])
Y_test = np.array(test_data['y_scaled'])
print('================================')
print('X_train: ', X_train.shape)
print('X_train_readability: ', X_train_readability.shape)
print('X_train_ling: ', X_train_linguistic_features.shape)
print('Y_train: ', Y_train.shape)
print('================================')
print('X_dev: ', X_dev.shape)
print('X_dev_readability: ', X_dev_readability.shape)
print('X_dev_ling: ', X_dev_linguistic_features.shape)
print('Y_dev: ', Y_dev.shape)
print('================================')
print('X_test: ', X_test.shape)
print('X_test_readability: ', X_test_readability.shape)
print('X_test_ling: ', X_test_linguistic_features.shape)
print('Y_test: ', Y_test.shape)
print('================================')
model = build_AES_aug_multitask(len(word_vocab), max_essay_len, configs,
embed_table, Y_train.shape[1])
dev_features_list = [X_dev]
test_features_list = [X_test]
evaluator = AllAttEvaluator(test_prompt_id, dev_data['prompt_ids'],
test_data['prompt_ids'], dev_features_list,
test_features_list, Y_dev, Y_test)
evaluator.evaluate(model, -1, print_info=True)
for ii in range(epochs):
print('Epoch %s/%s' % (str(ii + 1), epochs))
start_time = time.time()
model.fit(X_train,
Y_train,
batch_size=batch_size,
epochs=1,
verbose=0,
shuffle=True)
tt_time = time.time() - start_time
print("Training one epoch in %.3f s" % tt_time)
evaluator.evaluate(model, ii + 1)
evaluator.print_final_info()
