def main(config):
if config.validate:
output_len = config.output_seq_length
config = pickle.load(
open('saved_models/' + config.model_name + '/config.p', "rb"))
config.validate = True
config.output_seq_length = output_len
config.num_layers = 6
print(config)
elif config.backtest:
config = pickle.load(
open('saved_models/' + config.model_name + '/config.p', "rb"))
config.backtest = True
print(config)
t1 = time.time()
data_folder = os.path.abspath(os.path.abspath("../../../../")) + '/data/'
dataset = load_data(data_folder, config)
t2 = time.time()
print('Finished loading the dataset: ' + str(t2 - t1) + ' sec \n')
model = PricePredictor(config, dataset)
if config.validate:
model._validate(steps=config.output_seq_length, epoch=40)
# model._make_figs(steps = config.output_seq_length, epoch=40)
if config.backtest:
model._backtest2(epoch=180)
else:
model._train()
def main(config):
if config.validate:
config = pickle.load(
open('saved_models/' + config.model_name + '/config.p', "rb"))
config.validate = True
config.simulate = False
print(config)
t1 = time.time()
data_folder = os.path.abspath(os.path.abspath("../../../")) + '/data/'
dataset = load_data(data_folder, config)
t2 = time.time()
print('Finished loading the dataset: ' + str(t2 - t1) + ' sec \n')
model = PricePredictor(config, dataset)
if config.validate:
# model._validate( epoch=70)
# model._make_figs(epoch=70)
model._make_figs2(epoch=70)
elif config.tsne:
model._tsne(epoch=70)
else:
model._train()
def main(config):
if config.validate:
output_len = config.output_seq_length
l = config.num_layers
loss = config.loss
sl = config.num_stochastic_layers
config = pickle.load( open( 'saved_models/'+config.model_name+'/config.p', "rb" ))
config.validate = True
config.simulate = False
config.output_seq_length = output_len
config.num_layers = l
config.num_stochastic_layers = sl
config.loss = loss
print(config)
t1 = time.time()
data_folder = os.path.abspath(os.path.abspath("../../../../"))+'/data/'
dataset = load_data(data_folder, config)
t2 = time.time()
print('Finished loading the dataset: ' + str(t2-t1) +' sec \n')
model = PricePredictor(config, dataset)
if config.validate:
# model._make_figs(steps = config.output_seq_length, epoch=200)
# model._validate(steps = config.output_seq_length, epoch=160)
model._backtest(epoch=160)
else:
model._train()
def main(config):
if config.validate:
output_len = config.output_seq_length
file_path = config.file_path
seed = config.seed
loss = config.loss
target = config.target
config = pickle.load(
open('saved_models/' + config.model_name + '/config.p', "rb"))
config.validate = True
config.file_path = file_path
config.output_seq_length = output_len
config.seed = seed
config.loss = loss
config.backtest_target = 'close'
config.target = 'NDX'
print(config)
t1 = time.time()
data_folder = os.path.abspath(os.path.abspath("../../../../")) + '/data/'
dataset = load_data(data_folder, config)
t2 = time.time()
print('Finished loading the dataset: ' + str(t2 - t1) + ' sec \n')
model = PricePredictor(config, dataset)
if config.validate:
model._backtest(epoch=150)
# model._validate(steps = config.output_seq_length, epoch=150)
else:
model._train()
def main(config):
if config.validate:
output_len = config.output_seq_length
file_path = config.file_path
seed = config.seed
loss = config.loss
l = config.num_layers
sl = config.num_stochastic_layers
config = pickle.load(
open('saved_models/' + config.model_name + '/config.p', "rb"))
config.validate = True
config.file_path = file_path
config.output_seq_length = output_len
config.seed = seed
config.loss = loss
config.num_layers = l
config.num_stochastic_layers = sl
config.backtest_target = 'close_btc'
config.target = 'lr_btc'
config.model_name = 'vaegan_mv_hour_'
print(config)
elif config.backtest:
config = pickle.load(
open('saved_models/' + config.model_name + '/config.p', "rb"))
config.backtest = True
config.validate = False
print(config)
t1 = time.time()
data_folder = os.path.abspath(os.path.abspath("../../../../")) + '/data/'
dataset = load_data(data_folder, config)
t2 = time.time()
print('Finished loading the dataset: ' + str(t2 - t1) + ' sec \n')
model = PricePredictor(config, dataset)
if config.validate:
model._validate(steps=config.output_seq_length, epoch=500)
# model._make_figs(steps = config.output_seq_length, epoch=200)
# model._backtest(epoch=500)
else:
model._train()
data = DataPreparator(path=dataset_path,
col_names=column_names,
sep=separator,
na_val=na_value,
comment=comment,
result_col=result_column_name)
data.dataset["Origin"] = data.dataset["Origin"].map({
1: "USA",
2: "Europe",
3: "Japan"
})
data.get_dummies()
predictor = PricePredictor(data)
history = predictor.history()
loss, mae, mse = predictor.evaluate()
test_predictions = predictor.predict()
errors = test_predictions - predictor.test_labels
train_val = _smooth(history.history['mae'], std=2)
value = _smooth(history.history['val_mae'], std=2)
fig, (axs1, axs2, axs3) = plt.subplots(1, 3)
axs1.plot(history.epoch, train_val, label="Помилка %")
axs1.plot(history.epoch, value, "--", label="Помилка значення")
axs1.set(xlabel="Епоха", ylabel="Абсолютне значення помилки")
axs1.legend()
def build_model(schema, mappings, trie, args):
import tensorflow as tf
from cocoa.model.word_embedder import WordEmbedder
from cocoa.model.encdec import BasicEncoder, BasicDecoder, Sampler
from price_predictor import PricePredictor
from encdec import BasicEncoderDecoder, PriceDecoder, PriceEncoder, ContextDecoder, AttentionDecoder, LM, SlotFillingDecoder, ContextEncoder, TrieDecoder, ClassifyDecoder, CandidateSelector, IRSelector
from ranker import IRRanker, CheatRanker, EncDecRanker, SlotFillingRanker
from context_embedder import ContextEmbedder
from preprocess import markers
from cocoa.model.sequence_embedder import get_sequence_embedder
check_model_args(args)
tf.reset_default_graph()
tf.set_random_seed(args.random_seed)
with tf.variable_scope('GlobalDropout'):
if args.test:
keep_prob = tf.constant(1.)
else:
# When test on dev set, we need to feed in keep_prob = 1.0
keep_prob = tf.placeholder_with_default(tf.constant(1. -
args.dropout),
shape=[],
name='keep_prob')
vocab = mappings['vocab']
pad = vocab.to_ind(markers.PAD)
# Word embeddings
word_embeddings = None
context_word_embeddings = None
if args.pretrained_wordvec is not None:
word_embeddings = vocab.load_embeddings(args.pretrained_wordvec,
args.word_embed_size)
if args.context:
context_word_embeddings = mappings['kb_vocab'].load_embeddings(
args.pretrained_wordvec, args.word_embed_size)
with tf.variable_scope('EncoderWordEmbedder'):
encoder_word_embedder = WordEmbedder(vocab.size, args.word_embed_size,
word_embeddings, pad)
with tf.variable_scope('DecoderWordEmbedder'):
decoder_word_embedder = WordEmbedder(vocab.size, args.word_embed_size,
word_embeddings, pad)
if args.decoding[0] == 'sample':
sample_t = float(args.decoding[1])
sampler = Sampler(sample_t, trie=trie)
else:
raise ('Unknown decoding method')
opts = vars(args)
opts['vocab_size'] = vocab.size
opts['keep_prob'] = keep_prob
opts['embed_size'] = args.rnn_size
encoder_seq_embedder = get_sequence_embedder(args.encoder, **opts)
decoder_seq_embedder = get_sequence_embedder(args.decoder, **opts)
if args.context is not None:
context_opts = dict(opts)
context_opts['vocab_size'] = mappings['kb_vocab'].size
context_opts['embed_size'] = args.context_size
with tf.variable_scope('ContextWordEmbedder'):
context_word_embedder = WordEmbedder(context_opts['vocab_size'],
context_opts['embed_size'],
context_word_embeddings,
pad=pad)
with tf.variable_scope('CategoryWordEmbedder'):
category_word_embedder = WordEmbedder(mappings['cat_vocab'].size,
10,
pad=pad)
context_seq_embedder = get_sequence_embedder(args.context_encoder,
**context_opts)
context_embedder = ContextEmbedder(mappings['cat_vocab'].size,
context_word_embedder,
category_word_embedder,
context_seq_embedder, pad)
if args.predict_price:
price_predictor = PricePredictor(args.price_predictor_hidden_size,
args.price_hist_len, pad)
def get_decoder(args):
prompt_len = 2 # <role> <category>
if args.decoder == 'rnn':
if args.context is not None:
decoder = ContextDecoder(decoder_word_embedder,
decoder_seq_embedder,
context_embedder,
args.context,
pad,
keep_prob,
vocab.size,
sampler,
args.sampled_loss,
args.tied,
prompt_len=prompt_len)
else:
decoder = BasicDecoder(decoder_word_embedder,
decoder_seq_embedder,
pad,
keep_prob,
vocab.size,
sampler,
args.sampled_loss,
args.tied,
prompt_len=prompt_len)
else:
decoder = AttentionDecoder(decoder_word_embedder,
decoder_seq_embedder,
pad,
keep_prob,
vocab.size,
sampler,
args.sampled_loss,
context_embedder=context_embedder,
attention_memory=args.attention_memory,
prompt_len=prompt_len)
if args.predict_price:
decoder = PriceDecoder(decoder, price_predictor)
if args.slot_filling:
decoder = SlotFillingDecoder(decoder)
# Retrieval-based models
if args.model == 'selector':
decoder = ClassifyDecoder(decoder)
#decoder = TrieDecoder(decoder)
return decoder
def get_encoder(args):
if args.num_context > 0:
encoder = ContextEncoder(encoder_word_embedder,
encoder_seq_embedder, args.num_context,
pad, keep_prob)
else:
encoder = BasicEncoder(encoder_word_embedder, encoder_seq_embedder,
pad, keep_prob)
if args.predict_price:
encoder = PriceEncoder(encoder, price_predictor)
return encoder
if args.model == 'encdec' or args.ranker == 'encdec':
decoder = get_decoder(args)
encoder = get_encoder(args)
model = BasicEncoderDecoder(encoder,
decoder,
pad,
keep_prob,
stateful=args.stateful)
elif args.model == 'selector':
decoder = get_decoder(args)
encoder = get_encoder(args)
model = CandidateSelector(encoder, decoder, pad, keep_prob)
elif args.model == 'ir':
model = IRSelector()
elif args.model == 'lm':
decoder = get_decoder(args)
model = LM(decoder, pad)
elif args.model is not None:
raise ValueError('Unknown model')
if args.ranker == 'cheat':
model = CheatRanker()
if args.ranker == 'ir':
model = IRRanker()
elif args.ranker == 'encdec':
model = EncDecRanker(model, args.temperature)
elif args.ranker == 'sf':
model = SlotFillingRanker(model)
return model