def start_rnn():
load_data = LoadData()
process_data = ProcessDataRNN(sales=load_data.make_dataset(
categorize=CATEGORIZE, start=START, upper=UPPER))
# xt, yt, xv, yv, xe, ye = process_data.run()
xv, yv, xe, ye = process_data.run()
if not ONLY_LOAD_MODEL:
ckpt = ModelCheckpoint(OUTPUT_MODEL,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min')
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.2,
patience=5,
min_lr=MIN_LR,
verbose=1)
es = EarlyStopping(monitor='val_loss', patience=3)
# net = make_model(len(process_data.FEATS))
net = make_model_rnn(process_data.Z.shape[2])
print(net.summary())
n_slices = LEN // 28
brks = np.array(
[LEN - (n_slices - i) * 28 for i in range(n_slices + 1)])
brks = brks[brks >= max(LAGS) + 28]
print("#" * 30)
print(LEN, process_data.C.shape, process_data.Z.shape)
print(brks)
print(process_data.C.min(), process_data.ys.min(),
process_data.Z[:, 66:].min())
print("#" * 30)
net.fit_generator(DataGenerator(
(process_data.C, process_data.Z, process_data.ys),
brks[:-1],
batch_size=BATCH_SIZE),
epochs=EPOCH,
validation_data=(xv, yv),
callbacks=[ckpt, reduce_lr, es])
# net.fit(xt, yt, batch_size=BATCH_SIZE, epochs=EPOCH, validation_data=(xv, yv), callbacks=[ckpt, reduce_lr, es])
# nett = make_model(len(process_data.FEATS))
nett = make_model_rnn(process_data.Z.shape[2])
nett.load_weights(OUTPUT_MODEL)
pv = nett.predict(xv, batch_size=BATCH_SIZE, verbose=1)
pe = nett.predict(xe, batch_size=BATCH_SIZE, verbose=1)
print("Eva result:", nett.evaluate(xv, yv, batch_size=BATCH_SIZE))
# pv = pv.reshape((-1, 28, 9))
# pe = pe.reshape((-1, 28, 9))
sv = process_data.sv.reshape((-1, 28))
se = process_data.se.reshape((-1, 28))
# Yv = yv.reshape((-1, 28))
return process_data, yv, pv, pe, sv, se
def start_cnn():
load_data = LoadData()
process_data = ProcessDataCNN(sales=load_data.make_dataset(
categorize=CATEGORIZE, start=START, upper=UPPER))
xt, yt, xv, yv, xe, ye = process_data.run()
if not ONLY_LOAD_MODEL:
ckpt = ModelCheckpoint(OUTPUT_MODEL,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min')
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.2,
patience=5,
min_lr=MIN_LR,
verbose=1)
es = EarlyStopping(monitor='val_loss', patience=3)
net = make_model_cnn(len(process_data.FEATS))
plot_model(net, to_file='model.png')
print(net.summary())
# exit()
net.fit(xt,
yt,
batch_size=BATCH_SIZE,
epochs=EPOCH,
validation_data=(xv, yv),
callbacks=[reduce_lr, ckpt, es])
nett = make_model_cnn(len(process_data.feats_list))
nett.load_weights(OUTPUT_MODEL)
pv = nett.predict(xv, batch_size=BATCH_SIZE, verbose=1)
pe = nett.predict(xe, batch_size=BATCH_SIZE, verbose=1)
print("Eva result:", nett.evaluate(xv, yv, batch_size=BATCH_SIZE))
pv = pv.reshape((-1, 28, 9))
pe = pe.reshape((-1, 28, 9))
sv = process_data.sv.reshape((-1, 28))
se = process_data.se.reshape((-1, 28))
Yv = yv.reshape((-1, 28))
return process_data, Yv, pv, pe, sv, se