y=make_data('sin', T, [f, 2 * f, 3 * f]))
#base.plt.plot_data(trainData.x, trainData.y, title='Training Data')
#%%
epochs = 1000
base.train_network(x=trainData.x,
y=trainData.y,
plotLoss=True,
epochs=epochs,
eta=0.1,
monitorY=True)
#base.plt.plot_history(base.yHist)
#%%
base = RNN_Manager(RNN.load('50hidden_trained.pkl'))
Wx, Wh, Wy = base.rnn.get_weights()
#%%
#base.plt.plot_loss(base.loss, ax=plt.gca())
#%%
base.set_test_input(trainData.x, name='Training input 1')
base.plot_feedforward()
##base.plot_hidden()
#%%
lmbda = 0.9
try:
base.rnn.reg = HessEWC(lmbda, base.rnn, trainData, H)
except:
base.rnn.reg = HessEWC(lmbda, base.rnn, trainData)
eta = 0.1
seq_length = 25
h = 1e-4
n_epoch = 20
# np.random.seed(400) # TODO: remove
# compare_gradients()
RNN = RNN(K, m, eta, seq_length, init='xavier')
save = True
smooth_loss = -1
step = -1
last_epoch = 0
if save:
smooth_loss, step, last_epoch = RNN.load()
print('last smooth_loss: %f \t last step: %d \t last epoch: %d' %
(smooth_loss, step, last_epoch))
synth = RNN.synthesize(make_one_hot([char_to_ind['.']], K), 1000)
text = ""
for column in synth.T:
text += ind_to_char[np.argmax(column)]
print(text.encode('ascii', 'ignore').decode('ascii'))
exit()
losses = []
f = open(
'synthesized-' + str(
datetime.datetime.fromtimestamp(
time.time()).strftime('%Y-%m-%d %H:%M:%S')), 'w+')
from RNN import RNN
import numpy as np
import json
vocab_size = 2575
dm = data_manager(vocab_size=vocab_size)
for i in glob("../data/*"):
dm.add_data(i)
word_to_index, index_to_word = dm.get_indices()
model = RNN(word_to_index, index_to_word, word_dim=vocab_size)
model.load("models/model.data.npz")
sentence = []
all_sents = []
for i in range(100):
sentence = model.create_sentence()
all_sents.append(" ".join(sentence).replace(".", ".</br>").replace(
",", ",</br>"))
jsobj = dict()
jsobj["text"] = " ".join(all_sents)
print(json.dumps(jsobj))
