def test():
minibatch_size = 100
num_words = 20
tuplesize = 3
num_visible = tuplesize*num_words
num_hidden = 30
codec = Codec(tuplesize, num_words)
tuples = codec.tuples
words = codec.words
encoded = codec.tuples_to_matrix()
(num_data, _) = encoded.shape
print(words)
print('data count: ', num_data)
rbm = RBM(num_visible = num_visible,
num_hidden = num_hidden,
minibatch_size = minibatch_size)
id_indices = numpy.random.randint(low=0, high=num_data, size=minibatch_size)
input_data = T.constant(encoded[id_indices])
#print(input_data)
#print(rbm.propup(input_data).eval())
h1samples = rbm.sample_h_given_v(input_data).eval()
#print(h1samples)
#print(rbm.propdown(h1samples).eval())
v2samples = rbm.sample_v_given_h(h1samples).eval()
#print(v2samples)
(W,H,V) = rbm.contrastive_divergence_1(input_data)
#print(W.eval())
#print(H.eval())
#print(V.eval())
xvis = T.fmatrix('xvis')
h1samples = rbm.sample_h_given_v(xvis)
v2samples = rbm.sample_v_given_h(h1samples)
sample_vhv = theano.function([xvis], v2samples)
example_indices = numpy.random.randint(low=0, high=num_data, size=minibatch_size)
example_input_data = encoded[example_indices]
num_examples = min(10, minibatch_size)
def show_examples():
rec = sample_vhv(example_input_data)
for example in range(num_examples):
print('input words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if example_input_data[example, t*num_words + idx]])
print('reconstructed words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if rec[example, t*num_words + idx]])
vis = T.fmatrix('vis')
train = rbm.cd1_fun(vis)
draw = VisualizeWeights('Minibatches', rbm, tuplesize, words, num_hidden)
for epoch in range(1000):
show_examples()
all_vdiffs = numpy.zeros(num_visible)
print('epoch ', epoch)
numpy.random.shuffle(encoded)
for minibatch in range(num_data // minibatch_size):
mb_start = minibatch * minibatch_size;
mb_end = mb_start + minibatch_size;
input_data_indices = numpy.arange(mb_start, mb_end)
encoded_input = encoded[input_data_indices]
input_data = encoded_input
vdiffs = train(input_data)
all_vdiffs = all_vdiffs + numpy.abs(vdiffs)
print('reconstruction error: ', numpy.sum(all_vdiffs) * minibatch_size)
print(numpy.ndarray.astype(rbm.weights.get_value()*100, numpy.int32))
print(numpy.ndarray.astype(rbm.vbias.get_value()*100, numpy.int32))
print(numpy.ndarray.astype(rbm.hbias.get_value()*100, numpy.int32))
draw.epoch_finished(epoch)
def test():
num_words = 20
tuplesize = 3
num_visible = tuplesize*num_words
num_hidden = 30
codec = Codec(tuplesize, num_words)
tuples = codec.tuples
words = codec.words
encoded = codec.tuples_to_matrix()
(num_data, _) = encoded.shape
print('data count: ', num_data)
rbm = RBM1(num_visible = num_visible,
num_hidden = num_hidden)
input_data = T.constant(encoded[2])
#print(pairs[2])
#print(encoded[2])
#print(input_data)
#print(rbm.propup(input_data).eval())
h1samples = rbm.sample_h_given_v(input_data).eval()
#print(h1samples)
#print(rbm.propdown(h1samples).eval())
v2samples = rbm.sample_v_given_h(h1samples).eval()
#print(v2samples)
(W,H,V) = rbm.contrastive_divergence_1(input_data)
#print(W.eval())
xvis = T.fvector('xvis')
h1samples = rbm.sample_h_given_v(xvis)
v2samples = rbm.sample_v_given_h(h1samples)
sample_vhv = theano.function([xvis], v2samples)
num_examples = 20
example_indices = numpy.random.randint(low=0, high=num_data, size=num_examples)
def show_examples():
for example in example_indices:
dat = encoded[example]
v2samples = sample_vhv(dat)
print('input words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if encoded[example, t*num_words + idx]])
print('reconstructed words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if v2samples[t*num_words + idx]])
print('')
def report_hidden():
weights = rbm.weights.get_value()
for h in range(num_hidden):
print('hidden ', h)
for block in range(tuplesize):
for word in range(num_words):
w = weights[block*num_words+word, h]
if w > 0.5:
print(' %2i %8s %4.1f' % (block, words[word], w))
vis = T.fvector('vis')
train = rbm.cd1_fun(vis)
input_data = numpy.reshape(encoded[2],
num_visible)
train(input_data)
print(rbm.weights.get_value())
draw = VisualizeWeights('', rbm, tuplesize, words, num_hidden)
for epoch in range(500):
show_examples()
all_vdiffs = numpy.zeros(num_visible)
print('epoch ', epoch)
for i in range(num_data):
input_data = numpy.reshape(encoded[i],
num_visible)
vdiffs = train(input_data)
all_vdiffs = all_vdiffs + numpy.abs(vdiffs)
print('reconstruction error: ', numpy.sum(all_vdiffs))
print(T.cast(rbm.weights.get_value()*100, 'int32').eval())
draw.epoch_finished(epoch)
report_hidden()
def test():
minibatch_size = 100
num_words = 40
tuplesize = 5
num_visible = tuplesize*num_words
num_hidden = 140
codec = Codec(tuplesize, num_words)
tuples = codec.tuples
words = codec.words
encoded = codec.tuples_to_matrix()
(num_data, _) = encoded.shape
print(words)
print('data count: ', num_data)
rbm = RBM(num_visible = num_visible,
num_hidden = num_hidden,
minibatch_size = minibatch_size,
venabledp=1.0,
henabledp=0.7)
id_indices = numpy.random.randint(low=0, high=num_data, size=minibatch_size)
input_data = T.constant(encoded[id_indices])
#print(input_data)
#print(rbm.propup(input_data).eval())
#h1samples = rbm.sample_h_given_v(input_data).eval()
#print(h1samples)
#print(rbm.propdown(h1samples).eval())
#v2samples = rbm.sample_v_given_h(h1samples).eval()
#print(v2samples)
#(W,H,V) = rbm.contrastive_divergence_1(input_data)
#print(W.eval())
#print(H.eval())
#print(V.eval())
all_h_enabled = numpy.ones(num_hidden)
all_v_enabled = numpy.ones(num_visible)
xvis = T.fmatrix('xvis')
h1samples = rbm.sample_h_given_v(xvis, all_h_enabled)
v2samples = rbm.sample_v_given_h(h1samples, all_v_enabled)
sample_vhv = theano.function([xvis], v2samples)
example_indices = numpy.random.randint(low=0, high=num_data, size=minibatch_size)
example_input_data = encoded[example_indices]
num_examples = min(10, minibatch_size)
def show_examples():
rec = sample_vhv(example_input_data)
for example in range(num_examples):
print('input words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if example_input_data[example, t*num_words + idx]])
print('reconstructed words:',
[(t+1, words[idx])
for t in range(tuplesize)
for idx in range(num_words)
if rec[example, t*num_words + idx]])
def report_hidden():
weights = rbm.weights.get_value()
for h in range(num_hidden):
print('hidden ', h)
for block in range(tuplesize):
for word in range(num_words):
w = weights[block*num_words+word, h]
if w > 0.5:
print(' %2i %8s %4.1f' % (block, words[word], w))
vis = T.fmatrix('vis')
train = rbm.cd1_fun(vis)
draw = VisualizeWeights('Dropout (vp:%4.2f, hp:%4.2f)' % (rbm.venabledp, rbm.henabledp),
rbm, tuplesize, words, num_hidden,
num_visible)
for epoch in range(1000):
show_examples()
all_vdiffs = numpy.zeros(num_visible)
print('epoch ', epoch)
numpy.random.shuffle(encoded)
for minibatch in range(num_data // minibatch_size):
mb_start = minibatch * minibatch_size;
mb_end = mb_start + minibatch_size;
input_data_indices = numpy.arange(mb_start, mb_end)
encoded_input = encoded[input_data_indices]
input_data = encoded_input
(vdiffs, venabled, henabled) = train(input_data)
all_vdiffs = all_vdiffs + numpy.abs(vdiffs)
#print('venabled', venabled)
#print('henabled', henabled)
print('reconstruction error: ', numpy.sum(all_vdiffs) * minibatch_size)
#print(numpy.ndarray.astype(rbm.weights.get_value()*100, numpy.int32))
#print(numpy.ndarray.astype(rbm.vbias.get_value()*100, numpy.int32))
#print(numpy.ndarray.astype(rbm.hbias.get_value()*100, numpy.int32))
draw.epoch_finished(epoch)
report_hidden()
