rbmobject1.restore_weights('./out/rbmw1.chp')
rbmobject2.restore_weights('./out/rbmw2.chp')
iterations = len(trX) / FLAGS.batchsize
# Train First RBM
print 'first rbm: ', FLAGS.epochs, ' epochs of ', iterations, ' iterations'
for i in range(FLAGS.epochs):
print 'epoch=', i
for j in range(iterations):
batch_xs, batch_ys = mnist.train.next_batch(FLAGS.batchsize)
_, batch_err = rbmobject1.partial_fit(batch_xs)
rbmobject1.save_weights('./out/rbmw1.chp')
save_image("out/1rbm.jpg", rbmobject1.n_w.T, (28, 28), (100, 100))
show_image("out/1rbm.jpg")
rbmobject1.end_training()
#
# Train Second RBM
rbmobject1.restore_weights('./out/rbmw1.chp')
# improves init
w1 = rbmobject1.getOWeight()
rbmobject2.setOWeight(np.transpose(w1))
print 'second rbm: ', FLAGS.epochs, ' epochs of ', iterations, ' iterations'
for i in range(FLAGS.epochs):
print 'epoch=', i
for j in range(iterations):
batch_xs, batch_ys = mnist.train.next_batch(FLAGS.batchsize)
# Transform features with first rbm for second rbm
batch_xs = rbmobject1.sample_v2h(batch_xs)
_, batch_err = rbmobject2.partial_fit(batch_xs)
rbmobject2.save_weights('./out/rbmw2.chp')
