def __init__(self, v, h, init=True):
"""
We currently support only CD1
"""
self.v = v; self.h = h
if init:
self.w = [bias_mat(std_mat(v,h))]
def __init__(self, v, h, init=True):
"""
We currently support only CD1
"""
self.v = v
self.h = h
if init:
self.w = [bias_mat(std_mat(v, h))]
from mats.bias_mat import bias_mat
from mats.std_mat import std_mat
import data.bouncing_balls as b
res = 20
n_balls = 3
T = 100
dat = lambda : b.bounce_vec(res,n=3,T=100)
v=res**2
h=100
LR = .02
W = .02 * bias_mat(std_mat(v,h))
n_cd = 5
def grad(W, x):
return rbm.rbm_grad_cd(W, x, n_cd)
def loss(W, x):
return rbm.rbm_grad_cd(W, x, n_cd)[1]
from pylab import rand
from trainers.std_trainer3 import std_trainer
t = std_trainer(name='r1',
path='rbm/r_data',
W = W,
from p8.rnn_trbm.rnn_trbm import rnn_trbm
from p8.rnn_trbm.trbm import trbm
res = 30
n_balls = 3
T = 100
dat = lambda : b.bounce_vec(res,n=3,T=100)
v=res**2
h=400
NUM_ITER = 10**5
def LR(x): return .01 * (1- float(x)/NUM_ITER)
VH = bias_mat(std_mat(v,h))
HH = bias_mat(std_mat(h,h))
CD_n = 5
choice = input('press 0 for RNN-TRBM, 1 for plain TRBM\n...')
W0 = .005*rnn_trbm(VH, HH, CD_n = CD_n)
W1 = .005* trbm(VH, HH, CD_n = CD_n)
W = [W0, W1][choice]
name = 'r_balls_' + ['rnn-trbm', 'plain-trbm'][choice]
def grad(W, x): return W.grad(x)
def loss(W, x): return W.grad(x)[1]
from trainers.std_trainer3 import std_trainer
res = 30
n_balls = 3
T = 100
dat = lambda: b.bounce_vec(res, n=3, T=100)
v = res**2
h = 400
NUM_ITER = 10**5
def LR(x):
return .01 * (1 - float(x) / NUM_ITER)
VH = bias_mat(std_mat(v, h))
HH = bias_mat(std_mat(h, h))
CD_n = 5
choice = input('press 0 for RNN-TRBM, 1 for plain TRBM\n...')
W0 = .005 * rnn_trbm(VH, HH, CD_n=CD_n)
W1 = .005 * trbm(VH, HH, CD_n=CD_n)
W = [W0, W1][choice]
name = 'r_balls_' + ['rnn-trbm', 'plain-trbm'][choice]
def grad(W, x):
return W.grad(x)
import rbm
from mats.bias_mat import bias_mat
from mats.std_mat import std_mat
v=4
batch_size=20
W = bias_mat(std_mat(v,5))
def grad(W, x):
return rbm.rbm_grad_exact(W, x)
def loss(W, x):
return rbm.rbm_grad_exact(W, x)[1]
from pylab import rand
def data_fn():
return (rand(batch_size,v)<.5).astype('f')
from trainers.std_trainer3 import std_trainer
t = std_trainer(name='r1',
path='rbm/r_data',
W = W,
unnorm_grad_fn = grad,
unnorm_valid_loss = loss,
data_fn = data_fn,
valid_data_fn = data_fn,
num_iter = 1000)
from mats.bias_mat import bias_mat
from mats.std_mat import std_mat
import data.bouncing_balls as b
res = 20
n_balls = 3
T = 100
dat = lambda: b.bounce_vec(res, n=3, T=100)
v = res**2
h = 100
LR = .02
W = .02 * bias_mat(std_mat(v, h))
n_cd = 5
def grad(W, x):
return rbm.rbm_grad_cd(W, x, n_cd)
def loss(W, x):
return rbm.rbm_grad_cd(W, x, n_cd)[1]
from pylab import rand
from trainers.std_trainer3 import std_trainer