class PolicyDropoutLayer: def __init__(self, n_in, n_out, block_size, activation, do_dropout=False): self.block_size = block_size self.nblocks = n_out / block_size self.do_dropout = do_dropout assert n_out % block_size == 0 self.h = HiddenLayer(n_in, n_out, activation) shared.bind("reinforce") self.d = HiddenLayer(n_in, self.nblocks, T.nnet.sigmoid) shared.bind("default") def __call__(self, x, xmask=None): probs = self.d(x) * 0.98 + 0.01 mask = srng.uniform(probs.shape) < probs print xmask mask.name = "mask!" masked = self.h.activation(sparse_dot(x, xmask, self.h.W, mask, self.h.b, self.block_size)) if not "this is the equivalent computation in theano": h = self.h(x) if self.do_dropout: h = h * (srng.uniform(h.shape) < 0.5) h_r = h.reshape([h.shape[0], self.nblocks, self.block_size]) masked = h_r * mask.dimshuffle(0,1,'x') masked = masked.reshape(h.shape) self.sample_probs = T.prod(mask*probs+(1-probs)*(1-mask), axis=1) self.probs = probs return masked, mask
class PolicyDropoutLayer: def __init__(self, n_in, n_out, block_size, activation, do_dropout=False, reinforce_params="reinforce", default_params="default"): self.block_size = block_size self.nblocks = n_out / block_size self.do_dropout = do_dropout assert n_out % block_size == 0 self.h = HiddenLayer(n_in, n_out, activation) shared.bind(reinforce_params) self.d = HiddenLayer(n_in, self.nblocks, T.nnet.sigmoid) shared.bind(default_params) def __call__(self, x, xmask=None): probs = self.d(x) * 0.98 + 0.01 mask = srng.uniform(probs.shape) < probs print xmask mask.name = "mask!" masked = self.h.activation(sparse_dot(x, xmask, self.h.W, mask, self.h.b, self.block_size)) if not "this is the equivalent computation in theano": h = self.h(x) if self.do_dropout: h = h * (srng.uniform(h.shape) < 0.5) h_r = h.reshape([h.shape[0], self.nblocks, self.block_size]) masked = h_r * mask.dimshuffle(0,1,'x') masked = masked.reshape(h.shape) self.sample_probs = T.prod(mask*probs+(1-probs)*(1-mask), axis=1) self.probs = probs return masked, mask
class PolicyDropoutLayer: def __init__(self, n_in, n_out, block_size, activation, rate, do_dropout=False): self.rate = rate self.block_size = block_size self.nblocks = n_out / block_size self.do_dropout = do_dropout assert n_out % block_size == 0 self.h = HiddenLayer(n_in, n_out, activation) def __call__(self, x, xmask=None): mask = srng.uniform((x.shape[0],self.nblocks)) < self.rate masked = self.h.activation(sparse_dot(x, xmask, self.h.W, mask, self.h.b, self.block_size)) return masked, mask
class PolicyDropoutLayer: def __init__(self, n_in, n_out, block_size, activation, rate, do_dropout=False): self.rate = rate self.block_size = block_size self.nblocks = n_out / block_size self.do_dropout = do_dropout assert n_out % block_size == 0 self.h = HiddenLayer(n_in, n_out, activation) def __call__(self, x, xmask=None): mask = srng.uniform((x.shape[0], self.nblocks)) < self.rate masked = self.h.activation( sparse_dot(x, xmask, self.h.W, mask, self.h.b, self.block_size)) return masked, mask