From 4c5f52b63851f0abd1900b24321fa9692e823071 Mon Sep 17 00:00:00 2001 From: Alex Auvolat Date: Fri, 19 Jun 2015 14:38:58 -0400 Subject: GFGUR trains (a bit) --- gfgru.py | 120 ++++++++++++++++++++++++++++++++++++++++++--------------------- 1 file changed, 80 insertions(+), 40 deletions(-) (limited to 'gfgru.py') diff --git a/gfgru.py b/gfgru.py index 8f05d46..ac4c481 100644 --- a/gfgru.py +++ b/gfgru.py @@ -2,8 +2,8 @@ import theano from theano import tensor import numpy -from blocks.algorithms import Momentum, AdaDelta, RMSProp -from blocks.bricks import Tanh, Logistic, Softmax, Rectifier, Linear, MLP, Initializable, Identity +from blocks.algorithms import Momentum, AdaDelta, RMSProp, Adam +from blocks.bricks import Activation, Tanh, Logistic, Softmax, Rectifier, Linear, MLP, Initializable, Identity from blocks.bricks.base import application, lazy from blocks.bricks.recurrent import BaseRecurrent, recurrent from blocks.initialization import IsotropicGaussian, Constant @@ -13,44 +13,51 @@ from blocks.filter import VariableFilter from blocks.roles import WEIGHT, INITIAL_STATE, add_role from blocks.graph import ComputationGraph, apply_noise, apply_dropout +class TRectifier(Activation): + @application(inputs=['input_'], outputs=['output']) + def apply(self, input_): + return tensor.switch(input_ > 1, input_, 0) + # An epoch will be composed of 'num_seqs' sequences of len 'seq_len' # divided in chunks of lengh 'seq_div_size' num_seqs = 10 seq_len = 2000 -seq_div_size = 200 +seq_div_size = 100 io_dim = 256 recurrent_blocks = [ # (256, Tanh(), [2048], [Rectifier()]), - (512, Tanh(), [], []), - (512, Tanh(), [1024], [Rectifier()]), - (512, Tanh(), [], []), +# (512, Rectifier(), [1024], [Rectifier()]), + (512, Tanh(), [2048], [TRectifier()]), + (512, Tanh(), [2048], [TRectifier()]), + (512, Tanh(), [2048], [TRectifier()]), # (2, Tanh(), [2], [Rectifier()]), # (2, Tanh(), [], []), ] -control_hidden = [512] -control_hidden_activations = [Tanh()] +control_hidden = [1024] +control_hidden_activations = [Rectifier()] output_hidden = [1024] output_hidden_activations = [Rectifier()] -weight_noise_std = 0.02 -recurrent_dropout = 0.5 -control_dropout = 0.5 +weight_noise_std = 0.05 + +recurrent_h_dropout = 0 +control_h_dropout = 0 +output_h_dropout = 0.5 -step_rule = 'adadelta' +step_rule = 'adam' learning_rate = 0.1 -momentum = 0.9 +momentum = 0.99 -param_desc = '%s,c%s,o%s-n%s-d%s,%s-%dx%d(%d)-%s' % ( +param_desc = '%s,c%s,o%s-n%s-d%s,%s,%s-%s' % ( repr(map(lambda (a, b, c, d): (a, c), recurrent_blocks)), repr(control_hidden), repr(output_hidden), repr(weight_noise_std), - repr(recurrent_dropout), repr(control_dropout), - num_seqs, seq_len, seq_div_size, + repr(recurrent_h_dropout), repr(control_h_dropout), repr(output_h_dropout), step_rule ) @@ -68,10 +75,14 @@ elif step_rule == 'adadelta': step_rule = AdaDelta() elif step_rule == 'momentum': step_rule = Momentum(learning_rate=learning_rate, momentum=momentum) +elif step_rule == 'adam': + step_rule = Adam() else: assert(False) + + class GFGRU(BaseRecurrent, Initializable): def __init__(self, input_dim, recurrent_blocks, control_hidden, control_hidden_activations, **kwargs): super(GFGRU, self).__init__(**kwargs) @@ -94,7 +105,7 @@ class GFGRU(BaseRecurrent, Initializable): self.hidden_total_dim = sum(x for (x, _, _, _) in self.recurrent_blocks) # control block - self.cblocklen = len(self.recurrent_blocks) + 3 + self.cblocklen = len(self.recurrent_blocks) + 2 control_idim = self.hidden_total_dim + self.input_dim control_odim = len(self.recurrent_blocks) * self.cblocklen @@ -111,19 +122,19 @@ class GFGRU(BaseRecurrent, Initializable): idim = self.input_dim + self.hidden_total_dim if i > 0: idim = idim + self.recurrent_blocks[i-1][0] - rgate = MLP(dims=[self.hidden_total_dim, self.hidden_total_dim], - activations=[logistic], - name='rgate%d'%i) + idims = [idim] + hdim if hdim == []: inter = Identity() else: inter = MLP(dims=idims, activations=hact, name='inter%d'%i) - zgate = MLP(dims=[idims[-1], dim], activations=[logistic], name='zgate%d'%i) + + rgate = MLP(dims=[idims[-1], dim], activations=[logistic], name='rgate%d'%i) nstate = MLP(dims=[idims[-1], dim], activations=[act], name='nstate%d'%i) - for brick in [rgate, inter, zgate, nstate]: + + for brick in [inter, rgate, nstate]: self.children.append(brick) - self.blocks.append((rgate, inter, zgate, nstate)) + self.blocks.append((inter, rgate, nstate)) # init state zeros self.init_states_names = [] @@ -144,6 +155,17 @@ class GFGRU(BaseRecurrent, Initializable): return self.init_states_dict[name].shape.eval() return super(GFGRU, self).get_dim(name) + def recurrent_h_dropout_vars(self, cg): + ret = [] + for (inter, rgate, nstate) in self.blocks: + ret = ret + VariableFilter(name='input_', + bricks=inter.linear_transformations + rgate.linear_transformations + nstate.linear_transformations + )(cg) + return ret + + def control_h_dropout_vars(self, cg): + return VariableFilter(name='input_', bricks=self.control.linear_transformations)(cg) + @recurrent(sequences=['inputs'], contexts=[]) def apply(self, inputs=None, **kwargs): states = [kwargs[i] for i in self.init_states_names] @@ -154,23 +176,24 @@ class GFGRU(BaseRecurrent, Initializable): control_v = self.control.apply(concat_input_states) new_states = [] - for i, (rgate, inter, zgate, nstate) in enumerate(self.blocks): + for i, (inter, rgate, nstate) in enumerate(self.blocks): controls = control_v[:, i * self.cblocklen:(i+1) * self.cblocklen] - rgate_v = rgate.apply(concat_states) r_inputs = tensor.concatenate([s * controls[:, j][:, None] for j, s in enumerate(states)], axis=1) - r_inputs = r_inputs * (1 - rgate_v * controls[:, -1][:, None]) more_inputs = [inputs] if i > 0: - more_inputs = more_inputs + [new_states[-1]] + more_inputs.append(new_states[-1]) inter_inputs = tensor.concatenate([r_inputs] + more_inputs, axis=1) inter_v = inter.apply(inter_inputs) - zgate_v = zgate.apply(inter_v) + + rgate_v = rgate.apply(inter_v) nstate_v = nstate.apply(inter_v) - zctl = zgate_v * controls[:, -2][:, None] + controls[:, -3][:, None] - nstate_v = zctl * nstate_v + (1 - zctl) * states[i] + rctl = controls[:, -1][:, None] * rgate_v + uctl = controls[:, -2][:, None] + nstate_v = uctl * nstate_v + (1 - rctl) * states[i] + new_states.append(nstate_v) return new_states @@ -206,16 +229,24 @@ class Model(): hidden_total_dim = sum(x for (x, _, _, _) in recurrent_blocks) - prev_states = theano.shared(numpy.zeros((num_seqs, hidden_total_dim)).astype(theano.config.floatX), + prev_states_dict = {} + for i, (dim, _, _, _) in enumerate(recurrent_blocks): + prev_state = theano.shared(numpy.zeros((num_seqs, dim)).astype(theano.config.floatX), name='states_save') - states = [x.dimshuffle(1, 0, 2) for x in gfgru.apply(in_onehot.dimshuffle(1, 0, 2), states=prev_states)] - states = tensor.concatenate(states, axis=2) - new_states = states[:, -1, :] + prev_states_dict['init_state_%d'%i] = prev_state + + states = [x.dimshuffle(1, 0, 2) for x in gfgru.apply(in_onehot.dimshuffle(1, 0, 2), **prev_states_dict)] + + self.states = [] + for i, _ in enumerate(recurrent_blocks): + self.states.append((prev_states_dict['init_state_%d'%i], states[i][:, -1, :])) + + states_concat = tensor.concatenate(states, axis=2) out_mlp = MLP(dims=[hidden_total_dim] + output_hidden + [io_dim], activations=output_hidden_activations + [None], name='output_mlp') - states_sh = states.reshape((inp.shape[0]*inp.shape[1], hidden_total_dim)) + states_sh = states_concat.reshape((inp.shape[0]*inp.shape[1], hidden_total_dim)) out = out_mlp.apply(states_sh).reshape((inp.shape[0], inp.shape[1], io_dim)) @@ -230,8 +261,8 @@ class Model(): # Initialize all bricks for brick in [gfgru, out_mlp]: - brick.weights_init = IsotropicGaussian(0.1) - brick.biases_init = Constant(0.) + brick.weights_init = IsotropicGaussian(0.01) + brick.biases_init = Constant(0.001) brick.initialize() # Apply noise and dropout @@ -239,8 +270,18 @@ class Model(): if weight_noise_std > 0: noise_vars = VariableFilter(roles=[WEIGHT])(cg) cg = apply_noise(cg, noise_vars, weight_noise_std) - # if i_dropout > 0: - # cg = apply_dropout(cg, hidden[1:], i_dropout) + if recurrent_h_dropout > 0: + dv = gfgru.recurrent_h_dropout_vars(cg) + print "Recurrent H dropout on", len(dv), "vars" + cg = apply_dropout(cg, dv, recurrent_h_dropout) + if control_h_dropout > 0: + dv = gfgru.control_h_dropout_vars(cg) + print "Control H dropout on", len(dv), "vars" + cg = apply_dropout(cg, dv, control_h_dropout) + if output_h_dropout > 0: + dv = VariableFilter(name='input_', bricks=out_mlp.linear_transformations)(cg) + print "Output H dropout on", len(dv), "vars" + cg = apply_dropout(cg, dv, output_h_dropout) [cost_reg, error_rate_reg] = cg.outputs @@ -251,5 +292,4 @@ class Model(): self.out = out self.pred = pred - self.states = [(prev_states, new_states)] -- cgit v1.2.3