From c49e498aa046c363cc785b3f4f34b6d45e62a195 Mon Sep 17 00:00:00 2001 From: Connor Olding Date: Tue, 11 Apr 2017 10:32:48 +0000 Subject: [PATCH] add dropout and deterministic predictions --- optim_nn_core.py | 46 +++++++++++++++++++++++++++++++++++----------- optim_nn_mnist.py | 29 +++++++++++++++++------------ 2 files changed, 52 insertions(+), 23 deletions(-) diff --git a/optim_nn_core.py b/optim_nn_core.py index ce48733..86c1f69 100644 --- a/optim_nn_core.py +++ b/optim_nn_core.py @@ -374,6 +374,9 @@ class Layer: def forward(self, X): raise NotImplementedError("unimplemented", self) + def forward_deterministic(self, X): + return self.forward(X) + def backward(self, dY): raise NotImplementedError("unimplemented", self) @@ -390,11 +393,13 @@ class Layer: self.parents.append(parent) # TODO: better names for these (still) - - def _propagate(self, edges): + def _propagate(self, edges, deterministic): if not self.unsafe: assert len(edges) == 1, self - return self.forward(edges[0]) + if deterministic: + return self.forward_deterministic(edges[0]) + else: + return self.forward(edges[0]) def _backpropagate(self, edges): if len(edges) == 1: @@ -437,7 +442,7 @@ class Layer: for k, w in self.weights.items(): w.allocate(ins, outs, allocator=allocator) - def propagate(self, values): + def propagate(self, values, deterministic): if not self.unsafe: assert self.parents, self edges = [] @@ -447,7 +452,7 @@ class Layer: if not self.unsafe: self.validate_input(X) edges.append(X) - Y = self._propagate(edges) + Y = self._propagate(edges, deterministic) if not self.unsafe: self.validate_output(Y) return Y @@ -525,7 +530,7 @@ class ConstAffine(Layer): return dY * self.a class Sum(Layer): - def _propagate(self, edges): + def _propagate(self, edges, deterministic): return np.sum(edges, axis=0) def _backpropagate(self, edges): @@ -546,6 +551,23 @@ class ActivityRegularizer(Layer): def backward(self, dY): return dY + self.reg.backward(self.X) +class Dropout(Layer): + def __init__(self, dropout=0.0): + super().__init__() + self.p = _f(1 - dropout) + assert 0 <= self.p <= 1 + + def forward(self, X): + self.mask = (np.random.rand(*X.shape) < self.p) / self.p + return X * self.mask + + def forward_deterministic(self, X): + #self.mask = _1 + return X + + def backward(self, dY): + return dY * self.mask + # Activation Layers {{{2 class Sigmoid(Layer): # aka Logistic @@ -710,13 +732,13 @@ class Model: nodes.append(node) return nodes - def forward(self, X): + def forward(self, X, deterministic=False): values = dict() input_node = self.ordered_nodes[0] output_node = self.ordered_nodes[-1] - values[input_node] = input_node._propagate(np.expand_dims(X, 0)) + values[input_node] = input_node._propagate(np.expand_dims(X, 0), deterministic) for node in self.ordered_nodes[1:]: - values[node] = node.propagate(values) + values[node] = node.propagate(values, deterministic) return values[output_node] def backward(self, error): @@ -861,7 +883,7 @@ class Ritual: # i'm just making up names at this point self.learner.batch(b / batch_count) if test_only: - predicted = self.model.forward(batch_inputs) + predicted = self.model.forward(batch_inputs, deterministic=True) else: predicted = self.learn(batch_inputs, batch_outputs) self.update() @@ -873,6 +895,7 @@ class Ritual: # i'm just making up names at this point losses.append(batch_loss) cumsum_loss += batch_loss + # NOTE: this can use the non-deterministic predictions. fixme? batch_mloss = self.measure(predicted, batch_outputs) if np.isnan(batch_mloss): raise Exception("nan") @@ -915,7 +938,7 @@ class Ritual: # i'm just making up names at this point self.learner.batch(b / batch_count) if test_only: - predicted = self.model.forward(batch_inputs) + predicted = self.model.forward(batch_inputs, deterministic=True) else: predicted = self.learn(batch_inputs, batch_outputs) self.update() @@ -927,6 +950,7 @@ class Ritual: # i'm just making up names at this point losses.append(batch_loss) cumsum_loss += batch_loss + # NOTE: this can use the non-deterministic predictions. fixme? batch_mloss = self.measure(predicted, batch_outputs) if np.isnan(batch_mloss): raise Exception("nan") diff --git a/optim_nn_mnist.py b/optim_nn_mnist.py index 78d3f33..7749786 100755 --- a/optim_nn_mnist.py +++ b/optim_nn_mnist.py @@ -26,6 +26,7 @@ if use_emnist: reg = None final_reg = None + dropout = None actreg_lamb = None load_fn = None @@ -53,6 +54,7 @@ else: reg = L1L2(3.2e-5, 3.2e-4) final_reg = L1L2(3.2e-5, 1e-3) + dropout = 0.10 actreg_lamb = None # 1e-3 load_fn = None @@ -89,14 +91,17 @@ def get_mnist(fn='mnist.npz'): inputs, outputs, valid_inputs, valid_outputs = get_mnist(fn) -def actreg(y): - if not actreg_lamb: - return y - lamb = actreg_lamb # * np.prod(y.output_shape) - reg = SaturateRelu(lamb) - act = ActivityRegularizer(reg) - reg.lamb_orig = reg.lamb # HACK - return y.feed(act) +def regulate(y): + if actreg_lamb: + assert type(activation) == Relu, type(activation) + lamb = actreg_lamb # * np.prod(y.output_shape) + reg = SaturateRelu(lamb) + act = ActivityRegularizer(reg) + reg.lamb_orig = reg.lamb # HACK + y = y.feed(act) + if dropout: + y = y.feed(Dropout(dropout)) + return y x = Input(shape=inputs.shape[1:]) y = x @@ -104,7 +109,7 @@ y = x y = y.feed(Reshape(new_shape=(mnist_dim, mnist_dim))) for i in range(n_denses): if i > 0: - y = actreg(y) + y = regulate(y) y = y.feed(activation()) y = y.feed(Denses(new_dims[0], axis=0, init=init_he_normal, reg_w=reg, reg_b=reg)) @@ -113,11 +118,11 @@ for i in range(n_denses): y = y.feed(Flatten()) for i in range(n_dense): if i > 0: - y = actreg(y) + y = regulate(y) y = y.feed(activation()) y = y.feed(Dense(y.output_shape[0], init=init_he_normal, reg_w=reg, reg_b=reg)) -y = actreg(y) +y = regulate(y) y = y.feed(activation()) y = y.feed(Dense(mnist_classes, init=init_glorot_uniform, @@ -162,7 +167,7 @@ def measure_error(quiet=False): loss, mloss, _, _ = ritual.test_batched(inputs, outputs, bs, return_losses='both') c = Confidence() - predicted = ritual.model.forward(inputs) + predicted = ritual.model.forward(inputs, deterministic=True) confid = c.forward(predicted) if not quiet: