def xw_plus_b(x, weights, biases, name=None): """ Computes matmul(x, weights) + biases. Args: x: a 2D tensor. Dimensions typically: batch, in_units weights: a 2D tensor. Dimensions typically: in_units, out_units biases: a 1D tensor. Dimensions: out_units name: A name for the operation (optional). If not specified "xw_plus_b" is used. Returns: A 2-D Tensor computing matmul(x, weights) + biases. Dimensions typically: batch, out_units. """ if weights.shape is None: raise ValueError('weights must have a valid shape.') else: if len(weights.shape) != 2: raise ValueError('weights must be a 2D Tensor') if biases.shape is None: raise ValueError('biases must a have a valid shape.') else: if len(biases.shape) != 1: raise ValueError('biases must be a 1D Tensor') if weights.shape[1] != biases.shape[0]: raise ValueError('the shape of weights and biaes are incompatible.') return ops.InnerProduct([x, weights, biases], num_output=weights.shape[1], TransW=False)
def _create_graph(self): self.x = Tensor(shape=[None, self.img_channels, self.img_height, self.img_width]).Variable() self.y_r = Tensor(shape=[None], name='Yr').Variable() # As implemented in A3C paper self.n1 = ops.Relu(ops.Conv2D([self.x] + self.weight_bias(), kernel_size=8, stride=4, num_output=16)) self.n2 = ops.Relu(ops.Conv2D([self.n1] + self.weight_bias(), kernel_size=4, stride=2, num_output=32)) self.action_index = Tensor(shape=[None, self.num_actions]).Variable() self.d1 = ops.Relu(ops.InnerProduct([self.n2] + self.weight_bias(), num_output=256)) self.logits_v = ops.InnerProduct([self.d1] + self.weight_bias(), num_output=1) self.cost_v = ops.L2Loss([self.y_r, self.logits_v]) self.logits_p = ops.InnerProduct([self.d1] + self.weight_bias(), num_output=self.num_actions) if Config.USE_LOG_SOFTMAX: raise NotImplementedError() else: self.softmax_p = ops.Softmax(self.logits_p) self.selected_action_prob = ops.Sum(self.softmax_p * self.action_index, axis=1) self.cost_p_1 = ops.Log(ops.Clip(self.selected_action_prob, self.log_epsilon, None)) * \ (self.y_r - ops.StopGradient(self.logits_v)) self.cost_p_2 = ops.Sum(ops.Log(ops.Clip(self.softmax_p, self.log_epsilon, None)) * self.softmax_p, axis=1) * (-self.beta) self.cost_p_1_agg = ops.Sum(self.cost_p_1) self.cost_p_2_agg = ops.Sum(self.cost_p_2) self.cost_p = -(self.cost_p_1_agg + self.cost_p_2_agg) self.cost_all = self.cost_p + self.cost_v if Config.DUAL_RMSPROP: raise NotImplementedError() else: if Config.USE_GRAD_CLIP: self.opt = updaters.RMSPropUpdater(decay=Config.RMSPROP_DECAY, eps=Config.RMSPROP_EPSILON, clip_gradient=Config.GRAD_CLIP_NORM) else: self.opt = updaters.RMSPropUpdater(decay=Config.RMSPROP_DECAY, eps=Config.RMSPROP_EPSILON) grads = T.grad(self.cost_all, self.network_params) for p, g in zip(self.network_params, grads): self.opt.append((p, g), lr_mult=1.0)
def xw_plus_b(x, weights, biases, name=None): if weights.shape is None: raise ValueError('weights must have a valid shape.') else: if len(weights.shape) != 2: raise ValueError('weights must be a 2D Tensor') if biases.shape is None: raise ValueError('biases must a have a valid shape.') else: if len(biases.shape) != 1: raise ValueError('biases must be a 1D Tensor') if weights.shape[1] != biases.shape[0]: raise ValueError( 'the shape of weights and biaes are incompatible.') return ops.InnerProduct([x, weights, biases], num_output=weights.shape[1], TransW=False)
def Setup(self, bottom): super(InnerProductLayer, self).Setup(bottom) return ops.InnerProduct( bottom + [blob['data'] for blob in self._blobs], **self._param)