def _setup_functions(self): self._activation_func = nnprocessors.build_activation(self.activation) self._softmax_func = nnprocessors.build_activation('softmax') top_rep, self.output_func = self._recursive_func() self.monitors.append(("top_rep<0.1", 100 * (abs(top_rep) < 0.1).mean())) self.monitors.append(("top_rep<0.9", 100 * (abs(top_rep) < 0.9).mean())) self.monitors.append(("top_rep:mean", abs(top_rep).mean()))
def _setup_functions(self): self._activation_func = nnprocessors.build_activation(self.activation) self._softmax_func = nnprocessors.build_activation('softmax') self.hidden_func, self.output_func = self._recurrent_func() self.monitors.append( ("h<0.1", 100 * (abs(self.hidden_func[-1]) < 0.1).mean())) self.monitors.append( ("h<0.9", 100 * (abs(self.hidden_func[-1]) < 0.9).mean()))
def _setup_functions(self): self._activation_func = nnprocessors.build_activation(self.activation) self._softmax_func = nnprocessors.build_activation('softmax') top_rep, self.output_func = self._recursive_func() self.monitors.append( ("top_rep<0.1", 100 * (abs(top_rep) < 0.1).mean())) self.monitors.append( ("top_rep<0.9", 100 * (abs(top_rep) < 0.9).mean())) self.monitors.append(("top_rep:mean", abs(top_rep).mean()))
def _setup_functions(self): self._assistive_params = [] self._activation_func = nnprocessors.build_activation(self.activation) self._softmax_func = nnprocessors.build_activation('softmax') top_rep, self.output_func = self._recursive_func() # self.predict_func, self.predict_updates = self._encode_func() self.monitors.append(("top_rep<0.1", 100 * (abs(top_rep) < 0.1).mean())) self.monitors.append(("top_rep<0.9", 100 * (abs(top_rep) < 0.9).mean())) self.monitors.append(("top_rep:mean", abs(top_rep).mean()))
def _setup_functions(self): self._assistive_params = [] self._activation_func = nnprocessors.build_activation(self.activation) self._softmax_func = nnprocessors.build_activation('softmax') top_rep, self.output_func = self._recursive_func() # self.predict_func, self.predict_updates = self._encode_func() self.monitors.append(("top_rep<0.1", 100 * (abs(top_rep) < 0.1).mean())) self.monitors.append(("top_rep<0.9", 100 * (abs(top_rep) < 0.9).mean())) self.monitors.append(("top_rep:mean", abs(top_rep).mean()))
def _setup_functions(self): self._assistive_params = [] self._activation_func = nnprocessors.build_activation(self.activation) self._softmax_func = nnprocessors.build_activation('softmax') self.hidden_func, self.output_func, recurrent_updates = self._recurrent_func() self.predict_func, self.predict_updates = self._predict_func() self.monitors.append(("hh<0.1", 100 * (abs(self.hidden_func[-1]) < 0.1).mean())) self.monitors.append(("hh<0.9", 100 * (abs(self.hidden_func[-1]) < 0.9).mean())) self.updates.extend(recurrent_updates.items()) if self.update_h0: self.updates.append((self.h0, ifelse(T.eq(self._vars.k[-1], 0), self.init_h, self.hidden_func[-1]))) self.params.extend(self._assistive_params)
def _setup_functions(self): self._tanh = nnprocessors.build_activation('tanh') self._sigmoid = nnprocessors.build_activation('sigmoid') self._softmax = nnprocessors.build_activation('softmax') [self.output_func, self.hidden_func, self.memory_func], recurrent_updates = self._recurrent_func() self.predict_func, self.predict_updates = self._predict_func() self.monitors.append(("last_h<0.1", 100 * (abs(self.hidden_func[-1]) < 0.1).mean())) self.monitors.append(("last_h<0.9", 100 * (abs(self.hidden_func[-1]) < 0.9).mean())) self.monitors.append(("c<0.1", 100 * (abs(self.memory_func[-1]) < 0.1).mean())) self.monitors.append(("c<0.9", 100 * (abs(self.memory_func[-1]) < 0.9).mean())) self.updates.extend(recurrent_updates) if self.update_h0: self.updates.append((self.h0, ifelse(T.eq(self._vars.k[-1], 0), self.init_h, self.hidden_func[-1]))) self.updates.append((self.c0, ifelse(T.eq(self._vars.k[-1], 0), self.init_c, self.memory_func[-1])))
def _setup_functions(self): if self.shared_bias: self._vars.update_if_not_existing(self.shared_bias, self.B) bias = self.B if not self.shared_bias else self._vars.get( self.shared_bias) if self.disable_bias: bias = 0 self._activation_func = nnprocessors.build_activation(self.activation) self.preact_func = T.dot(self.x, self.W) + bias self.output_func = nnprocessors.add_noise( self._activation_func(self.preact_func), self.noise, self.dropouts)
def _setup_functions(self): if self.shared_bias: self._vars.update_if_not_existing(self.shared_bias, self.B) bias = self.B if not self.shared_bias else self._vars.get(self.shared_bias) if self.disable_bias: bias = 0 self._activation_func = nnprocessors.build_activation(self.activation) self.preact_func = T.dot(self.x, self.W) + bias self.output_func = nnprocessors.add_noise( self._activation_func(self.preact_func), self.noise, self.dropouts)
def _setup_functions(self): self._assistive_params = [] self._activation_func = nnprocessors.build_activation(self.activation) self.output_func = self._output_func()
def _setup_functions(self): self._assistive_params = [] self._activation_func = nnprocessors.build_activation(self.activation) self.output_func = self._output_func()
def _setup_functions(self): self._activation_func = nnprocessors.build_activation(self.activation) self._softmax_func = nnprocessors.build_activation('softmax') self.hidden_func, self.output_func = self._recurrent_func() self.monitors.append(("h<0.1", 100 * (abs(self.hidden_func[-1]) < 0.1).mean())) self.monitors.append(("h<0.9", 100 * (abs(self.hidden_func[-1]) < 0.9).mean()))