NetworkLstmLayer

class NetworkLstmLayer.RecurrentLayer(reverse=False, truncation=-1, compile=True, projection=0, sampling=1, **kwargs)[source]
recurrent = True[source]
layer_class = 'recurrent'[source]
compile()[source]
create_recurrent_weights(n, m)[source]
class NetworkLstmLayer.LstmLayer(n_out, sharpgates='none', **kwargs)[source]
layer_class = 'lstm'[source]
class NetworkLstmLayer.OptimizedLstmLayer(n_out, sharpgates='none', encoder=None, n_dec=0, **kwargs)[source]
layer_class = 'lstm_opt'[source]
get_branching()[source]
get_energy()[source]
make_constraints()[source]
class NetworkLstmLayer.SimpleLstmLayer(n_out, sharpgates='none', encoder=None, n_dec=0, **kwargs)[source]
layer_class = 'lstm_simple'[source]
NetworkLstmLayer.make_lstm_step(n_cells, W_re, W_out_proj=None, W_re_proj=None, W_peep_i=None, W_peep_f=None, W_peep_o=None, grad_clip=None, CI=None, CO=None, G=None)[source]
NetworkLstmLayer.lstm(z, i, W_re, W_out_proj=None, W_re_proj=None, W_peep_i=None, W_peep_f=None, W_peep_o=None, CI=None, CO=None, G=None, grad_clip=None, direction=1)[source]
class NetworkLstmLayer.Lstm2Layer(n_out, n_cells=None, n_proj=None, peepholes=False, direction=1, activation=None, grad_clip=None, truncation=None, **kwargs)[source]
recurrent = True[source]
layer_class = 'lstm2'[source]
class NetworkLstmLayer.Lstm3Layer(n_out, direction=1, grad_clip=None, **kwargs)[source]

Like lstm2 but even simpler.

recurrent = True[source]
layer_class = 'lstm3'[source]
class NetworkLstmLayer.LayerNormLstmLayer(n_out, direction=1, grad_clip=None, **kwargs)[source]

Layer Normalization, https://arxiv.org/abs/1607.06450

recurrent = True[source]
layer_class = 'ln_lstm'[source]
class NetworkLstmLayer.NativeLstmLayer(n_out, direction=1, truncation=None, **kwargs)[source]
recurrent = True[source]
layer_class = 'native_lstm'[source]
class NetworkLstmLayer.GenericLstmLayer(n_out, sublayer, out_sublayer=None, n_cells=None, activation=None, direction=1, grad_clip=None, truncation=None, **kwargs)[source]

LSTM implementation which allows a custom input+recurrent function (n_in + n_out -> n_cells * 4) and a custom output function (n_cells -> n_out) which is identity by default. You specify it as a sub layer.

recurrent = True[source]
layer_class = 'generic_lstm'[source]
class NetworkLstmLayer.AssociativeLstmLayer(n_out, n_copies, activation='tanh', direction=1, grad_clip=None, **kwargs)[source]

Associative Long Short-Term Memory http://arxiv.org/abs/1602.03032

recurrent = True[source]
layer_class = 'associative_lstm'[source]
class NetworkLstmLayer.LstmHalfGatesLayer(n_out, direction=1, activation='tanh', grad_clip=None, **kwargs)[source]
recurrent = True[source]
layer_class = 'lstm_half_gates'[source]
class NetworkLstmLayer.LstmProjGatesLayer(n_out, n_gate_proj, direction=1, activation='relu', grad_clip=None, **kwargs)[source]
recurrent = True[source]
layer_class = 'lstm_proj_gates'[source]
class NetworkLstmLayer.LstmComplexLayer(n_out, direction=1, activation='tanh', use_complex='1:1:1:1', grad_clip=None, **kwargs)[source]
recurrent = True[source]
layer_class = 'lstm_complex'[source]
class NetworkLstmLayer.ActLstmLayer(n_out, n_max_calc_steps=10, time_penalty=0.01, time_penalty_type='linear_p', total_halt_penalty=0.0, total_halt_penalty_type='inv', direction=1, eps=0.01, grad_clip=None, unroll_inner_scan=False, **kwargs)[source]

Adaptive Computation Time for Recurrent Neural Networks, Graves

recurrent = True[source]
layer_class = 'act_lstm'[source]
class NetworkLstmLayer.GRULayer(n_out, encoder=None, mode='cho', n_dec=0, **kwargs)[source]
layer_class = 'gru'[source]
class NetworkLstmLayer.SRULayer(n_out, encoder=None, psize=0, pact='relu', pdepth=1, carry_time=False, n_dec=0, **kwargs)[source]
layer_class = 'sru'[source]
class NetworkLstmLayer.SRALayer(n_out, encoder=None, psize=0, pact='relu', pdepth=1, n_dec=0, **kwargs)[source]
layer_class = 'sra'[source]