From b0ad9c907422e1256bc5ae6881913f71cd9d4aed Mon Sep 17 00:00:00 2001 From: caoying03 Date: Wed, 28 Jun 2017 10:12:23 +0800 Subject: [PATCH] enable intializing memory state for lstmemory_group. --- .../paddle/trainer_config_helpers/layers.py | 51 ++++++++++--------- .../paddle/trainer_config_helpers/networks.py | 48 ++++++++++------- 2 files changed, 57 insertions(+), 42 deletions(-) diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 84ed1607730..a601d5c84ad 100755 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -1149,10 +1149,10 @@ def pooling_layer(input, @layer_support(DROPOUT) def lstmemory(input, name=None, + size=None, reverse=False, act=None, gate_act=None, - size=None, state_act=None, bias_attr=None, param_attr=None, @@ -1194,6 +1194,8 @@ def lstmemory(input, :param name: The lstmemory layer name. :type name: basestring + :param size: DEPRECATED. size of the lstm cell + :type size: int :param input: input layer name. :type input: LayerOutput :param reverse: is sequence process reversed or not. @@ -1220,15 +1222,15 @@ def lstmemory(input, assert state_act.support_hppl assert act.support_hppl assert input.size is not None and input.size % 4 == 0 + if size is not None: if input.size / 4 == size: plog = logger.warning else: plog = logger.fatal - - plog("NOTE: The lstmemory layer[%s]'s size is set by previous input " - "layer. The lstm size should be equal with input layer size/4. The" - " size which is set explicitly will be ignored." % name) + plog("size of lstmemory layer: %s is automatically set to " + "size of input layer / 4. The parameter size passing to " + "this layer is ignored." % (name)) Layer( name=name, @@ -1255,11 +1257,11 @@ def lstmemory(input, @wrap_name_default("gru") @layer_support(DROPOUT) def grumemory(input, + size=None, name=None, reverse=False, act=None, gate_act=None, - size=None, bias_attr=None, param_attr=None, layer_attr=None): @@ -1318,6 +1320,8 @@ def grumemory(input, :type name: None|basestring :param input: input layer. :type input: LayerOutput. + :param size: DEPRECATED. size of the gru cell + :type size: int :param reverse: Whether sequence process is reversed or not. :type reverse: bool :param act: activation type, TanhActivation by default. This activation @@ -1334,9 +1338,6 @@ def grumemory(input, :type param_attr: ParameterAttribute|None|False :param layer_attr: Extra Layer attribute :type layer_attr: ExtraLayerAttribute|None - :param size: Stub parameter of size, but actually not used. If set this size - will get a warning. - :type size: None :return: LayerOutput object. :rtype: LayerOutput """ @@ -1348,9 +1349,9 @@ def grumemory(input, plog = logger.warning else: plog = logger.fatal - plog("NOTE: the gru memory layer's size is set by previous input layer," - " and should be input size / 3. Set size explicitly will be " - "ignored.") + plog("size of grumemory layer: %s is automatically set to " + "size of input layer / 3. The parameter size passing to this " + "layer is ignored." % (name)) Layer( name=name, @@ -2524,8 +2525,8 @@ def img_cmrnorm_layer(input, @wrap_bias_attr_default() -@wrap_param_attr_default(default_factory=lambda _: ParamAttr(initial_mean=1.0, - initial_std=0.)) +@wrap_param_attr_default( + default_factory=lambda _: ParamAttr(initial_mean=1.0, initial_std=0.)) @wrap_act_default(act=ReluActivation()) @wrap_name_default("batch_norm") @layer_support(DROPOUT) @@ -3013,25 +3014,25 @@ def lstm_step_layer(input, bias_attr=None, layer_attr=None): """ - LSTM Step Layer. It used in recurrent_group. The lstm equations are shown - as follow. + LSTM Step Layer. This function is used only in recurrent_group. + The lstm equations are shown as follows. .. math:: - i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i) + i_t & = \\sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + W_{c_i}c_{t-1} + b_i) - f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f) + f_t & = \\sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + W_{c_f}c_{t-1} + b_f) - c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c) + c_t & = f_tc_{t-1} + i_t tanh (W_{x_c}x_t+W_{h_c}h_{t-1} + b_c) - o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o) + o_t & = \\sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + W_{c_o}c_t + b_o) h_t & = o_t tanh(c_t) The input of lstm step is :math:`Wx_t + Wh_{t-1}`, and user should use :code:`mixed_layer` and :code:`full_matrix_projection` to calculate these - input vector. + input vectors. The state of lstm step is :math:`c_{t-1}`. And lstm step layer will do @@ -3042,14 +3043,14 @@ def lstm_step_layer(input, ... - This layer contains two outputs. Default output is :math:`h_t`. The other - output is :math:`o_t`, which name is 'state' and can use + This layer has two outputs. Default output is :math:`h_t`. The other + output is :math:`o_t`, whose name is 'state' and can use :code:`get_output_layer` to extract this output. :param name: Layer's name. :type name: basestring - :param size: Layer's size. NOTE: lstm layer's size, should be equal as - :code:`input.size/4`, and should be equal as + :param size: Layer's size. NOTE: lstm layer's size, should be equal to + :code:`input.size/4`, and should be equal to :code:`state.size`. :type size: int :param input: input layer. :math:`Wx_t + Wh_{t-1}` diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py index 67154a8d7d3..0d730e09951 100755 --- a/python/paddle/trainer_config_helpers/networks.py +++ b/python/paddle/trainer_config_helpers/networks.py @@ -614,6 +614,7 @@ def simple_lstm(input, @wrap_name_default('lstm_unit') def lstmemory_unit(input, + memory_boot=None, name=None, size=None, param_attr=None, @@ -626,9 +627,9 @@ def lstmemory_unit(input, lstm_layer_attr=None, get_output_layer_attr=None): """ - Define calculations that a LSTM unit performs in a single time step. - This function itself is not a recurrent layer, so that it can not be - directly applied to sequence input. This function is always used in + Define calculations that a LSTM unit performs during a single time step. + This function itself is not a recurrent layer, so it can not be + directly used to process sequence inputs. This function is always used in recurrent_group (see layers.py for more details) to implement attention mechanism. @@ -638,13 +639,13 @@ def lstmemory_unit(input, .. math:: - i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i) + i_t & = \\sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + W_{c_i}c_{t-1} + b_i) - f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f) + f_t & = \\sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + W_{c_f}c_{t-1} + b_f) - c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c) + c_t & = f_tc_{t-1} + i_t tanh (W_{x_c}x_t+W_{h_c}h_{t-1} + b_c) - o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o) + o_t & = \\sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + W_{c_o}c_t + b_o) h_t & = o_t tanh(c_t) @@ -661,6 +662,8 @@ def lstmemory_unit(input, :param input: input layer name. :type input: LayerOutput + :param memory_boot: the initialization state of the LSTM cell. + :type memory_boot: LayerOutput | None :param name: lstmemory unit name. :type name: basestring :param size: lstmemory unit size. @@ -692,7 +695,8 @@ def lstmemory_unit(input, assert input.size % 4 == 0 size = input.size / 4 out_mem = memory(name=name, size=size) - state_mem = memory(name="%s_state" % name, size=size) + state_mem = memory( + name="%s_state" % name, size=size, boot_layer=memory_boot) with mixed_layer( name="%s_input_recurrent" % name, @@ -726,6 +730,7 @@ def lstmemory_unit(input, def lstmemory_group(input, size=None, name=None, + memory_boot=None, reverse=False, param_attr=None, act=None, @@ -737,7 +742,7 @@ def lstmemory_group(input, lstm_layer_attr=None, get_output_layer_attr=None): """ - lstm_group is a recurrent layer group version of Long Short Term Memory. It + lstm_group is a recurrent_group version of Long Short Term Memory. It does exactly the same calculation as the lstmemory layer (see lstmemory in layers.py for the maths) does. A promising benefit is that LSTM memory cell states, or hidden states in every time step are accessible to the @@ -748,8 +753,8 @@ def lstmemory_group(input, NOTE: In PaddlePaddle's implementation, the following input-to-hidden multiplications: - :math:`W_{xi}x_{t}` , :math:`W_{xf}x_{t}`, - :math:`W_{xc}x_t`, :math:`W_{xo}x_{t}` are not done in lstmemory_unit to + :math:`W_{x_i}x_{t}` , :math:`W_{x_f}x_{t}`, + :math:`W_{x_c}x_t`, :math:`W_{x_o}x_{t}` are not done in lstmemory_unit to speed up the calculations. Consequently, an additional mixed_layer with full_matrix_projection must be included before lstmemory_unit is called. @@ -765,8 +770,10 @@ def lstmemory_group(input, :param input: input layer name. :type input: LayerOutput - :param name: lstmemory group name. + :param name: name of the lstmemory group. :type name: basestring + :param memory_boot: the initialization state of LSTM cell. + :type memory_boot: LayerOutput | None :param size: lstmemory group size. :type size: int :param reverse: is lstm reversed @@ -798,6 +805,7 @@ def lstmemory_group(input, def __lstm_step__(ipt): return lstmemory_unit( input=ipt, + memory_boot=memory_boot, name=name, size=size, mixed_bias_attr=mixed_bias_attr, @@ -819,6 +827,7 @@ def lstmemory_group(input, @wrap_name_default('gru_unit') def gru_unit(input, + memory_boot=None, size=None, name=None, gru_bias_attr=None, @@ -829,8 +838,8 @@ def gru_unit(input, naive=False): """ Define calculations that a gated recurrent unit performs in a single time - step. This function itself is not a recurrent layer, so that it can not be - directly applied to sequence input. This function is almost always used in + step. This function itself is not a recurrent layer, so it can not be + directly used to process sequence inputs. This function is always used in the recurrent_group (see layers.py for more details) to implement attention mechanism. @@ -838,6 +847,8 @@ def gru_unit(input, :param input: input layer name. :type input: LayerOutput + :param memory_boot: the initialization state of the LSTM cell. + :type memory_boot: LayerOutput | None :param name: name of the gru group. :type name: basestring :param size: hidden size of the gru. @@ -856,7 +867,7 @@ def gru_unit(input, if size is None: size = input.size / 3 - out_mem = memory(name=name, size=size) + out_mem = memory(name=name, size=size, boot_layer=memory_boot) if naive: __step__ = gru_step_naive_layer @@ -878,6 +889,7 @@ def gru_unit(input, @wrap_name_default('gru_group') def gru_group(input, + memory_boot=None, size=None, name=None, reverse=False, @@ -888,7 +900,7 @@ def gru_group(input, gru_layer_attr=None, naive=False): """ - gru_group is a recurrent layer group version of Gated Recurrent Unit. It + gru_group is a recurrent_group version of Gated Recurrent Unit. It does exactly the same calculation as the grumemory layer does. A promising benefit is that gru hidden states are accessible to the user. This is especially useful in attention model. If you do not need to access @@ -908,6 +920,8 @@ def gru_group(input, :param input: input layer name. :type input: LayerOutput + :param memory_boot: the initialization state of the LSTM cell. + :type memory_boot: LayerOutput | None :param name: name of the gru group. :type name: basestring :param size: hidden size of the gru. @@ -929,6 +943,7 @@ def gru_group(input, def __gru_step__(ipt): return gru_unit( input=ipt, + memory_boot=memory_boot, name=name, size=size, gru_bias_attr=gru_bias_attr, @@ -1083,7 +1098,6 @@ def simple_gru2(input, return grumemory( name=name, - size=size, input=m, reverse=reverse, bias_attr=gru_bias_attr, -- GitLab