def __init__(self, name, input_shape, output_dim, hidden_dim, hidden_nonlinearity=tf.nn.relu, lstm_layer_cls=L.LSTMLayer, output_nonlinearity=None, input_var=None, input_layer=None, forget_bias=1.0, use_peepholes=False, layer_args=None): with tf.variable_scope(name): if input_layer is None: l_in = L.InputLayer(shape=(None, None) + input_shape, input_var=input_var, name="input") else: l_in = input_layer l_step_input = L.InputLayer(shape=(None,) + input_shape, name="step_input") # contains previous hidden and cell state l_step_prev_state = L.InputLayer(shape=(None, hidden_dim * 2), name="step_prev_state") if layer_args is None: layer_args = dict() l_lstm = lstm_layer_cls(l_in, num_units=hidden_dim, hidden_nonlinearity=hidden_nonlinearity, hidden_init_trainable=False, name="lstm", forget_bias=forget_bias, cell_init_trainable=False, use_peepholes=use_peepholes, **layer_args) l_lstm_flat = L.ReshapeLayer( l_lstm, shape=(-1, hidden_dim), name="lstm_flat" ) l_output_flat = L.DenseLayer( l_lstm_flat, num_units=output_dim, nonlinearity=output_nonlinearity, name="output_flat" ) l_output = L.OpLayer( l_output_flat, op=lambda flat_output, l_input: tf.reshape(flat_output, tf.stack((tf.shape(l_input)[0], tf.shape(l_input)[1], -1))), shape_op=lambda flat_output_shape, l_input_shape: (l_input_shape[0], l_input_shape[1], flat_output_shape[-1]), extras=[l_in], name="output" ) l_step_state = l_lstm.get_step_layer(l_step_input, l_step_prev_state, name="step_state") l_step_hidden = L.SliceLayer(l_step_state, indices=slice(hidden_dim), name="step_hidden") l_step_cell = L.SliceLayer(l_step_state, indices=slice(hidden_dim, None), name="step_cell") l_step_output = L.DenseLayer( l_step_hidden, num_units=output_dim, nonlinearity=output_nonlinearity, W=l_output_flat.W, b=l_output_flat.b, name="step_output" ) self._l_in = l_in self._hid_init_param = l_lstm.h0 self._cell_init_param = l_lstm.c0 self._l_lstm = l_lstm self._l_out = l_output self._l_step_input = l_step_input self._l_step_prev_state = l_step_prev_state self._l_step_hidden = l_step_hidden self._l_step_cell = l_step_cell self._l_step_state = l_step_state self._l_step_output = l_step_output self._hidden_dim = hidden_dim
def __init__(self, name, input_shape, output_dim, input_var=None, input_layer=None, qmdp_param=None): with tf.variable_scope(name): hidden_dim = qmdp_param['grid_n'] * qmdp_param['grid_m'] if input_layer is None: l_in = L.InputLayer(shape=(None, None) + input_shape, input_var=input_var, name="input") else: l_in = input_layer l_step_input = L.InputLayer(shape=(None, ) + input_shape, name="step_input") l_step_prev_state = L.InputLayer(shape=(None, hidden_dim), name="step_prev_state") hidden_dim = qmdp_param['grid_n'] * qmdp_param['grid_m'] l_gru = FilterLayer(l_in, qmdp_param, name="qmdp_filter") l_gru_flat = L.ReshapeLayer(l_gru, shape=(-1, hidden_dim), name="gru_flat") l_output_flat = PlannerLayer(l_gru_flat, qmdp_param, name="output_flat") l_output = L.OpLayer( l_output_flat, op=lambda flat_output, l_input: tf.reshape( flat_output, tf.stack( (tf.shape(l_input)[0], tf.shape(l_input)[1], -1))), shape_op=lambda flat_output_shape, l_input_shape: (l_input_shape[0], l_input_shape[1], flat_output_shape[-1]), extras=[l_in], name="output") l_step_state = l_gru.get_step_layer(l_step_input, l_step_prev_state, name="step_state") l_step_hidden = l_step_state l_step_output = l_output_flat.get_step_layer(l_step_hidden, name="step_output") self._l_in = l_in self._hid_init_param = l_gru.h0 self._l_gru = l_gru self._l_output_flat = l_output_flat self._l_out = l_output self._l_step_input = l_step_input self._l_step_prev_state = l_step_prev_state self._l_step_hidden = l_step_hidden self._l_step_state = l_step_state self._l_step_output = l_step_output self._hidden_dim = hidden_dim
def __init__(self, name, input_shape, output_dim, hidden_dim, hidden_nonlinearity=tf.nn.relu, gru_layer_cls=L.GRULayer, output_nonlinearity=None, input_var=None, input_layer=None, layer_args=None): with tf.variable_scope(name): if input_layer is None: l_in = L.InputLayer(shape=(None, None) + input_shape, input_var=input_var, name="input") else: l_in = input_layer l_step_input = L.InputLayer(shape=(None,) + input_shape, name="step_input") l_step_prev_state = L.InputLayer(shape=(None, hidden_dim), name="step_prev_state") if layer_args is None: layer_args = dict() l_gru = gru_layer_cls(l_in, num_units=hidden_dim, hidden_nonlinearity=hidden_nonlinearity, hidden_init_trainable=False, name="gru", **layer_args) l_gru_flat = L.ReshapeLayer( l_gru, shape=(-1, hidden_dim), name="gru_flat" ) l_output_flat = L.DenseLayer( l_gru_flat, num_units=output_dim, nonlinearity=output_nonlinearity, name="output_flat" ) l_output = L.OpLayer( l_output_flat, op=lambda flat_output, l_input: tf.reshape(flat_output, tf.stack((tf.shape(l_input)[0], tf.shape(l_input)[1], -1))), shape_op=lambda flat_output_shape, l_input_shape: (l_input_shape[0], l_input_shape[1], flat_output_shape[-1]), extras=[l_in], name="output" ) l_step_state = l_gru.get_step_layer(l_step_input, l_step_prev_state, name="step_state") l_step_hidden = l_step_state l_step_output = L.DenseLayer( l_step_hidden, num_units=output_dim, nonlinearity=output_nonlinearity, W=l_output_flat.W, b=l_output_flat.b, name="step_output" ) self._l_in = l_in self._hid_init_param = l_gru.h0 self._l_gru = l_gru self._l_out = l_output self._l_step_input = l_step_input self._l_step_prev_state = l_step_prev_state self._l_step_hidden = l_step_hidden self._l_step_state = l_step_state self._l_step_output = l_step_output self._hidden_dim = hidden_dim
def __init__(self, name, input_shape, output_dim, hidden_dims, hidden_nonlinearity=tf.nn.relu, output_nonlinearity=None, input_var=None, input_layer=None): with tf.variable_scope(name): if input_layer is None: l_in = L.InputLayer(shape=(None, None) + input_shape, input_var=input_var, name="input") else: l_in = input_layer l_step_input = L.InputLayer(shape=(None, ) + input_shape, name="step_input") l_step_prev_hiddens = [ L.InputLayer(shape=(None, hidden_dim), name="step_prev_hidden%i" % i) for i, hidden_dim in enumerate(hidden_dims) ] # Build the unrolled GRU network, which operates laterally, then # vertically below = l_in l_grus = [] for i, hidden_dim in enumerate(hidden_dims): l_gru = L.GRULayer(below, num_units=hidden_dim, hidden_nonlinearity=hidden_nonlinearity, hidden_init_trainable=False, name="gru%i" % i) l_grus.append(l_gru) below = l_gru # Convert final hidden layer to flat representation l_gru_flat = L.ReshapeLayer(l_grus[-1], shape=(-1, hidden_dims[-1]), name="gru_flat") l_output_flat = L.DenseLayer(l_gru_flat, num_units=output_dim, nonlinearity=output_nonlinearity, name="output_flat") l_output = L.OpLayer( l_output_flat, op=lambda flat_output, l_input: tf.reshape( flat_output, tf.pack((tf.shape(l_input)[0], tf.shape(l_input)[1], -1))), shape_op=lambda flat_output_shape, l_input_shape: (l_input_shape[0], l_input_shape[1], flat_output_shape[-1]), extras=[l_in], name="output") # Build a single step of the GRU network, which operates vertically # and is replicated laterally below = l_step_input l_step_hiddens = [] for i, (l_gru, prev_hidden) in enumerate(zip(l_grus, l_step_prev_hiddens)): l_step_hidden = L.GRUStepLayer([below, prev_hidden], "step_hidden%i" % i, l_gru) l_step_hiddens.append(l_step_hidden) below = l_step_hidden l_step_output = L.DenseLayer(l_step_hiddens[-1], num_units=output_dim, nonlinearity=output_nonlinearity, W=l_output_flat.W, b=l_output_flat.b, name="step_output") self._l_in = l_in self._hid_inits = [l_gru.h0 for l_gru in l_grus] self._l_grus = l_grus self._l_out = l_output self._l_step_input = l_step_input self._l_step_prev_hiddens = l_step_prev_hiddens self._l_step_hiddens = l_step_hiddens self._l_step_output = l_step_output