def test_write(self): graph = tf.Graph() with graph.as_default(): with tf.Session(graph=graph) as session: mem = Memory(4, 5, 2, 1) M, u, p, L, ww, rw, r = session.run(mem.init_memory()) key = np.random.uniform(0, 1, (1, 5, 1)).astype(np.float32) strength = np.random.uniform(0, 1, (1, 1)).astype(np.float32) free_gates = np.random.uniform(0, 1, (1, 2)).astype(np.float32) write_gate = np.random.uniform(0, 1, (1, 1)).astype(np.float32) allocation_gate = np.random.uniform(0, 1, (1, 1)).astype(np.float32) write_vector = np.random.uniform(0, 1, (1, 5)).astype(np.float32) erase_vector = np.zeros((1, 5)).astype(np.float32) u_op, ww_op, M_op, L_op, p_op = mem.write( M, u, rw, ww, p, L, key, strength, free_gates, allocation_gate, write_gate, write_vector, erase_vector) session.run(tf.initialize_all_variables()) u, ww, M, L, p = session.run([u_op, ww_op, M_op, L_op, p_op]) self.assertEqual(u.shape, (1, 4)) self.assertEqual(ww.shape, (1, 4)) self.assertEqual(M.shape, (1, 4, 5)) self.assertEqual(L.shape, (1, 4, 4)) self.assertEqual(p.shape, (1, 4))
def test_construction(self): interface = DNC.interface( read_keys=None, read_strengths=None, write_key=np.random.uniform(0, 1, (3, 9, 1)).astype(np.float32), write_strength=np.random.uniform(0, 1, (3, 1)).astype(np.float32), erase_vector=tf.convert_to_tensor( np.zeros((3, 9)).astype(np.float32)), write_vector=tf.convert_to_tensor( np.random.uniform(0, 1, (3, 9)).astype(np.float32)), free_gates=np.random.uniform(0, 1, (3, 5)).astype(np.float32), allocation_gate=np.random.uniform(0, 1, (3, 1)).astype(np.float32), write_gate=np.random.uniform(0, 1, (3, 1)).astype(np.float32), read_modes=None, ) memory = Memory(13, 9, 5) memory_state = memory.get_initial_state(batch_size=3) usage, write_weighting, memory, link_matrix, precedence = memory.write( memory_state, interface) self.assertEqual(usage.shape, (3, 13)) self.assertEqual(write_weighting.shape, (3, 13)) self.assertEqual(memory.shape, (3, 13, 9)) self.assertEqual(link_matrix.shape, (3, 13, 13)) self.assertEqual(precedence.shape, (3, 13))
def test_construction(self): interface = DNC.interface( read_keys=None, read_strengths=None, write_key=np.random.uniform(0, 1, (3, 9, 1)).astype(np.float32), write_strength=np.random.uniform(0, 1, (3, 1)).astype(np.float32), erase_vector=tf.convert_to_tensor( np.zeros((3, 9)).astype(np.float32)), write_vector=tf.convert_to_tensor( np.random.uniform(0, 1, (3, 9)).astype(np.float32)), free_gates=np.random.uniform(0, 1, (3, 5)).astype(np.float32), allocation_gate=np.random.uniform(0, 1, (3, 1)).astype(np.float32), write_gate=np.random.uniform(0, 1, (3, 1)).astype(np.float32), read_modes=None, ) memory = Memory(13, 9, 5) memory_state = memory.initial_state(3) write_op = memory.write(memory_state, interface) init_op = tf.global_variables_initializer() with self.test_session() as session: init_op.run() usage, write_weighting, memory, link_matrix, precedence = session.run( write_op) self.assertEqual(usage.shape, (3, 13)) self.assertEqual(write_weighting.shape, (3, 13)) self.assertEqual(memory.shape, (3, 13, 9)) self.assertEqual(link_matrix.shape, (3, 13, 13)) self.assertEqual(precedence.shape, (3, 13))
class DNC: def __init__(self, controller_class, input_size, output_size, max_sequence_length, memory_words_num=256, memory_word_size=64, memory_read_heads=4, batch_size=128): """ constructs a complete DNC architecture as described in the DNC paper http://www.nature.com/nature/journal/vaop/ncurrent/full/nature20101.html Parameters: ----------- controller_class: BaseController a concrete implementation of the BaseController class input_size: int the size of the input vector output_size: int the size of the output vector max_sequence_length: int the maximum length of an input sequence memory_words_num: int the number of words that can be stored in memory memory_word_size: int the size of an individual word in memory memory_read_heads: int the number of read heads in the memory batch_size: int the size of the data batch """ self.input_size = input_size self.output_size = output_size self.max_sequence_length = max_sequence_length self.words_num = memory_words_num self.word_size = memory_word_size self.read_heads = memory_read_heads self.batch_size = batch_size self.memory = Memory(self.words_num, self.word_size, self.read_heads, self.batch_size) self.controller = controller_class(self.input_size, self.output_size, self.read_heads, self.word_size, self.batch_size) # input data placeholders self.input_data = tf.placeholder(tf.float32, [None, None, chunk_size], name='input') self.target_output = tf.placeholder(tf.float32, [None, None, output_size], name='targets') #self.input_data = tf.placeholder(tf.float32, [batch_size, None, input_size], name='input') #self.target_output = tf.placeholder(tf.float32, [batch_size, None, output_size], name='targets') self.sequence_length = tf.placeholder(tf.int32, name='sequence_length') self.build_graph() def _step_op(self, step, memory_state, controller_state=None): """ performs a step operation on the input step data Parameters: ---------- step: Tensor (batch_size, input_size) memory_state: Tuple a tuple of current memory parameters controller_state: Tuple the state of the controller if it's recurrent Returns: Tuple output: Tensor (batch_size, output_size) memory_view: dict """ last_read_vectors = memory_state[6] pre_output, interface, nn_state = None, None, None if self.controller.has_recurrent_nn: pre_output, interface, nn_state = self.controller.process_input( step, last_read_vectors, controller_state) else: pre_output, interface = self.controller.process_input( step, last_read_vectors) usage_vector, write_weighting, memory_matrix, link_matrix, precedence_vector = self.memory.write( memory_state[0], memory_state[1], memory_state[5], memory_state[4], memory_state[2], memory_state[3], interface['write_key'], interface['write_strength'], interface['free_gates'], interface['allocation_gate'], interface['write_gate'], interface['write_vector'], interface['erase_vector']) read_weightings, read_vectors = self.memory.read( memory_matrix, memory_state[5], interface['read_keys'], interface['read_strengths'], link_matrix, interface['read_modes'], ) return [ # report new memory state to be updated outside the condition branch memory_matrix, usage_vector, precedence_vector, link_matrix, write_weighting, read_weightings, read_vectors, self.controller.final_output(pre_output, read_vectors), interface['free_gates'], interface['allocation_gate'], interface['write_gate'], # report new state of RNN if exists nn_state[0] if nn_state is not None else tf.zeros(1), nn_state[1] if nn_state is not None else tf.zeros(1) ] def _loop_body(self, time, memory_state, outputs, free_gates, allocation_gates, write_gates, read_weightings, write_weightings, usage_vectors, controller_state): """ the body of the DNC sequence processing loop Parameters: ---------- time: Tensor outputs: TensorArray memory_state: Tuple free_gates: TensorArray allocation_gates: TensorArray write_gates: TensorArray read_weightings: TensorArray, write_weightings: TensorArray, usage_vectors: TensorArray, controller_state: Tuple Returns: Tuple containing all updated arguments """ step_input = self.unpacked_input_data.read(time) output_list = self._step_op(step_input, memory_state, controller_state) # update memory parameters new_controller_state = tf.zeros(1) new_memory_state = tuple(output_list[0:7]) new_controller_state = LSTMStateTuple(output_list[11], output_list[12]) outputs = outputs.write(time, output_list[7]) # collecting memory view for the current step free_gates = free_gates.write(time, output_list[8]) allocation_gates = allocation_gates.write(time, output_list[9]) write_gates = write_gates.write(time, output_list[10]) read_weightings = read_weightings.write(time, output_list[5]) write_weightings = write_weightings.write(time, output_list[4]) usage_vectors = usage_vectors.write(time, output_list[1]) return (time + 1, new_memory_state, outputs, free_gates, allocation_gates, write_gates, read_weightings, write_weightings, usage_vectors, new_controller_state) def build_graph(self): """ builds the computational graph that performs a step-by-step evaluation of the input data batches """ self.unpacked_input_data = dnc.utility.unpack_into_tensorarray( self.input_data, 1, self.sequence_length) outputs = tf.TensorArray(tf.float32, self.sequence_length, name='outputs') free_gates = tf.TensorArray(tf.float32, self.sequence_length, name='free_gates') allocation_gates = tf.TensorArray(tf.float32, self.sequence_length, name='allocation_gates') write_gates = tf.TensorArray(tf.float32, self.sequence_length, name='write_gates') read_weightings = tf.TensorArray(tf.float32, self.sequence_length, name='read_weightings') write_weightings = tf.TensorArray(tf.float32, self.sequence_length, name='write_weightings') usage_vectors = tf.TensorArray(tf.float32, self.sequence_length, name='usage_vectors') controller_state = self.controller.get_state( ) if self.controller.has_recurrent_nn else (tf.zeros(1), tf.zeros(1)) memory_state = self.memory.init_memory() if not isinstance(controller_state, LSTMStateTuple): controller_state = LSTMStateTuple(controller_state[0], controller_state[1]) final_results = None with tf.variable_scope("sequence_loop") as scope: time = tf.placeholder(dtype=tf.int32, name='time') final_results = tf.while_loop( cond=lambda time, *_: time < self.sequence_length, body=self._loop_body, loop_vars=(time, memory_state, outputs, free_gates, allocation_gates, write_gates, read_weightings, write_weightings, usage_vectors, controller_state), parallel_iterations=32, swap_memory=False) dependencies = [] if self.controller.has_recurrent_nn: dependencies.append(self.controller.update_state(final_results[9])) with tf.control_dependencies(dependencies): self.packed_output = dnc.utility.pack_into_tensor(final_results[2], axis=1) self.packed_memory_view = { 'free_gates': dnc.utility.pack_into_tensor(final_results[3], axis=1), 'allocation_gates': dnc.utility.pack_into_tensor(final_results[4], axis=1), 'write_gates': dnc.utility.pack_into_tensor(final_results[5], axis=1), 'read_weightings': dnc.utility.pack_into_tensor(final_results[6], axis=1), 'write_weightings': dnc.utility.pack_into_tensor(final_results[7], axis=1), 'usage_vectors': dnc.utility.pack_into_tensor(final_results[8], axis=1) } def get_outputs(self): """ returns the graph nodes for the output and memory view Returns: Tuple outputs: Tensor (batch_size, time_steps, output_size) memory_view: dict """ return self.packed_output, self.packed_memory_view def save(self, session, ckpts_dir, name): """ saves the current values of the model's parameters to a checkpoint Parameters: ---------- session: tf.Session the tensorflow session to save ckpts_dir: string the path to the checkpoints directories name: string the name of the checkpoint subdirectory """ checkpoint_dir = os.path.join(ckpts_dir, name) if not os.path.exists(checkpoint_dir): os.makedirs(checkpoint_dir) tf.train.Saver(tf.trainable_variables()).save( session, os.path.join(checkpoint_dir, 'model.ckpt')) def restore(self, session, ckpts_dir, name): """ session: tf.Session the tensorflow session to restore into ckpts_dir: string the path to the checkpoints directories name: string the name of the checkpoint subdirectory """ tf.train.Saver(tf.trainable_variables()).restore( session, os.path.join(ckpts_dir, name, 'model.ckpt'))