DCRNN/model/dcrnn_model.py

107 lines
5.4 KiB
Python

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
from tensorflow.contrib import legacy_seq2seq
from lib.metrics import masked_mse_loss, masked_mae_loss, masked_rmse_loss
from model.dcrnn_cell import DCGRUCell
from model.tf_model import TFModel
class DCRNNModel(TFModel):
def __init__(self, is_training, config, scaler=None, adj_mx=None):
super(DCRNNModel, self).__init__(config, scaler=scaler)
batch_size = int(config.get('batch_size'))
max_diffusion_step = int(config.get('max_diffusion_step', 2))
cl_decay_steps = int(config.get('cl_decay_steps', 1000))
filter_type = config.get('filter_type', 'laplacian')
horizon = int(config.get('horizon', 1))
input_dim = int(config.get('input_dim', 1))
loss_func = config.get('loss_func', 'MSE')
max_grad_norm = float(config.get('max_grad_norm', 5.0))
num_nodes = int(config.get('num_nodes', 1))
num_rnn_layers = int(config.get('num_rnn_layers', 1))
output_dim = int(config.get('output_dim', 1))
rnn_units = int(config.get('rnn_units'))
seq_len = int(config.get('seq_len'))
use_curriculum_learning = bool(config.get('use_curriculum_learning', False))
assert input_dim == output_dim, 'input_dim: %d != output_dim: %d' % (input_dim, output_dim)
# Input (batch_size, timesteps, num_sensor, input_dim)
self._inputs = tf.placeholder(tf.float32, shape=(batch_size, seq_len, num_nodes, input_dim), name='inputs')
# Labels: (batch_size, timesteps, num_sensor, output_dim)
self._labels = tf.placeholder(tf.float32, shape=(batch_size, horizon, num_nodes, output_dim), name='labels')
GO_SYMBOL = tf.zeros(shape=(batch_size, num_nodes * input_dim))
cell = DCGRUCell(rnn_units, adj_mx, max_diffusion_step=max_diffusion_step, num_nodes=num_nodes,
filter_type=filter_type)
cell_with_projection = DCGRUCell(rnn_units, adj_mx, max_diffusion_step=max_diffusion_step, num_nodes=num_nodes,
num_proj=output_dim, filter_type=filter_type)
encoding_cells = [cell] * num_rnn_layers
decoding_cells = [cell] * (num_rnn_layers - 1) + [cell_with_projection]
encoding_cells = tf.contrib.rnn.MultiRNNCell(encoding_cells, state_is_tuple=True)
decoding_cells = tf.contrib.rnn.MultiRNNCell(decoding_cells, state_is_tuple=True)
global_step = tf.train.get_or_create_global_step()
# Outputs: (batch_size, timesteps, num_nodes, output_dim)
with tf.variable_scope('DCRNN_SEQ'):
inputs = tf.unstack(tf.reshape(self._inputs, (batch_size, seq_len, num_nodes * input_dim)), axis=1)
labels = tf.unstack(tf.reshape(self._labels, (batch_size, horizon, num_nodes * output_dim)), axis=1)
labels.insert(0, GO_SYMBOL)
loop_function = None
if is_training:
if use_curriculum_learning:
def loop_function(prev, i):
c = tf.random_uniform((), minval=0, maxval=1.)
threshold = self._compute_sampling_threshold(global_step, cl_decay_steps)
result = tf.cond(tf.less(c, threshold), lambda: labels[i], lambda: prev)
return result
else:
# Return the output of the model.
def loop_function(prev, _):
return prev
_, enc_state = tf.contrib.rnn.static_rnn(encoding_cells, inputs, dtype=tf.float32)
outputs, final_state = legacy_seq2seq.rnn_decoder(labels, enc_state, decoding_cells,
loop_function=loop_function)
# Project the output to output_dim.
outputs = tf.stack(outputs[:-1], axis=1)
self._outputs = tf.reshape(outputs, (batch_size, horizon, num_nodes, output_dim), name='outputs')
preds = self._outputs[..., 0]
labels = self._labels[..., 0]
null_val = config.get('null_val', 0.)
self._mae = masked_mae_loss(self._scaler, null_val)(preds=preds, labels=labels)
if loss_func == 'MSE':
self._loss = masked_mse_loss(self._scaler, null_val)(preds=self._outputs, labels=self._labels)
elif loss_func == 'MAE':
self._loss = masked_mae_loss(self._scaler, null_val)(preds=self._outputs, labels=self._labels)
elif loss_func == 'RMSE':
self._loss = masked_rmse_loss(self._scaler, null_val)(preds=self._outputs, labels=self._labels)
else:
self._loss = masked_mse_loss(self._scaler, null_val)(preds=self._outputs, labels=self._labels)
if is_training:
optimizer = tf.train.AdamOptimizer(self._lr)
tvars = tf.trainable_variables()
grads = tf.gradients(self._loss, tvars)
grads, _ = tf.clip_by_global_norm(grads, max_grad_norm)
self._train_op = optimizer.apply_gradients(zip(grads, tvars), global_step=global_step, name='train_op')
self._merged = tf.summary.merge_all()
@staticmethod
def _compute_sampling_threshold(global_step, k):
"""
Computes the sampling probability for scheduled sampling using inverse sigmoid.
:param global_step:
:param k:
:return:
"""
return tf.cast(k / (k + tf.exp(global_step / k)), tf.float32)