import os import time import numpy as np import torch from lib import utils from model.pytorch.dcrnn_model import EncoderModel, DecoderModel device = torch.device("cuda" if torch.cuda.is_available() else "cpu") class DCRNNSupervisor: def __init__(self, adj_mx, **kwargs): self._kwargs = kwargs self._data_kwargs = kwargs.get('data') self._model_kwargs = kwargs.get('model') self._train_kwargs = kwargs.get('train') self.max_grad_norm = self._train_kwargs.get('max_grad_norm', 1.) # logging. self._log_dir = self._get_log_dir(kwargs) log_level = self._kwargs.get('log_level', 'INFO') self._logger = utils.get_logger(self._log_dir, __name__, 'info.log', level=log_level) # data set self._data = utils.load_dataset(**self._data_kwargs) self.standard_scaler = self._data['scaler'] self.num_nodes = int(self._model_kwargs.get('num_nodes', 1)) self.input_dim = int(self._model_kwargs.get('input_dim', 1)) self.seq_len = int(self._model_kwargs.get('seq_len')) # for the encoder self.output_dim = int(self._model_kwargs.get('output_dim', 1)) self.cl_decay_steps = int(self._model_kwargs.get('cl_decay_steps', 1000)) self.use_curriculum_learning = bool( self._model_kwargs.get('use_curriculum_learning', False)) self.horizon = int(self._model_kwargs.get('horizon', 1)) # for the decoder # setup model self.encoder_model = EncoderModel(True, adj_mx, **self._model_kwargs) self.decoder_model = DecoderModel(True, adj_mx, **self._model_kwargs) @staticmethod def _get_log_dir(kwargs): log_dir = kwargs['train'].get('log_dir') if log_dir is None: batch_size = kwargs['data'].get('batch_size') learning_rate = kwargs['train'].get('base_lr') max_diffusion_step = kwargs['model'].get('max_diffusion_step') num_rnn_layers = kwargs['model'].get('num_rnn_layers') rnn_units = kwargs['model'].get('rnn_units') structure = '-'.join( ['%d' % rnn_units for _ in range(num_rnn_layers)]) horizon = kwargs['model'].get('horizon') filter_type = kwargs['model'].get('filter_type') filter_type_abbr = 'L' if filter_type == 'random_walk': filter_type_abbr = 'R' elif filter_type == 'dual_random_walk': filter_type_abbr = 'DR' run_id = 'dcrnn_%s_%d_h_%d_%s_lr_%g_bs_%d_%s/' % ( filter_type_abbr, max_diffusion_step, horizon, structure, learning_rate, batch_size, time.strftime('%m%d%H%M%S')) base_dir = kwargs.get('base_dir') log_dir = os.path.join(base_dir, run_id) if not os.path.exists(log_dir): os.makedirs(log_dir) return log_dir def train(self, **kwargs): kwargs.update(self._train_kwargs) return self._train(**kwargs) def _train_one_batch(self, inputs, labels, batches_seen, encoder_optimizer, decoder_optimizer, criterion): """ :param inputs: shape (seq_len, batch_size, num_sensor, input_dim) :param labels: shape (horizon, batch_size, num_sensor, input_dim) :param encoder_optimizer: :param decoder_optimizer: :param criterion: minimize this criterion :return: loss? """ encoder_optimizer.zero_grad() decoder_optimizer.zero_grad() batch_size = inputs.size(1) inputs = inputs.view(self.seq_len, batch_size, self.num_nodes * self.input_dim) labels = labels[..., :self.output_dim].view(self.horizon, batch_size, self.num_nodes * self.output_dim) loss = 0 encoder_hidden_state = None for t in range(self.seq_len): _, encoder_hidden_state = self.encoder_model.forward(inputs[t], encoder_hidden_state) self._logger.info("Encoder complete, starting decoder") go_symbol = torch.zeros((batch_size, self.num_nodes * self.output_dim)) decoder_hidden_state = encoder_hidden_state decoder_input = go_symbol outputs = [] for t in range(self.horizon): decoder_output, decoder_hidden_state = self.decoder_model.forward(decoder_input, decoder_hidden_state) decoder_input = decoder_output outputs.append(decoder_output) if self.use_curriculum_learning: # todo check for is_training (pytorch way?) c = np.random.uniform(0, 1) if c < self._compute_sampling_threshold(batches_seen): decoder_input = labels[t] loss += criterion(self.standard_scaler.inverse_transform(decoder_output), self.standard_scaler.inverse_transform(labels[t])) self._logger.info("Decoder complete, starting backprop") loss.backward() # gradient clipping - this does it in place torch.nn.utils.clip_grad_norm_(self.encoder_model.parameters(), self.max_grad_norm) torch.nn.utils.clip_grad_norm_(self.decoder_model.parameters(), self.max_grad_norm) encoder_optimizer.step() decoder_optimizer.step() outputs = torch.stack(outputs) return outputs.view(self.horizon, batch_size, self.num_nodes, self.output_dim), loss.item() def _train(self, base_lr, steps, patience=50, epochs=100, min_learning_rate=2e-6, lr_decay_ratio=0.1, log_every=10, save_model=1, test_every_n_epochs=10, **kwargs): # steps is used in learning rate - will see if need to use it? encoder_optimizer = torch.optim.Adam(self.encoder_model.parameters(), lr=base_lr) decoder_optimizer = torch.optim.Adam(self.encoder_model.parameters(), lr=base_lr) criterion = torch.nn.L1Loss() # mae loss batches_seen = 0 self._logger.info('Start training ...') for epoch_num in range(epochs): train_iterator = self._data['train_loader'].get_iterator() losses = [] start_time = time.time() for _, (x, y) in enumerate(train_iterator): x = torch.from_numpy(x).float() y = torch.from_numpy(y).float() self._logger.debug("X: {}".format(x.size())) self._logger.debug("y: {}".format(y.size())) x = x.permute(1, 0, 2, 3) y = y.permute(1, 0, 2, 3) output, loss = self._train_one_batch(x, y, batches_seen, encoder_optimizer, decoder_optimizer, criterion) losses.append(loss) batches_seen += 1 end_time = time.time() if epoch_num % log_every == 0: message = 'Epoch [{}/{}] ({}) train_mae: {:.4f}, val_mae: {:.4f} ' \ 'lr:{:.6f} {:.1f}s'.format(epoch_num, epochs, batches_seen, np.mean(losses), 0.0, 0.0, (end_time - start_time)) self._logger.info(message) def _compute_sampling_threshold(self, batches_seen): return self.cl_decay_steps / ( self.cl_decay_steps + np.exp(batches_seen / self.cl_decay_steps))