class RNN(CustomModel):
def define_layers(self, params):
self.output_dim = params['output_dim']
self.rnn = RNNUnrollSubmodel(model=self, gru_hidden_sizes=params['hidden_sizes'],
std_hidden_sizes=[], n_out = params['output_dim'],
recurrent_type=params['recurrent_type'])
std_layers = {'bistablernn': self.rnn}
return {**std_layers}
@tf.function
def cross_entropy(self, samples, **kwargs):
observations = samples['groundtruth']
out = self.rnn.forward(samples['input'])
return {'cross-entropy': tf.nn.softmax_cross_entropy_with_logits(labels=observations, logits=out)}
@tf.function
def accuracy(self, samples):
observations = samples['groundtruth']
out = self.rnn.forward(samples['input'])
return {'accuracy': tf.metrics.categorical_accuracy(y_true=observations, y_pred=out)}
def pred(self, samples):
out = self.rnn.forward(samples['input'])
return {'pred':out}
def create_optimizer(self, params):
self.opt = tf.optimizers.Adam(learning_rate=params['lr'])
def define_layers(self, params):
self.output_dim = params['output_dim']
self.rnn = RNNUnrollSubmodel(model=self, gru_hidden_sizes=params['hidden_sizes'],
std_hidden_sizes=[], n_out = params['output_dim'],
recurrent_type=params['recurrent_type'])
std_layers = {'bistablernn': self.rnn}
return {**std_layers}
class RNN(CustomModel):
def define_layers(self, params):
self.output_dim = params['output_dim']
self.rnn = RNNUnrollSubmodel(model=self,
gru_hidden_sizes=params['hidden_sizes'],
std_hidden_sizes=[],
n_out=params['output_dim'],
recurrent_type=params['recurrent_type'])
std_layers = {'rnn': self.rnn}
return {**std_layers}
def mse(self, samples, **kwargs):
observations = samples['output']
out = self.rnn.forward(samples['input'])
return {'MSE': tf.square(observations - out)}
def pred(self, samples):
out = self.rnn.forward(samples['input'])
return {'pred': out}
def create_optimizer(self, params):
self.opt = tf.optimizers.Adam(learning_rate=params['lr'])