def tf_nse(true, pred, name='NSE'):
""" Nash-Sutcliff efficiency to be used as loss function. It is subtracted from one before being returned"""
neum = tf.reduce_sum(tf.square(tf.subtract(pred, true)))
denom = tf.reduce_sum(
tf.square(tf.subtract(true, tf.math.reduce_mean(true))))
const = tf.constant(1.0, dtype=tf.float32)
_nse = tf.subtract(const, tf.math.divide(neum, denom), name=name)
return tf.subtract(const, _nse, name=name + '_LOSS')
def val_step(_val_x, complete_y):
masks = []
y_trues = []
for out in range(self.outs):
mask = tf.greater(tf.reshape(complete_y[out], (-1,)), 0.0) # # # (batch_size,)
masks.append(mask)
y_trues.append(complete_y[out][mask])
predictions = self.k_model(_val_x, training=False) # predictions for this minibatch
val_step_losses = {}
for out in range(self.outs):
y_obj = predictions[out][masks[out]]
val_step_losses['_' + str(out)] = keras.backend.mean(loss_fn(y_trues[out], y_obj))
# Compute the loss value for this minibatch.
_x = tf.stack(list(val_step_losses.values()))
_x = tf.boolean_mask(_x, self.val_outs)
loss_val = tf.reduce_sum(tf.boolean_mask(_x, tf.math.is_finite(_x)))
val_step_losses.update({'loss': loss_val})
return val_step_losses
def train_step(train__x, complete_y):
#skip_flag = False
with tf.GradientTape() as taape:
_masks = []
_y_trues = []
for out in range(self.outs):
mask = tf.greater(tf.reshape(complete_y[out], (-1,)), 0.0) # # # (batch_size,)
_masks.append(mask)
_y_trues.append(complete_y[out][mask])
_predictions = self.k_model(train__x, training=True) # predictions for this minibatch
losses = {}
for out in range(self.outs):
y_obj = _predictions[out][_masks[out]]
#if len(y_obj) < 1:
# skip_flag = True
losses['_' + str(out)] = keras.backend.mean(loss_fn(_y_trues[out], y_obj))
# Compute the loss value for this minibatch.
# loss_val = tf.reduce_sum(list(losses.values()))
_x = tf.stack(list(losses.values()))
_x = tf.boolean_mask(_x, self.tr_outs)
loss_val = tf.reduce_sum(tf.boolean_mask(_x, tf.math.is_finite(_x)))
losses.update({'loss': float(loss_val)})
_grads = taape.gradient(loss_val, self.k_model.trainable_weights) # list
grads = [tf.clip_by_value(g, -1.0, 1.0) for g in _grads]
# Run one step of gradient descent by updating the value of the variables to minimize the loss.
optimizer.apply_gradients(zip(grads, self.k_model.trainable_weights))
return losses#, skip_flag