Add early stop to autoencoder

dfencoder/autoencoder.py

@@ -165,6 +165,7 @@ def __init__(self,
  progress_bar=True,
  n_megabatches=1,
  scaler='standard',
+ patience=5,
  preset_cats=None,
  *args,
  **kwargs):
@@ -226,6 +227,8 @@ def __init__(self,
  self.project_embeddings = project_embeddings
 
  self.scaler = scaler
+
+ self.patience = patience
 
  self.n_megabatches = n_megabatches
 
@@ -656,6 +659,9 @@ def fit(self, df, epochs=1, val=None):
  n_updates = len(df) // self.batch_size
  if len(df) % self.batch_size > 0:
  n_updates += 1
+ last_loss = 5000
+
+ count_es = 0
  for i in range(epochs):
  self.train()
  if self.verbose:
@@ -694,6 +700,28 @@ def fit(self, df, epochs=1, val=None):
  _, _, _, net_loss = self.compute_loss(num, bin, cat, slc_out, _id=True)
  id_loss.append(net_loss)
 
+ # Early stopping
+ current_net_loss = net_loss
+ if self.verbose:
+ print('The Current Net Loss:', current_net_loss)
+
+ if current_net_loss > last_loss:
+ count_es += 1
+ if self.verbose:
+ print('Early stop count:', count_es)
+
+ if count_es >= self.patience:
+ if self.verbose:
+ print('Early stopping: early stop count({}) >= patience({})'.format(count_es, self.patience))
+ break
+
+ else:
+ if self.verbose:
+ print('Set count for earlystop: 0')
+ count_es = 0
+
+ last_loss = current_net_loss
+
  self.logger.end_epoch()
  # if self.project_embeddings:
  # self.logger.show_embeddings(self.categorical_fts)
@@ -1050,4 +1078,4 @@ def get_results(self, df, return_abs=False):
  pdf['max_abs_z'] = combined_loss.max(dim=1)[0].cpu().numpy()
  pdf['mean_abs_z'] = combined_loss.mean(dim=1).cpu().numpy()
 
- return pdf
+ return pdf