1. Add col_name as default args in feature_preprocess (Fix #105)

2. Update TransAct with adjust_mask

fuxictr/datasets/criteo.py

@@ -22,7 +22,7 @@
 
 
 class CustomizedFeatureProcessor(FeatureProcessor):
- def convert_to_bucket(self, col_name=None):
+ def convert_to_bucket(self, col_name):
  def _convert_to_bucket(value):
  if value > 2:
  value = int(np.floor(np.log(value) ** 2))

fuxictr/datasets/kkbox.py

@@ -21,12 +21,12 @@
 
 
 class CustomizedFeatureProcessor(FeatureProcessor):
- def extract_country_code(self, col_name=None):
+ def extract_country_code(self, col_name):
  return pl.col(col_name).apply(lambda isrc: isrc[0:2] if not pl.is_null(isrc) else "")
 
- def bucketize_age(self, col_name=None):
+ def bucketize_age(self, col_name):
  def _bucketize(age):
- if pd.isnull(age):
+ if pl.is_null(age):
  return ""
  else:
  age = float(age)

fuxictr/preprocess/feature_processor.py

@@ -96,8 +96,12 @@ def preprocess(self, ddf):
  if col.get("preprocess"):
  preprocess_args = re.split(r"\(|\)", col["preprocess"])
  preprocess_fn = getattr(self, preprocess_args[0])
+ if len(preprocess_args) == 1:
+ preprocess_args = [name] # use col_name as args when not being explicitly set
+ else:
+ preprocess_args = preprocess_args[1:-1]
  ddf = ddf.with_columns(
- preprocess_fn(*preprocess_args[1:-1])
+ preprocess_fn(*preprocess_args)
  .alias(name)
  .cast(self.dtype_dict[name])
  )
@@ -364,5 +368,5 @@ def save_vocab(self, vocab_file):
  with open(vocab_file, "w") as fd:
  fd.write(json.dumps(vocab, indent=4))
 
- def copy_from(self, col_name):
- return pl.col(col_name)
+ def copy_from(self, src_col):
+ return pl.col(src_col)

model_zoo/TransAct/src/TransAct.py

@@ -220,7 +220,7 @@ def forward(self, target_emb, sequence_emb, time_interval_seq=None, mask=None):
  concat_seq_emb = torch.cat([sequence_emb,
  target_emb.unsqueeze(1).expand(-1, seq_len, -1)], dim=-1)
  # get sequence mask (1's are masked)
- key_padding_mask = mask
+ key_padding_mask = self.adjust_mask(mask) # keep the last dim
  if self.use_time_window_mask and self.training:
  rand_time_window_ms = random.randint(0, self.time_window_ms)
  time_window_mask = (time_interval_seq < rand_time_window_ms)
@@ -235,7 +235,6 @@ def forward(self, target_emb, sequence_emb, time_interval_seq=None, mask=None):
  output_concat.append(tfmr_out[:, -self.first_k_cols:].flatten(start_dim=1))
  if self.concat_max_pool:
  # Apply max pooling to the transformer output
- key_padding_mask = self.adjust_mask(key_padding_mask) # keep the last dim
  tfmr_out = tfmr_out.masked_fill(
  key_padding_mask.unsqueeze(-1).repeat(1, 1, tfmr_out.shape[-1]), -1e9
  )