add mean pooling to laser transformer

examples/laser/laser_src/laser_transformer.py

@@ -29,6 +29,8 @@
 
 logger = logging.getLogger(__name__)
 
+LASER_EMBED_DIM = 1024
+
 @register_model("laser_transformer")
 class LaserTransformerModel(FairseqEncoderDecoderModel):
     """Train Transformer for LASER task
@@ -66,7 +68,7 @@ def add_args(parser):
         )
         parser.add_argument(
             "--sentemb-criterion",
-            choices=["maxpool", "cls"],
+            choices=["maxpool", "meanpool", "cls"],
             help="How to build sentence embeddings?",
         )
         parser.add_argument(
@@ -126,7 +128,6 @@ def __init__(self, sentemb_criterion, *args, **kwargs):
         tasks = [
             task.split(":")[0] for task in namespace.student_teacher_config.split(",")
         ]
-        laser_embed_dim = 1024
         # if we have a masking task, then add a linear layer projecting from embed_dim to vocab_size
         if "mask" in tasks:
             self.project_vocabulary = nn.Linear(
@@ -143,9 +144,9 @@ def __init__(self, sentemb_criterion, *args, **kwargs):
             self.activation_fn = utils.get_activation_fn("tanh")
             self.layer_norm = LayerNorm(namespace.encoder_embed_dim)
         # if the embed_dim is different, then add a linear layer projecting from embed_dim to the Laser one (1024)
-        elif namespace.encoder_embed_dim != laser_embed_dim:
+        elif namespace.encoder_embed_dim != LASER_EMBED_DIM:
             self.output_projection = nn.Linear(
-                namespace.encoder_embed_dim, laser_embed_dim, bias=False
+                namespace.encoder_embed_dim, LASER_EMBED_DIM, bias=False
             )
             nn.init.normal_(
                 self.output_projection.weight,
@@ -191,6 +192,9 @@ def forward(
         encoder_out = super().forward(src_tokens, src_lengths)
 
         x = encoder_out["encoder_out"][0]  # T x B x D
+
+        if self.output_projection: # project to LASER's embed dim
+            x = self.output_projection(x)
 
         # project masked tokens only if performing MLM task
         if isinstance(masked_tokens, torch.Tensor):
@@ -204,25 +208,26 @@ def forward(
             return [x]  # MLM criterion takes first element of list as logits
         else:
             # if not MLM task return sentence embedding
-
-            if self.output_projection:
-                x = self.output_projection(x)
-
-            padding_mask = src_tokens.eq(self.padding_idx).t().unsqueeze(-1)
 
-            if padding_mask.any() and self.sentemb_criterion == "maxpool":
-                x = x.float().masked_fill_(padding_mask, float("-inf")).type_as(x)
+            padding_mask = src_tokens.eq(self.padding_idx).t().unsqueeze(-1)                
 
             if self.sentemb_criterion == "cls":
                 # determine location of 'cls' e.g. due to left-padding 'cls' may be different each sent
                 cls_indices = src_tokens.eq(self.dictionary.bos()).t()  # T x B
                 sentemb = x[cls_indices, :]
             # Build the sentence embedding by max-pooling over the encoder outputs
             elif self.sentemb_criterion == "maxpool":
+                if padding_mask.any():
+                    x = x.float().masked_fill_(padding_mask, float("-inf")).type_as(x)
                 sentemb = x.max(dim=0)[0]
+            # Build the sentence embedding by mean-pooling over the encoder outputs
+            elif self.sentemb_criterion == "meanpool":
+                if padding_mask.any():
+                    x = x.float().masked_fill_(padding_mask, 0.).type_as(x)
+                sentemb = x.sum(dim=0) / (~padding_mask).sum(dim=0)
             else:
                 raise Exception(
-                    "Please provide a sentence embedding option from [cls|maxpool]"
+                    "Please provide a sentence embedding option from [cls|maxpool|meanpool]"
                 )
 
             return {"sentemb": sentemb}  # B x D