Follow-ups from initial mypy PR (facebookresearch#25)

Summary:
Pull Request resolved: facebookresearch#25

A couple follow-ups from facebookresearch#22:
- Add TODO for type checking FLAVA code
- Add a couple other type hints missed by mypy in clip_text_encoder (due to https://mypy.readthedocs.io/en/stable/common_issues.html#no-errors-reported-for-obviously-wrong-code)
- Some changes to weighted_embedding_encoder to silence mypy errors

Test Plan:
```
$ python -m pytest -v
...
====================================================================================== short test summary info =======================================================================================
FAILED test/transforms/test_clip_transform.py::TestCLIPTransform::test_clip_multi_transform - requests.exceptions.MissingSchema: Invalid URL '/data/home/ebs/torchmultimodal/torchmultimodal/test/a...
FAILED test/transforms/test_clip_transform.py::TestCLIPTransform::test_clip_single_transform - requests.exceptions.MissingSchema: Invalid URL '/data/home/ebs/torchmultimodal/torchmultimodal/test/...
================================================================== 2 failed, 48 passed, 3 skipped, 26 warnings in 76.48s (0:01:16) ===================================================================

```
(Failures are pre-existing and will be fixed with a separate commit)

```
$ mypy
Success: no issues found in 29 source files
```

Reviewed By: langong347

Differential Revision: D35617460

Pulled By: ebsmothers

fbshipit-source-id: ab2bf0ec7f85fa700287787728aa40443b9c7a0c

Author: ebsmothers

mypy.ini

@@ -13,6 +13,7 @@ allow_redefinition = True
 namespace_packages = True
 install_types = True
 
+# TODO (T116951827): Remove after fixing FLAVA type check errors
 exclude = models/flava.py|modules/losses/flava.py
 
 [mypy-PIL.*]

torchmultimodal/modules/encoders/clip_text_encoder.py

@@ -61,7 +61,7 @@ def __init__(
         if use_clip_init:
             self.initialize_parameters()
 
-    def initialize_parameters(self):
+    def initialize_parameters(self) -> None:
         # Initialize token and positional embeddings
         nn.init.normal_(
             self.encoder.token_embedding.weight, std=self.TOKEN_EMBEDDING_INIT_STD
@@ -85,7 +85,7 @@ def initialize_parameters(self):
         # Initialize projection
         nn.init.normal_(self.projection.weight, std=self.width ** -0.5)
 
-    def build_attention_mask(self):
+    def build_attention_mask(self) -> torch.Tensor:
         mask = torch.full((self.context_length, self.context_length), True).triu(1)
         return mask.to(device=None, dtype=torch.bool)
 

torchmultimodal/modules/encoders/weighted_embedding_encoder.py

@@ -4,7 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from typing import Tuple, Union, Callable
+from typing import Callable, Tuple, Union
 
 import torch
 from torch import nn, Tensor
@@ -16,7 +16,7 @@ class WeightedEmbeddingEncoder(nn.Module):
 
     Args:
         embedding (nn.Embedding): embedding module
-        pooling_function (Callable[[Tensor], Union[Tensor, Tuple]]): pooling function to combine the weighted embeddings,\
+        pooling_function (Callable[[Tensor, int], Union[Tensor, Tuple]]): pooling function to combine the weighted embeddings,\
         example: torch.sum function should return a tensor or namedtuple containing the tensor in the values field like torch.max
         pooling_dim (int) : dimension along which the pooling function is applied
         use_hash (bool): if hashing based on embedding vocab size if applied to input
@@ -31,7 +31,7 @@ class WeightedEmbeddingEncoder(nn.Module):
     def __init__(
         self,
         embedding: nn.Embedding,
-        pooling_function: Callable[[Tensor], Union[Tensor, Tuple]],
+        pooling_function: Callable[[Tensor, int], Union[Tensor, Tuple]],
         pooling_dim: int = 1,
         use_hash: bool = False,
     ) -> None:
@@ -67,12 +67,12 @@ def forward(self, x: Tensor) -> Tensor:
 
         weighted_embeddings = self.embedding(index) * weights.unsqueeze(-1)
 
-        pooled_embeddings = self.pooling_function(
-            weighted_embeddings, dim=self.pooling_dim
-        )
-        if not isinstance(pooled_embeddings, Tensor):
+        pooled_embeddings = self.pooling_function(weighted_embeddings, self.pooling_dim)
+        if isinstance(pooled_embeddings, Tensor):
+            output: Tensor = pooled_embeddings
+        else:
             assert hasattr(
                 pooled_embeddings, "values"
             ), "pooled embeddings should be Tensor or tuple with values field as Tensor"
-            pooled_embeddings = pooled_embeddings.values
-        return pooled_embeddings
+            output = pooled_embeddings.values  # type: ignore
+        return output