Fix typos and improve language

lib/bumblebee/diffusion/layers.ex

@@ -63,7 +63,7 @@ defmodule Bumblebee.Diffusion.Layers do
 
     if rem(embedding_size, 2) != 0 do
       raise ArgumentError,
-            "expected embedding size to an even number, but got: #{inspect(embedding_size)}"
+            "expected embedding size to be an even number, but got: #{inspect(embedding_size)}"
     end
 
     half_size = div(embedding_size, 2)

lib/bumblebee/huggingface/transformers/model.ex

@@ -33,11 +33,11 @@ defprotocol Bumblebee.HuggingFace.Transformers.Model do
         ...
       }
 
-  The mapping should always use longest names, that is, depending on
+  The mapping should always use the longest names, that is, depending on
   the architecture, the PyTorch layer name could be either
   `"bert.embeddings.word_embeddings"` or `"embeddings.word_embeddings"`.
   The longer version should generally be used. Prefixes are removed/added
-  as necessary, so loading partial models supported automatically.
+  as necessary, so loading partial models is supported automatically.
 
   The layer names may include simple substitutions, useful for lists
   of layers:

lib/bumblebee/layers.ex

@@ -682,7 +682,7 @@ defmodule Bumblebee.Layers do
   end
 
   @doc """
-  Splits the input node with shape `{bach_size, sequence_length, 2}` into
+  Splits the input node with shape `{batch_size, sequence_length, 2}` into
   two nodes with shape `{batch_size, sequence_length}`.
   """
   def split_pair(%Axon{} = x) do
@@ -783,7 +783,7 @@ defmodule Bumblebee.Layers do
   end
 
   @doc """
-  Adds a layer that that computes cosine similarity between the inputs.
+  Adds a layer that computes cosine similarity between the inputs.
   """
   def cosine_similarity(x, y) do
     Axon.layer(&cosine_similarity_impl/3, [x, y], op_names: :cosine_similarity)

test/bumblebee/audio/whisper_test.exs

@@ -1,4 +1,4 @@
-defmodule Bumblebee.Text.WhisperTest do
+defmodule Bumblebee.Audio.WhisperTest do
   use ExUnit.Case, async: true
 
   import Bumblebee.TestHelpers