Support String Output for BytePairTokenizer (keras-team#438)

* Support string output for BytePairTokenizer

* Add unit test

* Minor edit

keras_nlp/tokenizers/byte_pair_tokenizer.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-""" Byte-pair encoder implementation.
+"""Byte-pair encoder implementation.
 
 This file implements the same logic as openai BPE:
 https://github.com/openai/gpt-2/blob/master/src/encoder.py,
@@ -159,12 +159,12 @@ def create_static_hashtable(keys, values, default):
 class BytePairTokenizer(tokenizer.Tokenizer):
     """Bype-pair encoding tokenizer layer.
 
-    This BPE tokenizer provides the same funtionality as official GPT2
+    This BPE tokenizer provides the same functionality as the official GPT-2
     tokenizer. Given the same `vocabulary` which maps tokens to ids, and `merges`
     which describes BPE merge rules, it should provide the same output
-    as openai implementation (https://github.com/openai/gpt-2/blob/master/src/encoder.py).
-    Different from openai, this implementation is graph-compatible, so you can
-    use it within a tf.data pipeline.
+    as OpenAI implementation (https://github.com/openai/gpt-2/blob/master/src/encoder.py).
+    Different from OpenAI, this implementation is graph-compatible, so you can
+    use it within a `tf.data` pipeline.
 
     If input is a batch of strings (rank > 0):
     By default, the layer will output a `tf.RaggedTensor` where the last
@@ -187,7 +187,7 @@ class BytePairTokenizer(tokenizer.Tokenizer):
 
     Examples:
 
-    Use in-momery vocabulary and merge list.
+    Use in-memory vocabulary and merge list.
 
     >>> vocab = {"butter": 1, "fly": 2}
     >>> merge = ["b u", "t t", "e r", "bu tt", "butt er", "f l", "fl y"]
@@ -244,7 +244,7 @@ def __init__(
             kwargs["dtype"] = tf.int32
         else:
             dtype = tf.dtypes.as_dtype(kwargs["dtype"])
-            if not dtype.is_integer:
+            if not dtype.is_integer and dtype != tf.string:
                 raise ValueError(
                     "Output dtype must be an integer type or a string. "
                     f"Received: `dtype={dtype}`"
@@ -484,28 +484,29 @@ def process_unseen_tokens():
             lambda: cache_lookup,
         )
 
-        # Encode merged tokens.
-        tokenized_words = tf.strings.split(tokenized_words, sep=" ")
-        encoding = self.token_to_id_map.lookup(tokenized_words)
+        tokens = tf.strings.split(tokenized_words, sep=" ")
+        if self.compute_dtype != tf.string:
+            # Encode merged tokens.
+            tokens = self.token_to_id_map.lookup(tokens)
 
         # Unflatten to match input.
-        encoding = tf.RaggedTensor.from_row_splits(
-            encoding.flat_values,
-            tf.gather(encoding.row_splits, token_row_splits),
+        tokens = tf.RaggedTensor.from_row_splits(
+            tokens.flat_values,
+            tf.gather(tokens.row_splits, token_row_splits),
         )
 
         # Convert to a dense output if `sequence_length` is set.
         if self.sequence_length:
-            output_shape = encoding.shape.as_list()
+            output_shape = tokens.shape.as_list()
             output_shape[-1] = self.sequence_length
-            encoding = encoding.to_tensor(shape=output_shape)
+            tokens = tokens.to_tensor(shape=output_shape)
 
         # Convert to a dense output if input in scalar
         if scalar_input:
-            encoding = tf.squeeze(encoding, 0)
-            tf.ensure_shape(encoding, shape=[self.sequence_length])
+            tokens = tf.squeeze(tokens, 0)
+            tf.ensure_shape(tokens, shape=[self.sequence_length])
 
-        return encoding
+        return tokens
 
     def detokenize(self, inputs):
         if not isinstance(inputs, (tf.Tensor, tf.RaggedTensor)):

keras_nlp/tokenizers/byte_pair_tokenizer_test.py

@@ -52,6 +52,20 @@ def test_tokenize_list_input(self):
         encoded = self.tokenizer(input_data)
         self.assertAllEqual(encoded, expected)
 
+    def test_tokenize_string_output(self):
+        input_data = ["quick brown fox.", "slow black bear."]
+        tokenizer = BytePairTokenizer(
+            vocabulary=VOCAB_PATH, merges=MERGE_PATH, dtype=tf.string
+        )
+        call_output = tokenizer(input_data)
+        expected = tf.ragged.constant(
+            [
+                ["quick", "Ġbrown", "Ġfox", "."],
+                ["slow", "Ġblack", "Ġbear", "."],
+            ]
+        )
+        self.assertAllEqual(call_output, expected)
+
     def test_tokenize_scalar_input(self):
         input_data = "brown."
         encoded = self.tokenizer.tokenize(input_data)