correct beam search sampling

src/transformers/modeling_tf_utils.py

@@ -760,9 +760,14 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
         ]
 
         # scores for each sentence in the beam
-        beam_scores_begin = tf.zeros((batch_size, 1), dtype=tf.float32)
-        beam_scores_end = tf.zeros((batch_size, num_beams - 1), dtype=tf.float32) * 1e-9
-        beam_scores = tf.reshape(tf.concat([beam_scores_begin, beam_scores_end], -1), (batch_size * num_beams,))
+        if do_sample is False:
+            beam_scores_begin = tf.zeros((batch_size, 1), dtype=tf.float32)
+            beam_scores_end = tf.zeros((batch_size, num_beams - 1), dtype=tf.float32) * 1e-9
+            beam_scores = tf.concat([beam_scores_begin, beam_scores_end], -1)
+        else:
+            beam_scores = tf.zeros((batch_size, num_beams), dtype=tf.float32)
+
+        beam_scores = tf.reshape(beam_scores, (batch_size * num_beams,))
 
         # cache compute states
         past = None
@@ -790,23 +795,24 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                 # Temperature (higher temperature => more likely to sample low probability tokens)
                 if temperature != 1.0:
                     next_token_logits = next_token_logits / temperature
+
+                scores = tf.nn.log_softmax(next_token_logits, axis=-1)  # (batch_size * num_beams, vocab_size)
+                _scores = scores + tf.broadcast_to(
+                    beam_scores[:, None], (batch_size * num_beams, vocab_size)
+                )  # (batch_size * num_beams, vocab_size)
+
                 # Top-p/top-k filtering
-                next_token_logits = tf_top_k_top_p_filtering(
-                    next_token_logits, top_k=top_k, top_p=top_p, min_tokens_to_keep=2
+                _scores = tf_top_k_top_p_filtering(
+                    _scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2
                 )  # (batch_size * num_beams, vocab_size)
                 # Sample 2 next tokens for each beam (so we have some spare tokens and match output of greedy beam search)
+                _scores = tf.reshape(_scores, (batch_size, num_beams * vocab_size))
+
                 next_tokens = tf.random.categorical(
-                    next_token_logits, dtype=tf.int32, num_samples=2
-                )  # (batch_size * num_beams, vocab_size)
+                    _scores, dtype=tf.int32, num_samples=2 * num_beams
+                )  # (batch_size, 2 * num_beams)
                 # Compute next scores
-                scores = tf.nn.log_softmax(next_token_logits, axis=-1)  # (batch_size * num_beams, vocab_size)
-                _scores = tf.gather(scores, next_tokens, batch_dims=1)  # (batch_size * num_beams, 2)
-                next_scores = _scores + tf.broadcast_to(
-                    beam_scores[:, None], (batch_size * num_beams, 2)
-                )  # (batch_size * num_beams, 2)
-                # Match shape of greedy beam search
-                next_tokens = tf.reshape(next_tokens, (batch_size, 2 * num_beams))  # (batch_size, 2 * num_beams)
-                next_scores = tf.reshape(next_scores, (batch_size, 2 * num_beams))  # (batch_size, 2 * num_beams)
+                next_scores = tf.gather(_scores, next_tokens, batch_dims=1)  # (batch_size, 2 * num_beams)
             else:
                 # do greedy beam search
                 scores = tf.nn.log_softmax(next_token_logits, axis=-1)  # (batch_size * num_beams, vocab_size)