default to codebook fps 75 more fixes

Signed-off-by: Paarth Neekhara <[email protected]>
paarthneekhara · Dec 14, 2023 · 05c16d3 · 05c16d3
1 parent 4e366f2
commit 05c16d3
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/nemo/collections/nlp/data/language_modeling/megatron/t5_speechlm_dataset.py b/nemo/collections/nlp/data/language_modeling/megatron/t5_speechlm_dataset.py
@@ -290,7 +290,7 @@ def load_data(self, dataset):
 
             if doc["answer_type"] in ["SPEECH", "AUDIOCODEC"]:
                 assert "answer_duration" in doc, f"answer_duration key not in document {doc}"
-                approx_answer_len = doc["answer_duration"] * self.codebook_fps
+                approx_answer_len = doc["answer_duration"] * (self.codebook_fps + 1)
                 if self.seq_pattern == "delay_parallel":
                     # In delay parallel, there is padding so add 8 frames
                     approx_answer_len = approx_answer_len + 8

diff --git a/nemo/collections/nlp/models/language_modeling/megatron_t5_speechlm_model.py b/nemo/collections/nlp/models/language_modeling/megatron_t5_speechlm_model.py
@@ -1107,7 +1107,7 @@ def build_virtual_prompt_dataset(
             cross_attention_epsilon=self.cfg.data.get('cross_attention_epsilon', 0.0),
             lm_vocab_size=self.lm_vocab_size,
             num_speech_codebooks=self.num_speech_codebooks,
-            codebook_fps=self.cfg.data.get('codebook_fps', 76),
+            codebook_fps=self.cfg.data.get('codebook_fps', 75),
         )
 
         rank = parallel_state.get_data_parallel_rank()