From 1cdb3cc28d4d89a576cc06e5cd8eb80da1f3a3aa Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 26 四月 2024 11:27:39 +0800
Subject: [PATCH] Dev gzf exp (#1665)
---
funasr/datasets/sense_voice_datasets/datasets.py | 56 +++++++++++++++++++++++++++++++++++++++++++++++++++++++-
1 files changed, 55 insertions(+), 1 deletions(-)
diff --git a/funasr/datasets/sense_voice_datasets/datasets.py b/funasr/datasets/sense_voice_datasets/datasets.py
index 5468ea6..6d9b035 100644
--- a/funasr/datasets/sense_voice_datasets/datasets.py
+++ b/funasr/datasets/sense_voice_datasets/datasets.py
@@ -1,3 +1,5 @@
+import logging
+
import torch
import random
@@ -46,6 +48,9 @@
self.float_pad_value = float_pad_value
self.sos = kwargs.get("sos", "<|startoftranscript|>")
self.eos = kwargs.get("eos", "<|endoftext|>")
+ self.batch_size = kwargs.get("batch_size")
+ self.batch_type = kwargs.get("batch_type")
+ self.prompt_ids_len = 0
def get_source_len(self, index):
item = self.index_ds[index]
@@ -69,6 +74,9 @@
speech, speech_lengths = extract_fbank(
data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
) # speech: [b, T, d]
+
+ if speech_lengths > self.batch_size:
+ return None
speech = speech.permute(0, 2, 1)
target = item["target"]
if self.preprocessor_text:
@@ -80,9 +88,12 @@
prompt = f"{self.sos}{task}{text_language}"
prompt_ids = self.tokenizer.encode(prompt, allowed_special="all")
prompt_ids_len = len(prompt_ids) - 1 # [sos, task]
+ self.prompt_ids_len = prompt_ids_len
target_ids = self.tokenizer.encode(target, allowed_special="all")
target_ids_len = len(target_ids) + 1 # [lid, text]
+ if target_ids_len > 200:
+ return None
eos = self.tokenizer.encode(self.eos, allowed_special="all") # [eos]
@@ -95,23 +106,38 @@
target_mask = (
[0] * (prompt_ids_len) + [1] * (target_ids_len) + [1]
) # [sos, task, lid, text, eos]: [0, 0, 1, 1, 1]
+ target_mask_lengths = len(target_mask)
target_mask = torch.tensor(target_mask, dtype=torch.float32)
-
+ target_mask_lengths = torch.tensor([target_mask_lengths], dtype=torch.int32)
return {
"speech": speech[0, :, :],
"speech_lengths": speech_lengths,
"text": text,
"text_lengths": text_lengths,
"target_mask": target_mask,
+ "target_mask_lengths": target_mask_lengths,
}
def collator(self, samples: list = None):
outputs = {}
for sample in samples:
+ if sample is None:
+ continue
for key in sample.keys():
if key not in outputs:
outputs[key] = []
outputs[key].append(sample[key])
+
+ if len(outputs) < 1:
+ logging.info(f"ERROR: data is empty!")
+ outputs = {
+ "speech": torch.rand((10, 128), dtype=torch.float32),
+ "speech_lengths": torch.tensor([10], dtype=torch.int32),
+ "text": torch.tensor([58836], dtype=torch.int32),
+ "text_lengths": torch.tensor([1], dtype=torch.int32),
+ "target_mask": torch.tensor([[0] * (self.prompt_ids_len) + [1] * (1) + [1]]),
+ }
+ return outputs
for key, data_list in outputs.items():
if isinstance(data_list[0], torch.Tensor):
@@ -124,4 +150,32 @@
outputs[key] = torch.nn.utils.rnn.pad_sequence(
data_list, batch_first=True, padding_value=pad_value
)
+
+ if self.batch_type != "example":
+ for i in range(3):
+ outputs = self._filter_badcase(outputs, i=i)
+
+ return outputs
+
+ def _filter_badcase(self, outputs, i=0):
+ b, t, _ = outputs["speech"].shape
+
+ if b * t > self.batch_size * 1.25:
+ beg = torch.randint(0, 2, ()).item()
+ if b < 2:
+ beg = 0
+ logging.info(
+ f"Warning, b * t: {b * t} > {self.batch_size}, drop half data {i}th, beg:{beg}"
+ )
+ for key, data_list in outputs.items():
+ outputs[key] = outputs[key][beg : beg + b : 2]
+
+
+ speech_lengths_max = outputs["speech_lengths"].max().item()
+ outputs["speech"] = outputs["speech"][:, :speech_lengths_max, :]
+ text_lengths_max = outputs["text_lengths"].max().item()
+ outputs["text"] = outputs["text"][:, :text_lengths_max]
+ target_mask_lengths_max = outputs["target_mask_lengths"].max().item()
+ outputs["target_mask"] = outputs["target_mask"][:, :target_mask_lengths_max]
+
return outputs
--
Gitblit v1.9.1