| | |
|---|
| | | self.batch_size = kwargs.get("batch_size") |
|---|
| | | self.batch_type = kwargs.get("batch_type") |
|---|
| | | self.prompt_ids_len = 0 |
|---|
| | | self.retry = kwargs.get("retry", 10) |
|---|
| | | self.retry = kwargs.get("retry", 100) |
|---|
| | | |
|---|
| | | self.permute = False |
|---|
| | | from funasr.frontends.whisper_frontend import WhisperFrontend |
|---|
| | |
|---|
| | | if badcase_flag: |
|---|
| | | continue |
|---|
| | | input_ids = torch.tensor(input_ids, dtype=torch.int64) # [: self.max_token_length] |
|---|
| | | attention_mask = torch.tensor([len(input_ids)], dtype=torch.int32) |
|---|
| | | attention_mask = torch.tensor([1] * len(input_ids), dtype=torch.int32) |
|---|
| | | labels = torch.tensor(labels, dtype=torch.int64) # [: self.max_token_length] |
|---|
| | | |
|---|
| | | fbank = speech[0, :, :] |
|---|
| | |
|---|
| | | |
|---|
| | | if self.batch_type != "example": |
|---|
| | | b, t = outputs["input_ids"].shape |
|---|
| | | if b * t > self.batch_size * self.batch_size_scale_ratio_max: |
|---|
| | | beg = torch.randint(0, 2, ()).item() |
|---|
| | | if b < 2: |
|---|
| | | beg = 0 |
|---|
| | | if b > 1 and b * t > self.batch_size * self.batch_size_scale_ratio_max: |
|---|
| | | logging.info( |
|---|
| | | f"Warning, b * t: {b * t} > {self.batch_size}, b: {b}, t: {t}, drop half data {idx}th, beg:{beg}" |
|---|
| | | f"Warning, b*t: {b}*{t}={b * t} > batch_size*relax: {self.batch_size_scale_ratio_max}*{self.batch_size}={self.batch_size_scale_ratio_max*self.batch_size}, drop last data" |
|---|
| | | ) |
|---|
| | | samples = samples[beg : beg + b : 2] |
|---|
| | | samples = samples[:-1] |
|---|
| | | continue |
|---|
| | | |
|---|
| | | break |
|---|
