| | |
|---|
| | | from distutils.version import LooseVersion |
|---|
| | | |
|---|
| | | from funasr.register import tables |
|---|
| | | from funasr.losses.label_smoothing_loss import ( |
|---|
| | | LabelSmoothingLoss, # noqa: H301 |
|---|
| | | ) |
|---|
| | | from funasr.utils import postprocess_utils |
|---|
| | | from funasr.metrics.compute_acc import th_accuracy |
|---|
| | | from funasr.models.paraformer.model import Paraformer |
|---|
| | |
|---|
| | | from funasr.models.transformer.utils.add_sos_eos import add_sos_eos |
|---|
| | | from funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_list |
|---|
| | | from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank |
|---|
| | | import pdb |
|---|
| | | |
|---|
| | | |
|---|
| | | if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"): |
|---|
| | | from torch.cuda.amp import autocast |
|---|
| | |
|---|
| | | if self.crit_attn_weight > 0: |
|---|
| | | self.attn_loss = torch.nn.L1Loss() |
|---|
| | | self.crit_attn_smooth = crit_attn_smooth |
|---|
| | | |
|---|
| | | |
|---|
| | | def forward( |
|---|
| | | self, |
|---|
| | |
|---|
| | | loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device) |
|---|
| | | return loss, stats, weight |
|---|
| | | |
|---|
| | | |
|---|
| | | def _calc_att_clas_loss( |
|---|
| | | self, |
|---|
| | | encoder_out: torch.Tensor, |
|---|
| | |
|---|
| | | |
|---|
| | | return loss_att, acc_att, cer_att, wer_att, loss_pre, loss_ideal |
|---|
| | | |
|---|
| | | |
|---|
| | | def sampler(self, encoder_out, encoder_out_lens, ys_pad, ys_pad_lens, pre_acoustic_embeds, contextual_info): |
|---|
| | | tgt_mask = (~make_pad_mask(ys_pad_lens, maxlen=ys_pad_lens.max())[:, :, None]).to(ys_pad.device) |
|---|
| | | ys_pad = ys_pad * tgt_mask[:, :, 0] |
|---|
| | |
|---|
| | | sematic_embeds = pre_acoustic_embeds.masked_fill(~input_mask_expand_dim, 0) + ys_pad_embed.masked_fill( |
|---|
| | | input_mask_expand_dim, 0) |
|---|
| | | return sematic_embeds * tgt_mask, decoder_out * tgt_mask |
|---|
| | | |
|---|
| | | |
|---|
| | | def cal_decoder_with_predictor(self, encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, hw_list=None, |
|---|
| | | clas_scale=1.0): |
|---|
| | |
|---|
| | | |
|---|
| | | return results, meta_data |
|---|
| | | |
|---|
| | | |
|---|
| | | def generate_hotwords_list(self, hotword_list_or_file, tokenizer=None, frontend=None): |
|---|
| | | def load_seg_dict(seg_dict_file): |
|---|
| | | seg_dict = {} |
|---|
| | |
|---|
| | | hotword_list = None |
|---|
| | | return hotword_list |
|---|
| | | |
|---|
| | | def export( |
|---|
| | | self, |
|---|
| | | **kwargs, |
|---|
| | | ): |
|---|
| | | if 'max_seq_len' not in kwargs: |
|---|
| | | kwargs['max_seq_len'] = 512 |
|---|
| | | from .export_meta import export_rebuild_model |
|---|
| | | models = export_rebuild_model(model=self, **kwargs) |
|---|
| | | return models |
|---|
