From 1b9ac4f7a2eb600e5b769e7294cb4a8f4ec31b63 Mon Sep 17 00:00:00 2001 From: TeaPoly Date: Fri, 2 Dec 2022 12:00:22 +0800 Subject: [PATCH] Fix some issue to make batch inference easy for predictor and decoder. --- funasr/bin/asr_inference_paraformer.py | 2 +- funasr/models/predictor/cif.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/funasr/bin/asr_inference_paraformer.py b/funasr/bin/asr_inference_paraformer.py index ed75010d1..179a62bb3 100755 --- a/funasr/bin/asr_inference_paraformer.py +++ b/funasr/bin/asr_inference_paraformer.py @@ -189,7 +189,7 @@ class Speech2Text: predictor_outs = self.asr_model.calc_predictor(enc, enc_len) pre_acoustic_embeds, pre_token_length = predictor_outs[0], predictor_outs[1] - pre_token_length = torch.tensor([pre_acoustic_embeds.size(1)], device=pre_acoustic_embeds.device) + pre_token_length = pre_token_length.long() decoder_outs = self.asr_model.cal_decoder_with_predictor(enc, enc_len, pre_acoustic_embeds, pre_token_length) decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1] diff --git a/funasr/models/predictor/cif.py b/funasr/models/predictor/cif.py index 819970870..cf60eafd1 100644 --- a/funasr/models/predictor/cif.py +++ b/funasr/models/predictor/cif.py @@ -147,7 +147,7 @@ class CifPredictorV2(nn.Module): b, t, d = hidden.size() tail_threshold = self.tail_threshold tail_threshold = torch.tensor([tail_threshold], dtype=alphas.dtype).to(alphas.device) - tail_threshold = torch.reshape(tail_threshold, (1, 1)) + tail_threshold = tail_threshold.unsqueeze(0).repeat(b, 1) alphas = torch.cat([alphas, tail_threshold], dim=1) zeros = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device) hidden = torch.cat([hidden, zeros], dim=1)