fix loss normalization for ddp training

haoneng.lhn

2023-12-08 acb9a0fec8d8a4dabeedcbb8e08c26f66d7083f0

 funasr/bin/asr_inference_launch.py

@@ -48,13 +48,13 @@
from funasr.utils.types import str2triple_str
from funasr.utils.types import str_or_none
from funasr.utils.vad_utils import slice_padding_fbank
from funasr.utils.speaker_utils import (check_audio_list, 
                                        sv_preprocess, 
                                        sv_chunk, 
                                        CAMPPlus, 
                                        extract_feature, 
from funasr.utils.speaker_utils import (check_audio_list,
                                        sv_preprocess,
                                        sv_chunk,
                                        extract_feature,
                                        postprocess,
                                        distribute_spk)
import funasr.modules.cnn as sv_module
from funasr.build_utils.build_model_from_file import build_model_from_file
from funasr.utils.cluster_backend import ClusterBackend
from funasr.utils.modelscope_utils import get_cache_dir
@@ -818,7 +818,15 @@
        format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
    )

    sv_model_file = asr_model_file.replace("model.pb", "campplus_cn_common.bin")
    sv_model_config_path = asr_model_file.replace("model.pb", "sv_model_config.yaml")
    if not os.path.exists(sv_model_config_path):
        sv_model_config = {'sv_model_class': 'CAMPPlus','sv_model_file': 'campplus_cn_common.bin', 'models_config': {}}
    else:
        with open(sv_model_config_path, 'r') as f:
            sv_model_config = yaml.load(f, Loader=yaml.FullLoader)
    if sv_model_config['models_config'] is None:
        sv_model_config['models_config'] = {}
    sv_model_file = asr_model_file.replace("model.pb", sv_model_config['sv_model_file'])

    if param_dict is not None:
        hotword_list_or_file = param_dict.get('hotword')
@@ -944,9 +952,15 @@
            #####  speaker_verification  #####
            ##################################
            # load sv model
            sv_model_dict = torch.load(sv_model_file, map_location=torch.device('cpu'))
            sv_model = CAMPPlus()
            if ngpu > 0:
                sv_model_dict = torch.load(sv_model_file)
                sv_model = getattr(sv_module, sv_model_config['sv_model_class'])(**sv_model_config['models_config'])
                sv_model.cuda()
            else:
                sv_model_dict = torch.load(sv_model_file, map_location=torch.device('cpu'))
                sv_model = getattr(sv_module, sv_model_config['sv_model_class'])(**sv_model_config['models_config'])
            sv_model.load_state_dict(sv_model_dict)
            print(f'load sv model params: {sv_model_file}')
            sv_model.eval()
            cb_model = ClusterBackend()
            vad_segments = []
@@ -956,24 +970,31 @@
                ed = int(vadsegment[1]) / 1000
                vad_segments.append(
                    [st, ed, audio[int(st * 16000):int(ed * 16000)]])
            check_audio_list(vad_segments)
            # sv pipeline
            segments = sv_chunk(vad_segments)
            embeddings = []
            for s in segments:
                #_, embs = self.sv_pipeline([s[2]], output_emb=True)
                # embeddings.append(embs)
                wavs = sv_preprocess([s[2]])
                # embs = self.forward(wavs)
                embs = []
                for x in wavs:
                    x = extract_feature([x])
                    embs.append(sv_model(x))
                embs = torch.cat(embs)
                embeddings.append(embs.detach().numpy())
            embeddings = np.concatenate(embeddings)
            labels = cb_model(embeddings)
            sv_output = postprocess(segments, vad_segments, labels, embeddings)
            audio_dur = check_audio_list(vad_segments)
            if audio_dur > 5:
                # sv pipeline
                segments = sv_chunk(vad_segments)
                embeddings = []
                for s in segments:
                    #_, embs = self.sv_pipeline([s[2]], output_emb=True)
                    # embeddings.append(embs)
                    wavs = sv_preprocess([s[2]])
                    # embs = self.forward(wavs)
                    embs = []
                    for x in wavs:
                        x = extract_feature([x])
                        if ngpu > 0:
                            x = x.cuda()
                        embs.append(sv_model(x))
                    embs = torch.cat(embs)
                    embeddings.append(embs.cpu().detach().numpy())
                embeddings = np.concatenate(embeddings)
                labels = cb_model(embeddings)
                sv_output = postprocess(segments, vad_segments, labels, embeddings)
            else:
                # fake speaker res for too shot utterance
                sv_output = [[0.0, vadsegments[-1][-1]/1000.0, 0]]
                logging.warning("Too short utterence found: {}, return default speaker results.".format(keys))

            speech, speech_lengths = batch["speech"], batch["speech_lengths"]