Dev gzf exp (#1654)

zhifu gao

2024-04-24 861147c7308b91068ffa02724fdf74ee623a909e

 funasr/models/paraformer_streaming/export_meta.py

@@ -17,12 +17,12 @@
        predictor_class = tables.predictor_classes.get(kwargs["predictor"]+"Export")
        model.predictor = predictor_class(model.predictor, onnx=is_onnx)


        decoder_class = tables.decoder_classes.get(kwargs["decoder"]+"Export")
        model.decoder = decoder_class(model.decoder, onnx=is_onnx)
        
        from funasr.utils.torch_function import sequence_mask
        model.make_pad_mask = sequence_mask(kwargs['max_seq_len'], flip=False)

    model.make_pad_mask = sequence_mask(kwargs["max_seq_len"], flip=False)
        
        model.forward = types.MethodType(export_forward, model)
        model.export_dummy_inputs = types.MethodType(export_dummy_inputs, model)
@@ -54,6 +54,7 @@
    
    import copy
    import types

    encoder_model = copy.copy(model)
    decoder_model = copy.copy(model)
    
@@ -99,33 +100,24 @@


def export_encoder_input_names(self):
    return ['speech', 'speech_lengths']
    return ["speech", "speech_lengths"]


def export_encoder_output_names(self):
    return ['enc', 'enc_len', 'alphas']
    return ["enc", "enc_len", "alphas"]


def export_encoder_dynamic_axes(self):
    return {
        'speech': {
            0: 'batch_size',
            1: 'feats_length'
        "speech": {0: "batch_size", 1: "feats_length"},
        "speech_lengths": {
            0: "batch_size",
        },
        'speech_lengths': {
            0: 'batch_size',
        "enc": {0: "batch_size", 1: "feats_length"},
        "enc_len": {
            0: "batch_size",
        },
        'enc': {
            0: 'batch_size',
            1: 'feats_length'
        },
        'enc_len': {
            0: 'batch_size',
        },
        'alphas': {
            0: 'batch_size',
            1: 'feats_length'
        },
        "alphas": {0: "batch_size", 1: "feats_length"},
    }


@@ -141,7 +133,9 @@
    acoustic_embeds_len: torch.Tensor,
    *args,
):
    decoder_out, out_caches = self.decoder(enc, enc_len, acoustic_embeds, acoustic_embeds_len, *args)
    decoder_out, out_caches = self.decoder(
        enc, enc_len, acoustic_embeds, acoustic_embeds_len, *args
    )
    sample_ids = decoder_out.argmax(dim=-1)
    
    return decoder_out, sample_ids, out_caches