Dev gzf exp (#1654)

zhifu gao

2024-04-24 861147c7308b91068ffa02724fdf74ee623a909e

 funasr/models/contextual_paraformer/export_meta.py

@@ -11,9 +11,7 @@


class ContextualEmbedderExport2(ContextualEmbedderExport):
    def __init__(self,
                 model,
                 **kwargs):
    def __init__(self, model, **kwargs):
        super().__init__(model)
        self.embedding = model.bias_embed
        model.bias_encoder.batch_first = False
@@ -39,21 +37,32 @@
    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.feats_dim = 560

    import copy

    backbone_model = copy.copy(model)

    # backbone
    backbone_model.forward = types.MethodType(export_backbone_forward, backbone_model)
    backbone_model.export_dummy_inputs = types.MethodType(export_backbone_dummy_inputs, backbone_model)
    backbone_model.export_input_names = types.MethodType(export_backbone_input_names, backbone_model)
    backbone_model.export_output_names = types.MethodType(export_backbone_output_names, backbone_model)
    backbone_model.export_dynamic_axes = types.MethodType(export_backbone_dynamic_axes, backbone_model)
    backbone_model.export_dummy_inputs = types.MethodType(
        export_backbone_dummy_inputs, backbone_model
    )
    backbone_model.export_input_names = types.MethodType(
        export_backbone_input_names, backbone_model
    )
    backbone_model.export_output_names = types.MethodType(
        export_backbone_output_names, backbone_model
    )
    backbone_model.export_dynamic_axes = types.MethodType(
        export_backbone_dynamic_axes, backbone_model
    )
    backbone_model.export_name = types.MethodType(export_backbone_name, backbone_model)
    
    return backbone_model, embedder_model


def export_backbone_forward(
            self,
@@ -73,36 +82,32 @@

    return decoder_out, pre_token_length


def export_backbone_dummy_inputs(self):
    speech = torch.randn(2, 30, self.feats_dim)
    speech_lengths = torch.tensor([6, 30], dtype=torch.int32)
    bias_embed = torch.randn(2, 1, 512)
    return (speech, speech_lengths, bias_embed)


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


def export_backbone_output_names(self):
    return ['logits', 'token_num']
    return ["logits", "token_num"]


def export_backbone_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',
        },
        'bias_embed': {
            0: 'batch_size',
            1: 'num_hotwords'
        },
        'logits': {
            0: 'batch_size',
            1: 'logits_length'
        },
        "bias_embed": {0: "batch_size", 1: "num_hotwords"},
        "logits": {0: "batch_size", 1: "logits_length"},
    }
    

def export_backbone_name(self):
    return 'model.onnx'
    return "model.onnx"