[Quantization] model quantization for inference

wanchen.swc

2023-03-06 69ccdd35cda4c8482e189fa350fbcb83997872f2

[Quantization] model quantization for inference

 funasr/export/export_model.py

@@ -15,7 +15,9 @@
# assert torch_version > 1.9

class ASRModelExportParaformer:
    def __init__(self, cache_dir: Union[Path, str] = None, onnx: bool = True):
    def __init__(
        self, cache_dir: Union[Path, str] = None, onnx: bool = True, quant: bool = True
    ):
        assert check_argument_types()
        self.set_all_random_seed(0)
        if cache_dir is None:
@@ -28,6 +30,7 @@
        )
        print("output dir: {}".format(self.cache_dir))
        self.onnx = onnx
        self.quant = quant
        

    def _export(
@@ -56,6 +59,28 @@
        print("output dir: {}".format(export_dir))


    def _torch_quantize(self, model):
        from torch_quant.module import ModuleFilter
        from torch_quant.observer import HistogramObserver
        from torch_quant.quantizer import Backend, Quantizer
        from funasr.export.models.modules.decoder_layer import DecoderLayerSANM
        from funasr.export.models.modules.encoder_layer import EncoderLayerSANM
        module_filter = ModuleFilter(include_classes=[EncoderLayerSANM, DecoderLayerSANM])
        module_filter.exclude_op_types = [torch.nn.Conv1d]
        quantizer = Quantizer(
            module_filter=module_filter,
            backend=Backend.FBGEMM,
            act_ob_ctr=HistogramObserver,
        )
        model.eval()
        calib_model = quantizer.calib(model)
        # run calibration data
        # using dummy inputs for a example
        dummy_input = model.get_dummy_inputs()
        _ = calib_model(*dummy_input)
        quant_model = quantizer.quantize(model)
        return quant_model

    def _export_torchscripts(self, model, verbose, path, enc_size=None):
        if enc_size:
            dummy_input = model.get_dummy_inputs(enc_size)
@@ -65,6 +90,12 @@
        # model_script = torch.jit.script(model)
        model_script = torch.jit.trace(model, dummy_input)
        model_script.save(os.path.join(path, f'{model.model_name}.torchscripts'))

        if self.quant:
            quant_model = self._torch_quantize(model)
            model_script = torch.jit.trace(quant_model, dummy_input)
            model_script.save(os.path.join(path, f'{model.model_name}_quant.torchscripts'))


    def set_all_random_seed(self, seed: int):
        random.seed(seed)
@@ -107,17 +138,27 @@

        # model_script = torch.jit.script(model)
        model_script = model #torch.jit.trace(model)
        model_path = os.path.join(path, f'{model.model_name}.onnx')

        torch.onnx.export(
            model_script,
            dummy_input,
            os.path.join(path, f'{model.model_name}.onnx'),
            model_path,
            verbose=verbose,
            opset_version=14,
            input_names=model.get_input_names(),
            output_names=model.get_output_names(),
            dynamic_axes=model.get_dynamic_axes()
        )

        if self.quant:
            from onnxruntime.quantization import QuantType, quantize_dynamic
            quant_model_path = os.path.join(path, f'{model.model_name}_quant.onnx')
            quantize_dynamic(
                model_input=model_path,
                model_output=quant_model_path,
                weight_type=QuantType.QUInt8,
            )


if __name__ == '__main__':
@@ -126,10 +167,12 @@
    model_path = sys.argv[1]
    output_dir = sys.argv[2]
    onnx = sys.argv[3]
    quant = sys.argv[4]
    onnx = onnx.lower()
    onnx = onnx == 'true'
    quant = quant == 'true'
    # model_path = 'damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'
    # output_dir = "../export"
    export_model = ASRModelExportParaformer(cache_dir=output_dir, onnx=onnx)
    export_model = ASRModelExportParaformer(cache_dir=output_dir, onnx=onnx, quant=quant)
    export_model.export(model_path)
    # export_model.export('/root/cache/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')
    # export_model.export('/root/cache/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')

 funasr/export/models/modules/encoder_layer.py

@@ -16,6 +16,7 @@
        self.feed_forward = model.feed_forward
        self.norm1 = model.norm1
        self.norm2 = model.norm2
        self.in_size = model.in_size
        self.size = model.size

    def forward(self, x, mask):
@@ -23,13 +24,12 @@
        residual = x
        x = self.norm1(x)
        x = self.self_attn(x, mask)
        if x.size(2) == residual.size(2):
        if self.in_size == self.size:
            x = x + residual
        residual = x
        x = self.norm2(x)
        x = self.feed_forward(x)
        if x.size(2) == residual.size(2):
            x = x + residual
        x = x + residual

        return x, mask


 funasr/export/models/modules/multihead_att.py

@@ -64,6 +64,21 @@
        return self.linear_out(context_layer)  # (batch, time1, d_model)


def preprocess_for_attn(x, mask, cache, pad_fn):
    x = x * mask
    x = x.transpose(1, 2)
    if cache is None:
        x = pad_fn(x)
    else:
        x = torch.cat((cache[:, :, 1:], x), dim=2)
        cache = x
    return x, cache


import torch.fx
torch.fx.wrap('preprocess_for_attn')


class MultiHeadedAttentionSANMDecoder(nn.Module):
    def __init__(self, model):
        super().__init__()
@@ -73,16 +88,7 @@
        self.attn = None

    def forward(self, inputs, mask, cache=None):
        # b, t, d = inputs.size()
        # mask = torch.reshape(mask, (b, -1, 1))
        inputs = inputs * mask

        x = inputs.transpose(1, 2)
        if cache is None:
            x = self.pad_fn(x)
        else:
            x = torch.cat((cache[:, :, 1:], x), dim=2)
            cache = x
        x, cache = preprocess_for_attn(inputs, mask, cache, self.pad_fn)
        x = self.fsmn_block(x)
        x = x.transpose(1, 2)

@@ -232,4 +238,4 @@
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(new_context_layer_shape)
        return self.linear_out(context_layer)  # (batch, time1, d_model)