[Mxnet-1397] Support symbolic api for requantize and dequantize #14749

shoubhik · 2019-04-19T23:06:56Z

Description

If I have a pre-qauntized model from another framework, e.g Tensorflow or Pyrtorch and want to use Mxnet as inference engine, I would like to be able to set the int8 weights, scales and shifts manullay instead of Mxnet converting a model for me. I would like to do so in a nn.HybridBlock. For example I can create a quantized convolution network as below

class QuantizedConv(nn.HybridBlock):
    def __init__(self, 
                 kernel_size, 
                 strides, 
                 num_filter, 
                 kernel_shape,
                 activation=None,
                 dilation=(1, 1),
                 padding=(0, 0),
                 groups=1,
                 use_bias=True,
                 layout='NCHW',
                 adj=None,
                 in_channels=0,
                 op_name='Convolution',
                 weight_initializer=None,
                 weight_min_initializer=None,
                 weight_max_initializer=None,
                 bias_initializer='zeros',
                 weight_shape=None,
                 **kwargs):
        super(QuantizedConv, self).__init__(**kwargs)
        self._kwargs = {
                'kernel': kernel_size, 'stride': strides, 'dilate': dilation,
                'pad': padding, 'num_filter': num_filter, 'num_group': groups,
                'no_bias': not use_bias, 'layout': layout}
        if adj is not None:
            self._kwargs['adj'] = adj
        dshape = [0]*(len(kernel_size) + 2)
        dshape[layout.find('N')] = 1
        dshape[layout.find('C')] = in_channels
        if weight_shape is None:
            self.wshapes = _infer_weight_shape(op_name, dshape, self._kwargs)
        else:
            self.wshapes = weight_shape
        
        self.weight = self.params.get('weight', shape=kernel_shape,
                                          init=weight_initializer,
                                          allow_deferred_init=True)
        self.weight_min = self.params.get('weight_min', shape=(1,),
                                         init=weight_min_initializer,
                                         allow_deferred_init=True)
        self.weight_max = self.params.get('weight_max', shape=(1,),
                                         init=weight_max_initializer,
                                         allow_deferred_init=True)
        
        if use_bias:
            self.bias = self.params.get('bias', shape=self.wshapes[2],
                                            init=bias_initializer,
                                            allow_deferred_init=True)
            self.bias_min = self.params.get('bias_min', shape=(1,),
                                            init=bias_initializer,
                                            allow_deferred_init=True)
            self.bias_max = self.params.get('bias_max', shape=(1,),
                                            init=bias_initializer,
                                            allow_deferred_init=True)
        else:
            self.bias = None
            self.bias_min = None
            self.bias_max = None

        self.kernel_size = kernel_size
        self.strides = strides
        self.num_filter = num_filter
        # for string reprentation
        self.layout = layout
        
        if activation is not None:
            self.act = Activation(activation, prefix=activation+'_')
        else:
            self.act = None

        
    def hybrid_forward(self, F, x, weight, weight_min, weight_max, bias=None, 
                       bias_min=None, bias_max=None):
        if not isinstance(weight, mxnet.ndarray.ndarray.NDArray): 
            print('F={}'.format(F))
            weight_temp = weight.astype(dtype='int8')
            print(weight)
            print(weight_temp)
            print(F.contrib.requantize(weight_temp, weight.min(), weight.max()))
        q_inputs, q_inputs_min, q_inputs_max = F.contrib.quantize(
            data=x, 
            min_range=x.min(), 
            max_range=x.max(),
            out_type='uint8')
        q_output, q_output_min, q_output_max = F.contrib.quantized_conv(
                       data=q_inputs, 
                       weight=weight.astype(dtype ='int8'), 
                       bias=bias.astype(dtype='int8'), 
                       min_data=q_inputs_min, 
                       max_data=q_inputs_max, 
                       min_weight=weight_min,
                       max_weight=weight_max,
                       min_bias=bias_min,
                       max_bias=bias_max, 
                       kernel=self.kernel_size,
                       stride=self.strides,
                       num_filter=self.num_filter, 
                       layout=self.layout,
                       cudnn_off=True, 
                       cudnn_tune='off',
                       name='fwd')
        q_8_out, q_8_out_min, q_8_out_max = F.contrib.requantize(
            data=q_output.astype('int32'), 
            min_range=q_output_min, 
            max_range=q_output_max,
            name='fwd', 
        )
        act = F.contrib.dequantize(
            data=q_8_out, 
            min_range=q_8_out_min, 
            max_range=q_8_out_max,
            name='fwd')
        if self.act is not None:
            act = self.act(act)
        return act
    
    def __repr__(self):
        s = '{name}({mapping}, kernel_size={kernel}, stride={stride}'
        len_kernel_size = len(self._kwargs['kernel'])
        if self._kwargs['pad'] != (0,) * len_kernel_size:
            s += ', padding={pad}'
        if self._kwargs['dilate'] != (1,) * len_kernel_size:
            s += ', dilation={dilate}'
        if hasattr(self, 'out_pad') and self.out_pad != (0,) * len_kernel_size:
            s += ', output_padding={out_pad}'.format(out_pad=self.out_pad)
        if self._kwargs['num_group'] != 1:
            s += ', groups={num_group}'
        if self.weight is None:
            s += ', bias=False'
        if self.act:
            s += ', {}'.format(self.act)
        s += ')'
        shape = self.weight.shape
        return s.format(name=self.__class__.__name__,
                        mapping='{0} -> {1}'.format(shape[1] if shape[1] else None, shape[0]),
                        **self._kwargs)

I can the later set the weight and ranges from my saved model. Currently F.contrib.requantize(....) and F.contrib.dequantize(....) fail when F is a symbol, i.e, after calling hybradize() on the network. For more detailed error example please look at the jira. In this CR I am enabaling requantize and dequantize to be called from symbolic API.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
Changes are complete (i.e. I finished coding on this PR)
All changes have test coverage:
Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
Code is well-documented:
For user-facing API changes, API doc string has been updated.
For new C++ functions in header files, their functionalities and arguments are documented.
For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

Enable mx.sym.contrib.dequantize to be called directly, tests, (and when applicable, API doc)
Enable mx.sym.contrib.requantize to be called directly, tests, (and when applicable, API doc)