Add a hack to fix saturation issue on AVX2 AVX512 (open-mmlab#3)

Add the function that tries to find the first convolution in backbone and
changes its export function to make export in int7 (instead of int8)
to avoid the saturation issue.

At the moment the trick works on Mobilenet-s only, for other backbones
it changes nothing

mmdet/integration/nncf/compression.py

@@ -1,5 +1,6 @@
 import pathlib
 import torch
+from functools import partial
 
 from mmdet.utils import get_root_logger
 from .utils import (check_nncf_is_enabled, get_nncf_version, is_nncf_enabled,
@@ -153,9 +154,54 @@ def dummy_forward(model):
                                                       nncf_config,
                                                       dummy_forward_fn=dummy_forward,
                                                       resuming_state_dict=resuming_state_dict)
+    model = change_export_func_first_conv(model)
+
     return compression_ctrl, model
 
 
+def change_export_func_first_conv(model):
+    ''' To avoid saturation issue
+    At the moment works only for mobilenet
+    '''
+
+    def run_hacked_export_quantization(self, x):
+        from nncf.quantization.layers import (
+            ExportQuantizeToFakeQuantize, ExportQuantizeToONNXQuantDequant,
+            QuantizerExportMode, get_scale_zp_from_input_low_input_high)
+        from nncf.utils import no_jit_trace
+        with no_jit_trace():
+            input_range = abs(self.scale) + self.eps
+            # todo: take bias into account during input_low/input_high calculation
+            input_low = input_range * self.level_low / self.level_high
+            input_high = input_range
+
+            if self._export_mode == QuantizerExportMode.ONNX_QUANTIZE_DEQUANTIZE_PAIRS:
+                y_scale, y_zero_point = get_scale_zp_from_input_low_input_high(self.level_low,
+                                                                               self.level_high,
+                                                                               input_low,
+                                                                               input_high)
+
+        if self._export_mode == QuantizerExportMode.ONNX_QUANTIZE_DEQUANTIZE_PAIRS:
+            return ExportQuantizeToONNXQuantDequant.apply(x, y_scale, y_zero_point)
+        if self._export_mode == QuantizerExportMode.FAKE_QUANTIZE:
+            x = x / 2.0
+            return ExportQuantizeToFakeQuantize.apply(x, self.levels, input_low, input_high, input_low * 2,
+                                                      input_high * 2)
+        raise RuntimeError
+
+    logger = get_root_logger()
+    orig_model = model.get_nncf_wrapped_model()
+    try:
+        # pylint: disable=protected-access
+        module_ = orig_model.backbone.features.init_block.conv.pre_ops._modules['0']
+    except AttributeError as e:
+        logger.info(f'Cannot change an export function for the first Conv due  {e}')
+        return model
+    module_.op.run_export_quantization = partial(run_hacked_export_quantization, module_.op)
+    logger.info('Change an export function for the first Conv to avoid saturation issue on AVX2, AVX512')
+    return model
+
+
 def get_uncompressed_model(module):
     if not is_nncf_enabled():
         return module