!6774 mindir_mode weight_min_max

Merge pull request !6774 from baiyangfan/mindir_mode

mindspore/train/quant/quant.py

@@ -342,6 +342,14 @@ class ExportToQuantInferNetwork:
         network = self._convert_quant2deploy(network)
         return network
 
+    def statistic_weight(self, weight):
+        out_nums = np.shape(weight)[0]
+        sta_metric = np.zeros((out_nums, 2), dtype=np.float32)
+        for num in range(out_nums):
+            sta_metric[num, 0] = np.min(weight[num])
+            sta_metric[num, 1] = np.max(weight[num])
+        return np.mean(sta_metric[:, 1]).tolist(), np.mean(sta_metric[:, 0]).tolist()
+
     def _get_quant_block(self, cell_core, activation, fake_quant_a_out):
         """convet network's quant subcell to deploy subcell"""
         # Calculate the scale and zero point
@@ -357,14 +365,12 @@ class ExportToQuantInferNetwork:
         param_dict["mean"] = self.mean
         param_dict["std_dev"] = self.std_dev
         param_dict["symmetric"] = fake_quant_a_out.symmetric
-        if self.is_mindir:
-            scale_w, zp_w, param_dict["filter_maxq"], param_dict["filter_minq"] = \
-                quant_utils.scale_zp_max_min_from_fake_quant_cell(cell_core.fake_quant_weight, np_type)
-            scale_a_out, _, param_dict["output_maxq"], param_dict["output_minq"] = \
-                quant_utils.scale_zp_max_min_from_fake_quant_cell(fake_quant_a_out, np_type)
-        else:
-            scale_w, zp_w = quant_utils.scale_zp_from_fake_quant_cell(cell_core.fake_quant_weight, np_type)
-            scale_a_out, _ = quant_utils.scale_zp_from_fake_quant_cell(fake_quant_a_out, np_type)
+
+        scale_w, zp_w, _, _ = \
+            quant_utils.scale_zp_max_min_from_fake_quant_cell(cell_core.fake_quant_weight, np_type)
+        scale_a_out, _, param_dict["output_maxq"], param_dict["output_minq"] = \
+            quant_utils.scale_zp_max_min_from_fake_quant_cell(fake_quant_a_out, np_type)
+
         info = self.quant_info_table.get(w_minq_name, None)
         if info:
             fack_quant_a_in_op, minq_name = info
@@ -403,6 +409,8 @@ class ExportToQuantInferNetwork:
             weight, bias = quant_utils.fold_batchnorm(weight, cell_core)
         elif isinstance(cell_core, quant.Conv2dBnWithoutFoldQuant):
             weight, bias = quant_utils.without_fold_batchnorm(weight, cell_core)
+        if self.is_mindir:
+            param_dict["filter_maxq"], param_dict["filter_minq"] = self.statistic_weight(weight)
         weight_b = weight
         bias_b = bias
         # apply the quant
@@ -467,6 +475,9 @@ class ExportToQuantInferNetwork:
             elif isinstance(subcell, _AddFakeQuantAfterSubCell):
                 op = subcell.subcell
                 if op.name in ConvertToQuantNetwork.__quant_op_name__ and isinstance(op, ops.Primitive):
+                    if self.is_mindir:
+                        op.add_prim_attr('output_maxq', Tensor(subcell.fake_quant_act.maxq.data.asnumpy()))
+                        op.add_prim_attr('output_minq', Tensor(subcell.fake_quant_act.minq.data.asnumpy()))
                     network.__delattr__(name)
                     network.__setattr__(name, op)
                     change = True

mindspore/train/quant/quant_utils.py

@@ -120,31 +120,6 @@ def weight2int(data, scale, zero_point):
 
     return np.round((data / scale) + zero_point)
 
-
-def scale_zp_from_fake_quant_cell(cell, data_type):
-    r"""
-    Get calculate quantization params for scale and zero point From `FakeQuantWithMinMax`.
-
-    Args:
-        cell (Cell): `mindspore.nn.layer.FakeQuantWithMinMax`
-        data_type (numpy type): Can ben `numpy.int8` or `numpy.uint8`.
-
-    Returns:
-        scale (numpy.ndarray): quantization param.
-        zero point (numpy.ndarray): quantization param.
-    """
-    minq = cell.minq.data.asnumpy()
-    maxq = cell.maxq.data.asnumpy()
-    op = cell.fake_quant_infer
-
-    scale, zp = cal_quantization_params(
-        minq, maxq, data_type,
-        num_bits=op.num_bits,
-        symmetric=op.symmetric,
-        narrow_range=op.narrow_range)
-    return scale, zp
-
-
 def scale_zp_max_min_from_fake_quant_cell(cell, data_type):
     """Get calculate quantization params for scale, zero point, max and min from `FakeQuantWithMinMax`."""
     minq = cell.minq.data.asnumpy()