!4199 [MS][QUANT] mindspore model zoo example for hand make quant graph

Merge pull request !4199 from chenzhongming/quant

mindspore/nn/layer/quant.py

@@ -396,7 +396,7 @@ class FakeQuantWithMinMax(Cell):
 
 class Conv2dBnFoldQuant(Cell):
     r"""
-    2D convolution with BatchNormal op folded layer.
+    2D convolution with BatchNormal op folded construct.
 
     This part is a more detailed overview of Conv2d op.
 
@@ -434,10 +434,9 @@ class Conv2dBnFoldQuant(Cell):
         Tensor of shape :math:`(N, C_{out}, H_{out}, W_{out})`.
 
     Examples:
-        >>> batchnorm_quant = nn.Conv2dBnFoldQuant(1, 6, kernel_size= (2, 2), stride=(1, 1), pad_mode="valid",
-        >>>                                           dilation=(1, 1))
-        >>> input_x = Tensor(np.random.randint(-2, 2, (2, 1, 1, 3)), mindspore.float32)
-        >>> result = batchnorm_quant(input_x)
+        >>> conv2d_bn = nn.Conv2dBnFoldQuant(1, 6, kernel_size=(2, 2), stride=(1, 1), pad_mode="valid")
+        >>> x = Tensor(np.random.randint(-2, 2, (2, 1, 1, 3)), mindspore.float32)
+        >>> y = conv2d_bn(x)
     """
 
     def __init__(self,
@@ -508,7 +507,7 @@ class Conv2dBnFoldQuant(Cell):
             channel_axis = 0
         self.weight = Parameter(initializer(weight_init, weight_shape), name='weight')
 
-        # initialize batchnorm Parameter
+        # initialize BatchNorm Parameter
         self.gamma = Parameter(initializer(gamma_init, [out_channels]), name='gamma')
         self.beta = Parameter(initializer(beta_init, [out_channels]), name='beta')
         self.moving_mean = Parameter(initializer(mean_init, [out_channels]), name='moving_mean', requires_grad=False)
@@ -583,7 +582,7 @@ class Conv2dBnFoldQuant(Cell):
 
 class Conv2dBnWithoutFoldQuant(Cell):
     r"""
-    2D convolution + batchnorm without fold with fake quant op layer.
+    2D convolution + batchnorm without fold with fake quant construct.
 
     This part is a more detailed overview of Conv2d op.
 
@@ -617,10 +616,9 @@ class Conv2dBnWithoutFoldQuant(Cell):
         Tensor of shape :math:`(N, C_{out}, H_{out}, W_{out})`.
 
     Examples:
-        >>> conv2d_quant = nn.Conv2dQuant(1, 6, kernel_size=(2, 2), stride=(1, 1), pad_mode="valid",
-        >>>                               dilation=(1, 1))
-        >>> input_x = Tensor(np.random.randint(-2, 2, (2, 1, 1, 3)), mstype.float32)
-        >>> result = conv2d_quant(input_x)
+        >>> conv2d_quant = nn.Conv2dBnWithoutFoldQuant(1, 6, kernel_size=(2, 2), stride=(1, 1), pad_mode="valid")
+        >>> x = Tensor(np.random.randint(-2, 2, (2, 1, 1, 3)), mstype.float32)
+        >>> y = conv2d_quant(x)
     """
 
     def __init__(self,
@@ -687,7 +685,7 @@ class Conv2dBnWithoutFoldQuant(Cell):
                                                      quant_delay=quant_delay)
         self.has_bn = validator.check_bool("has_bn", has_bn)
         if has_bn:
-            self.batchnorm = BatchNorm2d(out_channels)
+            self.batchnorm = BatchNorm2d(out_channels, eps=eps, momentum=momentum)
 
     def construct(self, x):
         weight = self.fake_quant_weight(self.weight)
@@ -740,10 +738,9 @@ class Conv2dQuant(Cell):
         Tensor of shape :math:`(N, C_{out}, H_{out}, W_{out})`.
 
     Examples:
-        >>> conv2d_quant = nn.Conv2dQuant(1, 6, kernel_size= (2, 2), stride=(1, 1), pad_mode="valid",
-        >>>                               dilation=(1, 1))
-        >>> input_x = Tensor(np.random.randint(-2, 2, (2, 1, 1, 3)), mindspore.float32)
-        >>> result = conv2d_quant(input_x)
+        >>> conv2d_quant = nn.Conv2dQuant(1, 6, kernel_size= (2, 2), stride=(1, 1), pad_mode="valid")
+        >>> x = Tensor(np.random.randint(-2, 2, (2, 1, 1, 3)), mindspore.float32)
+        >>> y = conv2d_quant(x)
     """
 
     def __init__(self,

mindspore/train/quant/quant.py

@@ -473,7 +473,7 @@ def export(network, *inputs, file_name, mean=127.5, std_dev=127.5, file_format='
 
 def convert_quant_network(network,
                           bn_fold=False,
-                          freeze_bn=0,
+                          freeze_bn=10000,
                           quant_delay=(0, 0),
                           num_bits=(8, 8),
                           per_channel=(False, False),

model_zoo/official/cv/mobilenetv2_quant/src/mobilenetV2_quant.py

+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""MobileNetV2 Quant model define"""
+
+import mindspore.nn as nn
+from mindspore.ops import operations as P
+
+__all__ = ['mobilenetV2_quant']
+
+_quant_delay = 200
+_ema_decay = 0.999
+_symmetric = False
+_per_channel = False
+
+
+def _make_divisible(v, divisor, min_value=None):
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    # Make sure that round down does not go down by more than 10%.
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+class GlobalAvgPooling(nn.Cell):
+    """
+    Global avg pooling definition.
+
+    Args:
+
+    Returns:
+        Tensor, output tensor.
+
+    Examples:
+        >>> GlobalAvgPooling()
+    """
+
+    def __init__(self):
+        super(GlobalAvgPooling, self).__init__()
+        self.mean = P.ReduceMean(keep_dims=False)
+
+    def construct(self, x):
+        x = self.mean(x, (2, 3))
+        return x
+
+
+class ConvBNReLU(nn.Cell):
+    """
+    Convolution/Depthwise fused with Batchnorm and ReLU block definition.
+
+    Args:
+        in_planes (int): Input channel.
+        out_planes (int): Output channel.
+        kernel_size (int): Input kernel size.
+        stride (int): Stride size for the first convolutional layer. Default: 1.
+        groups (int): channel group. Convolution is 1 while Depthiwse is input channel. Default: 1.
+
+    Returns:
+        Tensor, output tensor.
+
+    Examples:
+        >>> ConvBNReLU(16, 256, kernel_size=1, stride=1, groups=1)
+    """
+
+    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
+        super(ConvBNReLU, self).__init__()
+        padding = (kernel_size - 1) // 2
+        conv = nn.Conv2dBnFoldQuant(in_planes, out_planes, kernel_size, stride,
+                                    pad_mode='pad', padding=padding, quant_delay=_quant_delay, group=groups,
+                                    per_channel=_per_channel, symmetric=_symmetric)
+        layers = [conv, nn.ReLU()]
+        self.features = nn.SequentialCell(layers)
+        self.fake = nn.FakeQuantWithMinMax(ema=True, ema_decay=_ema_decay, min_init=0, quant_delay=_quant_delay)
+
+    def construct(self, x):
+        output = self.features(x)
+        output = self.fake(output)
+        return output
+
+
+class InvertedResidual(nn.Cell):
+    """
+    Mobilenetv2 residual block definition.
+
+    Args:
+        inp (int): Input channel.
+        oup (int): Output channel.
+        stride (int): Stride size for the first convolutional layer. Default: 1.
+        expand_ratio (int): expand ration of input channel
+
+    Returns:
+        Tensor, output tensor.
+
+    Examples:
+        >>> ResidualBlock(3, 256, 1, 1)
+    """
+
+    def __init__(self, inp, oup, stride, expand_ratio):
+        super(InvertedResidual, self).__init__()
+        assert stride in [1, 2]
+
+        hidden_dim = int(round(inp * expand_ratio))
+        self.use_res_connect = stride == 1 and inp == oup
+
+        layers = []
+        if expand_ratio != 1:
+            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
+        layers.extend([
+            # dw
+            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
+            # pw-linear
+            nn.Conv2dBnFoldQuant(hidden_dim, oup, kernel_size=1, stride=1, pad_mode='pad', padding=0, group=1,
+                                 per_channel=_per_channel, symmetric=_symmetric, quant_delay=_quant_delay),
+            nn.FakeQuantWithMinMax(ema=True, ema_decay=_ema_decay, quant_delay=_quant_delay)
+        ])
+        self.conv = nn.SequentialCell(layers)
+        self.add = P.TensorAdd()
+        self.add_fake = nn.FakeQuantWithMinMax(ema=True, ema_decay=_ema_decay, quant_delay=_quant_delay)
+
+    def construct(self, x):
+        identity = x
+        x = self.conv(x)
+        if self.use_res_connect:
+            x = self.add(identity, x)
+            x = self.add_fake(x)
+        return x
+
+
+class MobileNetV2Quant(nn.Cell):
+    """
+    MobileNetV2Quant architecture.
+
+    Args:
+        class_num (Cell): number of classes.
+        width_mult (int): Channels multiplier for round to 8/16 and others. Default is 1.
+        has_dropout (bool): Is dropout used. Default is false
+        inverted_residual_setting (list): Inverted residual settings. Default is None
+        round_nearest (list): Channel round to . Default is 8
+    Returns:
+        Tensor, output tensor.
+
+    Examples:
+        >>> MobileNetV2Quant(num_classes=1000)
+    """
+
+    def __init__(self, num_classes=1000, width_mult=1.,
+                 has_dropout=False, inverted_residual_setting=None, round_nearest=8):
+        super(MobileNetV2Quant, self).__init__()
+        block = InvertedResidual
+        input_channel = 32
+        last_channel = 1280
+        # setting of inverted residual blocks
+        self.cfgs = inverted_residual_setting
+        if inverted_residual_setting is None:
+            self.cfgs = [
+                # t, c, n, s
+                [1, 16, 1, 1],
+                [6, 24, 2, 2],
+                [6, 32, 3, 2],
+                [6, 64, 4, 2],
+                [6, 96, 3, 1],
+                [6, 160, 3, 2],
+                [6, 320, 1, 1],
+            ]
+
+        # building first layer
+        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
+        self.out_channels = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
+        self.input_fake = nn.FakeQuantWithMinMax(ema=True, ema_decay=_ema_decay, quant_delay=_quant_delay)
+        features = [ConvBNReLU(3, input_channel, stride=2)]
+        # building inverted residual blocks
+        for t, c, n, s in self.cfgs:
+            output_channel = _make_divisible(c * width_mult, round_nearest)
+            for i in range(n):
+                stride = s if i == 0 else 1
+                features.append(block(input_channel, output_channel, stride, expand_ratio=t))
+                input_channel = output_channel
+        # building last several layers
+        features.append(ConvBNReLU(input_channel, self.out_channels, kernel_size=1))
+        # make it nn.CellList
+        self.features = nn.SequentialCell(features)
+        # mobilenet head
+        head = ([GlobalAvgPooling(),
+                 nn.DenseQuant(self.out_channels, num_classes, has_bias=True, per_channel=_per_channel,
+                               symmetric=_symmetric, quant_delay=_quant_delay),
+                 nn.FakeQuantWithMinMax(ema=True, ema_decay=_ema_decay)] if not has_dropout else
+                [GlobalAvgPooling(),
+                 nn.Dropout(0.2),
+                 nn.DenseQuant(self.out_channels, num_classes, has_bias=True, per_channel=_per_channel,
+                               symmetric=_symmetric, quant_delay=_quant_delay),
+                 nn.FakeQuantWithMinMax(ema=True, ema_decay=_ema_decay, quant_delay=_quant_delay)])
+        self.head = nn.SequentialCell(head)
+
+    def construct(self, x):
+        x = self.input_fake(x)
+        x = self.features(x)
+        x = self.head(x)
+        return x
+
+
+def mobilenetV2_quant(**kwargs):
+    """
+    Constructs a MobileNet V2 model
+    """
+    return MobileNetV2Quant(**kwargs)