def __init__(self, num_class=10, channel=1):
super(LeNet5, self).__init__()
self.num_class = num_class
self.conv1 = nn.Conv2dBnFoldQuant(channel,
6,
5,
pad_mode='valid',
per_channel=True,
quant_delay=900)
self.conv2 = nn.Conv2dBnFoldQuant(6,
16,
5,
pad_mode='valid',
per_channel=True,
quant_delay=900)
self.fc1 = nn.DenseQuant(16 * 5 * 5,
120,
per_channel=True,
quant_delay=900)
self.fc2 = nn.DenseQuant(120, 84, per_channel=True, quant_delay=900)
self.fc3 = nn.DenseQuant(84,
self.num_class,
per_channel=True,
quant_delay=900)
self.relu = nn.ActQuant(nn.ReLU(), per_channel=False, quant_delay=900)
self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = nn.Flatten()
def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
super(ConvBNReLU, self).__init__()
padding = (kernel_size - 1) // 2
conv = Conv2dBnFoldQuant(in_planes, out_planes, kernel_size, stride, pad_mode='pad', padding=padding,
group=groups, fake=_fake, quant_config=_quant_config)
layers = [conv, nn.ActQuant(nn.ReLU())] if _fake else [conv, nn.ReLU()]
self.features = nn.SequentialCell(layers)
def __init__(self, in_planes, out_planes, kernel_size=3, stride=1):
super(FirstQuantLayer, self).__init__()
padding = (kernel_size - 1) // 2
in_channels = in_planes
out_channels = out_planes
conv_inner = nn.Conv2dBnFoldQuantOneConv(in_channels,
out_channels,
kernel_size=kernel_size,
stride=stride,
pad_mode='pad',
padding=padding,
quant_config=quant_config,
quant_dtype=QuantDtype.INT8)
activation = nn.ActQuant(activation=nn.ReLU6(),
quant_config=quant_config,
quant_dtype=QuantDtype.INT8)
self.features = nn.SequentialCell([conv_inner, activation])
def __init__(self,
in_planes,
out_planes,
kernel_size=3,
stride=1,
groups=1):
super(ConvBNReLU, self).__init__()
padding = (kernel_size - 1) // 2
conv = Conv2dBatchNormQuant(in_planes,
out_planes,
kernel_size,
stride,
pad_mode='pad',
padding=padding,
group=groups,
fake=_fake,
per_channel=_per_channel,
symmetric=_symmetric)
layers = [conv, nn.ActQuant(nn.ReLU())] if _fake else [conv, nn.ReLU()]
self.features = nn.SequentialCell(layers)
def __init__(self, num_class=10, channel=1):
super(LeNet5, self).__init__()
self.num_class = num_class
self.qconfig = create_quant_config(per_channel=(True, False),
symmetric=(True, False))
self.conv1 = nn.Conv2dQuant(channel,
6,
5,
pad_mode='valid',
quant_config=self.qconfig,
quant_dtype=QuantDtype.INT8)
self.conv2 = nn.Conv2dQuant(6,
16,
5,
pad_mode='valid',
quant_config=self.qconfig,
quant_dtype=QuantDtype.INT8)
self.fc1 = nn.DenseQuant(16 * 5 * 5,
120,
quant_config=self.qconfig,
quant_dtype=QuantDtype.INT8)
self.fc2 = nn.DenseQuant(120,
84,
quant_config=self.qconfig,
quant_dtype=QuantDtype.INT8)
self.fc3 = nn.DenseQuant(84,
self.num_class,
quant_config=self.qconfig,
quant_dtype=QuantDtype.INT8)
self.relu = nn.ActQuant(nn.ReLU(),
quant_config=self.qconfig,
quant_dtype=QuantDtype.INT8)
self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = nn.Flatten()
