def __init__(self, optimized: bool = False):
super(ToyClassifier, self).__init__()
block_params = [(3, 18, 1), (18, 36, 2), (36, 74, 1), (74, 146, 2),
(146, 290, 1), (290, 578, 2), (578, 1154, 1),
(1154, 1154, 2)]
if optimized:
blocks = [
OptimizedConvBNReLU(in_channels=3,
out_channels=_make_divisible(
block_params[0][1], 8),
stride=block_params[0][2])
]
for in_channels, out_channels, stride in block_params[1:]:
blocks.append(
OptimizedConvBNReLU(
in_channels=_make_divisible(in_channels, 8),
out_channels=_make_divisible(out_channels, 8),
stride=stride))
in_features = _make_divisible(1154, 8)
else:
blocks = [
ConvBNReLU(in_channels=in_channels,
out_channels=out_channels,
stride=stride)
for (in_channels, out_channels, stride) in block_params
]
in_features = 1154
self.blocks = nn.ModuleList(blocks)
self.pooling = nn.AdaptiveAvgPool2d(1)
self.classifier = nn.Conv2d(in_channels=in_features,
out_channels=1000,
kernel_size=(1, 1))
self.quant = torch.quantization.QuantStub()
self.dequant = torch.quantization.DeQuantStub()
def adjust_channels(channels: int,
width_mult: float,
min_value: Optional[int] = None) -> int:
return _make_divisible(channels * width_mult, 8, min_value)
def adjust_channels(channels: int, width_mult: float):
return _make_divisible(channels * width_mult, 8)
def __init__(self, input_channels: int, squeeze_factor: int = 4):
super().__init__()
squeeze_channels = _make_divisible(input_channels // squeeze_factor, 8)
self.fc1 = nn.Conv2d(input_channels, squeeze_channels, 1)
self.fc2 = nn.Conv2d(squeeze_channels, input_channels, 1)