|
567]
elif groups == 2:
out_channels = [-1, 24, 200, 400, 800]
elif groups == 3:
out_channels = [-1, 24, 240, 480, 960]
elif groups == 4:
out_channels = [-1, 24, 272, 544, 1088]
elif groups == 8:
out_channels = [-1, 24, 384, 768, 1536]
data = shuffleUnit(data, out_channels[stage - 1], out_channels[stage],
'concat', groups, grouped_conv)
for i in range(stage_repeats[stage - 2]):
data = shuffleUnit(data, out_channels[stage], out_channels[stage],
'add', groups, True)
return data
def get_shufflenet(num_classes=10):
data = mx.sym.var('data')
data = mx.sym.Convolution(data=data, num_filter=24,
kernel=(3, 3), stride=(2, 2), pad=(1, 1))
data = mx.sym.Pooling(data=data, kernel=(3, 3), pool_type='max',
stride=(2, 2), pad=(1, 1))
data = make_stage(data, 2)
data = make_stage(data, 3)
data = make_stage(data, 4)
data = mx.sym.Pooling(data=data, kernel=(1, 1), global_pool=True, pool_type='avg')
data = mx.sym.flatten(data=data)
data = mx.sym.FullyConnected(data=data, num_hidden=num_classes)
out = mx.sym.SoftmaxOutput(data=data, name='softmax')
return out
这两个函数可以直接得到作者在论文中的表：
图7
结果比较
论文后面用了种实验证明这两个技术的有效性，且证实了ShuffleNet的优秀，这里就不细说，看论文后面的表就能一目了然。
|