a=featuremap*1.5，像这样的简单的乘法或者加法该怎么办？

[dreamhighchina]

因为这个1.5并不是什么blob，所以在 layer = caffe_net.Layer_param(name=layer_name, type='Eltwise', bottom=[log.blobs(input), log.blobs(args[0])], top=top_blobs)的时候，会出现 log.blobs(args[0])keyerror的原因，这个该怎么解决？？？

[xxradon]

那就是用torch.mul()之类的函数替换*

[Guanmoyu]

你可以在重载的_mul函数里加个判断分支，如果为常数，重载为Scale层

[Guanmoyu]

x * 1.5这种转换到caffe，我这边有三种思路，假设x维度为(n,c,h,w)，第一种把1.5先expand as成相同维度的blob，每个值都是1.5，然后用eltwise实现，需要重载一下expand_as；第二种，把1.5先expand as (n,c,1,1)的blob，每个值都是1.5，然后用双bottom的scale实现，这种稍微绕了一下；第三种，可以将scale的weight设为1.5，具体的可以参考_instance_norm的实现，应该也是可以实现的，不过我没试过。

[dreamhighchina]

我觉得还是expand的方式比较好哈哈，回头试试，因为对caffe框架不熟悉，谢谢大佬指导。

[baoxin1100]

x * 1.5这种，可以修改pytorch_to_caffe.py的_mul函数，如下，添加对int值或者float值的判断，用Scale完成乘法/加法/减法操作。关键在layer.scale_param(scale=scale, bias=None)这一句，如果是乘法，就设置scale，bias设为None或者0，同理如果是加减法，scale要设置为1，bias设置为具体的数值。注意这里的scale和bias的形状是featuremap通道的个数，即scale.shape[0]==featuremap.shape[1]，可以看下面代码， args[0]就是1.5，要把1.5转为通道长度的ndarray。

def _mul(input, *args):
    x = raw__mul__(input, *args)
    if not NET_INITTED or log.blobs(input) is None:
        return x
    if log.blobs(args[0]) is not None and input.shape == args[0].shape:
        layer_name = log.add_layer(name='mul')
        top_blobs = log.add_blobs([x], name='mul_blob')
        layer = caffe_net.Layer_param(name=layer_name, type='Eltwise',
                                      bottom=[log.blobs(input), log.blobs(args[0])], top=top_blobs)
        layer.param.eltwise_param.operation = 0  # product is 0
        log.cnet.add_layer(layer)
    elif isinstance(args[0], int) or isinstance(args[0], float):
        layer_name = log.add_layer(name='mul')
        top_blobs = log.add_blobs([x], name='mul_blob')
        layer = caffe_net.Layer_param(name=layer_name, type='Scale',
                                      bottom=[log.blobs(input)], top=top_blobs)
        scale = np.array(args[0],dtype=np.float32).squeeze().repeat(input.shape[1], 0)
        layer.scale_param(scale=scale, bias=None)
        layer.add_data(scale)
        log.cnet.add_layer(layer)

同理，加法_add和减法_sub也可添加相应的判断，代码如下： 加法：

def _add(input, *args):
    x = raw__add__(input, *args)
    if not NET_INITTED or log.blobs(input) is None:
        return x
    if log.blobs(args[0]) is not None and input.shape == args[0].shape:
        layer_name = log.add_layer(name='add')
        top_blobs = log.add_blobs([x], name='add_blob')
        layer = caffe_net.Layer_param(name=layer_name, type='Eltwise',
                                      bottom=[log.blobs(input),log.blobs(args[0])], top=top_blobs)
        layer.param.eltwise_param.operation = 1 # sum is 1
        log.cnet.add_layer(layer)
    elif isinstance(args[0], int) or isinstance(args[0], float):
        layer_name = log.add_layer(name='sub')
        top_blobs = log.add_blobs([x], name='sub_blob')
        layer = caffe_net.Layer_param(name=layer_name, type='Scale',
                                      bottom=[log.blobs(input)], top=top_blobs)
        bias = np.array(args[0],dtype=np.float32).squeeze().repeat(input.shape[1], 0)
        scale = np.ones(bias.shape, dtype=np.float32)
        layer.scale_param(scale=scale, bias=bias)
        layer.add_data(scale, bias)
        log.cnet.add_layer(layer)

减法：

def _sub(input, *args):
    x = raw__sub__(input, *args)
    if not NET_INITTED or log.blobs(input) is None:
        return x
    if log.blobs(args[0]) is not None and input.shape == args[0].shape:
        layer_name = log.add_layer(name='sub')
        top_blobs = log.add_blobs([x], name='sub_blob')
        layer = caffe_net.Layer_param(name=layer_name, type='Eltwise',
                                      bottom=[log.blobs(input),log.blobs(args[0])], top=top_blobs)
        layer.param.eltwise_param.operation = 1 # sum is 1
        layer.param.eltwise_param.coeff.extend([1., -1.])
        log.cnet.add_layer(layer)
    elif isinstance(args[0], int) or isinstance(args[0], float):
        layer_name = log.add_layer(name='add')
        top_blobs = log.add_blobs([x], name='add_blob')
        layer = caffe_net.Layer_param(name=layer_name, type='Scale',
                                      bottom=[log.blobs(input)], top=top_blobs)
        bias = np.array(args[0],dtype=np.float32).squeeze().repeat(input.shape[1], 0)
        scale = np.ones(bias.shape, dtype=np.float32)
        layer.scale_param(scale=scale, bias=bias)
        layer.add_data(scale, -bias)
        log.cnet.add_layer(layer)