您好，网络的最后一层是卷积层，请问如果按照下面这种方式去处理最后一层，有什么问题吗？
class FinalConvLayerObserver(Meltable):

def __init__(self, conv2d):
    super(FinalLayerObserver, self).__init__()
    assert isinstance(conv2d, nn.Conv2d)
    self.conv2d = conv2d

    self.in_mask =nn.init.constant_(conv2d.weight, 0).to('cpu')# torch.zeros(conv2d.weight,0).to('cpu')
    self.f_hook = conv2d.register_forward_hook(self._forward_hook)


def _forward_hook(self, m, _in, _out):
    x = _in[0]
    self.in_mask += x.data.abs().cpu().sum(0, keepdim=True).view(-1)

def forward(self, x):
    return self.conv2d(x)

def melt(self):
    with torch.no_grad():
        replacer = nn.conv2d(int((self.in_mask != 0).sum()), self.conv2d.weight).to(
            self.conv2d.weight.device)

        replacer.weight.set_(self.conv2d.weight[:, self.in_mask != 0])

        replacer.bias.set_(self.conv2d.bias)
    return replacer

Dear author,
GBN elegantly relieve restraint from the shortcut in ResNet. It gets good results on resnet56. I really like your work.But I encountered problems while experimenting on rennet50,
Firstly, the BASE_FLOPS I calculate is 4135.790M instead of your 4142.714M,
In addition, the test accuracy of the first Tock is lower than yours.
it's appreciated if you can provide your pruned model or give some suggestions

剪枝完保存的pth文件加载时权重名都变了，还有对应的通道数，这个咋看，怎么正确加载剪枝完的pth模型呢

Unexpected key(s) in state_dict: "bn1.g", "bn1.area", "bn1.score", "bn1.bn_mask", "bn1.bn.weight", "bn1.bn.bias", "bn1.bn.running_mean", "bn1.bn.running_var", "bn1.bn.num_batches_tracked", "layer1.0.bn1.g", "layer1.0.bn1.area", "layer1.0.bn1.score", "layer1.0.bn1.bn_mask", "layer1.0.bn1.bn.weight", "layer1.0.bn1.bn.bias", "layer1.0.bn1.bn.running_mean", "layer1.0.bn1.bn.running_var", "layer1.0.bn1.bn.num_batches_tracked", "layer1.0.bn2.g", "layer1.0.bn2.area", "layer1.0.bn2.score", "layer1.0.bn2.bn_mask", "layer1.0.bn2.bn.weight", "layer1.0.bn2.bn.bias", "layer1.0.bn2.bn.running_mean", "layer1.0.bn2.bn.running_var", "layer1.0.bn2.bn.num_batches_tracked", "layer1.0.shortcut.1.g", "layer1.0.shortcut.1.area", "layer1.0.shortcut.1.score", "layer1.0.shortcut.1.bn_mask", "layer1.0.shortcut.1.bn.weight", "layer1.0.shortcut.1.bn.bias", "layer1.0.shortcut.1.bn.running_mean", "layer1.0.shortcut.1.bn.running_var", "layer1.0.shortcut.1.bn.num_batches_tracked", "layer1.1.bn1.g", "layer1.1.bn1.area", "layer1.1.bn1.score", "layer1.1.bn1.bn_mask", "layer1.1.bn1.bn.weight",

Hi，
我自己的模型最后一层不是linear，而是自己构建的其他模块名字。我应该怎么替换它呢？
如果我直接注释这一行代码的话，会出现如下错误：
ndexError: index 0 is out of bounds for dimension 0 with size 0
谢谢

Hello, I have a question about the way you counted FLOPs.
According to your code prune/utils.py, you used thop.
However, the code is recently corrected that it actually measured MACs rather than FLOPs.
(Please refer to Lyken17/pytorch-OpCounter#37
and Lyken17/pytorch-OpCounter@41fd65c)

In this case, are the FLOPs values in your paper still valid?

Thank you.
Jung

def forward(self, x):
    x = self.conv1(x)
    x = self.bn1(x)
    x = self.relu(x)
    x = self.maxpool(x)

    x = self.layer1(x)
    x = self.layer2(x)
    x = self.layer3(x)
    x = self.layer4(x)

    x = self.deconv_layers(x) #自己定义的一个模块
    #print("x:",x)

    ret = {}
    
    ret['hm'] = self.hm(x) #输出1
    ret['reg'] = self.reg(x)#输出2
    ret['wh'] = self.wh(x)#输出3
   return [ret]

如上，我应该如何修改代码才能让最后一层只改变输入大小？谢谢
输出1，2，3是我模型输出的三个分支
output= {'hm': 13, 'wh': 2, 'reg': 2}

Hello,

I am getting an error on mismatched Batch normalization layer when pruning resnet50 for imagenet dataset.

Could you please provide Resnet50 model definition and how it is used for imagenet dataset pruning?

Hi, I tried to modify your code by using the imagenet's resnet50 on cifar10. I followed the vgg16 example and changed it to resnet50. However, I got the following error when running the pruning, CUDA_VISIBLE_DEVICES=0 python ./run/resnet-50/resnet50_prune_demo.py --config ./run/resnet-50/prune.json.

File "/home/ccng/anaconda3/envs/gatedeco/lib/python3.7/site-packages/torch/nn/modules/module.py", line 591, in getattr
type(self).name, name))
AttributeError: 'Resnet50' object has no attribute 'classifier'

May I know how to resolve this error? Thanks.

When i use my own resnet56 model, it got a very poor performance at the start Tock step.
I use SGD to train my model with lr, moment , weight decay set to 0.1, 0.9, 1e-4. The learning rate is divide by 0.1 at epoch 80 and 120. The training setting is same as the one you mentioned in your paper.


During training process, when the lr is decayed to 0.01, the model can get at least 92.8% test acc. But your first tock step shows that the acc is lower than 91% when the lr is 0.01, I think it is impossible.

Hi，how to compress own model use this alg?

Thanks

(｡･∀･)ﾉﾞ嗨，感谢您的工作，是非常简单有效的方法！

ECCV2018有一篇图森科技的工作，也是用一个可训练的标量去衡量通道重要性，文章在此Data-Driven Sparse Structure Selection for Deep Neural Networks，不知道您有没有与此文章做过比较。

Dear author,

Thank for this impressive piece of work.
How to implement convolution layer with cardinality ?

in universal.py we have:

def melt(self):
        if self.conv.groups == 1:
            groups = 1
        elif self.conv.groups == self.conv.out_channels:
            groups = int((self.out_mask != 0).sum())
        else:
            assert False

        replacer = nn.Conv2d(
            in_channels = int((self.in_mask != 0).sum()),
            out_channels = int((self.out_mask != 0).sum()),
            kernel_size = self.conv.kernel_size,
            stride = self.conv.stride,
            padding = self.conv.padding,
            dilation = self.conv.dilation,
            groups = groups,
            bias = (self.conv.bias is not None)
        ).to(self.conv.weight.device)

        with torch.no_grad():
            if self.conv.groups == 1:
                replacer.weight.set_(self.conv.weight[self.out_mask != 0][:, self.in_mask != 0])
            else:
                replacer.weight.set_(self.conv.weight[self.out_mask != 0])
            if self.conv.bias is not None:
                replacer.bias.set_(self.conv.bias[self.out_mask != 0])
        return replacer

if the convolution layer have a cardinality (like in several modern model), we get assert False.
How to implement cardinality for convolution layers ?

Hi,

Thank you for your excellent work.

When running your "Command line (VGG-16) demo", I get a JSON encoder error in logger.py at line 63 under the save_network function - "TypeError: Object of type 'Config' is not JSON serializable". Can you please suggest a fix to this issue?

PS- I am using python 3.6.10 and PyTorch 1.0.1.post2.

Best,
Niam

class Conv2dObserver(Meltable):
def init(self, conv):
super(Conv2dObserver, self).init()
assert isinstance(conv, nn.Conv2d)
self.conv = conv
self.in_mask = torch.zeros(conv.in_channels).to('cpu')
self.out_mask = torch.zeros(conv.out_channels).to('cpu')
self.f_hook = conv.register_forward_hook(self._forward_hook)

def extra_repr(self):
    return '(%d, %d) -> (%d, %d)' % (self.conv.in_channels, self.conv.out_channels, int((self.in_mask != 0).sum()), int((self.out_mask != 0).sum()))

def _forward_hook(self, m, _in, _out):
    x = _in[0]
    self.in_mask += x.data.abs().sum(2, keepdim=True).sum(3, keepdim=True).cpu().sum(0, keepdim=True).view(-1)

def _backward_hook(self, grad):
    self.out_mask += grad.data.abs().sum(2, keepdim=True).sum(3, keepdim=True).cpu().sum(0, keepdim=True).view(-1)
    new_grad = torch.ones_like(grad)
    return new_grad

def forward(self, x):
    output = self.conv(x)
    noise = torch.zeros_like(output).normal_()
    output = output + noise
    if self.training:
        output.register_hook(self._backward_hook)
    return output

def melt(self):
    if self.conv.groups == 1:
        groups = 1
    elif self.conv.groups == self.conv.out_channels:
        groups = int((self.out_mask != 0).sum())
    else:
        assert False

    print("in_channels:",int((self.in_mask != 0).sum()))
    print("out_channels:", int((self.out_mask != 0).sum()))
    print("kernel_size:", self.conv.kernel_size)
    print("stride:", self.conv.stride)
    print("padding:", self.conv.padding)
    print("dilation:", self.conv.dilation)
    print("groups:", groups)
    print("bias:", (self.conv.bias is not None))

    replacer = nn.Conv2d(
        in_channels = int((self.in_mask != 0).sum()),
        out_channels = int((self.out_mask != 0).sum()),
        kernel_size = self.conv.kernel_size,
        stride = self.conv.stride,
        padding = self.conv.padding,
        dilation = self.conv.dilation,
        groups = groups,
       bias = (self.conv.bias is not None)
    ).to(self.conv.weight.device)

    with torch.no_grad():
        if self.conv.groups == 1:
            replacer.weight.set_(self.conv.weight[self.out_mask != 0][:, self.in_mask != 0])
        else:
            replacer.weight.set_(self.conv.weight[self.out_mask != 0])
        if self.conv.bias is not None:
            replacer.bias.set_(self.conv.bias[self.out_mask != 0])
    return replacer

@classmethod
def transform(cls, net):
    r = []
    def _inject(modules):
        keys = modules.keys()
        for k in keys:
            if len(modules[k]._modules) > 0:
                _inject(modules[k]._modules)
            if isinstance(modules[k], nn.Conv2d):
                modules[k] = Conv2dObserver(modules[k])
                r.append(modules[k])
    _inject(net._modules)
    return r

Thank you for your great work!
May I ask why Meltable. observe should be called before melt_all and finetuning the net.

If I want to prune my model with object detection models, what should I pay attention to? What should I change?

Hello,

I am using 2 V100 GPUs. I trying to apply the pruning technique on pre-trained resnet50 model for imagenet dataset.

The pruning is running. However, it is using only 1 of the two GPUs. I am running the python script in the following way (I made myself the resnet50_prune.py without changing any of the pruning code).
CUDA_VISIBLE_DEVICES=0,1 python ./run/resnet-50/resnet50_prune.py

As title,

Thank you for work and sharing this code.

I was wondering the meaning of new_grad = torch.ones_like(grad) in the following code in universal.py.

    def _backward_hook(self, grad):
        self.out_mask += grad.data.abs().sum(2, keepdim=True).sum(3, keepdim=True).cpu().sum(0, keepdim=True).view(-1)
        new_grad = torch.ones_like(grad)
        return new_grad

I run the examples ,
for point in [i for i in list(flops_save_points)]:
if flops_ratio <= point:
torch.save(pack.net.module.state_dict(), './logs/fashion_ticktock/%s.pth' % str(point)),

But ,seved model size never change to small.

感谢youzhonghui提供如此好的剪枝方法，请问Conv2dObserver里，out_mask是通过_backward_hook累加梯度得到的，后面的通道选择上需要选取self.out_mask != 0的通道，由于梯度是float类型，这样子的话累加和怎么等于0呢？谢谢