hi，thanks for your best contributions! When i converted the trained model on cifar10, i find the converted model get more larger,e.g. 10723->24418, how can i converted the model right, or the fact is that.

the train arguments on cifar10 is : --model condensenetv2 -b 64 -j 12 --data cifar10 --stages 14-14-14 --growth 8-16-32

thanks!

Awesome job! Recently, I have read your paper and have a question about the way you prune the weight. In the paper, you write as following

M^g_{i,j} is set to zero for all j in g-th group for each pruned output feature map i.

Maybe it should be implemented as self._mask[d, i:i+d_in, :, :].fill_(0) rather than https://github.com/jianghaojun/CondenseNetV2/blob/c771957cb8fe466d0ecbafe9060e4c342a33fc4d/utils/layers.py#L247

Is there anything wrong of my understanding?

您好，请问我能用作者训练好imagenet 模型进行迁移学习吗，我并没有发现该模型，因为我简单的把该结构拿出来进行图像检索，具体操作为去掉最后的分类层，添加l连接层(1024,128)发现结果并不是很好，大概是我做错了，而且我发现收敛的很慢，运行同样的epoch和损失函数，该网络65个epoch时只有50%,而imagenet预训练下的alexnet都有70%，实验设备有限，我希望能快速得到结果，希望作者能给合适imagenet模型，邮箱是[email protected],万分感谢。这是我修改的地方

class CondenseNetV2(nn.Module): def init(self, args, bit):

    super(CondenseNetV2, self).__init__()

    self.stages = args.stages
    self.growth = args.growth
    assert len(self.stages) == len(self.growth)
    self.args = args
    self.progress = 0.0
    if args.dataset in ['cifar10', 'cifar100']:
        self.init_stride = 1
        self.pool_size = 8
    else:
        self.init_stride = 2
        self.pool_size = 7

    self.features = nn.Sequential()
    ### Initial nChannels should be 3
    self.num_features = 2 * self.growth[0]
    ### Dense-block 1 (224x224)
    self.features.add_module('init_conv', nn.Conv2d(3, self.num_features,
                                                    kernel_size=3,
                                                    stride=self.init_stride,
                                                    padding=1,
                                                    bias=False))
    for i in range(len(self.stages)):
        activation = 'HS' if i >= args.HS_start_block else 'ReLU'
        use_se = True if i >= args.SE_start_block else False
        ### Dense-block i
        self.add_block(i, activation, use_se)

#只修改了这里 #self.fc = nn.Linear(self.num_features, args.fc_channel) #self.fc_act = HS() self.hash = nn.Linear(self.num_features, bit) self.fc_act = HS() ### Classifier layer #self.classifier = nn.Linear(args.fc_channel, args.num_classes) self._initialize() #************************

def cdnv2_b(self,args, bit):#这里没改，就加个bit args.stages = '2-4-6-8-6' args.growth = '6-12-24-48-96' print('Stages: {}, Growth: {}'.format(args.stages, args.growth)) args.stages = list(map(int, args.stages.split('-'))) args.growth = list(map(int, args.growth.split('-'))) args.condense_factor = 6 args.trans_factor = 6 args.group_1x1 = 6 args.group_3x3 = 6 args.group_trans = 6 args.bottleneck = 4 args.last_se_reduction = 16 args.HS_start_block = 2 args.SE_start_block = 3 args.fc_channel = 1024 return CondenseNetV2(args, bit)

#************************************************args参数只传了这些，其他都没改 import os import argparse import warnings warnings.filterwarnings("ignore") parser = argparse.ArgumentParser(description='PyTorch Condensed Convolutional Networks') args, unknown = parser.parse_known_args() args.dataset = 'cifar100' args.num_classes = 100 cdnv2_b(args, 64):

size mismatch for features.denseblock_1.denselayer_1.conv_1.conv.weight: copying a param with shape torch.Size([32, 2, 1, 1]) from checkpoint, the shape in current model is torch.Size([32, 16, 1, 1]).

Sfr-shufflenetv2 is mentioned in the paper, but I can't find its code. Is Sfr-shufflenetv2 add or concat 1 * 1 convolution and the other half channel of the input? And the paper mentioned that you shuffle the channel after using the SFR module in condensenetv2, but I didn't see the use of channel shuffle in the corresponding condensenetv2.py file, so channel shuffle is not necessary, right? thanks!!