Guangyu Guo (郭广宇)
Guangyu Guo (郭广宇)
> ``` > % Matlab Code > % FLOPs for nn.Conv3d, without bias > A_size = [num_out/group, prod(kernel)*num_input/group]; % weight > C_size = [prod(kernel)*num_input/group, out_h*out_w*out_t]; % im2col > flops =...
I have fix the problem by using vlfeat toolbox, http://www.vlfeat.org/index.html, vl_colsubset is a function of this toolbox.
@sudonto After download and install the vlfeat toolbox under the st-resnet directory, I add `run(fullfile(fileparts(mfilename('fullpath')), ... 'vlfeat','toolbox', 'vl_setup.m')) ;` to the STMulNet.m before runing it
抱歉，这一部分有个师弟正在做自己的新工作，代码得等他的论文完成之后开源
你好，这里主要是因为论文和代码表述方式不一样 代码里面self.image_size = cfg.DATA.CROP_SIZE是高分辨率图像的大小；公式里面H和W是低分辨率图像的大小，所以KW和KH是高分辨率的大小，论文中的Algorithm 1会表达的比较明确。
224/3=74.6 向上取整是75 这里的S的计算公式有误，因该是对空间维度每个边单独计算S，我稍后仔细核对后在readme里面勘误。红色文字是通过代码debug得到的真实数值，是没有问题的。
Initialization in tools_1_ts/train_isrd_5runs.py: Line 41-49 Model file: ib/models/hrir.py
For CNN, return the original feature For ViT, in lib/models/vit_pixel.py, Line 264-267 and Line 273-278 ``` if return_feat: feats = [] patch_tokens = x.permute(0, 2, 1).view(x.size(0), x.size(2), int(math.sqrt(x.size(1))), int(math.sqrt(x.size(1)))) feats.append(patch_tokens)...