AliceH1226的博客

推理还cuda out of memory 是有问题的，要加上不计算梯度的with torch.no_grad()

python3.8 conda 安装dlib

sgdr = CosineAnnealingLR_with_Restart(optimizer, T_max=config.cycle_inter, T_mult=1, model=net, ...

from sklearn.model_selection import train_test_splitimport osimgs_dir = r"img"train_dir = r"train"valid_dir = r"test"image_list = os.listdir(imgs_dir)X_train, X_test = train_test_split(image_list, test_size=0.2)with open(train_dir, 'a') as f: for

FL(pt)=−αt(1−pt)γlog⁡(pt),pt=p if y=1,else (1−p)FL(p_t) = - \alpha_t (1-p_t)^{\gamma} \log{(p_t)}, p_t=p \ if \ y=1,else \ (1-p)FL(pt​)=−αt​(1−pt​)γlog(pt​),pt​=p if y=1,else (1−p)论文中α=0.25\alpha=0.25α=0.25, γ=2\gamma =2γ=2时表

GDθt=θt−1−α ∂Loss∂θ\theta_t=\theta_{t-1}-\alpha\ \frac{\partial Loss}{\partial \theta}θt​=θt−1​−α ∂θ∂Loss​有三种，sgd,bgd,mini-batch gd,s,b, gd分别指stochastic, batch, gradient descent。区别分别是每一步计算一项，一小批，全部数据。在实现的时候，dataloader的batchsize分别是1，num_mini_bat

same convstride为1：3,1,x,x1,1,0,1减半convstride为2：3,2,x,x1,2,0,1

递归版：rl−1=sl∗kl+kl−slr_{l-1}=s_l*k_l+k_l-s_lrl−1​=sl​∗kl​+kl​−sl​迭代版：r0=∑l=1L((kl−1)∏i=1l−1si)+1r_0=\sum_{l=1}^{L}((k_l-1)\prod_{i=1}^{l-1}s_i)+1r0​=∑l=1L​((kl​−1)∏i=1l−1​si​)+1如果是空洞卷积，只需要改一下klk_lkl​，把它改成α(kl−1)+1\alpha(k_l-1)+1α(kl​−1)+1，α\alphaα是dila.

关于kernelbasic block 的kernel都是3，有两组kernel。bottle net block 的kenel是1,3，1，有三组kernel。关于channelbasic block 是 inchan, outchan, outchan, outchan。bottle net block 是 inchan, outchan, outchan, outchan, outchan, outchan*expansion关于downsample两者都是，最后一组cbr，看实际情况在