CV baseline之VGG
作业题
1个7x7卷积核所需参数量：7x7xCxC=49C2
3个3x3卷积核所需参数量：3x(3x3xCxC)=27C2
参数减少比：（49-27）/49=44%
1）采用小卷积核可以获得高精度
achieve better accuracy. For instance, the best-performing submissions to the ILSVRC- 2013 utilised smaller receptive window size and smaller stride of the first convolutional layer.
2）采用多尺度及稠密预测可以获得高精度
Another line of improvements dealt with training and testing the network densely over the whole image and over multiple scales.
3）1x1卷积可以认为是线性变换，同时增加非线性层
In one of the configuration we also utilise 1x1 convolution filters, which can be seen as a linear transformation of the input channels(followed by non-linearity)
4）填充大小准则：保持卷积后特征图分辨率不变
the spatial padding of conv, layer input is such tha the spatial resolution is preserved after convolution.
5）LRN对精度无提升
such normalization does not improve the performance on the ILSVRC dataset, but leads to increased memory consumption and computation time.
6）Xavier初始化可达到较好效果
It is worth nothing that after the paper submission we found that it is possible to initialise the weights without pretraining by using random initialisation procedure.
7）S远大于224，图片可能仅包含物体的一部分
S>>224, the crop will correspond to a small part of the image, containing a small object or an object part.
8）大尺度模型采用小尺度模型初始化，可加快收敛
To speed-up training of the S=384 network, it was initialised with the weights pretrained with S=256, and we used a smaller initial learning rate of 0.001.
9）物体尺寸不一，采用多尺度训练，可以提高精度
Since objects in images can be of different size, multi scale training is beneficial to take this into account during training.
10）multi crop存在重复计算，因而低效
there is no need to sample multiple crops at test time, which is less efficient as it requires network recomputation for each crop.
11）multi crop可看成dense的补充，因而它们边界处理有所不同
Also, multi-crop evaluation is complementary to dense evaluation due to different convolution boundary conditions.
12）小而深的卷积网络优于大而浅的卷积网络
which confirms that a deep net with small filters outperforms a shallow net with larger filters.
13）尺度扰动对训练和测试阶段都有帮助
The results, indicate that scale iittering at test time leads to better performance.
14）scheduler=torch.optimize.lr_scheduler.ReduceLROnPlateau(patience=5):当参数不再变化时再减小学习率
