Only save model trainable parameters, torch.save(model.state_dict())
params(counts) Conv2d: Cin×Cout×K×K BatchNorm2d: 2N (N is conv filters) LeakyReLU: 0 Upsample: 0
MACs & FLOPs
FLOPs is abbreviation of floating operations which includes mul / add / div ... etc.
MACs stands for multiply–accumulate operation that performs a <- a + (b x c).
As shown in the text, one MACs has one mul and one add. That is why in many places FLOPs is nearly two times as MACs.
When comparing MACs /FLOPs, we want the number to be implementation-agnostic and as general as possible. Therefore in THOP, we only consider the number of multiplications and ignore all other operations.
PS: The FLOPs is approximated by multiplying two.原始代码中 BatchNorm2d 尺寸为 2N, 从保存的模型来看，应为 4N, 分别是 (weight, bias, running_mean, running_var)
yolov3, input(3, 416, 416)
================================================================
Total params: 61,949,149
Trainable params: 61,949,149
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 1.98
Forward/backward pass size (MB): 901.03
Params size (MB): 236.32
Estimated Total Size (MB): 1139.33
----------------------------------------------------------------网络结构
shortcut:
相应层直接叠加
route:
相应层进行cat操作，在filter的维度上进行叠加
num_classes:
模型最后一个 conv，filter 数量由 anchor_num*(num_class+5) 决定。tensorflow 由输入参数部分修改，pytorch 由 config 中的 crate_custom_model.sh 决定。
output:
x,y: (sigmoid(x/y) + grid_x/y)*stride.
x,y在sigmoid后，位于[0,1]之间，grid_size由输出的feature_map中的维度决定。grid_x/y其实就是[0,1,2，...13),
[0,1,2...26), [0,1,2...52). 原图中的网格映射到feature map中的一个点。只需判断，x/y偏移了多少，
就是在对应的网格中进行了相应的偏移。当x/y为0.5时，此时预测的框中心与anchor重合。
w/h: exp(w/h) * anchor_w/h * stride.
这里anchor_w/h 是 anchor_w/h / stride后的。
当w/h位于0时，exp(w/h)位于1，此时预测的w,h结果正好与anchor的尺寸一致。
pred_conf/pred_cls: sigmoid(pred_conf/pred_cls)统一在(0, 1)之间.
loss
metrics
训练流程
pytorch-yolov3
- 修改 torchsummary, BatchNorm2d 参数量为 4×N
- 模型 flops 统计
- mixup
- test
- evaluate
- train
pytorch-ssd
- SSD
- test
- evaluate
- train
pytorch-anchor-free
https://github.com/eriklindernoren/PyTorch-YOLOv3
https://github.com/sksq96/pytorch-summary
https://github.com/Lyken17/pytorch-OpCounter
https://github.com/qfgaohao/pytorch-ssd
https://zhuanlan.zhihu.com/p/33544892