author : oukohou
time : 2019-09-26 16:44:48
email : [email protected]

A pytorch reimplementation of SSR-Net.

the official keras version is here: SSR-Net

results on MegaAge_Asian datasets:

[^1]: train from scratch, use MSEloss;
[^2]: use pretrianed my implementation_v1, use L1Loss.

Note:

• This SSR-Net model can't fit big learning rate, learning rate should be smaller than 0.002. otherwise the model will very likely always output 0, me myself suspects this is because of the utilizing Tanh as activation function.
• And also: Batchsize could severely affect the results. A set of tested params can be :

  batch_size = 50
  input_size = 64
  num_epochs = 90
  learning_rate = 0.001 # originally 0.001
  weight_decay = 1e-4 # originally 1e-4
  augment = False
  optimizer_ft = optim.Adam(params_to_update, lr=learning_rate, weight_decay=weight_decay)
  criterion = nn.L1Loss()
  lr_scheduler = optim.lr_scheduler.StepLR(optimizer_ft, step_size=30, gamma=0.1)
  

• The dataset preprocess is quite easy. For MegaAsian datasets, you can use the ./datasets/read_megaasina_data.py directly; for other datasets, just generate a pandas csv file in format like:

  filename,age
  1.jpg,23
  ...
  

is OK. But also, remember to change the ./datasets/read_imdb_data.py accordingly.

onnxruntime C++ implementation

thanks to DefTruth 's implementation here: How to convert SSRNet to ONNX and implements with onnxruntime c++.

another small note:

my reading understanding of SSRNet can be found:

• on my blog site here:论文阅读-年龄估计_SSRNet
• or on zhihu here: 论文阅读-年龄估计_SSRNet.

which was written in Chinese.