yaojunr / futurelab-ai-contest

2018 Futurelab AI Contest

In brief

The model is based on MobileNetV2 and NasNet. It is the ensemble of these two models with concatenation of these two feature maps.
Parameters：

lr:0.001(every epoch decay 0.98)
weight-decay:0.001
momentum:0.99
batch_size:64
epoch:100

The MobileNetV2 network structure is given by this, and the pretrained model can be downloaded by this. We remove the parameters of the final fully connected layer and remove the prefix module in the state_dict.

The pretrained NasNet model, which can be downloaded by this, is processed in the same way as MobileNetV2.

Software requirement

• Ubuntu 16.04
• CUDA 8
• Cudnn v6+
• python 3.5
  • pytorch 0.4.0
  • torchvision 0.2.0
  • numpy 1.13.3
  • Pillow 5.1.0
  • tensorboard_logger 0.1.0（optional）

Deploy

• Train

  Dataset processing: split the sample dataset into training set and validation set with the default ratio 10:1：

  python3 datautil.py <sample csv file>

  training：

  python3 baseline.py [--val <val csv file>] <train csv file>

  Important: the path of the data folder and the path of the csv file should be in the same folder

• Test

  python3 test.py  --test_model <model to be tested> --data <data folder>