Awesome

Implementation of Network Slimming

Implementation of "Learning Efficient Convolutional Networks through Network Slimming".

Requirements

• Python 2.7
• Chainer 2.0.0
• Cupy 1.0.0
• matplotlib

Example

Train model

$ python src/train.py model/base.json -p vgg_1 -g 0 --no-valid-data

Prune

$ python src/prune.py model/vgg_1.model model/base.json model/vgg_2_org.model model/vgg_2.json 0.5

Fine tune

$ python src/train.py model/vgg_2.json -g 0 -m model/vgg_2_org.model --lambda-value 0

How to use

Train

$ python src/train.py <structure_path> -g <gpu_id> -m <model_path> -b <batch_size> -p <prefix> --optimizer <optimizer> --epoch <epoch_num> --lr-decay-epoch <lr_decay>

While training the following files are saved.

• model parameter file <prefix>.model
• loss and error log file <prefix>_log.csv
• loss curve image <prefix>_loss.png
• error curve image <prefix>_error.png

Parameters:

• <structure_path>: Required
  Model structure JSON file path (e.g. model/base.json)
• -g (--gpu) <int>: Optional
  GPU device ID. Negative value indicates CPU (default: -1)
• -m (--model) <model name>: Optional
  Model file path (default: None)
• -b (--batch_size) <int>: Optional
  Mini batch size (default: 64)
• -p (--prefix) <str>: Optional
  Prefix of saved model file. (default: file name of structure JSON except extension)
• --base <int>: Optional
  Base size of model (default: 64)
• --epoch <int>: Optional
  Training epoch (default: 160)
• --save-epoch <int>: Optional
  Epoch interval to save model parameter file. 0 indicates model paramete is not saved at fixed intervals. Note that the best accuracy model is always saved even if this parameter is 0. (default: 0)
• --optimizer <str>: Optional
  Optimizer name (sgd or adam, default: sgd)
• --lr <float>: Optional
  Initial learning rate for SGD (default: 0.1)
• --alpha <float>: Optional
  Initial alpha for Adam (default: 0.001)
• --lr-decay-epoch <int>: Optional
  Epoch interval to decay learning rate. Learning rate is decay to 1/10 at this intervals. (default: 80,120)
• --weight-decay <float>: Optional
  Weight decay (default: 0.0001)
• --lambda-value <float>: Optional
  Regularization factor for gamma of batch normalization. It is recommended to set 0 for fine tuning. (default: 0.0001)
• --no-valid-data: Optional
  If set validation data is not used (i.e. train data is not separated to train/validation).
• --seed <int>: Optional
  Random seed (default: 1)

Example:

$ python src/train.py model/base.json -g 0 -b 64 -p vgg_1 --optimizer sgd --epoch 160 --lr-decay-epoch 80,120

Prune

$ python src/prune.py <src_model_path> <src_structure_path> <dest_model_path> <dest_structure_path> <prune_ratio>

Parameters:

• <src_model_path(str): Required
  Source model file path
• <src_structure_path>(str): Required
  Source structure JSON file path
• <dest_model_path(str): Required
  Destination model file path
• <dest_structure_path>(str): Required
  Destination structure JSON file path
• <prune ratio>(float): Required Prune ratio

Example:

$ python src/prune.py model/vgg_1.model model/base.json model/vgg_2_org.model model/vgg_2.json 0.5

License

MIT License