This implements weakly supervised learning of residual networks. This Torch implementation is based on fb.resnet.torch.
The download_pretrained_models script downloads pretrained models in data/pretrained_models directory.
th download_pretrained_models.lua
This implementation uses the following packages:
- torch
- nn
- cunn
- cudnn
- optim
- paths
- csvigo
- matio
You also need to install spatial-pooling.torch package to have spatial pooling modules.
To train ResNet-101 with WELDON pooling on VOC 2007 dataset, run main.lua with
th main.lua -optim sgd -LR 1e-2 -netType resnet101-weldon -batchSize 40 -imageSize 224 -data /path_dataset/VOCdevkit/VOC2007/ -dataset voc2007-cls -loss MultiLabel -train multilabel -k 15 -nEpochs 20
-
LR: initial learning rate. -
imageSize: size of the image. -
batchSize: number of images per batch -
k: number of regions for WELDON pooling. -
nEpochs: number of training epochs.
To train ResNet-101 with GlobalMaxPooling on VOC 2007 dataset, run main.lua with
th main.lua -optim sgd -LR 1e-2 -netType resnet101-gmp -batchSize 40 -imageSize 224 -data /path_dataset/VOCdevkit/VOC2007/ -dataset voc2007-cls -loss MultiLabel -train multilabel -k 15 -nEpochs 20
To train ResNet-101 with GlobalAveragePooling on VOC 2007 dataset, run main.lua with
th main.lua -optim sgd -LR 1e-2 -netType resnet101-gap -batchSize 40 -imageSize 224 -data /path_dataset/VOCdevkit/VOC2007/ -dataset voc2007-cls -loss MultiLabel -train multilabel -k 15 -nEpochs 20
MIT License