Skip to content

Latest commit

 

History

History
84 lines (63 loc) · 6.13 KB

README_IMAGENET.md

File metadata and controls

84 lines (63 loc) · 6.13 KB
Raw

Training & Evaluating on Antialiased models on Imagenet

We describe how to evaluate models for shift-invariance.

(0) Prepare ImageNet

(1) Evaluating models

We provide models with filter sizes 2,3,5 for AlexNet, VGG16, VGG16bn, ResNet18,34,50,101, DenseNet121, and MobileNetv2.

Evaluating accuracy

python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a alexnet_lpf --weights ./weights/alexnet_lpf3.pth.tar --gpu 0
python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a vgg16_lpf --weights ./weights/vgg16_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a vgg16_bn_lpf --weights ./weights/vgg16_bn_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a resnet18_lpf --weights ./weights/resnet18_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a resnet34_lpf --weights ./weights/resnet34_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a resnet50_lpf --weights ./weights/resnet50_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a resnet101_lpf --weights ./weights/resnet101_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a densenet121_lpf --weights ./weights/densenet121_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -e -f 3 -a mobilenet_v2_lpf --weights ./weights/mobilenet_v2_lpf3.pth.tar

Evaluating consistency

Same as above, but flag -es evaluates the shift-consistency -- how often two random 224x224 crops are classified the same.

python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a alexnet_lpf --weights ./weights/alexnet_lpf3.pth.tar --gpu 0
python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a vgg16_lpf --weights ./weights/vgg16_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a vgg16_bn_lpf --weights ./weights/vgg16_bn_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a resnet18_lpf --weights ./weights/resnet18_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a resnet34_lpf --weights ./weights/resnet34_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a resnet50_lpf --weights ./weights/resnet50_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a resnet101_lpf --weights ./weights/resnet101_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a densenet121_lpf --weights ./weights/densenet121_lpf3.pth.tar
python main.py --data /PTH/TO/ILSVRC2012 -es -b 8 -f 3 -a mobilenet_v2_lpf --weights ./weights/mobilenet_v2_lpf3.pth.tar

Some notes:

  • These line commands are very similar to the base PyTorch repository. We simply add suffix _lpf to the architecture and specify -f for filter size.
  • Substitute -f 3 and appropriate filepath for different filter sizes.
  • The example commands use our weights. You can them from your own training checkpoints by subsituting --weights PTH/TO/WEIGHTS for --resume PTH/TO/CHECKPOINT.

(2) Training antialiased models

The following commands train antialiased AlexNet, VGG16, VGG16bn, ResNet18,34,50, and Densenet121 models with filter size 5. Best checkpoint will be saved [[OUT_DIR]]/model_best.pth.tar.

python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a alexnet_lpf --out-dir alexnet_lpf3 --gpu 0 --lr .01
python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a vgg16_lpf --out-dir vgg16_lpf3 --lr .01 -b 128 -ba 2
python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a vgg16_bn_lpf --out-dir vgg16_bn_lpf3 --lr .05 -b 128 -ba 2
python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a resnet18_lpf --out-dir resnet18_lpf3
python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a resnet34_lpf --out-dir resnet34_lpf3
python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a resnet50_lpf --out-dir resnet50_lpf3
python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a resnet101_lpf --out-dir resnet101_lpf3
python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a densenet121_lpf --out-dir densenet121_lpf3 -b 128 -ba 2

Some notes:

  • As suggested by the official repository, AlexNet and VGG16 require lower learning rates of 0.01 (default is 0.1).
  • VGG16_bn also required a slightly lower learning rate of 0.05.
  • I train AlexNet on a single GPU (the network is fast, so preprocessing becomes the limiting factor if multiple GPUs are used).
  • MobileNet was trained with the training recipe from here, which is not currently supported in this repo.
  • Default batch size is 256. Some extra memory is added for the antialiasing layers, so the default batchsize may no longer fit in memory. To get around this, we simply accumulate gradients over 2 smaller batches -b 128 with flag --ba 2. You may find this useful, even for the default models, if you are training with smaller/fewer GPUs. It is not exactly identical to training with a large batch, as the batchnorm statistics will be computed with a smaller batch.

Checkpoint vs weights:

  • To resume training session, use flag --resume [[OUT_DIR]]/checkpoint_[[NUM]].pth.tar. This flag can be used instead of --weights in the evaluation scripts above.
  • Saved checkpoints include model weights and optimizer parameters. Also, if you trained with parallelization, then the weights/optimizer dicts will include parallelization. To strip optimizer parameters away and 'deparallelize' the model weights, run the following command (with appropriate substitution) afterwards:
python main.py --data /PTH/TO/ILSVRC2012 -f 3 -a resnet18_lpf --resume resnet18_lpf3/model_best.pth.tar --save_weights resnet18_lpf3/weights.pth.tar

I used this postprocessing step to provide the pretrained weights. As seen here, weights should be loaded before parallelizing the model. Meanwhile, the checkpoint is loaded after parallelizing the model.

(3) Results

Results are here.