Skip to content

Latest commit

 

History

History
103 lines (75 loc) · 2.69 KB

README.md

File metadata and controls

103 lines (75 loc) · 2.69 KB
Raw

Reproduced ResNet on CIFAR-10 and CIFAR-100 dataset.

contact: panyx0718 ([email protected])

Dataset:

https://www.cs.toronto.edu/~kriz/cifar.html

Related papers:

Identity Mappings in Deep Residual Networks

https://arxiv.org/pdf/1603.05027v2.pdf

Deep Residual Learning for Image Recognition

https://arxiv.org/pdf/1512.03385v1.pdf

Wide Residual Networks

https://arxiv.org/pdf/1605.07146v1.pdf

Settings:

  • Random split 50k training set into 45k/5k train/eval split.
  • Pad to 36x36 and random crop. Horizontal flip. Per-image whitenting.
  • Momentum optimizer 0.9.
  • Learning rate schedule: 0.1 (40k), 0.01 (60k), 0.001 (>60k).
  • L2 weight decay: 0.002.
  • Batch size: 128. (28-10 wide and 1001 layer bottleneck use 64)

Results:

![Precisions](g3doc/cifar_resnet.gif) ![Precisions Legends](g3doc/cifar_resnet_legends.gif)
CIFAR-10 Model Best Precision Steps
32 layer 92.5% ~80k
110 layer 93.6% ~80k
164 layer bottleneck 94.5% ~80k
1001 layer bottleneck 94.9% ~80k
28-10 wide 95% ~90k
CIFAR-100 Model Best Precision Steps
32 layer 68.1% ~45k
110 layer 71.3% ~60k
164 layer bottleneck 75.7% ~50k
1001 layer bottleneck 78.2% ~70k
28-10 wide 78.3% ~70k

Prerequisite:

  1. Install TensorFlow, Bazel.

  2. Download CIFAR-10/CIFAR-100 dataset.

curl -o cifar-10-binary.tar.gz https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz
curl -o cifar-100-binary.tar.gz https://www.cs.toronto.edu/~kriz/cifar-100-binary.tar.gz

How to run:

# cd to the your workspace.
# It contains an empty WORKSPACE file, resnet codes and cifar10 dataset.
ls -R
  .:
  cifar10  resnet  WORKSPACE

  ./cifar10:
  test.bin  train.bin  validation.bin

  ./resnet:
  BUILD  cifar_input.py  g3doc  README.md  resnet_main.py  resnet_model.py

# Build everything for GPU.
bazel build -c opt --config=cuda resnet/...

# Train the model.
bazel-bin/resnet/resnet_main --train_data_path=cifar10/train.bin \
                             --log_root=/tmp/resnet_model \
                             --train_dir=/tmp/resnet_model/train \
                             --dataset='cifar10' \
                             --num_gpus=1

# Evaluate the model.
# Avoid running on the same GPU as the training job at the same time,
# otherwise, you might run out of memory.
bazel-bin/resnet/resnet_main --eval_data_path=cifar10/test.bin \
                             --log_root=/tmp/resnet_model \
                             --eval_dir=/tmp/resnet_model/test \
                             --mode=eval \
                             --dataset='cifar10' \
                             --num_gpus=0