This repository contains the slim(tensorflow) implementation for the paper Densely Connected Convolutional Networks.There are two types of Densely Connected Convolutional Networks are available:

• DenseNet - without bottleneck layers
• DenseNet-BC - with bottleneck layers

Citation:

 @inproceedings{huang2017densely,
      title={Densely connected convolutional networks},
      author={Huang, Gao and Liu, Zhuang and van der Maaten, Laurens and Weinberger, Kilian Q },
      booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
      year={2017}
  }

The current version supports CIFAR-10 datasets. In order to be used for training a DENSENET model, the former need to be converted to TF-Records using the download_and_convert_data.py script (https://github.com/tensorflow/models/tree/master/slim).

The evaluation on Imagenet will be added soon.

The present TensorFlow implementation of Densenet models have the following performances:

The evaluation metrics should be reproducible by running the following command:

EVAL_DIR=./logs/
CHECKPOINT_PATH=./checkpoints/DENSENET_iter_.ckpt
python eval_ssd_network.py \
	--dataset_dir=${DATASET_DIR} \
	--checkpoint_path=${CHECKPOINT_FILE} \
	--dataset_name=cifar10 \
	--dataset_split_name=test \
	--model_name=densenet_40

The script train_densenet.py is in charged of training the network. Similarly to TF-Slim models, one can pass numerous options to the training process (dataset, optimiser, hyper-parameters, model, ...). In particular, it is possible to provide a checkpoint file which can be use as starting point in order to fine-tune a network.

For instance, one can train a DenseNet (L=40, k=12) on CIFAR-10 as following:

DATASET_DIR=./tfrecords
TRAIN_DIR=./logs/
python train_ssd_network.py \
     --train_dir=${TRAIN_DIR} \
    --dataset_name=cifar10 \
    --dataset_split_name=train \
    --preprocessing_name=cifarnet \
    --dataset_dir=${DATASET_DIR} \
    --model_name=densenet_40 \
    --save_summaries_secs=600 \
    --save_interval_secs=100 \
    --optimizer=adam \
    --learning_rate=0.1 \
    --batch_size=64 \
    --num_clones=4 \
    --num_classes=10 \
    --weight_decay=0.0001 \
    --log_every_n_steps=100 \
    --learning_rate_decay_type=densenetForCifar

Note that in addition to the training script flags, one may also want to experiment with data augmentation parameters (random cropping, resolution, ...) in preprocessing/cifarnet_preprocessing.py.

The num_clones parameter can be set to change the number of model clones to deploy. In our experiment environment (cudnn v5.1, CUDA 8.0, four K40m GPU), the code runs with speed 1.5iters/s when batch size is set to be 64. The hyperparameters are identical to the original [torch implementation] (https://github.com/liuzhuang13/DenseNet).

• Python 3
• TensorFlow >= 1.0

Please let me know if you have any questions!