• Fully utilize your cuda cores!
• Unlike CNN models using a lot of Conv1x1 to reduce model size and number of MACs, HarDNet mainly uses Conv3x3 (with only one Conv1x1 layer for each HarDNet block) to increase the computational density.
• Increased computational density changes a model from Memory-Bound to Compute-Bound

• k = growth rate (as in DenseNet)
• m = channel weighting factor (1.6~1.7)
• Conv3x3 for all layers (no bottleneck layer)
• Conv-BN-ReLU for all layers intead of BN-ReLU-Conv used in DenseNet
• See MIPT-Oulu/pytorch_bn_fusion to get rid of BatchNorm for inference.
• No global dense connection (input of a HarDBlk is NOT reused as a part of output)

• Enhanced local feature extraction to benefit the detection of small objects
• A transitional Conv1x1 layer is employed after each HarDNet block (HarDBlk)

* Inference time measured on an NVidia 1080ti with pytorch 1.1.0
300 iteraions of random 1024x1024 input images are averaged.

** Inference time measured on an NVidia Jetson nano with TensorRT
500 iteraions of random 320x320 input images are averaged.

Training prodedure is branched from https://github.com/pytorch/examples/tree/master/imagenet

Training:

python main.py -a hardnet68 [imagenet-folder with train and val folders]

arch = hardnet39ds | hardnet68ds | hardnet68 | hardnet85

Evaluating:

python main.py -a hardnet68 --pretrained -e [imagenet-folder with train and val folders]

for HarDNet85, please download pretrained weights from here

• epochs 150 ~ 250
• initial lr = 0.05
• batch size = 256
• weight decay = 6e-5
• cosine learning rate decay
• nestrov = True