/NTIRE2023-ESR-MFA

We placed 2nd in the Parameters and FLOPs sub-track of the NTIRE2023 ESR Challenge

Primary LanguagePythonMIT LicenseMIT

NTIRE 2023 Challenge on Efficient Super-Resolution @ CVPR 2023

📖 MFA: A Lightweight Multi-scale Feature Attention for Image Super-Resolution

We placed 2nd in the Parameters and FLOPs sub-track of the NTIRE2023 ESR Challenge.

An overview of the proposed MFA model

How to test the model?

  1. Run the [run.sh] script (./run.sh) 
    CUDA_VISIBLE_DEVICES=0 python test_demo.py --data_dir [path to your data dir] --save_dir [path to your save dir] --model_id 16
    • Be sure the change the directories --data_dir and --save_dir.

How to calculate the number of parameters, FLOPs, and activations

    from utils.model_summary import get_model_flops, get_model_activation
    from models.team16_MFA import MFA
    model = MFA(dim=33, n_blocks=8, ffn_scale=1.8, upscaling_factor=4)
    
    input_dim = (3, 256, 256)  # set the input dimension
    activations, num_conv = get_model_activation(model, input_dim)
    activations = activations / 10 ** 6
    print("{:>16s} : {:<.4f} [M]".format("#Activations", activations))
    print("{:>16s} : {:<d}".format("#Conv2d", num_conv))

    flops = get_model_flops(model, input_dim, False)
    flops = flops / 10 ** 9
    print("{:>16s} : {:<.4f} [G]".format("FLOPs", flops))

    num_parameters = sum(map(lambda x: x.numel(), model.parameters()))
    num_parameters = num_parameters / 10 ** 6
    print("{:>16s} : {:<.4f} [M]".format("#Params", num_parameters))

Results of NTIRE 2023 Efficient SR Challenge

Our team (NJUST_E) placed 2nd in the Parameters and FLOPs sub-track of the NTIRE 2023 ESR Challenge.