Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
TFPnP: Tuning-free Plug-and-Play Proximal Algorithms with Applications to Inverse Imaging Problems
Authors: Kaixuan Wei, Angelica Aviles-Rivero, Jingwei Liang, Ying Fu, Hua Huang, Carola-Bibiane Schönlieb
JMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate, through rigorous numerical and visual experiments, that the learned policy can customize parameters to different settings, and is often more efficient and effective than existing handcrafted criteria. |
| Researcher Affiliation | Academia | Kaixuan Wei EMAIL School of Computer Science and Technology Beijing Institute of Technology, Beijing, China Angelica Aviles-Rivero EMAIL Department of Pure Mathematics and Mathematical Statistics University of Cambridge, Cambridge, United Kingdom Jingwei Liang EMAIL Institute of Natural Sciences and School of Mathematical Sciences Shanghai Jiao Tong University, Shanghai, China Ying Fu EMAIL School of Computer Science and Technology Beijing Institute of Technology, Beijing, China Hua Huang EMAIL School of Artificial Intelligence Beijing Normal University, Beijing, China Carola-Bibiane Sch onlieb EMAIL Department of Applied Mathematics and Theoretical Physics University of Cambridge, Cambridge, United Kingdom |
| Pseudocode | Yes | Algorithm 1 Training Scheme |
| Open Source Code | Yes | 1. Our code and pretrained models are made publicly available at https://github.com/Vandermode/TFPn P |
| Open Datasets | Yes | To train the denoising network, we follow the common practice that uses 87,000 overlapping patches (with size 128 x 128) drawn from 400 images from the BSD dataset (Martin et al., 2001). For each patch, we add white Gaussian noise with noise level sampled from [1, 50]. The denoising networks are trained with 50 epochs using L1 loss and Adam optimizer (Kingma and Ba, 2014) with batch size 32. The base learning rate is set to 10^-4 and halved at epoch 30, then reduced to 10^-5 at epoch 40. To train the policy network (and an auxiliary value network only used during training), we use the 17,125 resized images with size 128 x 128 from the PASCAL VOC dataset (Everingham et al., 2014). Table 4: Quantitative results (PSNR) of different CS-MRI methods under various acceleration factors f and noise levels σn on two datasets. The best results are indicated in orange color and the second best results are denoted in blue color. 9. https://my.vanderbilt.edu/masi/ 50 medical images from MICCAI 2013 grand challenge dataset9. Our TFPn P algorithm yields considerably better results than state-of-the-arts (e.g., it outperforms the second-best method RPGD 0.4d B in terms of PSNR under 7.5% Gaussian noise setting), suggesting the effectiveness of our algorithm. The numerical and visual results are evaluated on 50 randomly selected lung CT images from the COVID-19 CT lung and infection segmentation dataset (Ma et al., 2020), shown in Table 5 and Figure 4 respectively. |
| Dataset Splits | No | The paper mentions using specific datasets for training and evaluation, e.g., "87,000 overlapping patches [...] from the BSD dataset" and "17,125 resized images [...] from the PASCAL VOC dataset" for training, and evaluating on "seven widely used medical images" and "50 medical images from MICCAI 2013 grand challenge dataset". However, it does not provide explicit training, validation, or test splits (e.g., percentages, exact counts, or specific predefined splits) for these datasets in the context of the experiments conducted in the paper. |
| Hardware Specification | No | The paper mentions "GPU runtime (millisecond)" in Table 2, implying the use of GPUs, but it does not specify any particular GPU models (e.g., NVIDIA A100, RTX 3090, Tesla V100), CPU models, or other specific hardware configurations used for running the experiments. |
| Software Dependencies | No | The paper mentions "Adam optimizer (Kingma and Ba, 2014)" and "Torch Radon library (Ronchetti, 2020)" which is integrated into "PyTorch". While these software components are named, specific version numbers for PyTorch, Adam (or its implementation), or Torch Radon are not provided, preventing a fully reproducible software environment description. |
| Experiment Setup | Yes | To train the denoising network, we follow the common practice that uses 87,000 overlapping patches (with size 128 x 128) drawn from 400 images from the BSD dataset (Martin et al., 2001). For each patch, we add white Gaussian noise with noise level sampled from [1, 50]. The denoising networks are trained with 50 epochs using L1 loss and Adam optimizer (Kingma and Ba, 2014) with batch size 32. The base learning rate is set to 10^-4 and halved at epoch 30, then reduced to 10^-5 at epoch 40. To train the policy network (and an auxiliary value network only used during training), we use the 17,125 resized images with size 128 x 128 from the PASCAL VOC dataset (Everingham et al., 2014). We use Adam optimizer with batch size 48 and 2500 training iterations. We start by setting the learning rates lθ, lφ for updating the policy network πθ and the value network V π φ to 1 x 10^-4 and 5 x 10^-5 respectively. Then we reduced these values to 5 x 10^-5 and 1 x 10^-5 respectively at training iteration 1600. The value network learning makes use of a target value network, which is a soft copy of the value network itself. The weight parameter β for softly updating the target value network is set to 10^-3 (see line 15 in Algorithm 1). In each training iteration, we alternate between collecting states (in a state buffer) from the environment with the current policy and updating the network parameters using policy gradients from batches sampled from the state buffer B. Ten gradient steps are performed at each training iteration. 4. The choice of the hyperparameters m, N and η is discussed in Section 5.1. Its value is set to 5 empirically. To avoid the optimization loops infinitely, the maximum time step N is set to 6 in our algorithm, leading to 30 iterations of the optimization at most. We set η = 0.05 in our algorithm4. |