GeNIe: Generative Hard Negative Images Through Diffusion
Authors: Soroush Abbasi Koohpayegani, Anuj Singh, Navaneet K L, Hamed Pirsiavash, Hadi J. Rad
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive experiments, in both few-shot and long-tail distribution settings, demonstrate the effectiveness of our novel augmentation method and its superior performance over the prior art. |
| Researcher Affiliation | Collaboration | 1University of California, Davis 2Delft University of Technology, The Netherlands 3Shell Global Solutions International B.V., Amsterdam, The Netherlands |
| Pseudocode | Yes | 3.1 Psuedocode of Ge NIe: As illustrated in Algorithm 2, we provide a detailed pytorch-style pseudocode for Ge NIe. |
| Open Source Code | Yes | Our code is available at: https://github.com/UCDvision/Ge NIe |
| Open Datasets | Yes | We conduct our few-shot experiments on two most commonly adopted few-shot classification datasets: mini-Imagenet (Ravi & Larochelle, 2017) and tiered-Imagenet (Ren et al., 2018). ... We perform experiments on Image Net-LT (Liu et al., 2019). ... We assess our method on several datasets: Food101 (Bossard et al., 2014) ... CUB200 (Wah et al., 2011) ... Cars196 (Krause et al., 2013) ... FGVC-Aircraft (Maji et al., 2013). |
| Dataset Splits | Yes | We follow the predominantly adopted settings of (Ravi & Larochelle, 2017; Chen et al., 2019a) where we split the entire dataset into 64 classes for training, 16 for validation and 20 for testing. ... We conduct our experiments on the two most commonly adopted settings: (5-way, 1-shot) and (5-way, 5-shot) classification settings. Following the literature, we sample 16-shots per class for the query set in both settings. ... Dataset: We perform experiments on Image Net-LT (Liu et al., 2019). ... Imagenet-LT classes can be divided into 3 groups: Few with less than 20 images, Med with 20 100 images, and Many with more than 100 images. |
| Hardware Specification | Yes | We use a single NVIDIA RTX 3090 for image generation. ... The runtime has been averaged over 10 different image-generations on an NVIDIA Tesla-V100 GPU with 16GB VRAM ... Training takes 2 hours on four NVIDIA RTX 3090 GPUs. ... Training takes 4 hours on 8 NVIDIA RTX 3090 GPUs. |
| Software Dependencies | No | We use Stable Diffusion 1.5 (Rombach et al., 2021a) as our base diffusion model. ... Algorithm 2: Py Torch-style Pseudocode of Ge NIe. ... pipe = Stable Diffusion Pipeline.from_pretrained("stable-diffusion-v1-5") scheduler = DPMSolver Multistep Scheduler.from_config(pipe.scheduler.config) |
| Experiment Setup | Yes | For Ge NIe-Ada in all scenarios, we utilize Ge NIe to generate augmentations from the noise level set {0.5, 0.6, 0.7, 0.8, 0.9}. ... We use r = 0.8 in our experiments. We generate 4 samples per class as augmentations in the 5-way, 1-shot setting and 20 samples per class as augmentations in the 5-way, 5-shot setting. ... NUM_INFERENCE_STEPS = 50 # Number of steps for reverse diffusion NUM_TRAIN_STEPS = 1000 # Number of steps for forward diffusion ... We optimize an embedding for 5000 iterations for each class in the dataset, followed by augmentation similar as the DAFusion method. ... We use the same hyperparameters as in (Xu et al., 2023) for finetuning: 100 epochs, lr = 0.008, batch size of 1024, Cut Mix and Mix Up for the data augmentation. |