Switched Flow Matching: Eliminating Singularities via Switching ODEs
Authors: Qunxi Zhu, Wei Lin
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of the newly proposed SFM through several numerical examples. We validate the effectiveness of the newly proposed SFM through extensive experiments on both synthetic and real-world datasets, achieving competitive or even better performance compared to existing methods, such as FM. |
| Researcher Affiliation | Academia | Qunxi Zhu 1 Wei Lin 1 2 3 4 1Research Institute of Intelligent Complex Systems, Fudan University, China. 2School of Mathematical Sciences, LMNS, and SCMS, Fudan University, China. 3State Key Laboratory of Medical Neurobiology and MOE Frontiers Center for Brain Science, Institutes of Brain Science, Fudan University, China. 4Shanghai Artificial Intelligence Laboratory, China. |
| Pseudocode | Yes | For a better understanding of the proposed SFM, we provide the main pseudocode as shown in Algorithm 1 as well as the pseudocode for constructing the switching coupling as shown in Algorithm 2 with the optional module for OT-SFM (see Algorithm 3). For inference, the pseudocode is displayed in Algorithm 4. |
| Open Source Code | Yes | The code is available at https://github.com/zhuqunxi/switched-flow-matching. |
| Open Datasets | Yes | CIFAR-10 dataset. Table 2 shows the image generation results of our SFM variants on the CIFAR-10 dataset. Figure 13. True samples from the source distribution (left, Gaussian mixture) and the target distribution (right, CIFAR-10 dataset). |
| Dataset Splits | No | The paper mentions using the CIFAR-10 dataset, which has standard splits, but does not explicitly state how the training, validation, and test splits were performed or used in their experiments, nor does it provide percentages or sample counts for these splits. |
| Hardware Specification | Yes | All our experiments were conducted on a single 11GB GTX 1080 Ti GPU. |
| Software Dependencies | No | The paper mentions using `torchdiffeq` package and `Tensor Flow-GAN library` but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | Adam optimizer with β1 = 0.9, β2 = 0.999, ϵ = 10 8, and no weight decay, channels = 128, depth = 2 , channels multiple = [1, 2, 2, 2], heads channels = 64, attention resolution = 16, dropout = 0.1, batch size per gpu = 128, gpus = 1, learning rate = 2 10 4, gradient clipping with norm = 1.0, exponential moving average weights with decay = 0.9999. For the synthetic experiments, we provide the detailed setup for different datasets (see Table 8). |