Bespoke Solvers for Generative Flow Models

Authors: Neta Shaul, Juan Perez, Ricky T. Q. Chen, Ali Thabet, Albert Pumarola, Yaron Lipman

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For example, a Bespoke solver for a CIFAR10 model produces samples with Fréchet Inception Distance (FID) of 2.73 with 10 NFE, and gets to 1% of the Ground Truth (GT) FID (2.59) for this model with only 20 NFE.
Researcher Affiliation Collaboration N. Shaul1 J. Perez2 R. T. Q. Chen3 A. Thabet2 A. Pumarola2 Y. Lipman3,1 1Weizmann Institute of Science 2Gen AI, Meta 3FAIR, Meta
Pseudocode Yes Algorithm 1 Numerical ODE solver.
Open Source Code No No explicit statement or link to the open-source code for the methodology described in this paper is provided. The paper cites a third-party tool's GitHub, but not its own implementation.
Open Datasets Yes Our method works with pre-trained models: we use the pre-trained CIFAR10 (Krizhevsky & Hinton, 2009) model of (Song et al., 2020b) with published weights from EDM (Karras et al., 2022). Additionally, we trained diffusion/flow models on the datasets: CIFAR10, AFHQ-256 (Choi et al., 2020a) and Image Net-64/128 (Deng et al., 2009).
Dataset Splits Yes We compute FID (Heusel et al., 2017) and validation RMSE (equation 6) is computed on a set of 10K fresh noise samples x0 p(x0); Figure 12 depicts an example of RMSE vs. training iterations for different n values. Unless otherwise stated, below we report results on best FID iteration and show samples on best RMSE validation iteration.
Hardware Specification No Table 5: Pre-trained models hyper-parameters. GPUs 8 8 64 64 64. Specific GPU models or processor types are not mentioned, only the number of GPUs used.
Software Dependencies No The paper mentions "Adam optimizer Kingma & Ba (2017)", "DOPRI5 method (Shampine, 1986)", and refers to "Chen, 2018" likely for `torchdiffeq`. However, no specific version numbers for these or other software components are provided.
Experiment Setup Yes Table 3: Hyper-parameters of Bespoke solvers training on CIFAR10/Image Net-64/Image Net-128/AFHQ 256. Total number of trajectories 72k 48k 48k 4k Batch size 12 8 8 1 Number of iterations 6k 6k 6k 4k