Meta-Reinforced Synthetic Data for One-Shot Fine-Grained Visual Recognition

Authors: Satoshi Tsutsui, Yanwei Fu, David Crandall

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The model is trained in an end-to-end manner, and our experiments demonstrate consistent improvement over baselines on one-shot fine-grained image classification benchmarks.
Researcher Affiliation Academia Satoshi Tsutsui Indiana University USA stsutsui@indiana.edu Yanwei Fu Fudan University China yanweifu@fudan.edu.cn David Crandall Indiana University USA djcran@indiana.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Further implementation details are available as supplemental source code.2 http://vision.soic.indiana.edu/metairnet/
Open Datasets Yes We use the fine-grained classification dataset of Caltech UCSD Birds (CUB) [32] for our main experiments, and another fine-grained dataset of North American Birds (NAB) [30] for secondary experiments.
Dataset Splits Yes For CUB, we use the same train/val/test split used in previous work [4], and for NAB we randomly split with a proportion of train:val:test = 2:1:1; see supplementary material for details.
Hardware Specification No The paper mentions 'an NVidia Titan Xp GPU' for a specific generation step during a pilot study (Section 3), but does not specify the hardware used for the main Meta IRNet experiments described in Section 5.1.
Software Dependencies No The paper mentions using 'Adam' optimizer and 'Res Net18' for image classification, but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow, CUDA versions).
Experiment Setup Yes We set λp = 0.1 and λz = 0.1, and perform 500 gradient descent updates with the Adam [18] optimizer with learning rate 0.01 for z and 0.0005 for the fully connected layers, to produce scale and shift parameters of the batch normalization layers. We train F and C with Adam with a default learning rate of 0.001.