Practical Integration via Separable Bijective Networks
Authors: Christopher M Bender, Patrick Emmanuel, Michael K. Reiter, Junier Oliva
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | All models were constructed using Py Torch (Paszke et al., 2019), trained using Py Torch Lightning (Falcon, 2019), utilized bijectors and distributions from Pyro (Bingham et al., 2018), and were trained using Adam (Kingma & Ba, 2015). We assess the integrable model s performance in a semi-supervised regime and against OOD examples. |
| Researcher Affiliation | Collaboration | Christopher M. Bender1,2, Patrick R. Emmanuel2, Michael K. Reiter3, Junier B. Oliva1 1Department of Computer Science, The University of North Carolina 2The Johns Hopkins University Applied Physics Laboratory 3Department of Computer Science, Duke University |
| Pseudocode | No | The paper describes methods through text and mathematical equations but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code utilized in this paper can be found at https://github.com/lupalab/sep_bij_nets. |
| Open Datasets | Yes | We test the impact of integrable models on OOD performance against several standard image classification datasets: MNIST (Deng, 2012), Fashion MNIST (FMNIST) (Xiao et al., 2017), SVHN (Netzer et al., 2011), and CIFAR10 (Krizhevsky et al., 2009). |
| Dataset Splits | No | Table 5 contains the validation standard accuracy and bits per dimension for all datasets with all integration regularizers. The paper refers to validation results and uses standard datasets, which often implies standard splits, but it does not explicitly state the specific percentages or counts for training, validation, and test sets used for these experiments. |
| Hardware Specification | No | The paper mentions software frameworks like PyTorch but does not provide specific hardware details such as GPU models, CPU types, or memory used for the experiments. |
| Software Dependencies | No | The paper mentions using PyTorch, PyTorch Lightning, and Pyro, but it does not specify concrete version numbers for these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | The learning rate is set to 0.001 with standard weight decay of 1e-4 over 20 epochs using Adam (Kingma & Ba, 2015). We utilize a batch size of 256, a learning rate of 1.5e-4 over the bijective layers, and a learning rate of 1.0e-3 over the separable layers using Adam (Kingma & Ba, 2015). Both learning rates are exponentially decayed at a rate of 0.75 per epoch. Standard weight decay is applied with a weight of 1e-4. The network is trained for 50 epochs. |