Deep Isometric Learning for Visual Recognition

Authors: Haozhi Qi, Chong You, Xiaolong Wang, Yi Ma, Jitendra Malik

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform extensive experiments with so designed simple networks on image classification using the Image Net dataset (Deng et al., 2009; Russakovsky et al., 2015). Our results show that an ISONet with more than 100 layers is trainable and can achieve surprisingly competitive performance. We evaluate the performance of R-ISONet for object detection and instance segmentation on the COCO dataset (Lin et al., 2014). 3. Experiments
Researcher Affiliation Academia Haozhi Qi 1 Chong You 1 Xiaolong Wang 1 2 Yi Ma 1 Jitendra Malik 1 1UC Berkeley 2UC San Diego.
Pseudocode No The paper describes methods and architectures but does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/Haozhi Qi/ISONet.
Open Datasets Yes We test the performance of ISONets on the ILSVRC2012 image classification dataset (Deng et al., 2009; Russakovsky et al., 2015).
Dataset Splits Yes The training set contains 1.28 million images from 1,000 categories. Evaluation is performed on the validation set which contains 50,000 images.
Hardware Specification No The paper mentions 'The training is performed on 8 GPUs', but does not provide specific hardware details such as GPU models, CPU models, or memory specifications.
Software Dependencies No The paper does not provide specific software names with version numbers (e.g., PyTorch 1.9, Python 3.8, CUDA 11.1).
Experiment Setup Yes All models are trained using SGD with weight decay 0.0001, momentum 0.9 and mini-batch size 256. The initial learning rate is 0.1 for R-ISONet and 0.02 for ISONet. The models are trained for 100 epochs with the learning rate subsequently divided by a factor of 10 at the 30th, 60th and 90th epochs. To stabilize training in the early stage, we perform linear scheduled warmup (Goyal et al., 2017) for the first 5 epochs for both our methods and baseline methods. The orthogonal regularization coefficient γ in (13) (when used) is 0.0001 except for ISONet-101 where a stronger regularization (0.0003) is needed.