Fixing the train-test resolution discrepancy
Authors: Hugo Touvron, Andrea Vedaldi, Matthijs Douze, Herve Jegou
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet50 trained on 128x128 images, and 79.8% with one trained at 224x224. A ResNeXt-101 32x48d pre-trained with weak supervision on 940 million 224x224 images and further optimized with our technique for test resolution 320x320 achieves 86.4% top-1 accuracy (top-5: 98.0%). |
| Researcher Affiliation | Industry | Hugo Touvron, Andrea Vedaldi, Matthijs Douze, Herv´e J´egou Facebook AI Research |
| Pseudocode | No | The paper describes the methods in text and does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | An open-source implementation of our method is available at https://github.com/facebookresearch/FixRes. |
| Open Datasets | Yes | We experiment on the ImageNet-2012 benchmark [29] |
| Dataset Splits | Yes | We experiment on the ImageNet-2012 benchmark [29], reporting validation performance as top-1 accuracy. To assess the significance of our results, we compute the standard deviation of the top-1 accuracy: we classify the validation images, split the set into 10 folds and measure the accuracy on 9 of them, leaving one out in turn. |
| Hardware Specification | Yes | We ran our training experiments on machines with 8 Tesla V100 GPUs and 80 CPU cores. |
| Software Dependencies | No | The paper mentions 'PyTorch library' and 'PyTorch hub repository' but does not specify version numbers for PyTorch or other software dependencies. |
| Experiment Setup | Yes | We train ResNet-50 with SGD with a learning rate of 0.1 B/256, where B is the batch size... The learning rate is divided by 10 every 30 epochs. With a Repeated Augmentation of 3, an epoch processes 5005 * 512/B batches... In the initial training, we use B = 512, 120 epochs and the default PyTorch data augmentation: horizontal flip, random resized crop (as in section 3) and color jittering. To finetune, the initial learning rate is 0.008 same decay, B = 512, 60 epochs. |