LoCo: Local Contrastive Representation Learning
Authors: Yuwen Xiong, Mengye Ren, Raquel Urtasun
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Aside from standard Image Net experiments, we also show results on complex downstream tasks such as object detection and instance segmentation directly using readout features. In this section, we conduct experiments to test the hypotheses we made in Section 4 and verify our design choices. |
| Researcher Affiliation | Collaboration | Yuwen Xiong Uber ATG University of Toronto yuwen@cs.toronto.edu Mengye Ren Uber ATG University of Toronto mren@cs.toronto.edu Raquel Urtasun Uber ATG University of Toronto urtasun@cs.toronto.edu |
| Pseudocode | No | The paper does not include any sections or figures explicitly labeled 'Pseudocode' or 'Algorithm', nor does it present structured steps in a code-like format. |
| Open Source Code | No | The paper mentions using 'mmdetection [10]' and refers to 'Py Torch' for supervised pre-training weights, but it does not provide a specific link or explicit statement that the code for the described methodology (Lo Co) is open-source or publicly available. |
| Open Datasets | Yes | Aside from standard Image Net experiments, we also show results on complex downstream tasks such as object detection and instance segmentation directly using readout features. We use Mask R-CNN [24] on Cityscapes [15] and COCO [37] to evaluate object detection and instance segmentation performance. |
| Dataset Splits | Yes | Following [23], we randomly sample 10k and 1k COCO images for training, namely COCO-10K and COCO-1K. These are 10% and 1% of the full COCO train2017 set. We report AP on the official val2017 set. |
| Hardware Specification | No | The paper discusses 'GPU memory' and 'memory saving ratio' but does not specify the exact hardware components (e.g., GPU models, CPU models, or memory amounts) used for conducting the experiments. |
| Software Dependencies | No | The paper mentions using 'LARS optimizer', 'SGD', 'Mask R-CNN', and 'mmdetection [10]', along with 'Py Torch' for pretrained weights, but it does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | Unless otherwise specified, we train with a batch size of 4096 using the LARS optimizer [58]. We train models 800 epochs to show that Lo Co can perform well on very long training schedules and match state-of-the-art performance; we use a learning rate of 4.8 with a cosine decay schedule without restart [38]; linear warm-up is used for the first 10 epochs. SGD without momentum is used as the optimizer for 100 training epochs with a learning rate of 30 and decayed by a factor of 10 at epoch 30, 60 and 90, the same procedure done in [22]. We train models for 60k iterations (96 epochs) on COCO-10K and 15k iterations (240 epochs) on COCO-1K with a batch size of 16. |