flava-full

模型介绍

FLAVA model was developed by the researchers at FAIR to understand if a single model can work across different modalities with a unified architecture. The model was pretrained solely using publicly available multimodal datasets containing 70M image-text pairs in total and thus fully reproducible. Unimodal datasets ImageNet and BookCorpus + CCNews were also used to provide unimodal data to the model. The model (i) similar to CLIP can be used for arbitrary image classification tasks in a zero-shot manner (ii) used for image or text retrieval in a zero-shot manner (iii) can also be fine-tuned for natural language understanding (NLU) tasks such as GLUE and vision-and-language reasoning tasks such as VQA v2. The model is able to use the data available as images, text corpus and image-text pairs. In the original paper, the authors evaluate FLAVA on 32 tasks from computer vision, NLU and vision-and-language domains and show impressive performance across the board scoring higher micro-average than CLIP while being open. The FLAVA model uses a ViT-B/32 transformer for both image encoder and text encoder. FLAVA also employs a multimodal encoder on top for multimodal tasks such as vision-and-language tasks (VQA) which is a 6-layer encoder. Each component of FLAVA model can be loaded individually from `facebook/flava-full` checkpoint. If you need complete heads used for pretraining, please use `FlavaForPreTraining` model class otherwise `FlavaModel` should suffice for most use case. This repository also contains code to pretrain the FLAVA model from scratch. FLAVA model supports vision, language and multimodal inputs. You can pass inputs corresponding to the domain you are concerned with to get losses and outputs related to that domain.

主要特点

特点摘要待依据来源补齐，不凭模型名称推断能力。

局限与注意事项

暂无足够来源支持具体缺点结论。请重点验证实际任务表现、时延、数据政策、许可与版本生命周期。

来源报价与费用分析

同一模型在不同渠道、地区和配置下可能采用不同价格。下表独立保留各渠道报价，不将渠道价格视为厂商官方统一价格。

公开评测与运行配置

暂无已采集的可核验公开评分。模型未上榜不代表能力较低。

常见问题

资料来源与可追溯记录

许可证：bsd-3-clause。模型介绍与第三方评测均以原来源的具体版本和方法为准。