wangchen1801

哈喽同学，模型包括query和support两个分支，整个模型是可以端到端直接训练的。support分支中，模型会为每个类别图像(support image)都生成一个prototype，它们用于计算meta loss(分类任务，交叉熵损失)。不同类别的prototypes并不会融合而成为一个。最后，这些prototypes与query分支融合，用于辅助目标检测任务。

FFA模块则是新加的一个模块，它可以额外地提取"fine-grained prototypes"，融合到query分支当中，这部分和meta loss没有关系。

支持集图片和查询集图片是同时放入模型的，两条分支平行地工作。按照我的理解，这里的元学习方法与一般深度学习方法，区别主要在模型结构上，并没有多个任务之说。如果需要了解具体的训练流程，可能得从代码中找一下了～

Hello, please use the same configuration as in this repository to reproduce the result, otherwise it may not achieve the reported accuracy. And the learning rate should also following the...

Hello, sorry for the late reply, I might not be able to respond promptly due to due to being busy with work. However, this issue did not occur in my...

Thanks! Feel free to ask if you still have any questions.

Hello, you might be referring to the inference.py in the [mmfewshot](https://github.com/open-mmlab/mmfewshot/blob/main/mmfewshot/detection/apis/inference.py). I haven't used this file before and can't provide instructions at the moment. However, I would recommend checking out...

It refers to different fewshot samples. See [issue#5](https://github.com/wangchen1801/FPD/issues/5) for more information.

Thank you for your interest in this work! Given a feature map of size (B,C,H,W), we can get the heatmap with following script: ```python import cv2 import torch.nn.functional as F...

We only need to fine-tuning the model on the new fewshot samples, i.e. perform fine-tuning train and testing.