久久久久久久久中文,蜜桃精品视频首页,婷婷无码久久久

四月了~2021的三分之一好快！加油哈！
大概接下來一個月保持兩更或一更了吧
今天記錄的還是針對跨域的論文閱讀

論文名稱：
《Cross-Domain Few-Shot Learning with Meta Fine-Tuning》
論文地址：https://arxiv.org/abs/2005.10544v1
論文代碼： https://github.com/johncai117/Meta-Fine-Tuning
本篇文章只記錄個人閱讀論文的筆記，具體翻譯、代碼等不展開，詳細(xì)可見上述的鏈接.

Background

1.Unfortunately, acquiring a large training data-set is costly due to the need for human annotation. Furthermore, when dealing with rare examples in medical images (e.g. rare diseases) or satellite images (e.g. oil spills), the ability to obtain labelled samples is limited.
（由于需要人工標(biāo)注，獲取大量的訓(xùn)練數(shù)據(jù)集較為昂貴。除此，在醫(yī)學(xué)圖像或者衛(wèi)星圖像這種較少樣本的數(shù)據(jù)集中，獲取帶有標(biāo)簽的樣本是有限的。）
----產(chǎn)生小樣本學(xué)習(xí)的原因（few-shot learning）
2.However, existing few-shot learning methods have been developed with the assumption that the training and test data-set arise from the same distribution. Domain shift would thus be an additional problem as it may prevent the robust transfer of features.
（指出小樣本學(xué)習(xí)中存在的domain shift即域遷移的問題，因為該問題的存在影響了小樣本學(xué)習(xí)的魯棒性）
3.the CVPR 2020 challenge has introduced a new benchmark that aims to test for generalization ability across a range of vastly different domains, with domains from nat?ural and medical images, domains without perspective, and domains without color
（CVPR2020引入了一個針對小樣本域適應(yīng)的數(shù)據(jù)集，詳細(xì)可參考之前的論文閱讀：http://www.itdecent.cn/p/e6dc55021885）
綜上，將該數(shù)據(jù)集作為域適應(yīng)的基準(zhǔn)數(shù)據(jù)集進(jìn)行探索。

Work

Main contribution:
1.Integration of fine-tuning into the episodic training process by exploiting a first-order MAML-based meta-learning algorithm
(通過利用基于一階MAML的元學(xué)習(xí)算法將微調(diào)集成到episodic訓(xùn)練過程中（以下稱為“元微調(diào)”）。
這樣做是為了使網(wǎng)絡(luò)學(xué)習(xí)一組初始權(quán)重，這些初始權(quán)重很容易在測試域的支持集上進(jìn)行微調(diào)。)
2.Integrates the Meta FineTuning algorithm into a Graph Neural Network that exploits the non-Euclidean structure of the relation between the support set and the query samples.
(將Meta FineTuning算法集成到圖神經(jīng)網(wǎng)絡(luò)中，該網(wǎng)絡(luò)利用支持集和查詢樣本之間關(guān)系的非歐幾里得結(jié)構(gòu)。)
3.implement data augmentation on the support set during fine-tuning, and achieve a further improvement in accuracy
(在微調(diào)期間對支持集執(zhí)行數(shù)據(jù)擴充，并進(jìn)一步提高準(zhǔn)確性)
4.combine the above method with a modified fine-tuning baseline method, and combine them into an ensemble to jointly make predictions.
(將上述方法與改進(jìn)的微調(diào)基線方法結(jié)合起來，然后將它們組合成一個整體以共同進(jìn)行預(yù)測)

Methodology

1.Graph Neural Networks

元學(xué)習(xí)模塊采用的是圖神經(jīng)網(wǎng)絡(luò)，在之前寫過的論文閱讀中《 Cross-domain few-shot classification
via learned feature-wise transformation》即出現(xiàn)過可參考（http://www.itdecent.cn/p/353cb4926278）
簡單來說，首先，使用線性層將維數(shù)為F的特征向量投影到較低維的空間dk上。然后圖卷積接收Signal S。
然后，使用線性運算對局部Signal執(zhí)行圖形卷積層GC(）。這將產(chǎn)生一個輸出 $x^{k + 1}$ = $GC(X^{k})$
為了學(xué)習(xí)邊特征，MLP會獲取圖中頂點的輸出向量之間的絕對差值
[用節(jié)點來存儲更新權(quán)重的,每一個節(jié)點都代表一幅輸入的圖像，而每個邊上的權(quán)重就表示兩幅圖之間的關(guān)系]

2.Meta Fine-Tuning
元微調(diào)的核心思想是，我們可以使用元學(xué)習(xí)來優(yōu)化的預(yù)訓(xùn)練模型，而不是找到一組旨在微調(diào)的權(quán)重初始化
To this end,we apply and adapt the first-order MAML algorithm and simulate the episodic training process. A first-order MAML algorithm can achieve comparable results with the second order algorithm at a lower computational cost
為此，我們應(yīng)用并修改了一階MAML算法，并模擬了訓(xùn)練過程。一階MAML算法可以較低的計算成本獲得與二階算法相當(dāng)?shù)慕Y(jié)果
該方法還可以應(yīng)用于任何主干深度的模型，并且可以凍結(jié)到任意數(shù)量的層。對于本文，我們凍結(jié)了最后一個網(wǎng)網(wǎng)塊

During step 1 (Meta Fine-Tuning), only support examples are used, and the first 8-layers are frozen. A linear classifier on the ResNet10 features is used to predict the support labels, and the last 2-layers are updated accordingly using CE Loss for 5 epochs.
在步驟1（Meta Fine-Tuning）中，僅使用支持示例，并且凍結(jié)了前8層。使用ResNet10功能上的線性分類器來預(yù)測支持標(biāo)簽，并使用5個CE loss相應(yīng)地更新最后2層。
At step 2, all layers are updated using the episodic training loss. At prediction stage on the test
domain, all layers in the ResNet10 are frozen in step 2
在步驟2，使用episodic 訓(xùn)練損失來更新所有層。在測試域的預(yù)測階段，在步驟2中凍結(jié)ResNet10中的所有層

3. Data Augmentation

對于訓(xùn)練期間的數(shù)據(jù)增強，我們使用代碼庫中的默認(rèn)參數(shù)。對于測試期間的數(shù)據(jù)增強，我們從支持圖像（我們知道的標(biāo)簽）中采樣17個附加圖像，并隨機執(zhí)行抖動、隨機作物和水平翻轉(zhuǎn)（如果適用）。在微調(diào)過程中，我們通過將模型更頻繁地暴露于原始圖像來對原始圖像進(jìn)行加權(quán)。在最后的預(yù)測階段，支持圖像和查詢圖像都只使用基本圖像（即中心作物）。

4.Combining Scores in the Ensemble
For our final submission results, we combine the predic?tions from the modified baseline fine-tuning model and the meta fine-tuning GNN model by normalizing the scores using a softmax function so that the scores from each model sum to 1 and are between 0 and 1, which ensures that each model is given equal weight in the prediction. Then we add them together and take argmax
通過使用softmax函數(shù)對分?jǐn)?shù)進(jìn)行歸一化，將修改后的基線微調(diào)模型和元微調(diào)GNN模型的預(yù)測結(jié)合起來，以使每個模型的分?jǐn)?shù)總和為1，且介于0之間和1，可確保在預(yù)測中為每個模型賦予相等的權(quán)重。然后我們將它們加在一起并取argmax
流程圖：

通過將一階MAML算法擴展到元微調(diào)，并將其與GNN、數(shù)據(jù)增強和集成方法相結(jié)合來實現(xiàn)的。
（有點集成大法的內(nèi)卷了吧）