ensemble

 

输入进去的不是预测结果，而是神经网络去掉softmax的最后一层的向量或者张量值，之前都是用的投票法，或者均值法

metric leanring：是去掉softmax全连接，直接输入到ml，但是也可以用全连接的结果结合别的网络模型输入，你说的双子，都是在最后一层下手脚

focal loss, arcface loss，shpoere loss都是这样操作

因此，Metric Learning作为经典识别网络的替代方案，可以很好地适应某些特定的图像识别场景。一种较好的做法，是丢弃经典神经网络最后的softmax层，改成直接输出一根feature vector，去特征库里面按照Metric Learning寻找最近邻的类别作为匹配项。

鲸鱼kernel 总览：

0 bounding-box

https://www.kaggle.com/phhasian0710/create-bounding-box-images-whale-recognition

keras

1 https://www.kaggle.com/suicaokhoailang/resnet50-bounding-boxes-0-628-lb(https://www.kaggle.com/suicaokhoailang/generating-whale-bounding-boxes/)

https://www.kaggle.com/frkhit/triplet-loss-resnet50-bounding-box-0-621

tensorflow

fastai  

2 fastai 生成npy，再几个平均相加来ensemble

https://www.kaggle.com/lichen368/ensembling-with-averaged-probabilities-0-701-lb （Resnet50<就是第一个>，Resnet34）

https://www.kaggle.com/axel81/ensemble-resnet50-resnet101-lb-0-702

 

2.2 csv 的ensemble  average-precision-metric

https://www.kaggle.com/matthewa313/ensembling-algorithm-for-average-precision-metric

直接是csv的前5个，不用任何库

https://www.kaggle.com/jionie/esemble-lb-0-842

https://www.kaggle.com/axel81/siamese-ensemble-of-ensemble-lb-0-824

https://www.kaggle.com/vanshjatana/let-s-blend-0-867

https://www.kaggle.com/axel81/weighted-average-ensemble-v-n-lb-0-776

https://www.kaggle.com/alokevil/whale-ensemble-lb-0-866

3  siamese

https://www.kaggle.com/martinpiotte/whale-recognition-model-with-score-0-78563

keras 

https://www.kaggle.com/seesee/siamese-pretrained-0-822

keras。180 epoch，自己的bounding box，做法（https://www.kaggle.com/suicaokhoailang/generating-whale-bounding-boxes）

4Triplet Neural Network

https://www.kaggle.com/ashishpatel26/triplet-loss-network-for-humpback-whale-prediction

据说比 siamese 更好，keras，Triplet loss

https://www.kaggle.com/iafoss/similarity-densenet121-0-805lb-kernel-time-limit

kernel 的时间显著了， designed a loss function，metric learning ，DenseNet169  比Resnet50 好，使用裁剪后的方形图像

 

5 数据处理，简单的CNN

https://www.kaggle.com/pestipeti/keras-cnn-starter

keras

https://www.kaggle.com/satian/keras-mobilenet-starter/comments

keras 0.313  mobilenet 

https://www.kaggle.com/artgor/pytorch-whale-identifier

pytorch

https://www.kaggle.com/jhonatansilva31415/whales-a-simple-guide

pytorch，数据处理很多

https://www.kaggle.com/stalkermustang/pytorch-pretraiedmodels-se-resnext101-baseline

pytorch se_resnet101 0.402

https://www.kaggle.com/danlester/whales-transfer-learning-from-resnet18-in-pytorch

pytorch resnet18 0.292

https://www.kaggle.com/suicaokhoailang/wip-resnet18-baseline-with-fastai-0-392-lb

fastai  Resnet18 0.392

https://www.kaggle.com/asanakoev/easy-peasy-resnet50-with-fastai-0-574-lb

fastai  resnet50 0.574

https://www.kaggle.com/ateplyuk/resnext50-no-0-crop-sz384-0-651

fastai  resnet50 0.651

https://www.kaggle.com/axel81/denset-121-sgd-with-restarts-lb-0-602

fastai  denset121 0.602