圖像分類：來自Kaggle中13個(gè)項(xiàng)目的Tips和Tricks

這些比賽是：

• Intel Image Classification：
  https://www.kaggle.com/puneet6060/intel-image-classification
• Recursion Cellular Image Classification：
  https://www.kaggle.com/c/recursion-cellular-image-classification
• SIIM-ISIC Melanoma Classification：
  https://www.kaggle.com/c/siim-isic-melanoma-classification
• APTOS 2019 Blindness Detection：
  https://www.kaggle.com/c/aptos2019-blindness-detection/notebooks
• Diabetic Retinopathy Detection：
  https://www.kaggle.com/c/diabetic-retinopathy-detection
• ML Project?—?Image Classification：
  https://www.kaggle.com/c/image-classification-fashion-mnist/notebooks
• Cdiscount’s Image Classification Challenge：

  https://www.kaggle.com/c/cdiscount-image-classification-challenge/notebooks

• Plant seedlings classifications：

  https://www.kaggle.com/c/plant-seedlings-classification/notebooks

• Aesthetic Visual Analysis：

  https://www.kaggle.com/c/aesthetic-visual-analysis/notebooks

我們會(huì)討論調(diào)試深度學(xué)習(xí)解決方案的三個(gè)主要方面：

• 數(shù)據(jù)

• 模型

• 損失函數(shù)

還有很多例子項(xiàng)目(和參考資料)供你參考。

數(shù)據(jù)

圖像預(yù)處理 + EDA

• Visualisation：

  https://www.kaggle.com/allunia/protein-atlas-exploration-and-baseline#Building-a-baseline-model-

• Dealing with Class imbalance：
  https://www.kaggle.com/rohandeysarkar/ultimate-image-classification-guide-2020
• Fill missing values (labels, features and, etc.)：
  https://www.kaggle.com/datafan07/analysis-of-melanoma-metadata-and-effnet-ensemble
• Normalisation?：
  https://www.kaggle.com/vincee/intel-image-classification-cnn-keras
• Pre-processing：

  https://www.kaggle.com/ratthachat/aptos-eye-preprocessing-in-diabetic-retinopathy#3.A-Important-Update-on-Color-Version-of-Cropping-&-Ben's-Preprocessing

數(shù)據(jù)增強(qiáng)

• Horizontal Flip：

  https://www.kaggle.com/datafan07/analysis-of-melanoma-metadata-and-effnet-ensemble

• Random Rotate and Random Dihedral：
  https://www.kaggle.com/iafoss/pretrained-resnet34-with-rgby-0-460-public-lb
• Hue, Saturation, Contrast, Brightness, Crop：
  https://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords
• Colour jitter：

  https://www.kaggle.com/nroman/melanoma-pytorch-starter-efficientnet

模型

開發(fā)一個(gè)基線

用Jeremy Howard的名言：“你應(yīng)該能夠在15分鐘內(nèi)使用50%或更少的數(shù)據(jù)集快速測(cè)試你是否正在朝著一個(gè)有希望的方向前進(jìn)，如果沒有，你必須重新考慮一切?！?span style="color: rgb(51, 51, 51);font-size: 17px;">

開發(fā)一個(gè)足夠大可以過擬合的模型

• 添加更多層

• 使用更好的結(jié)構(gòu)

• 更完善的訓(xùn)練流程

結(jié)構(gòu)

• Residual Networks

• Wide Residual Networks

• Inception

• EfficientNet

• Swish activation

• Residual Attention Network

訓(xùn)練過程

• Mixed-Precision Training

• Large Batch-Size Training

• Cross-Validation Set

• Weight Initialization

• Self-Supervised Training (Knowledge Distillation)

• Learning Rate Scheduler

• Learning Rate Warmup

• Early Stopping

• Differential Learning Rates

• Ensemble

• Transfer Learning

• Fine-Tuning

超參數(shù)調(diào)試

正則化

• Adding Dropout：
  https://www.kaggle.com/allunia/protein-atlas-exploration-and-baseline
• Adding or changing the position of Batch Norm：
  https://www.kaggle.com/allunia/protein-atlas-exploration-and-baseline
• Data augmentation：
  https://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords
• Mixup：
  https://arxiv.org/abs/1710.09412
• Weight regularization：
  https://www.kaggle.com/allunia/protein-atlas-exploration-and-baseline
• Gradient clipping：
  https://www.kaggle.com/allunia/protein-atlas-exploration-and-baseline

損失函數(shù)

這里是一些最流行的損失函數(shù)，與項(xiàng)目實(shí)例，你會(huì)發(fā)現(xiàn)一些技巧，以提高你的模型的能力：

• Label smoothing
• Focal loss
• SparseMax loss and Weighted cross-entropy
• BCE loss, BCE with logits loss and Categorical cross-entropy loss
• Additive Angular Margin Loss for Deep Face Recognition

評(píng)估 + 錯(cuò)誤分析

在這里，我們做消融研究，并分析我們的實(shí)驗(yàn)結(jié)果。我們確定了我們的模型的弱點(diǎn)和長(zhǎng)處，并確定了未來需要改進(jìn)的地方。在這個(gè)階段，你可以使用以下技術(shù)，并在鏈接的示例中查看它們是如何實(shí)現(xiàn)的：

• Tracking metrics and Confusion matrix：

  https://www.kaggle.com/vincee/intel-image-classification-cnn-keras

• Grad CAM：
  https://arxiv.org/pdf/1610.02391v1.pdf
• Test Time Augmentation (TTA)：

  https://www.kaggle.com/iafoss/pretrained-resnet34-with-rgby-0-460-public-lb

最后

有許多方法來調(diào)整你的模型，并且新的想法總是會(huì)出現(xiàn)。深度學(xué)習(xí)是一個(gè)快速發(fā)展的領(lǐng)域，沒有什么靈丹妙藥。我們必須做很多實(shí)驗(yàn)，足夠的試驗(yàn)和錯(cuò)誤會(huì)帶來突破。