EfficientDet

An EfficientDet implementation in TF2.0 based on the paper EfficientDet: Scalable and Efficient Object Detection on CVPR'20.

Our implement is under \ours directory.

Useful Reference

Tensorflow Custom layers

如何使用Keras fit和fit_generator（动手教程)

How to Train Your Keras Model (Dramatically Faster)

Distributed training with TensorFlow

Model Training Guide

Command flags

• --weighted-bifpn: Use Weighted BiFPN
• --snapshot: specify checkpoint or pretrained model, you can use imagenet as pretrained backbone
• --phi: determine which model to use, you can choose 0(EfficientDet D0) ~ 6(EfficientDet D6)
• --freeze-backbone: Fix backbone weight
• --steps: determine the number of steps in a epoch
• --epochs: training epochs
• pascal: use Pascal VOC dataset format and give dataset path to VOC2007 directory

Command

For example, train EfficicentDet D0 with ImageNet pretrained backbone and Weighted BiFPN

python3 train.py --snapshot imagenet --phi 1 --weighted-bifpn --gpu 3 --random-transform --compute-val-loss --freeze-backbone --batch-size 32 --steps 150 --epochs 100 pascal /opt/shared-disk2/sychou/comp2/VOCdevkit/VOC2007/

Model Evaluation Guide

Processes

因為檔案路徑(import utils)的關係，必須把Testing的檔案eval/common.py移到上一層資料夾，方可執行使用。並且注意檔案內以下幾個參數

1. phi: 選擇EfficientDet的backbone，必須和training時python3 train.py --phi參數選擇的相同
2. weighted_bifpn: 使否使用weighted BiFPN
3. PascalVocGenerator: 填上testing dataset的path，例如: '/opt/shared-disk2/sychou/comp2/VOCdevkit/test/VOC2007'
4. model_path: 填上訓練好的模型weights，例如: '/home/ccchen/sychou/comp2/comp2/efficent_series/EfficientDet/old_checkpoint/pascal_45_0.3772_0.3917.h5'

Evaluation Results

BUJO+ Environment:

• GPU: 2080Ti 10986MB
• Python 3.6.8 :: Anaconda, Inc.
• Tensorflow: 2.2.0, with GPU support
• Cuda compilation tools, release 10.0, V10.0.130

Here is the trained checkpoint.

Model Prediction Guide

Processes

1. 將 efficientdet.py 的 model_path 改成自己的 model，並將 phi 改成正確的 phi
2. predict.py 裡面有兩個段落，上面是可以偵測圖片的，下面是生成 test_prediction.txt
3. 執行 predict.py 生成 test_prediction.txt
4. 確保 txt2csv.py 跟 test_prediction.txt 在同個資料夾下，執行 txt2csv.py 產生最終的 output.csv

常見問題 :

1. 一般預測用 predict.py 的 51 行，comp2 用 predict.py 的 52 行
2. 轉檔部分請用 txt2csv.py 否則容易發生各種難以預料的環境問題 (切勿直接使用 evaluate.py 生成答案)

11/24 Training Checkpoints

• command:

python3 train.py --snapshot imagenet --phi 1 --weighted-bifpn --gpu 3 --random-transform --compute-val-loss --freeze-backbone --batch-size 32 --steps 150 --epochs 100 pascal /opt/shared-disk2/sychou/comp2/VOCdevkit/VOC2007/

• Backbone: EfficeintNet B1, phi 1

pascal_99_0.1957_0.1596.h5

The final result(Epoch 100) of the first time training. Training: mAp 0.97, Testing: mAp 0.74

pascal_98_0.1906_0.1555.h5

The last two result(Epoch 99) of the first time training. mAp 0.97 However, I forget to split train and test into different folders. It probably train on test dataset. It perhaps overfits.

---

11/25 Training Checkpoints

• Phi 6: Ran out of memory with batch size 8

• Phi 5: Ran out of memory with batch size 32

• Phi 3: Ran out of memory with batch size 16, start training successfully with batch size 8

• command:

python3 train.py --snapshot imagenet --phi 3 --weighted-bifpn --gpu 3 --random-transform --compute-val-loss --freeze-backbone --batch-size 8 --steps 150 --epochs 100 pascal /opt/shared-disk2/sychou/comp2/VOCdevkit/trainval/VOC2007/

• Backbone: EfficeintNet B3, phi 3

pascal_45_0.3772_0.3917.h5

Epoch 45, mAp 0.73. Train 7.5 hours

pascal_100_0.2464_0.2470.h5

Epoch 100, 0.9399, Train 15 hours

Testing mAP: 0.7704

11/27 Training Checkpoints

• phi 2: Just train 44 epoch

phi1-ep045-loss0.491-val_loss0.495.h5

Epoch 45, batch size 16, Train 9 hours, Kaggle score 0.79725

11/28 Training Checkpoints

• phi 1: Train 100 epoch(50 epoch train, 50 epoch fine tune)

phi1-Epoch100-Total_Loss0.4847-Val_Loss0.4685.h5

Train 100 epoch(50 epoch train, 50 epoch fine tune)

50 epoch train batch size 32

50 epoch fine tune batch size 4

Train 16 hours

11/29 Training Checkpoints

phi1-Epoch100-Total_Loss0.4847-Val_Loss0.4685.h5

Train 85 epoch(No fine tune), batch size 8

phi3-Epoch85-Total_Loss0.3117-Val_Loss0.3696.h5

Train 12 hours

phi1-Epoch176-Total_Loss0.6145-Val_Loss0.6303.h5

Train 176 epoch(050 epoch training, 50100 epoch fine tune, 100~176 epoch training), batch size 24

phi3-Epoch85-Total_Loss0.3117-Val_Loss0.3696.h5