修改TensorFlow Object Detection API

代碼倉庫：https://github.com/lijiancheng0614/tensorflow_object_detection

修改TensorFlow Object Detection API，添加一些方便使用或新的功能。

中文

使用時記得添加PYTHONPATH：

# From tensorflow/models/
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

以及編譯protobuf：

# From tensorflow/models/
protoc object_detection/protos/*.proto --python_out=.

在train_config中添加from_detection_checkpoint參數

增加加載模型所有參數和不加載模型的模式。

原來：

from_detection_checkpoint爲bool型：

• True: 加載detection模型參數中FeatureExtractor部分。

• False: 加載classification模型參數。

現在：

from_detection_checkpoint爲uint32型：

• 3: 不加載模型參數（適用於train from scratch的情況。）

• 2: 加載模型所有參數（適用於訓練中斷後重新加載checkpoint的情況。）

• 1: 加載detection模型參數中FeatureExtractor部分。

• 0: 加載classification模型參數。

修改文件：

• trainer.py

• core/model.py

• protos/train.proto

• meta_architectures/ssd_meta_arch.py

• meta_architectures/faster_rcnn_meta_arch.py

更新訓練時的summary

把histograms和first_clone_scope等summaries去掉。這樣訓練的event文件變小，方便tensorboard加載。

修改文件：

• trainer.py

在eval.py中添加gpu_allow_growth參數

在eval.py中添加gpu_allow_growth參數，默認爲True，即不佔用GPU全部內存，而是動態申請顯存。

修改文件：

• eval.py

• evaluator.py

• eval_util.py

在train.py中添加gpu_allow_growth參數

在train.py中添加gpu_allow_growth參數，默認爲True，即不佔用GPU全部內存，而是動態申請顯存。

修改文件：

• train.py

• trainer.py

在train.config中添加max_to_keep參數

在train_config中添加max_to_keep參數，默認爲5，即保留最後5個checkpoint。如爲0則保留所有的checkpoint。

修改文件：

• trainer.py

• train.proto

網絡添加FocalSigmoidClassificationLoss

model中的loss中的classification_loss可使用focal_sigmoid。

關於Focal loss，參考 https://arxiv.org/pdf/1708.02002.pdf

修改文件：

• core/losses.py

• builders/losses_builder.py

• protos/losses.proto

English

Remember update PYTHONPATH:

# From tensorflow/models/
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

And compile protobuf:

# From tensorflow/models/
protoc object_detection/protos/*.proto --python_out=.

Add from_detection_checkpoint parameter in train_config

Old:

bool from_detection_checkpoint

• True: the checkpoint was an object detection model that have the same parameters with the exception of the num_classes parameter.

• False: the checkpoint was a object classification model.

New:

uint32 from_detection_checkpoint

• 3: don’t load any variables.

• 2: load all variables.

• 1: load feature extractor variables from an object detection model, same as True.

• 0: load feature extractor variables from a object classification model, same as False.

Modified files:

• trainer.py

• core/model.py

• protos/train.proto

• meta_architectures/ssd_meta_arch.py

• meta_architectures/faster_rcnn_meta_arch.py

Remove some summaries when training

Remove summaries about histograms and first_clone_scope when training.

Modified files:

• trainer.py

Add gpu_allow_growth parameter in eval.py

Add gpu_allow_growth parameter in eval.py, default value is True which means attempting to allocate only as much GPU memory based on runtime allocations.

Modified files:

• eval.py

• evaluator.py

• eval_util.py

Add gpu_allow_growth parameter in train.py

Add gpu_allow_growth parameter in train.py, default value is True which means attempting to allocate only as much GPU memory based on runtime allocations.

Modified files:

• train.py

• trainer.py

Add max_to_keep parameter in train_config

Add max_to_keep parameter in train_config, default value is 5 which means the 5 most recent checkpoint files are kept. If 0, all checkpoint files are kept.

Modified files:

• trainer.py

• protos/train.proto

Add FocalSigmoidClassificationLoss in model

In config, model -> loss -> classification_loss can be focal_sigmoid, parameters: anchorwise_output, gamma.

Reference: https://arxiv.org/pdf/1708.02002.pdf

Modified files:

• core/losses.py

• builders/losses_builder.py

• protos/losses.proto