Retraining faster rcnn from object detection 2 api on a custom dataste

未关闭
  • 哪里不对？
#11,251 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
20/100
Issue 类型
缺陷
描述清晰度
需要澄清
活跃度
停滞
技术栈
python, tensorflow
领域
machine-learning

调研方向

从所引用的 Faster R-CNN 流水线配置以及 model_builder 和 config_util 入口点开始。检查 checkpoint 的恢复方式以及 classification head 的替换方式，然后运行所示的虚拟图像初始化和训练设置，以识别恢复变量、head 维度或损失行为方面的不一致。

由索引模型根据 Issue 内容生成。

描述

models:research type:feature

Dear all,

I am trying to retrain the classification head of faster rcnn while keeping other branches, I defined this code:

import tensorflow as tf
from object_detection.builders import model_builder
from object_detection.utils import config_util

Clear the current TensorFlow session

tf.keras.backend.clear_session()

print('Building model and restoring weights for fine-tuning...', flush=True)

Settings

num_classes = 2
pipeline_config = '/MyPath/models/research/object_detection/configs/tf2/faster_rcnn_resnet50_v1_640x640_coco17_tpu-8.config'
checkpoint_path = 'MyPath/models/research/object_detection/test_data2/checkpoint/faster_rcnn_resnet50_v1_640x640_coco17_tpu-8/checkpoint/ckpt-0'

Load pipeline config and build a detection model

configs = config_util.get_configs_from_pipeline_file(pipeline_config)
model_config = configs['model']

Update the model config to reflect the number of classes

model_config.faster_rcnn.num_classes = num_classes
model_config.faster_rcnn.feature_extractor.batch_norm_trainable = False

Build the detection model

detection_model = model_builder.build(model_config=model_config, is_training=False)

Restore the model weights from the checkpoint, excluding the classification and mask heads

ckpt = tf.compat.v2.train.Checkpoint(model=detection_model)
ckpt.restore(checkpoint_path).expect_partial()

Reinitialize the class prediction head to match the new number of classes

class_prediction_head = detection_model._mask_rcnn_box_predictor._class_prediction_head

Access the Dense layer responsible for class prediction

dense_layer = class_prediction_head._class_predictor_layers[1]

num_units = num_classes + 1 # Add 1 for the background class

Create a new Dense layer with the correct number of classes

new_dense_layer = tf.keras.layers.Dense(
units=num_units,
activation=dense_layer.activation,
use_bias=dense_layer.use_bias,
kernel_initializer=tf.keras.initializers.RandomNormal(stddev=0.01), # Initialize new layer weights randomly
bias_initializer=dense_layer.bias_initializer,
kernel_regularizer=dense_layer.kernel_regularizer,
bias_regularizer=dense_layer.bias_regularizer,
activity_regularizer=dense_layer.activity_regularizer,
kernel_constraint=dense_layer.kernel_constraint,
bias_constraint=dense_layer.bias_constraint,
name=dense_layer.name # Keep the same name
)

Replace the old Dense layer with the new one in the class prediction head

class_prediction_head._class_predictor_layers[1] = new_dense_layer

Run the model through a dummy image to ensure all variables are properly initialized

image, shapes = detection_model.preprocess(tf.zeros([1, 640, 640, 3]))
prediction_dict = detection_model.predict(image, shapes)
_ = detection_model.postprocess(prediction_dict, shapes)

print('Weights restored for box predictor! Classification head initialized with the new number of classes.')

When establishing the training, my loss is around the 1. I choose the following optimizer:

learning_rate = 0.000001
optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate, momentum=0.9)

Is the restoration of checkpoint correct? Can Any help to detect any inconsistencies in the definition of model.

Thank you

主要语言
Python
星标
77.6k
派生
44.8k
30 天内没有已合并 PR

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

tensorflow/models 的其他 Issue

查看 tensorflow/models 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。