如何基于TensorFlow Model Zoo预训练模型完成自定义数据集的目标检测迁移学习（非Keras实现，以Centernet_1024x1024为例）

Hey there! I’ve been working with TensorFlow’s Model Zoo for custom object detection without relying on Keras, so I can walk you through the exact process step by step. Here’s how to pull it off with a model like Centernet_1024x1024:

First, you’ll need to convert your dataset to TFRecord format—this is the standard input format for TensorFlow Object Detection API models (no Keras involved here).

• Use the official conversion scripts from the Object Detection API repo:
  • For Pascal VOC-style datasets: create_pascal_tf_record.py
  • For COCO-style datasets: create_coco_tf_record.py
• If your dataset uses a custom format, modify one of these scripts to parse your annotations and images into TFRecords.
• Create a label_map.pbtxt file that maps class IDs to names, like this:

  item {
    id: 1
    name: 'your_class_1'
  }
  item {
    id: 2
    name: 'your_class_2'
  }
  

Grab the pre-trained checkpoint from TensorFlow Model Zoo:

• Download the Centernet_1024x1024 package (look for the .tar.gz file containing the checkpoint)
• Extract it to a directory (e.g., ./pretrained_centernet/)—you’ll need the ckpt-xxxx files inside for fine-tuning.

Copy the default config file for your Centernet model (e.g., centernet_hg104_1024x1024_coco17_tpu-32.config from the Model Zoo or Object Detection API repo) and modify these critical sections:

• num_classes: Set this to the number of classes in your custom dataset.
• train_input_reader:
  • Update input_path to point to your training TFRecord file.
  • Update label_map_path to your label_map.pbtxt file.
• eval_input_reader: Do the same as above for your validation TFRecord.
• train_config:
  • Set fine_tune_checkpoint to the path of your pre-trained checkpoint (e.g., ./pretrained_centernet/ckpt-32768—note you don’t include the file extension).
  • Set fine_tune_checkpoint_type: "detection" (since we’re fine-tuning the full detection pipeline).
  • Adjust batch_size, learning_rate, and num_steps based on your hardware and dataset size.

Use the official model_main_tf2.py script (this is the non-Keras, native TensorFlow training entry point). Run this command in your terminal:

python model_main_tf2.py \
  --model_dir=./training_output/ \
  --pipeline_config_path=./centernet_custom.config \
  --num_train_steps=50000 \
  --alsologtostderr

• model_dir: Where training checkpoints and logs will be saved.
• pipeline_config_path: Path to your modified config file.

Track progress with TensorBoard (no Keras callbacks needed):

tensorboard --logdir=./training_output/

You’ll be able to view loss curves, evaluation metrics, and more in your browser.

Evaluate the Model #

Run evaluation using the same script, pointing to your training checkpoints:

python model_main_tf2.py \
  --model_dir=./training_output/ \
  --pipeline_config_path=./centernet_custom.config \
  --checkpoint_dir=./training_output/ \
  --alsologtostderr

Export for Inference #

Export the trained model as a saved model (native TensorFlow format, no Keras wrapper):

python exporter_main_v2.py \
  --input_type=image_tensor \
  --pipeline_config_path=./centernet_custom.config \
  --trained_checkpoint_dir=./training_output/ \
  --output_directory=./exported_centernet/

Load the exported model and run detections without any Keras APIs:

import tensorflow as tf
import numpy as np

# Load the saved model (native TensorFlow format)
detect_fn = tf.saved_model.load('./exported_centernet/saved_model')

# Load and preprocess an image
image_path = "test_image.jpg"
image = tf.io.read_file(image_path)
image = tf.image.decode_jpeg(image, channels=3)
input_tensor = tf.convert_to_tensor(image)
input_tensor = input_tensor[tf.newaxis, ...]  # Add batch dimension

# Run inference
detections = detect_fn(input_tensor)

# Process results
num_detections = int(detections.pop('num_detections'))
detections = {key: value[0, :num_detections].numpy()
              for key, value in detections.items()}
detections['num_detections'] = num_detections
detections['detection_classes'] = detections['detection_classes'].astype(np.int64)

# Example: Print detected classes and scores
for i in range(num_detections):
    print(f"Class: {detections['detection_classes'][i]}, Score: {detections['detection_scores'][i]:.2f}")

Quick Notes to Avoid Pitfalls #

• Make sure you’ve properly installed the TensorFlow Object Detection API (compile protobufs, add the repo paths to your PYTHONPATH).
• If you hit OOM errors, reduce the batch_size in your config or enable mixed precision training by adding use_mixed_precision: true to train_config.

Let me know if you run into specific issues with any step—I’ve debugged my fair share of these setups!

内容的提问来源于stack exchange，提问作者Julio Gomez