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

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

Start with the referenced Faster R-CNN pipeline config and the model_builder and config_util entry points. Check how the checkpoint is restored and how the classification head is replaced, then run the shown dummy-image initialization and training setup to identify inconsistencies in restored variables, head dimensions, or loss behavior.