- Requirements: TensorFlow 1.2 or above
- Running: Training: python CSAP --train Test: python CSAP --test
- Key parameters
CSAP.py
GO_ON_TRAINING: Whether to continue training from the last breakpoint, 1 is yes, 0 is no
LOAD_PREVIOUS_POS: Whether to load the coding information, if it is not the first run and the training data has not been changed, set it to 0, otherwise 1 for recoding
MASK_DATA_PATH: If the roi area information is included in the test image, you can import the file under the corresponding path and only retain the detection results in the area
var_list: The core parameters used for transfer learning in tf.train.AdamOptimizer. Single-step transfer learning is performed when it is set to self.detector.vars_d, and transitive transfer is performed when it is set to self.detector.vars_stage2
model_fpn.py
with tf.variable_scope(self.name) as scope: Fully use the training data under the current task for training
with tf.variable_scope('trains_1') as scope: Used as a single-step transfer learning and general analysis of the detection model features, only this part of the feature is trained
with tf.variable_scope('trains_2') as scope: The training part of transfer learning based on single-step transfer learning
Reference：Q. An, Z. Pan, H. You and Y. Hu, "Transitive Transfer Learning-Based Anchor Free Rotatable Detector for SAR Target Detection With Few Samples," in IEEE Access, vol. 9, pp. 24011-24025, 2021, doi: 10.1109/ACCESS.2021.3056663.