maskrcnn如何防止过拟合

修改文件mrcnn/model.py大约970行位置在全连接层加上dropout0.5# Classifier head mrcnn_class_logits KL.TimeDistributed(KL.Dense(num_classes), namemrcnn_class_logits)(shared) # 在全连接层加上dropout 0.5防止过拟合 2022-4-15 tjm mrcnn_class_logits KL.TimeDistributed(KL.Dropout(0.5), namemrcnn_class_logits_dropout)(mrcnn_class_logits) mrcnn_probs KL.TimeDistributed(KL.Activation(softmax), namemrcnn_class)(mrcnn_class_logits) # BBox head # [batch, num_rois, NUM_CLASSES * (dy, dx, log(dh), log(dw))] x KL.TimeDistributed(KL.Dense(num_classes * 4, activationlinear), namemrcnn_bbox_fc)(shared) x KL.TimeDistributed(KL.Dropout(0.5), namemrcnn_bbox_fc_dropout)(x)只有mrcnn中分类和框回归任务有全连接层所以这部分损失明显增大。试验下结果再考虑是否在卷积层加dropout测试的时候记得去掉dropout学习率自动衰减