train without pretrain weights。。。

[AiIsBetter]

I don't want to use the weight of the pre-training to train the model, where can I change； I modified ： # restorer, restore_ckpt = faster_rcnn.get_restorer() # if not restorer is None: # restorer.restore(sess, restore_ckpt) # print('restore model') added: summary_writer = tf.summary.FileWriter(summary_path, graph=sess.graph) saver = tf.train.Saver(max_to_keep=10) But the result is not correct, no target can be predicted, and then the loss is also down.I look forward to your reply. Thank you. @yangxue0827 .

[zhao34731]

Also I train from scratch, but the loss is nan for VOC dataset. Here is an example , I wonder whether the loss will be normal after training a long time. I chck the anchor sizes and stride. It seems fine for the voc dataset.

step160 image_name:b'000321.jpg' | rpn_loc_loss:0.005 | rpn_cla_loss:0.364 | rpn_total_loss:0.369 | fast_rcnn_loc_loss:0.000 | fast_rcnn_cla_loss:1.953 | fast_rcnn_total_loss:1.953 | total_loss:nan | per_cost_time:0.227s