JAVA

I have the same problem. How can I solve it?

The problem is"'float' object cannot be interpreted as an integer" coords_x = x / img_size------→coords_x = x // img_size coords_y = y/img_size------→coords_y = y//img_size max_counter = len(train_set)/batch_size------→max_counter = len(train_set)//batch_size python2...

The same problem.

The　ｓａｍｅ　ｐｒｏｂｌｅｍ，

也是纠结了半天，他这个代码的loss基本不可训练，mAP基本也就是0，训练不起来。不知道啥毛病。

CHECKPOINT_PATH=./checkpoints/ssd_300_vgg.ckpt checkpoint_path=./checkpoints/vgg_16.ckpt This is the essence.

The author can't explain it.

same problem.

训练70个epoch，train_first_stage_epochs=20、train_second_stage_epochs=50，第70个epoch的mAP=79.23%，第16个epochd mAP=80.91%，感觉训练多了反而没用，就修改了一个参数，batch_size=4;请仓主帮忙解答一下哦，谢谢 @YunYang1994

这个代码貌似存在问题 self.conv = slim.conv2d(conv, 3, [3, 3], rate=self.dilate[6], activation_fn=None, weights_regularizer=slim.l2_regularizer(self.hps.weight_decay_rate), scope='conv_%d'%(self.hps.num_conv), reuse=self.reuse) self.clear = self._image - self.conv 那么问题来了，计算损失的时候用self.conv 这个应该是噪声图像，计算结果的时候用model.clear，自相矛盾 content_cost = (1./self.hps.batch_size)*tf.nn.l2_loss(self.conv - self.label) tf_psnr = tf.image.psnr(labels, model.clear, 1.)