When I ran the code, it raised an error at line 'fcos .load_weight('./fcos/test-95778').
I wonder what file should I put under that path? is the 'weight' here means the pre-trained model? should I download a pre-trained model and if that what file type does it expect, like ckpt or pb?
Thanks!

How much is the loss about when the model converges？ I have trained for 69 epochs, but the loss doesn't seem to drop for a long time.
Thx!

Is there any evaluate program for evaluating the training result?

        if self.mode == 'train':
            total_loss = []
            # ground_truth [batch, num_box, ymin, ymax, xmin, xmax, classid]
            for i in range(self.batch_size):
                gt_i = self.ground_truth[i, ...]
                slice_index = tf.argmin(gt_i, axis=0)[0]
                gt_i = tf.gather(gt_i, tf.range(0, slice_index, dtype=tf.int64))
                gt_size = tf.sqrt(gt_i[..., 2] * gt_i[..., 3])
                g3 = tf.boolean_mask(gt_i, gt_size <= 64.)

#############
I found this part maybe different form origin paper. I dont't get the point of "slice_index".
For your implement, if gt_i contain just one bbox , “slice_index = tf.argmin(gt_i, axis=0)[0]” will be equal zero, so “gt_i ” in next line will gather nothing(for tf.range(0, 0)).
Or am I have some misunderstanding? THX

Dear author，thank you for sharing your reimplement of FCOS！
Can you share your trained weights on VOC ?

When I train my own data, the loss shows me nan, what is the problem?

Hi,
What were the mAP and bbox AP of this paritcular model ?
Thanks!