Training 200000iters on 3000 image dataset, but boxes had no control effect

[Hui-88]

Thank you for sharing. I have a question to ask you. I trained 200000 iters on my own dataset of 3000 images, with batch_size=2. The box coordinates in the tsv file are the top left and width height, but there was no layout control effect after training. Excuse me, is the dataset too small? Do I need to add iters. look forward to your reply！

[Hui-88]

[qinghew]

i have a similar situation. during training, my self.alpha_attn and self.alpha_dense are too small

[Hui-88]

@qinghew
Sorry, may I ask which .py file self.alpha_attn and self.alpha_dense are in ? I can't find them

[ldy674]

感谢您的分享。我有一个问题要问你。我在自己的 3000 张图像数据集上训练了 200000 个迭代器，其中 batch_size=2。tsv 文件中的 box 坐标是左上角和宽度高度，但训练后没有布局控制效果。请问，数据集是不是太小了？我需要添加迭代器吗？期待您的回复！

Hello, I'm working on reproducing GLIGEN with my own dataset, but I'm a bit confused about the handling of the dataset, can I ask you for some advice? What process do you follow to process your own dataset that can be used as an input to GLIGEN?