I trained your project, but it always output the same DSL content no matter what images are given

[sunnycomes]

Hi, there. When learning pix2code-template, I found that the trained web model can only output the same DSL content, that is to say, no matter what pictures are given, the output DSL is the same. Note in advance: since my GPU is only 8g, according to the original configuration of your project, my GPU memory is not enough, so I made some modifications during training, changed dtype (float32 - > float16), the model can be trained, but there are problems as I said above. Besides, after several epoches of learning, the loss is always around 0.2. Is this a normal situation? It will be great if you can help me out.

[luweiagi]

Hi, sunnycomes, I also meet the same problem, I user the original code tonybeltramelli/pix2code, different pic input, same dsl code output. My gpu is only 4g, so I change batch size from 64 to 2, and I do not change data type as you(float32 to float16). So I do not know what the reason is. It seems that the model itself has problem.