neka-nat
The results seems not good.
The result seems not good as the result in the paper.It there just need more training or other tricks?
Hello, where did you training the module, CPU or GPU? I can train it on CPU, but maybe my 2 8G Kingston RAMs can't offer enough storage capacity，only can train the module in Tc+Td epochs. So the result seems not good because my computer did not complete the training of the entire model.
Ooh,I just run the master branch code on my gpu.I ignored the tc and the td stage.I would try to run the animi branch today.
Yeah, the master branch code does not involve the tc and the td stage. I am looking forward to your new result on the animi branch. Could you please tell me the version of your GPU, CUDA, Tensorflow-gpu and Keras , My computer is nvidia GTX1080, CUDA8.0 Tensorflow-gpu(1.6.0) and Keras (2.1.5) ,but it doesn't work for the anim and separate_c_d branch.
I use M40,cuda8.0,Tensorflow-gpu（1.6.0）and Keras（2.0.8）and it run 100 epochs about 2.5 hours.
Oh, thanks, maybe my GPU is only 8G, I‘ll try to change the code for solving the problems.
@dongdong092 when you finished the training, how do you load the trained weight to predict the test sets. Please give me some advice, thanks!
I run the other branch .In the 20th opoches,it stops by the following error: print("%d [D loss: %e] [G mse: %e]" % (n, d_loss, g_loss)) TypeError: must be real number, not list
I also met this error，because the g_loss is not a real number, I print g_loss separately by the following code, which show g_loss is a list such as 'G mse: [0.0028160308, 0.0028160308, 1.1920933e-07]': print('%d G mse:' % n), print(g_loss), print("[D loss: %e]" %d_loss)
I train this model on Street View dataset,my result doesn't good as well. Do anybody achieve good result?
@xieenze I train the model on place2 dataset, the result looks good. But when I test images with arbitrary mask shape, the results are too bad. Do you test any images and how to do it ?
@dongdong092 im running very slow compared to your 100Ep @ 2.5Hours. I use batch size 16 on a TitanV GPU (and dual xeon cpus to spare) and Im getting around 20mins per Epoch... Im also only using the 35000 validation images from places...
