tensorfreitas / dcgan-for-bird-generation Goto Github PK

Hi
@Goldesel23

It is not working with tensorflow-gpu . If I uninstall tensorflow and install tensorflow-gpu, it throws error of no module named tensorflow. If I keep both, training goes with tensorflow cpu version.

How to resolve this. please help

Could you share the code for test ? I can not code correctly.
Thanks.

Hi

@Goldesel23
I would like to train with face images dataset. Can you please help me how to do with your repo. I want to use your repository and mention your repository in my work as citation .

Hi
@ Goldesel..

Can You please tell what have you done to make it work with tensorflow-gpu. Have you made any changes in the code or something with installation version. I have installed tensorflow-gpu 1.14. Still it does not work.

My tensoflow-gpu is successfully installed in virtual env with pip. Its my request not to close the issue before it is solved.

Hi

I would like to train dataset with K40 GPU. Is is possible with the your repository. I think tensorflow-gpu install will do? Or do You have some other dependencies or any particular version of tensorflow-gpu. I am having cuda-7.5 installed on Ubuntu 16.04 with K40 GPU.

Thanks in advance.

Hi there! Thanks for your code.
I am currently learning python and keras and found your implementation as just what I wanted to get into it.
I am trying to implement loading of models.
Your implementation saves thes whole model in hdf5 format.
I have searched through documentation about loading models back in found this

from keras.models import load_model

model.save('my_model.h5')  # creates a HDF5 file 'my_model.h5'
del model  # deletes the existing model

# returns a compiled model
# identical to the previous one
model = load_model('my_model.h5')

Now I am a Python beginner. Bloody beginner. I know that I should implement some argument handling for let's say batch_size, epochs, output_size, graph_plot and of course checkpoint directory so that the training can be resumed.
My problem is is don't know how. I studied carpedm's DCGAN for Tensorflow but I'm not able to transfer the load handling.
Any help would be greatly appreciated!

Hi

I want to train GAN with my own dataset and with different input image size. i.e. 300* 300 . But in the code if I am making changes to input size from 6464 to 300 300, it's showing error.

What should I do to get high resolution images of greater size?

Thanks in advance

Hi Tiago,

Thanks for sharing the code. I have two questions, thank you for your help.

1. If I would like to train the model using my own dataset, are there any requirements for the image data set itself, such as size, RGB/gray, etc?
2. After training the model, how to use the code to generate the new images?

Thanks

Hi @ Goldesel

I want to know the unit of loss while training the Model-

Batch 168/170 generator loss | discriminator loss : 3.7583103 | 0.9910032 - batch took 2.11966586113 s.
Batch 169/170 generator loss | discriminator loss : 2.2026653 | 0.93578804 - batch took 2.1271212101 s.
Batch 170/170 generator loss | discriminator loss : 2.7014318 | 0.7624327 - batch took 2.07665514946 s.

The loss 2.7014318 is in percentage I mean it is ~2 % loss ? Am I right.

Howdy @Goldesel23 ! I'm just learning about GANs, and reading about tuning the number of generator training steps per discriminator training steps (GpD). As you've written, can you please confirm that it's 2-to-1 here? That is, within each epoch, the discriminator trains for 1 batch_size then the generator follows by training for 2batch_size. Do I have that correct? I intend to adjust to allow for user-defined GpD, but wanted to make sure I've read your implementation correctly.

not able to get the attributes ??

where should i place or edit attribute.txt file to solve this error???

Hey
U did a great job. I am using ur code to solve some problems.
Particularly (dcgan) code.
And I am getting errors while changing image_shape.
So can u suggest me where to update in the code to generalise it for different shapes of images.
Such as (128,128,3).
If possible can u please forward me the generalised version.

Thanks

Hi
@Goldesel23 . I took celebA dataset. it has approx 10k categories. As it is Conditional GAN, I labelled all taken images as 1 means single labeled data is feeded for training. Does it has any problem in quality of results at the end as i dont train with multiple labels. My resullts are pretty good with this dataset but still little blur. Does it has anything to do with single labeled data taken?

Thanks in advance