Can't generate images from any text about tac-gan

[dashayushman]

@GaryLMS you are welcome. Closing this issue and good luck (y)

[dashayushman]

Hi @GaryLMS ,

could you provide the hyper-params that you are using while training and generating images? Because I used a fixed set of parameters for training and generating the images and it works fine.

PS: I have fixed some minor bugs though. Note that I have changed the model.py script. If you are using the old code then just checkout everything except for model.py

[GaryLMS]

Hello @dashayushman
Thanks for your reply!!!

I train the model again yesterday with step 1.4 command
$python train.py --t_dim=100 --image_size=128 --data_set=flowers --model_name=TAC_GAN --train=True --resume_model=True --z_dim=100 --n_classes=102 --epochs=400 --save_every=20 --caption_vector_length=4800 --batch_size=128

I just update the model.py you just modified, and I first encode the text into vectors by
$python encode_text.py --caption_file=Data/text.txt --data_dir=Data

and then use the command you just update
$python generate_images.py --data_set=flowers --checkpoints_dir=Data/training/TAC_GAN/checkpoints --images_per_caption=30 --data_dir=Data

now it comes

2017-07-28 21:18:56.734720: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Assign requires shapes of both tensors to match. lhs shape= [512] rhs shape= [100]
[[Node: save/Assign_13 = Assign[T=DT_FLOAT, _class=["loc:@d_embedding/bias"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/gpu:0"](d_embedding/bias, save/RestoreV2_13/_19)]]
2017-07-28 21:18:56.734810: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Assign requires shapes of both tensors to match. lhs shape= [512] rhs shape= [100]
[[Node: save/Assign_13 = Assign[T=DT_FLOAT, _class=["loc:@d_embedding/bias"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/gpu:0"](d_embedding/bias, save/RestoreV2_13/_19)]]
2017-07-28 21:18:56.734814: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Assign requires shapes of both tensors to match. lhs shape= [512] rhs shape= [100]
[[Node: save/Assign_13 = Assign[T=DT_FLOAT, _class=["loc:@d_embedding/bias"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/gpu:0"](d_embedding/bias, save/RestoreV2_13/_19)]]
2017-07-28 21:18:56.736200: W tensorflow/core/framework/op_kernel.cc:1158] Invalid argument: Assign requires shapes of both tensors to match. lhs shape= [512] rhs shape= [100]
....

Thank you!
P.S. I find you also update train.py, so should I re-train model to get correct result?

---

Oh, I think the problem is the command I use for training is
$python train.py --t_dim=100 --image_size=128 --data_set=flowers --model_name=TAC_GAN --train=True --resume_model=True --z_dim=100 --n_classes=102 --epochs=400 --save_every=20 --caption_vector_length=4800 --batch_size=128

set t_dim = 100, but the default t_dim in generate_images.py is 512(and you just modified to 256), so the size is mismatch, so I add the argument
$python generate_images.py --data_set=flowers --checkpoints_dir=Data/training/TAC_GAN/checkpoints --images_per_caption=30 --data_dir=Data --t_dim=100

now I can generate images successfully! Thank you.

[dashayushman]

@GaryLMS please retrain the model for some iterations and then use the following command for generating the images

$python generate_images.py --t_dim=100 --data_set=flowers --checkpoints_dir=Data/training/TAC_GAN/checkpoints --images_per_caption=30 --data_dir=Data

you can notice that I have added t_dim=100. I believe this is what is causing the dimension mismatch. As mentioned in the README file, whenever you change your hyperparameters, please use the same in all other scripts which need them.

[GaryLMS]

@dashayushman
Thank you! I also find the mismatch in t_dim, I have updated my comment, the problem has been solved!!!
Thanks for your reply!