Generated images don't correspond to caption

[shenkev]

Maybe I'm doing something stupid but essentially I followed your generating images (with pretrained model) steps:

python generate_thought_vectors.py --caption_file="Data/sample_captions.txt"

where sample_captions.txt has just these words: "the flower has yellow petals and the center of it is brown" (Is this format correct?)

python generate_images.py --model_path=Data/Models/latest_model_flowers_temp.ckpt --n_images=8

A single combined_image_0.jpg comes out in the val_samples folder but the images are DO NOT "have yellow petals and the center of it is brown". As far as I can tell, the model is not conditioning on the text properly. What am I doing wrong?

[paarthneekhara]

Can you try a simple caption like "A red flower" just to check if there is no blunder

[shenkev]

This is the output for "A red flower" with n=5:

[shenkev]

Maybe this is relevant?

I replaced ops.py according to the solution mentioned in this issue: #13

Then I changed sigmoid_cross_entropy_with_logits() to take the "logits" and "labels" arguments explicitly. i.e. sigmoid_cross_entropy_with_logits(logits=... , labels=... ) because it complained about passing arguments explicitly.

[paarthneekhara]

Well, there seems something wrong for sure. The output should not look messy like above. Did you train the model or use the pretrained one?

[shenkev]

I used the pretrained model. I mean the fact that I'm pointing model_path to latest_model_flowers_temp.ckpt should guarantee that model is loaded. Also the name of the file latest_model_flowers_temp.ckpt indicates this is the up-to-date pretrained model.

python generate_images.py --model_path=Data/Models/latest_model_flowers_temp.ckpt --n_images=8

[paarthneekhara]

Incase you are still facing troubles, try using a more recent implementation https://github.com/zsdonghao/text-to-image

[shenkev]

Thanks, I'll take a look

[chenxinjing]

@shenkev @paarthneekhara I also encountered the same problem, do you have a solution? I don't modify the parameters. I used tensorflow 1.3.

[Fuhongshuai]

I tried using a more recent implementation, but also encountered the same problem, do you have a solution?@shenkev @paarthneekhara

[silverlilin]

Incase you are still facing troubles, try using a more recent implementation https://github.com/zsdonghao/text-to-image

why do i detect the same result? The images obtained from different models are different.