octavian-ai / biggan-tpu-tensorflow Goto Github PK

When I run the code, it shows the following error:

Traceback (most recent call last):
File "main_tpu.py", line 62, in
main()
File "main_tpu.py", line 58, in main
get_estimator(args, gan, True))
File "/content/BigGAN-TPU-TensorFlow/main_loop.py", line 33, in run_main_loop
prefetch_inception_model()
File "/content/BigGAN-TPU-TensorFlow/inception_score.py", line 17, in prefetch_inception_model
calculate_inception_score(images)
File "/content/BigGAN-TPU-TensorFlow/inception_score.py", line 45, in calculate_inception_score
v_images = tf.contrib.gan.eval.preprocess_image(v_image_batch)
File "/tensorflow-1.15.2/python3.6/tensorflow_core/python/util/lazy_loader.py", line 63, in getattr
return getattr(module, item)
AttributeError: module 'tensorflow_core.contrib' has no attribute 'gan'

Hi,

I'm trying to train BigGAN on a dataset and see this is a recent project -- I have a few questions. Have you had success/progress with training on ImageNet so far? Are there any specific training scripts that are or are not currently working? What tpu-size are you using?

Thanks!

We run **'launch_train_tpu_sagan.sh' on V3 pod for 300k steps, but it seems to collapse from 10k.
And the generation quality is not good, i wonder wether you've managed to achieve or at least close to biggan's performance?
Here're some results:
After 100k

After 180k

After 300k

The function parse_tfrecord_progan returns the parsed images as a tf.float32 however the labels are returned as an int64, this causes the ConcatV2 to complain of varying types

2019-10-03 06:27:54  tensorflow[2134] ERROR Error recorded from training_loop: Tensors in list passed to 'values' of 'ConcatV2' Op have types [**float32**, **int64**] that don't all match.

ValueError: Tensor conversion requested dtype float32 for Tensor with dtype int64: 'Tensor("InfeedQueue/dequeue:1", shape=(256, 2), dtype=int64, device=/device:TPU_REPLICATED_CORE:0)'

Changing the dtype of the label to tf.float32 on the following line seems to fix the issue. Am I missing something here?