The codes of paper "Long Text Generation via Adversarial Training with Leaked Information" on AAAI 2018. Text generation using GAN and Hierarchical Reinforcement Learning.
I am a little confused to bleu calculation nltk.translate.bleu_score.sentence_bleu(reference, h, weight) in which reference is save/realtest_coco.txt and h is save/generator_sample.txt,
is that somewhat demands on realtest_coco.txt, size and contents?
Also , this sentence in the paper
In each step, it receives generator D’s high-level feature representation, e.g., the feature map of the CNN, and uses it to form the guiding goal for the W ORKER module in that timestep.
i am not sure whether it's "generator" or "discriminator".
btw, it's really a very nice job, so many thanks to you.
The details can be found in the "Image Coco" subsection in the paper, we randomly choose 80000 sentences from Coco dataset as the training dataset, then other 5000 sentences are also randomly chosen as the test dataset( save/realtest_coco.txt ).
CNN is discriminator and WORKER module belongs to the generator.
Thanks for your attention.
I hope the above can answer your questions.
i get it.
for '1', i've seen the details in the paper, as i am a beginner, i just, a little confused, no doubts :).
thanks for your answer and thank you so much..
I am a little confused to bleu calculation
nltk.translate.bleu_score.sentence_bleu(reference, h, weight)
in which reference issave/realtest_coco.txt
and h issave/generator_sample.txt
,is that somewhat demands on
realtest_coco.txt
, size and contents?Also , this sentence in the paper
i am not sure whether it's "generator" or "discriminator".
btw, it's really a very nice job, so many thanks to you.