draw
draw copied to clipboard
ericjang
TensorFlow Implementation of "DRAW: A Recurrent Neural Network For Image Generation"
Ln-44: e=tf.random_normal((batch_size,z_size), mean=0, stddev=1) # Qsampler noise I think it should be placed into the funcion sampleQ. Or else the inference will fail. However, when I made such a modification,...
 Though it works, the attention windows act not like in the paper, as shown in the above.
我需要训练类似这种图 
Add options for num_intra_threads/num_inter_threads. Please review.
FYI, README.md has a missing word: > You can visualize the results by running the script `python plot_data.py ` > > For example, > > `python `**plot_data.py**` myattn /tmp/draw/draw_data.npy`
According to [Auto-Encoding Variational Bayes](https://arxiv.org/abs/1312.6114) `kl_terms[t]=0.5*tf.reduce_sum(mu2+sigma2-2*logsigma,1)-T*.5` should be `0.5*tf.reduce_sum(mu2+sigma2-2*logsigma-1,1)` equals to `0.5*tf.reduce_sum(mu2+sigma2-2*logsigma,1)-z_size*.5`
It may not be a problem, but I am just curious about why x_hat (involving true data) is also used for prediction period. Because I think, after training, the model...
