pixel-rnn-tensorflow
pixel-rnn-tensorflow copied to clipboard
carpedm20
in progress
Hi All! I wonder how to acquire pretrained network or weights file to run the things solely to the inference? Somehow we have faced some difficulties with training the net...
Hi, @carpedm20 When I run the command: > $ python main.py --data=mnist --model=pixel_rnn Its output is as follows: > I tensorflow/stream_executor/dso_loader.cc:108] successfully opened CUDA library libcublas.so locally > I tensorflow/stream_executor/dso_loader.cc:108]...
`output_state_bw_with_last_zeros = tf.concat(1, [output_state_bw_except_last, dummy_zeros])` should be `output_state_bw_with_last_zeros = tf.concat(1, [dummy_zeros, output_state_bw_except_last])`
File: **utils.py** Function: **get_model_dir**
https://github.com/carpedm20/pixel-rnn-tensorflow/blob/1d98c28c62ed9c9584a342b63e4612dd9e5ae688/ops.py#L315 I have found at L315 that we have used the sigmoid function either to calculate the new inputs g, is that possible? Shouldn't we use the tanh instead? Thank...
I have about 15,000 images that are all 16x16 in size. They are jpeg formatted. How would I go about using them as training data? I could not find anything...
Hi , i also implement PixelCNN , and training same like your learning curve But i can't generate samples from zeros very well , can someone give me some advice...
In `network.py` ``` if conf.use_gpu: data_format = "NHWC" else: data_format = "NCHW" ``` I think `NCHW` is preferred for GPU, while for CPU, we should use `NHWC`.
Not really an issue (tell me a better place to put this plz), does anyone have a trained model I can use, perhaps trained on ImageNet or Cifar :)
