loliverhennigh / convolutional-lstm-in-tensorflow Goto Github PK

Hi, I have a question.

In paper, each gate has cell memory term with convolution.

But your code doesn't have.

I think i missed your intention of what it is, please let me know what I missed it.

Hi,
I got this error:
ValueError: Dimension 1 in both shapes must be equal, but are 12 and 8 for 'network/conv_lstm/BasicConvLSTMCell/Conv/concat' (op: 'ConcatV2') with input shapes: [?,12,12,4], [16,8,8,4], [].

Any suggestion on this ?

Thanks

Hi @loliverhennigh ,

Thanks a lot for your code. It's very useful. Now I have an issue about applying it on grayscale image. I change the input channel, however get the error: " Trying to share variable network/encode_1_conv/weights, but specified shape (3, 3, 3, 8) and found shape (3, 3, 1, 8)."

How does that happen? Where do I need to change when I want to apply on image with different channel.

Look forward to your reply!

The graph on the readme shows that as the steps increase, the error of the convolutional lstm increases faster than the convolutional model. Should the y axis be error or performance?

what is you input prixel' range , it is [0,255]or [0,1].I think your generate network the last layer should use sigmoid ,why you use linear?

Hi, @loliverhennigh ,

I got the following error. Any suggestion?

init network from scratch
WARNING:tensorflow:Passing a `GraphDef` to the SummaryWriter is deprecated. Pass a `Graph` object instead, such as `sess.graph`.
Traceback (most recent call last):
  File "main_conv_lstm.py", line 187, in <module>
    tf.app.run()
  File "/root/anaconda3/envs/tf_1.0/lib/python3.5/site-packages/tensorflow/python/platform/app.py", line 44, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "main_conv_lstm.py", line 184, in main
    train()
  File "main_conv_lstm.py", line 146, in train
    dat = generate_bouncing_ball_sample(FLAGS.batch_size, FLAGS.seq_length, 32, FLAGS.num_balls)
  File "main_conv_lstm.py", line 39, in generate_bouncing_ball_sample
    dat[i, :, :, :, :] = b.bounce_vec(32, num_balls, seq_length)
  File "/data/code/Convolutional-LSTM-in-Tensorflow/bouncing_balls.py", line 181, in bounce_vec
    x,v = bounce_n(T,n,r,m);
  File "/data/code/Convolutional-LSTM-in-Tensorflow/bouncing_balls.py", line 40, in bounce_n
    if r==None: r=array([4.0]*n)
ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

Hey Oliver,

Great job implementing conv lstm. Someone requested me to add this type of LSTM to my repo:

https://github.com/LeavesBreathe/tensorflow_with_latest_papers

Just wondering if we can copy and use your code in the repo (of course we would site you appropriately for creating this code)

Hi Oliver!

Thanks for the code. How can I modify the stride? I have increased the filter size, but would also control the way the filter goes over the input structure.

Best
Tessa

Hi Oliver,
Thanks for a nice implementation.
I had one problem in understanding main_conv_lstm.py. Shouldn't the hidden stated be equated to None before each iteration ( after line 143 in main_conv_lstm.py )?
If this is not done, I think the last hidden state of previous batch step will be provided as input as first hidden state to the next batch.

Thanks

Hi,
As I understood from the convolutional LSTM paper (https://arxiv.org/pdf/1506.04214.pdf), cell state C is used to compute the input, output and forget gates (Equation (2) in the paper). However, I couldn't find this in your implementation in the computation of i,f, and o. Only the input x and hidden state h are being passed to _conv_linear method. Can I know if this is intentional and what's the reason behind leaving C out of your equations?

Thanks
Shamit

Hi, just a quick question. How is the structure of the network in the main_conv_lstm.py??

If I understand correctly is that the network passed the image to the 4 layer of convolutional. Then to one layer of conv lstm. Then to 4 layer deconvolutional. Isn't it?

If yes then, how to use multi layer of this conv lstm?.

Thanks

This is the tensorboard view, the model(residual_model) diverge when the training step is about 151K, what should I do to prevent model is diverged? Thank you for your implementing the convlstm, It help me a lot!!