Reproducing this work

[CubicQubit]

Hi there,

I was wondering if the original authors have run into this error tensorflow/core/grappler/optimizers/meta_optimizer.cc:561] function_optimizer failed: Invalid argument: Input 0 of node zeros_like_72 was passed float from sequential/lstm/StatefulPartitionedCall:9 incompatible with expected variant. 2021-10-13 18:40:24.989984: E tensorflow/core/grappler/optimizers/meta_optimizer.cc:561] function_optimizer failed: Invalid argument: Input 0 of node zeros_like_72 was passed float from sequential/lstm/StatefulPartitionedCall:9 incompatible with expected variant.

I got this while running python3.7.10 and installed the requirements.txt. The program still runs and outputs normally but it cannot learn the toy examples for sines very well. I was wondering if the above error is causing problems as I'm trying to run cotgan on some time-series data.

Training curve:

Real samples: Plot the first 4 dimensions for 8 samples, with increasing frequency for each dimension.

Generated samples: It learned the range but not the frequency / seasonality

Config:

model_params:
  name: 'cotgan'
  reg_penalty: 10
  sinkhorn_eps: 100
  sinkhorn_l: 100
  nlstm: 1
  scaling_coef: 1

  z_dims_t: 10
  y_dims: 10

  g_state_size: 256
  d_state_size: 256
  g_filter_size: 32
  d_filter_size: 32
  disc_kernel_width: 5
  gen: 'lstm'

  batch_norm: True
  learning_rate: 0.001

[tianlinxu312]

We have never seen this error before. You may need to google around and see if you can find a solution.

Is this sine function the same as we used in the paper? We also ran the model on 1D sine waves (didn't report in the paper as it's too simple), which can be learnt easily.