Not Training/Sampling Properly

[nicholas-abad]

Hi,

I've tried training the default LSTM model and the RNN (with basic hyperparameters) on a scraped Friends script (which can be found here)

but for some reason, after 100 epochs in both cases, sampling results in the same exact script as previously. For example after training, a sample would give

Mr. Geller: 
I'm not gonna tell you what they spent on that wedding... 
but forty thousand dollars is a lot of money! 

Mrs. Geller: 
Well, at least she had the chance to leave a man at the altar... 

Monica: 
What's that supposed to mean? 

Mrs. Geller: 
Nothing! It's an expression.

but this is exactly the same text that's in the training input. This is how it is for all samples however.

When monitoring the NN, here's my command line inputs as well as a sample of my training, which actually "converges" extremely quickly:

➜  word-rnn-tensorflow git:(master) ✗ python train.py
/Users/.pyenv/versions/3.6.8/lib/python3.6/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
reading text file
2019-11-26 11:04:23.936040: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
0/25700 (epoch 0), train_loss = 7.081, time/batch = 0.401
model saved to save/model.ckpt
50/25700 (epoch 0), train_loss = 6.326, time/batch = 0.396
100/25700 (epoch 0), train_loss = 5.708, time/batch = 0.300
150/25700 (epoch 0), train_loss = 5.014, time/batch = 0.290
200/25700 (epoch 0), train_loss = 4.287, time/batch = 0.295

Would anybody happen to know why this would be the case?