Update readme.md to fix the issue #7040. (#7154)

* Update research/maskgan/README.md
Co-Authored-By: Andrew M Dai <andy.dai@gmail.com>

research/maskgan/README.md

@@ -19,7 +19,7 @@ Copy PTB data downloaded from the above tensorflow RNN tutorial to folder "/tmp/
 Make folder /tmp/pretrain-lm and copy checkpoints from above Tensorflow RNN tutorial under this folder.
 
 
-2. Run MaskGAN in MLE pretraining mode. If step 1 was not run, set
+2. Run MaskGAN in MLE pretraining mode. If step 1 was not run*, set
 `language_model_ckpt_dir` to empty.
 
 ```bash
@@ -87,6 +87,20 @@ python generate_samples.py \
  --number_epochs=4
 ```
 
+
+*  While trying to run Step 2, the following error appears:
+   NotFoundError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a Variable name or other graph    key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original    error:
+
+   Key critic/rnn/biases not found in checkpoint
+   [[node save/RestoreV2 (defined at train_mask_gan.py:431) ]]
+
+   This is an issue with seq2seq model because it uses the attention mechanism.
+   The issue arises if you saved the model with an earlier version (seq2seq is old) and restore with a recent one (saver.restore got updated).
+   The naming convention for LSTM parameters changed, e.g. cell_0/basic_lstm_cell/weights became cell_0/basic_lstm_cell/kernel.
+   Which is why you cannot restore them if you try to restore old checkpoints with recent TF.
+   The below script will help rename the variables and everything will work as expected.
+   https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/rnn/python/tools/checkpoint_convert.py
+
 ## Contact for Issues
 
 *   Liam Fedus, @liamb315 <liam.fedus@gmail.com>