Sketch-RNN ln7 Error: could not cast hparam

[antocreo]

I am trying to run the Sketch-RN demo Sketch_RNN.ipynb.
when I run:
[train_set, valid_set, test_set, hps_model, eval_hps_model, sample_hps_model] = load_env(data_dir, model_dir)

this is the error I get:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-7-c072ee5d7a3f> in <module>()
----> 1 [train_set, valid_set, test_set, hps_model, eval_hps_model, sample_hps_model] = load_env(data_dir, model_dir)

/Users/libphd/anaconda3/envs/magenta/lib/python2.7/site-packages/magenta/models/sketch_rnn/sketch_rnn_train.pyc in load_env(data_dir, model_dir)
     73   model_params = sketch_rnn_model.get_default_hparams()
     74   with tf.gfile.Open(os.path.join(model_dir, 'model_config.json'), 'r') as f:
---> 75     model_params.parse_json(f.read())
     76   return load_dataset(data_dir, model_params, inference_mode=True)
     77 

/Users/libphd/anaconda3/envs/magenta/lib/python2.7/site-packages/tensorflow/contrib/training/python/training/hparam.pyc in parse_json(self, values_json)
    585     """
    586     values_map = json.loads(values_json)
--> 587     return self.override_from_dict(values_map)
    588 
    589   def values(self):

/Users/libphd/anaconda3/envs/magenta/lib/python2.7/site-packages/tensorflow/contrib/training/python/training/hparam.pyc in override_from_dict(self, values_dict)
    537     """
    538     for name, value in values_dict.items():
--> 539       self.set_hparam(name, value)
    540     return self
    541 

/Users/libphd/anaconda3/envs/magenta/lib/python2.7/site-packages/tensorflow/contrib/training/python/training/hparam.pyc in set_hparam(self, name, value)
    499         raise ValueError(
    500             'Must pass a list for multi-valued parameter: %s.' % name)
--> 501       setattr(self, name, _cast_to_type_if_compatible(name, param_type, value))
    502 
    503   def parse(self, values):

/Users/libphd/anaconda3/envs/magenta/lib/python2.7/site-packages/tensorflow/contrib/training/python/training/hparam.pyc in _cast_to_type_if_compatible(name, param_type, value)
    173   # Avoid converting a number or string type to a boolean or vice versa.
    174   if issubclass(param_type, bool) != isinstance(value, bool):
--> 175     raise ValueError(fail_msg)
    176 
    177   # Avoid converting float to an integer (the reverse is fine).

ValueError: Could not cast hparam 'conditional' of type '<type 'bool'>' from value 1

Can't figure out the problem though...

[antocreo]

seems the problem is with TF 1.6
I've installed TF 1.2.1 and it works well...

[ShusenTang]

this problem will appear since TF 1.5.
the reason is tf can't recognise the 0/1 as the bool type da in the model's json file such as "site-packages/magenta/models/aaron_sheep/layer_norm/model_config.json".
If you don't want to install old version tensorflow, you should change all ''0/1'' into ''false/true" if the paramter type is bool.
for example, the changed model_config.json file should like this:

[antocreo]

@ShusenTang thanks for replying. I am using 1.2.1 at the moment and it works alright. I'll definitely do what you suggested if I switch to TF>1.5 :) cheers

[hardmaru]

Hi @ShusenTang

Thanks for the suggestion. When TF was in the 1.2 days the casting was limited so I used 0/1 but your suggestion works well for current TF versions (including 1.8)