Training rasa custom keras policy using rasa Agent

I have developed my custom keras policy using model functional API of keras. As rasa ‘keras policy’ has developed for Sequential model of keras, I have customized every functions and its arguments in my custom keras policy.

When I have imported my custom keras policy and tried to train using the command ‘rasa train core’, it throws me an error.

> async def train_core(
>     domain_file: Text = "domain.yml",
>     model_directory: Text = "models",
>     model_name: Text = "current",
>     training_data_file: Text = "data/",
> ):
>     agent = Agent(
>         domain_file,
>         policies=[
>             MemoizationPolicy(max_history=3),
>             MappingPolicy(),
>             Custom_KerasPolicy(batch_size=64, epochs=100, validation_split=0.2),
>         ],
>     )

I get the following error,

TypeError: Style() takes no arguments

Could I get help to solve this issue?

@sarankarthik could you share the full stack trace? And also the code of your custom keras policy, as well as the configuration you used in the config file?

@akelad Thanks for the Response! The above-mentioned Problem is due to a dependency prompt-toolkit, which I have installed later. But now, I get a new error as follows,

Traceback (most recent call last): File “c:\ica_trial\rasa\rasa\core\”, line 20, in policy_from_module_path return class_from_module_path(module_path, lookup_path=“rasa.core.policies”) File “c:\ica_trial\rasa\rasa\utils\”, line 174, in class_from_module_path m = importlib.import_module(module_name) File “c:\apps\sa2446\lib\”, line 127, in import_module return _bootstrap._gcd_import(name[level:], package, level) File “”, line 1006, in _gcd_import File “”, line 983, in _find_and_load File “”, line 967, in _find_and_load_unlocked File “”, line 677, in _load_unlocked File “”, line 728, in exec_module File “”, line 219, in _call_with_frames_removed File “C:\ICA\restaurantbot\”, line 47, in from Lstm import CUSTOM_LSTM File “C:\ICA\restaurantbot\”, line 16, in from keras.engine.base_layer import Layer, disable_tracking, InputSpec ImportError: cannot import name ‘disable_tracking’ from ‘keras.engine.base_layer’ (c:\apps\sa2446\lib\site-packages\keras\engine\

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File “c:\ica_trial\rasa\rasa\core\policies\”, line 316, in from_dict constr_func = registry.policy_from_module_path(policy_name) File “c:\ica_trial\rasa\rasa\core\”, line 22, in policy_from_module_path raise ImportError(“Cannot retrieve policy from path ‘{}’”.format(module_path)) ImportError: Cannot retrieve policy from path ‘custom_keras_policy.Custom_KerasPolicy’

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File “C:\Apps\sa2446\Scripts\”, line 11, in load_entry_point(‘rasa’, ‘console_scripts’, ‘rasa’)() File “c:\ica_trial\rasa\”, line 76, in main cmdline_arguments.func(cmdline_arguments) File “c:\ica_trial\rasa\rasa\cli\”, line 111, in train_core kwargs=kwargs, File “c:\ica_trial\rasa\rasa\”, line 252, in train_core kwargs=kwargs, File “c:\apps\sa2446\lib\asyncio\”, line 579, in run_until_complete return future.result() File “c:\ica_trial\rasa\rasa\”, line 308, in train_core_async kwargs=kwargs, File “c:\ica_trial\rasa\rasa\”, line 340, in _train_core_with_validated_data kwargs=kwargs, File “c:\ica_trial\rasa\rasa\core\”, line 42, in train policies = config.load(policy_config) File “c:\ica_trial\rasa\rasa\core\”, line 28, in load return PolicyEnsemble.from_dict(config_data) File “c:\ica_trial\rasa\rasa\core\policies\”, line 324, in from_dict “”.format(policy_name) rasa.core.policies.ensemble.InvalidPolicyConfig: Module for policy ‘custom_keras_policy.Custom_KerasPolicy’ could not be loaded. Please make sure the name is a valid policy.

This is my config file,

language: “en”


  • name: “WhitespaceTokenizer”

  • name: “RegexFeaturizer”

  • name: “CountVectorsFeaturizer”

  • name: “EmbeddingIntentClassifier”

    hidden_layers_sizes_a: [1024, 512] hidden_layers_sizes_b: [] batch_size: [256, 1054] epochs: 100 embed_dim: 20 mu_pos: 0.8 # should be 0.0 < … < 1.0 for ‘cosine’ mu_neg: -0.4 # should be -1.0 < … < 1.0 for ‘cosine’ similarity_type: “cosine” # string ‘cosine’ or ‘inner’ num_neg: 20 use_max_sim_neg: true

    C2: 0.0015 C_emb: 0.6 droprate: 0.25

    intent_tokenization_flag: true intent_split_symbol: “+”

    evaluate_every_num_epochs: 10 # small values may hurt performance evaluate_on_num_examples: 1000 # large values may hurt performance

  • name: “CRFEntityExtractor”

    BILOU_flag: true

    max_iterations: 50

    L1_c: 0.1

    L2_c: 0.1

  • name: “EntitySynonymMapper”


  • name: “custom_keras_policy.Custom_KerasPolicy” batch_size: 100 epochs: 100 validation_split: 0.2
  • name: MemoizationPolicy
  • name: MappingPolicy

My custom model_architecture has shown below,

def model_architecture(self, data_path) -> keras.models.Models:        

     def train(self, data_path):

            #Get encoder inputs, decoder inputs and decoder targets from each splitted data to train in progressive encoder
            en_in_data1, en_ti = EncoderData.encoder_data(self, data_path, column)
            de_in_data1, de_out_data1, de_ti = DecoderData.decoder_data(self, data_path, column)
            _, _,num_encoder_tokens, _ = EncoderData.data_prep(self, data_path, column)
            _, _,num_decoder_tokens, _, _ = DecoderData.data_prep(self, data_path, column)

            encoder_inputs = Input(shape=(None, num_encoder_tokens))
            encoder = LSTM(latent_dim, return_sequences = True, return_state=True)
            encoder1 = LSTM(latent_dim, return_state=True)
            en_out, e_h, e_c = encoder(encoder_inputs)
            e_s = [e_h, e_c]

            encoder_outputs, state_h, state_c = encoder1(en_out, initial_state=e_s)

            encoder_states = [state_h, state_c]
  "encoder output shape:{encoder_outputs}")

            decoder_inputs = Input(shape=(None, num_decoder_tokens))

            decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
            decoder_lstm1 = LSTM(latent_dim, return_sequences=True, return_state=True)
            decoder_outputs1, d_h, d_c = decoder_lstm(decoder_inputs,
            d_s = [d_h, d_c]
            decoder_outputs, de_h, de_c = decoder_lstm1(decoder_outputs1, initial_state=d_s)

            decoder_dense = Dense(num_decoder_tokens, activation='softmax')

            decoder_outputs = decoder_dense(decoder_outputs)
            #Define Model input and output shapes for initial column 
            model1 = Model([encoder_inputs, decoder_inputs], decoder_outputs)

            model1.compile(optimizer='rmsprop', loss='categorical_crossentropy',


            return model1

      model1 = train(self, data_path)

@akelad could you tell me how to integrate my custom keras policy into rasa?

It seems that Rasa agent starts to train the network only for the listed policies offered by Rasa.(rasa/ at master · RasaHQ/rasa · GitHub, rasa/ at master · RasaHQ/rasa · GitHub)

Please not the function ‘train’ in both the scripts. Could you help me in this situation?

@sarankarthik you just need to add it to your config.yml file and then it will also train the custom policy

@akelad, Thanks for the information. I have added it in my config.yml file and I am able to train the model. But when I chat, it throws me the following error,

    2019-12-06 15:14:05 INFO     root  - Starting Rasa server on http://localhost:5005
Using TensorFlow backend.
Bot loaded. Type a message and press enter (use '/stop' to exit): 
Your input ->  hello                                                                                                                                                                                              
2019-12-06 15:14:22 ERROR    asyncio  - Task exception was never retrieved
future: <Task finished coro=<configure_app.<locals>.run_cmdline_io() done, defined at C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\> exception=TimeoutError()>
Traceback (most recent call last):
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\", line 119, in run_cmdline_io
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\channels\", line 137, in record_messages
    async for response in bot_responses:
  File "c:\apps\sa2446\lib\site-packages\async_generator\", line 366, in step
    return await ANextIter(self._it, start_fn, *args)
  File "c:\apps\sa2446\lib\site-packages\async_generator\", line 205, in throw
    return self._invoke(self._it.throw, type, value, traceback)
  File "c:\apps\sa2446\lib\site-packages\async_generator\", line 209, in _invoke
    result = fn(*args)
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\channels\", line 102, in send_message_receive_stream
    async for line in resp.content:
  File "c:\apps\sa2446\lib\site-packages\aiohttp\", line 40, in __anext__
    rv = await self.read_func()
  File "c:\apps\sa2446\lib\site-packages\aiohttp\", line 329, in readline
    await self._wait('readline')
  File "c:\apps\sa2446\lib\site-packages\aiohttp\", line 297, in _wait
    await waiter
  File "c:\apps\sa2446\lib\site-packages\aiohttp\", line 585, in __exit__
    raise asyncio.TimeoutError from None
2019-12-06 15:14:22 ERROR    asyncio  - Task exception was never retrieved
future: <Task finished coro=<RestInput.on_message_wrapper() done, defined at C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\channels\> exception=TypeError("'>' not supported between instances of 'NoneType' and 'int'")>
Traceback (most recent call last):
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\channels\", line 387, in on_message_wrapper
    await on_new_message(message)
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\channels\", line 65, in handler
    await app.agent.handle_message(*args, **kwargs)
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\", line 488, in handle_message
    return await processor.handle_message(message)
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\", line 90, in handle_message
    await self._predict_and_execute_next_action(message, tracker)
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\", line 353, in _predict_and_execute_next_action
    action, policy, confidence = self.predict_next_action(tracker)
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\", line 180, in predict_next_action
    action_confidences, policy = self._get_next_action_probabilities(tracker)
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\", line 565, in _get_next_action_probabilities
    tracker, self.domain
  File "C:\Users\SA2446\AppData\Roaming\Python\Python37\site-packages\rasa\core\policies\", line 353, in probabilities_using_best_policy
    if (confidence, p.priority) > (max_confidence, best_policy_priority):
TypeError: '>' not supported between instances of 'NoneType' and 'int'

I have also modified the Prediction_action_probabilities() method according to my requirements, as I am using seq2seq model. I have acquired the current state of dialogue using rasa dialogue state tracker and I ffed it as inputs to get the states values of encoder, which I then feed it to decoder model to predict the output as follows,

states_value = encoder_model.predict(input_seq)   [equivalent to  'X = self.featurizer.create_X([tracker], domain)']

I feed this states value to decoder model,

output_tokens, h, c = decoder_model.predict([target_seq] + states_value) [equivalent to 'y_pred = self.model.predict(X, batch_size=1)']

I use the one hot encoded information of 3D vector of

np.zeros((len(texts), max_seq_length, num_of_features), dtype='float32')

Could I get help to solve this issue?

I mean it seems like your policy doesn’t have a priority from that error message? Did you make sure to define the priority property?

Thanks for the information. It works properly now.