search

allenai / deep_qa

A deep NLP library, based on Keras / tf, focused on question answering (but useful for other NLP too)
Apache License 2.0
404 stars 133 forks source link

Started implementing verb_semantics_model #366

Closed bhavanadalvi closed 7 years ago

bhavanadalvi commented 7 years ago

Yes, I did that explicitly and that error is gone. But there is some other problem while accessing index of verb_array_input:

Does this look familiar to you?

tests/models/sequence_tagging/verb_semantics_model_test.py:19:

tests/common/test_case.py:76: in ensure_model_trains_and_loads model.train() deep_qa/training/trainer.py:357: in train self._save_auxiliary_files() deep_qa/training/text_trainer.py:223: in _save_auxiliary_files super(TextTrainer, self)._save_auxiliary_files() deep_qa/training/trainer.py:619: in _save_auxiliary_files model_config = self.model.to_json() ../../anaconda/lib/python3.5/site-packages/keras/engine/topology.py:2546: in to_json model_config = self._updated_config() ../../anaconda/lib/python3.5/site-packages/keras/engine/topology.py:2513: in _updated_config config = self.get_config()

self = <deep_qa.training.models.DeepQaModel object at 0x129c48a58>

def get_config(self):
    config = {
        'name': self.name,
    }
    node_conversion_map = {}
    for layer in self.layers:
        if issubclass(layer.__class__, Container):
            # Containers start with a pre-existing node
            # linking their input to output.
            kept_nodes = 1
        else:
            kept_nodes = 0
        for original_node_index, node in enumerate(layer.inbound_nodes):
            node_key = layer.name + '_ib-' + str(original_node_index)
            if node_key in self.container_nodes:
                node_conversion_map[node_key] = kept_nodes
                kept_nodes += 1
    layer_configs = []
    for layer in self.layers:  # From the earliest layers on.
        layer_class_name = layer.__class__.__name__
        layer_config = layer.get_config()
        filtered_inbound_nodes = []
        for original_node_index, node in enumerate(layer.inbound_nodes):
            node_key = layer.name + '_ib-' + str(original_node_index)
            if node_key in self.container_nodes:
                # The node is relevant to the model:
                # add to filtered_inbound_nodes.
                if node.arguments:
                    try:
                        json.dumps(node.arguments)
                        kwargs = node.arguments
                    except TypeError:
                        warnings.warn(
                            'Layer ' + layer.name +
                            ' was passed non-serializable keyword

arguments: ' + str(node.arguments) + '. They will not be included ' 'in the serialized model (and thus will be missing ' 'at deserialization time).') kwargs = {} else: kwargs = {} if node.inbound_layers: node_data = [] for i in range(len(node.inbound_layers)): inbound_layer = node.inbound_layers[i] node_index = node.node_indices[i] tensor_index = node.tensor_indices[i] node_key = inbound_layer.name + '_ib-' + str(node_index) new_node_index = node_conversion_map.get(node_key, 0) node_data.append([inbound_layer.name, new_node_index, tensor_index, kwargs]) filtered_inbound_nodes.append(node_data) layer_configs.append({ 'name': layer.name, 'class_name': layer_class_name, 'config': layer_config, 'inbound_nodes': filtered_inbound_nodes, }) config['layers'] = layer_configs

    # Gather info about inputs and outputs.
    model_inputs = []
    for i in range(len(self.input_layers)):
        layer = self.input_layers[i]
        node_index = self.input_layers_node_indices[i]
        node_key = layer.name + '_ib-' + str(node_index)
      new_node_index = node_conversion_map[node_key]

E KeyError: 'verb_array_input_ib-0'

../../anaconda/lib/python3.5/site-packages/keras/engine/topology.py:2298: KeyError

Captured stdout call

Layer (type) Output Shape Param # Connected to Input mask Output mask

====================================================================================================================================================== word_array_input (InputLayer) (None, 8) 0 None None

word_embedding (TimeDistributedEmbeddin (None, 8, 6) 90 word_array_input[0][0] None Tensor("word_embedding/No

dropout_1 (Dropout) (None, 8, 6) 0 word_embedding[0][0] Tensor("word_embedding/No Tensor("word_embedding/No

bow_encoder_1 (BOWEncoder) (None, 6) 0 dropout_1[0][0] Tensor("word_embedding/No None

dense_2 (Dense) (None, 3) 21 bow_encoder_1[0][0] None None

time_distributed_1 (TimeDistributed) (None, 8, 3) 21 dropout_1[0][0] Tensor("word_embedding/No Tensor("word_embedding/No

Total params: 132

On Mon, May 22, 2017 at 1:55 PM, Mark Neumann notifications@github.com wrote:

@DeNeutoy commented on this pull request.

In deep_qa/models/sequence_tagging/verb_semantics_model.py https://github.com/allenai/deep_qa/pull/366#discussion_r117843333:

  • seq2seq_encoders, then predicts a tag at each index.
  • Parameters
  • num_stacked_rnns : int, optional (default: 1)
  • The number of seq2seq_encoders that we should stack on top of each other before
  • predicting tags.
  • instance_type : str
  • Specifies the particular subclass of TaggedSequenceInstance to use for loading data,
  • which in turn defines things like how the input data is formatted and tokenized.
  • """
  • def init(self, params: Params):
  • self.num_stacked_rnns = params.pop('num_stacked_rnns', 1)
  • instance_type = params.pop('instance_type', "VerbSemanticsInstance")
  • self.instance_type = concrete_instances[instance_type]
  • super(VerbSemanticsModel, self).init(params)

@bhavanadalvi https://github.com/bhavanadalvi: re matt's comment in his email, you need to add the following line here:

params.setdefault('validation_metric', 'val_loss')

This is because we are using some Keras functionality which allows us to save a model when some metric has fallen/improved. There is a default for this in keras if your model has a single output, but as yours has 2, you need to set it explicitly, before you call the super class.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/allenai/deep_qa/pull/366#pullrequestreview-39583775, or mute the thread https://github.com/notifications/unsubscribe-auth/AKc_2YmrclveRr-hoKBe2W_EDgRXATrcks5r8fYqgaJpZM4NivfU .

DeNeutoy commented 7 years ago

This is happening because the verb_array and entity_array are not part of the model - add them in somewhere and this should fix that.