This library is an implementation of Heirarchical Attention Networks for Document Classification (Yang etal,2015).Link. This is compatible with Keras and Tensorflow (keras version >=2.0.6). As the paper suggests, it uses hierarchical attention mechanism and capabilities of Word Encoder (including bi-directional Recurrent unit- GRU) ,Sentence Attention and Document Classification are addressed.
The library or the Layer is compatible with Tensorflow and Keras. Installation is carried out using the pip command as follows:
pip install MiniAttention==0.1For using inside the Jupyter Notebook or Python IDE (along with Keras layers):
import MiniAttention.MiniAttention as MAThe Layer takes as input a 3D Tensor with dimensions: (sample_size,steps,features) The Layer as an output produces a 2D Tensor with dimensions: (sample_size,features) This Layer can be used after the Keras.Embedding() Layer to provide a global attention using the features and Embedding weights.Additionally Embedding libraries like (Glove,Word2Vec,ELMO) can be used in the Embedding Layer. It can also be used in between and before LSTM/Bidirectional LSTM/ GRU and recurrent layers . In this context, it has a capability to cater to Sequential and Model (functional) types of Model architectures(specific for Keras). The functional (keras.models.Model) version is provided as follows:
inp=Input(shape=(inp_shape,))
z=Embedding(max_features,256)(inp)
z=MA.MiniAttentionBlock(keras.initializers.he_uniform,None,None,keras.regularizers.L2(l2=0.02),None,None,None,None,None)(z)
z=tf.keras.layers.Bidirectional(LSTM(128,recurrent_activation="relu",return_sequences=True))(z)
z=tf.keras.layers.Bidirectional(LSTM(64,recurrent_activation="relu",return_sequences=True))(z)
z=MA.MiniAttentionBlock(keras.initializers.he_uniform,None,None,keras.regularizers.L2(l2=0.02),None,None,None,None,None)(z)
z=keras.layers.Dense(64,activation="relu")(z)
z=keras.layers.Dense(64,activation="relu")(z)
z=keras.layers.Dense(1,activation="sigmoid")(z)
model=keras.models.Model(inputs=inp,outputs=z)
model.compile(loss="binary_crossentropy",metrics=['accuracy'],optimizer=keras.optimizers.Adagrad(learning_rate=1e-3))
model.summary()For Sequential Model (keras.models.Sequential):
model = Sequential()
model.add(Embedding(max_features,128,input_shape=(100,)))
model.add(MA.MiniAttentionBlock(None,None,None,None,None,None,None,None,None))
model.add(LSTM(128))
model.add(Dense(8,activation='relu'))
model.add(Dense(4,activation='sigmoid'))
model.compile(loss='binary_crossentropy',metrics=['accuracy'],optimizer='Adagrad')
model.summary()The arguments for the MiniAttentionBlock class include:
1.W_init: Weight Initializer - Compatible with keras.initializers
2.b_init: Bias Initializer - Compatible with keras.initializers
3.u_init: Output Initializer -Compatible with keras.initializers
4.W_reg: Weight Regularizer - Compatible with keras.regularizers
5.b_reg: Bias Regularizer - Compatible with keras.regularizers
6.u_reg: Output Regularizer -Compatible with keras.regularizers
7.W_const: Weight Constraint - Compatible with keras.constraints
8.b_const: Bias Constraint - Compatible with keras.constraints
9.u_const: Output Constraint -Compatible with keras.constraints
10.bias: Boolean - True/False- Whether to use bias (Optional)
There are 3 main functions inside the MiniAttentionBlock class. The "init" method is used for initializeing the weight , bias tensors for computation. The "attention_block" method is used for assigning the variables (tensors) and checking for the input tensor size. The "build_nomask" method is used for computing the attention modules.Uses tanh as the internal activation function with exponential normalization.Masking has not been added to the library yet.
For reference on how to use the library, a Jupyter Notebook sample is present in the repository: "MiniAttention_on_IMDB.ipynb".
This is a sample which uses this Layer with Keras.layers.Embedding() Layer in IMDB binary classification.It uses the default keras embedding which is followed from the official tutorial by Keras docs. Alternately,"Tensorboard-tfds-IMDB.py" contains a tensorboard demonstration.
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
MIT