README.md

Model_X

Model_X provides keras-layer like state-of-the-art general-purpose architectures for Natural Language Understanding (NLU) and Natural Language Generation (NLG).

Implementation

1. BiLSTM+BiGRU Architectures

a. BiLSTMGRUSpatialDropout1D

from model_X.bilstm_architectures import *
from model_X.dense_architectures import DenseLayerModel
from tensorflow.keras.layers import *
from tensorflow.keras.models import Model

input_shape = (100,)
model_input = Input(shape=input_shape)
bilstm_layers = BiLSTMGRUSpatialDropout1D(10, 100)(model_input)
dense_layers = DenseLayerModel()(bilstm_layers)
output = Dense(3, activation='softmax')(dense_layers)
full_model = Model(inputs=model_input, outputs=output)
print(full_model.summary())

b. BiLSTMGRUSelfAttention

from model_X.bilstm_architectures import *
from model_X.dense_architectures import DenseLayerModel
from tensorflow.keras.layers import *
from tensorflow.keras.models import Model

input_shape = (100,)
model_input = Input(shape=input_shape)
bilstm_layers = BiLSTMGRUAttention(10, 100)(model_input)
dense_layers = DenseLayerModel()(bilstm_layers)
output = Dense(3, activation='softmax')(dense_layers)
full_model = Model(inputs=model_input, outputs=output)
print(full_model.summary())

c. BiLSTMGRUMultiHeadAttention

from model_X.bilstm_architectures import *
from model_X.dense_architectures import DenseLayerModel
from tensorflow.keras.layers import *
from tensorflow.keras.models import Model

input_shape = (100,)
model_input = Input(shape=input_shape)
bilstm_layers = BiLSTMGRUMultiHeadAttention(10, 100)(model_input)
dense_layers = DenseLayerModel()(bilstm_layers)
output = Dense(3, activation='softmax')(dense_layers)
full_model = Model(inputs=model_input, outputs=output)
print(full_model.summary())

d. SplitBiLSTMGRUSpatialDropout1D

from model_X.bilstm_architectures import *
from model_X.dense_architectures import DenseLayerModel
from tensorflow.keras.layers import *
from tensorflow.keras.models import Model

input_shape = (100,)
model_input = Input(shape=input_shape)
bilstm_layers = SplitBiLSTMGRUSpatialDropout1D(10, 100)(model_input)
dense_layers = DenseLayerModel()(bilstm_layers)
output = Dense(3, activation='softmax')(dense_layers)
full_model = Model(inputs=model_input, outputs=output)
print(full_model.summary())

e. SplitBiLSTMGRU

from model_X.bilstm_architectures import *
from model_X.dense_architectures import DenseLayerModel
from tensorflow.keras.layers import *
from tensorflow.keras.models import Model

input_shape = (100,)
model_input = Input(shape=input_shape)
bilstm_layers = SplitBiLSTMGRU(10, 100)(model_input)
dense_layers = DenseLayerModel()(bilstm_layers)
output = Dense(3, activation='softmax')(dense_layers)
full_model = Model(inputs=model_input, outputs=output)
print(full_model.summary())

2. Dense Architectures

a. DenseLayerModel

from model_X.dense_architectures import DenseLayerModel
from tensorflow.keras.layers import *
from tensorflow.keras.models import Model

input_shape = (100,)
model_input = Input(shape=input_shape)
dense_layers = DenseLayerModel()(model_input)
output = Dense(3, activation='softmax')(dense_layers)
full_model = Model(inputs=model_input, outputs=output)
print(full_model.summary())

3 Transformer Architectures

a. VanillaTransformer

from transformers_architectures import *
from tensorflow.keras.layers import *
from tensorflow.keras.models import Model
import argparse

config = argparse.Namespace(vocab_size=1000,
                        embed_dim=512,
                        ff_dim=32,
                        num_heads=8,
                        rate=0.1,
                        maxlen=128)

inputs = tf.keras.layers.Input(shape=(config.maxlen,))
pooled_output,sequence_output = VanillaTransformer(config)(inputs)
output = Dense(3, activation='softmax')(pooled_output)
full_model = Model(inputs=model_input, outputs=output)
print(full_model.summary())