- Use data with feedbacks. You need two columns - text and tone mark
- Use function clean() to preprocess text and vectorization() to make a proper form of text for Tensorflow
df['Текст'] = clean(df['Текст'], normalize_form=True)
vectorize_layer = vectorization(df['Текст'], max_features=5000, max_sequence_length=128)- Use get_train_val_pipeline() to create Tensorflow datasets. You can choose split ration and batch_size, shuffling for both train and validation data
train_data, valid_data = get_train_val_pipeline(df['Текст'], df['positive'], test_size = 0.2, shuffle_train=True, shuffle_val=True,
val_batch_size=128, train_batch_size=128, shuffle_buffer_size = 10000)- Use create_model() to create model. You can define GPU or LSTM architecture of recurrent layers, number of units, bidirectional or not. Also it is possible to add Conv1D layer, number of filters and dropout to prevent over-fitting. Model uses Adam optimizer with option to set a learning rate
model = create_model(vectorize_layer, architecture='GRU', n_rnn_layers=3, rnn_units=256, loss=2e-5, embed_output_dim=256, conv1D=True,
conv1D_filters=32, bidirectional=False, dropout=0.5)Training result on 27000 feedbacks about Tinkoff bank. Result on the screen below. Model was trained on 1 GPU (Tesla K100) with 4 GRU-Bidirectional Layers. Training time with 200 epochs - near 3 hours
