import pandas as pd from model import FinalCRRNNNetwork from sklearn.model_selection import train_test_split from keras.callbacks import TensorBoard from time import time from keras.optimizers import Adam tensorboard = TensorBoard(log_dir="logs/{}".format(time()), histogram_freq=2, write_graph=True) train_df = pd.read_csv('dataset/train.csv') test_df = pd.read_csv('dataset/test.csv') train_input_seqs, train_target_seqs = train_df[['input', 'expected']].values.T network = FinalCRRNNNetwork(train_df, n_grams=1, optimizer=Adam(lr=0.0004), maxseq_len=512) network.get_model_summary() """ __________________________________________________________________________________________________ Layer (type) Output Shape Param # Connected to ================================================================================================== input_1 (InputLayer) (None, 512) 0 __________________________________________________________________________________________________ embedding_1 (Embedding) (None, 512, 128) 2816 input_1[0][0] __________________________________________________________________________________________________ conv1d_1 (Conv1D) (None, 512, 128) 82048 embedding_1[0][0] __________________________________________________________________________________________________ conv1d_2 (Conv1D) (None, 512, 128) 49280 embedding_1[0][0] __________________________________________________________________________________________________ concatenate_1 (Concatenate) (None, 512, 384) 0 conv1d_1[0][0] conv1d_2[0][0] embedding_1[0][0] __________________________________________________________________________________________________ bidirectional_1 (Bidirectional) (None, 512, 400) 702000 concatenate_1[0][0] __________________________________________________________________________________________________ concatenate_2 (Concatenate) (None, 512, 784) 0 bidirectional_1[0][0] concatenate_1[0][0] __________________________________________________________________________________________________ conv1d_3 (Conv1D) (None, 512, 500) 392500 concatenate_2[0][0] __________________________________________________________________________________________________ bidirectional_2 (Bidirectional) (None, 512, 400) 841200 conv1d_3[0][0] __________________________________________________________________________________________________ concatenate_3 (Concatenate) (None, 512, 800) 0 bidirectional_1[0][0] bidirectional_2[0][0] __________________________________________________________________________________________________ conv1d_4 (Conv1D) (None, 512, 500) 400500 concatenate_3[0][0] __________________________________________________________________________________________________ bidirectional_3 (Bidirectional) (None, 512, 400) 841200 conv1d_4[0][0] __________________________________________________________________________________________________ dense_1 (Dense) (None, 512, 256) 102656 bidirectional_3[0][0] __________________________________________________________________________________________________ dense_2 (Dense) (None, 512, 128) 32896 dense_1[0][0] __________________________________________________________________________________________________ time_distributed_1 (TimeDistrib (None, 512, 9) 1161 dense_2[0][0] ================================================================================================== Total params: 3,448,257 Trainable params: 3,448,257 Non-trainable params: 0 __________________________________________________________________________________________________ """ X_train, X_val, y_train, y_val = train_test_split(train_input_seqs, train_target_seqs, test_size = .1, random_state = 123456) network.fit(X_train, y_train, X_val, y_val, epochs=20, batch_size=32) from sklearn.model_selection import KFold kfold = KFold(n_splits=10, shuffle=True, random_state=123456) i = 0 for train, test in kfold.split(train_input_seqs.reshape(-1, 1), train_target_seqs.reshape(-1, 1)): network.fit(train_input_seqs[train], train_target_seqs[train], train_input_seqs[test], train_target_seqs[test], epochs=1, batch_size=96, tensorboard=tensorboard) network.save_model(path='FinalCRRNN_model_v2{}.h5'.format(i)) i += 1 """ Train on 5909 samples, validate on 657 samples Epoch 1/1 5909/5909 [==============================] - 386s 65ms/step - loss: 0.6430 - custome_accuracy: 0.6314 - val_loss: 0.7077 - val_custome_accuracy: 0.5952 Train on 5909 samples, validate on 657 samples Epoch 1/1 5909/5909 [==============================] - 387s 66ms/step - loss: 0.6443 - custome_accuracy: 0.6313 - val_loss: 0.6296 - val_custome_accuracy: 0.6306 Train on 5909 samples, validate on 657 samples Epoch 1/1 5909/5909 [==============================] - 384s 65ms/step - loss: 0.6426 - custome_accuracy: 0.6320 - val_loss: 0.6067 - val_custome_accuracy: 0.6460 Train on 5909 samples, validate on 657 samples Epoch 1/1 5909/5909 [==============================] - 383s 65ms/step - loss: 0.6382 - custome_accuracy: 0.6337 - val_loss: 0.6190 - val_custome_accuracy: 0.6413 Train on 5909 samples, validate on 657 samples Epoch 1/1 5909/5909 [==============================] - 384s 65ms/step - loss: 0.6287 - custome_accuracy: 0.6378 - val_loss: 0.6514 - val_custome_accuracy: 0.6374 Train on 5909 samples, validate on 657 samples Epoch 1/1 5909/5909 [==============================] - 386s 65ms/step - loss: 0.6259 - custome_accuracy: 0.6395 - val_loss: 0.6258 - val_custome_accuracy: 0.6482 Train on 5910 samples, validate on 656 samples Epoch 1/1 5910/5910 [==============================] - 381s 64ms/step - loss: 0.6236 - custome_accuracy: 0.6413 - val_loss: 0.6181 - val_custome_accuracy: 0.6487 Train on 5910 samples, validate on 656 samples Epoch 1/1 5910/5910 [==============================] - 380s 64ms/step - loss: 0.6203 - custome_accuracy: 0.6435 - val_loss: 0.5989 - val_custome_accuracy: 0.6523 Train on 5910 samples, validate on 656 samples Epoch 1/1 5910/5910 [==============================] - 377s 64ms/step - loss: 0.6152 - custome_accuracy: 0.6464 - val_loss: 0.5906 - val_custome_accuracy: 0.6580 Train on 5910 samples, validate on 656 samples Epoch 1/1 5910/5910 [==============================] - 381s 64ms/step - loss: 0.6073 - custome_accuracy: 0.6505 - val_loss: 0.6164 - val_custome_accuracy: 0.6468 """ network.save_model(path='FinalCRRNN_model_v2.h5') kaggle_result = network.output_kaggle_result(test_df.id, test_df.input, test_df.len) kaggle_result.to_csv('kaggle_output_FinalCRRNN.csv', index=False)