lstm

keras_lstm/predict.py

@@ -0,0 +1,87 @@
+import tensorflow as tf
+import pickle
+import os
+
+gpus = tf.config.experimental.list_physical_devices('GPU')
+for gpu in gpus:
+    tf.config.experimental.set_memory_growth(gpu, True)
+
+from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
+import numpy as np
+from tensorflow.keras.preprocessing.text import Tokenizer
+
+
+from tensorflow.keras.preprocessing.sequence import pad_sequences
+from tensorflow.keras.models import Sequential
+from tensorflow.keras.layers import Activation, Dropout, Dense
+from tensorflow.keras.layers import Embedding
+import sys
+from tensorflow.keras.layers import MaxPooling1D, GlobalMaxPooling2D, LSTM, Bidirectional
+from sklearn.utils import shuffle
+
+
+vocab_size = int(sys.argv[1])
+embedding_size = int(sys.argv[2])
+LSTM_SIZE = int(sys.argv[3])
+DROPOUT_LSTM = float(sys.argv[4])
+DROPOUT_REGULAR = float(sys.argv[5])
+BATCH_SIZE = int(sys.argv[6])
+checkpoints = sys.argv[7]
+#p1 = float(sys.argv[3])
+#p2 = float(sys.argv[4])
+
+FILE='2'
+
+train_year = [float(a.rstrip('\n')) for a in open(f'../train/expected{FILE}.tsv','r')]
+max_year = max(train_year)
+min_year = min(train_year)
+
+train_text_tokenized= pickle.load(open('train_text_30k_for_keras.pickle', 'rb'))
+maxlen = max([len(a) for a in train_text_tokenized])
+
+tokenizer = pickle.load(open('tokenizer.pickle','rb'))
+test_text_tokenized = [a.rstrip('\n') for a in open('../test-A/in.tsv', 'r')]
+test_text_tokenized = tokenizer.texts_to_sequences(test_text_tokenized)
+test_text_tokenized = pad_sequences(test_text_tokenized, padding='post', maxlen=maxlen)
+
+eval_year = [float(a.rstrip()) for a in open(f'../dev-0/expected{FILE}.tsv','r')]
+
+maxlen = max([len(a) for a in train_text_tokenized])
+
+
+model = Sequential()
+embedding_layer = Embedding(vocab_size, embedding_size, input_length=maxlen , trainable=True)
+model.add(embedding_layer)
+#model.add(Bidirectional(LSTM(LSTM_SIZE, dropout = DROPOUT_LSTM, return_sequences=True)))
+model.add(Bidirectional(LSTM(LSTM_SIZE, dropout = DROPOUT_LSTM)))
+model.add(Dropout(DROPOUT_REGULAR))
+model.add(Dense(1, activation='linear'))
+
+model_checkpoint = sorted(os.listdir(checkpoints))[-1]
+model.load_weights('checkpoints/saved-model-000006-0.02.hdf5','rb')
+
+scaler = pickle.load(open('minmaxscaler.pickle', 'rb'))
+
+# EVAL PREDS
+
+eval_text_tokenized = pickle.load(open('eval_text_30k_for_keras.pickle', 'rb'))
+eval_preds = scaler.inverse_transform(model.predict(eval_text_tokenized, batch_size=20))
+eval_preds = np.minimum(eval_preds, max_year)
+eval_preds = np.maximum(eval_preds, min_year)
+
+f = open(f'../dev-0/out{FILE}.tsv', 'w')
+for i in eval_preds:
+    f.write(str(i[0]) + '\n')
+f.close()
+
+# TEST PREDS
+
+test_preds = scaler.inverse_transform(model.predict(test_text_tokenized, batch_size=20))
+test_preds = np.minimum(test_preds, max_year)
+test_preds = np.maximum(test_preds, min_year)
+
+f = open(f'../test-A/out{FILE}.tsv', 'w')
+for i in test_preds:
+    f.write(str(i[0]) + '\n')
+f.close()
+

keras_lstm/run.sh

@@ -0,0 +1,2 @@
+python  train.py 50000 300 256 0.3 0.3 400 checkpoints/
+python  predict.py 50000 300 256 0.3 0.3 400 checkpoints/

keras_lstm/train.py

@@ -0,0 +1,88 @@
+import tensorflow as tf
+import pickle
+
+gpus = tf.config.experimental.list_physical_devices('GPU')
+for gpu in gpus:
+    tf.config.experimental.set_memory_growth(gpu, True)
+
+print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))
+
+from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
+import numpy as np
+from tensorflow.keras.preprocessing.text import Tokenizer
+
+from sklearn.preprocessing import MinMaxScaler
+
+
+from tensorflow.keras.preprocessing.sequence import pad_sequences
+from tensorflow.keras.models import Sequential
+from tensorflow.keras.layers import Activation, Dropout, Dense
+from tensorflow.keras.layers import Embedding
+import sys
+from tensorflow.keras.layers import MaxPooling1D, GlobalMaxPooling2D, LSTM, Bidirectional
+from sklearn.utils import shuffle
+
+
+vocab_size = int(sys.argv[1])
+embedding_size = int(sys.argv[2])
+LSTM_SIZE = int(sys.argv[3])
+DROPOUT_LSTM = float(sys.argv[4])
+DROPOUT_REGULAR = float(sys.argv[5])
+BATCH_SIZE = int(sys.argv[6])
+checkpoints = sys.argv[7]
+#p1 = float(sys.argv[3])
+#p2 = float(sys.argv[4])
+
+FILE='2'
+
+train_text = [a.rstrip('\n') for a in open('../train/in.tsv','r')]
+train_year = [float(a.rstrip('\n')) for a in open(f'../train/expected{FILE}.tsv','r')]
+
+
+tokenizer = Tokenizer(num_words=vocab_size)
+tokenizer.fit_on_texts(train_text)
+train_text_tokenized = tokenizer.texts_to_sequences(train_text)
+maxlen = 500
+train_text_tokenized = pad_sequences(train_text_tokenized, padding='post', maxlen=maxlen)
+pickle.dump(train_text_tokenized, open('train_text_30k_for_keras.pickle', 'wb'))
+pickle.dump(tokenizer, open('tokenizer.pickle', 'wb'))
+train_text_tokenized= pickle.load(open('train_text_30k_for_keras.pickle', 'rb'))
+
+
+
+eval_text_tokenized = [a.rstrip('\n') for a in open('../dev-0/in.tsv', 'r')]
+eval_text_tokenized = tokenizer.texts_to_sequences(eval_text_tokenized)
+eval_text_tokenized = pad_sequences(eval_text_tokenized, padding='post', maxlen=maxlen)
+pickle.dump(eval_text_tokenized, open('eval_text_30k_for_keras.pickle', 'wb'))
+eval_text_tokenized = pickle.load(open('eval_text_30k_for_keras.pickle', 'rb'))
+
+eval_year = [float(a.rstrip()) for a in open(f'../dev-0/expected{FILE}.tsv','r')]
+
+
+model = Sequential()
+embedding_layer = Embedding(vocab_size, embedding_size, input_length=maxlen , trainable=True)
+model.add(embedding_layer)
+#model.add(Bidirectional(LSTM(LSTM_SIZE, dropout = DROPOUT_LSTM, return_sequences=True)))
+model.add(Bidirectional(LSTM(LSTM_SIZE, dropout = DROPOUT_LSTM)))
+
+model.add(Dropout(DROPOUT_REGULAR))
+model.add(Dense(1, activation='linear'))
+
+train_text_tokenized, train_year = shuffle(train_text_tokenized, train_year)
+
+train_year = np.array(train_year).reshape(-1,1)
+eval_year = np.array(eval_year).reshape(-1,1)
+
+scaler = MinMaxScaler().fit(train_year)
+pickle.dump(scaler, open('minmaxscaler.pickle', 'wb'))
+
+train_year_scaled = scaler.transform(train_year)
+eval_year_scaled = scaler.transform(eval_year)
+
+
+
+filepath = "./" + checkpoints + "/saved-model-{epoch:06d}-{val_loss:.2f}.hdf5"
+checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_best_only=True)
+es = EarlyStopping(monitor='val_loss', patience = 70)
+model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), loss='mse', metrics=['mse'])
+history = model.fit(train_text_tokenized, train_year_scaled, batch_size=BATCH_SIZE, epochs=5000, verbose=1, validation_data = (eval_text_tokenized, eval_year_scaled), callbacks = [es, checkpoint])