challenging-america-geo-pre.../keras_lstm/train.py
2021-06-10 13:49:16 +02:00

89 lines
3.4 KiB
Python

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])