Fix dev-0 files, add gonito.yaml

gonito.yaml

@@ -0,0 +1,5 @@
+description: roberta base with regression layer on top lr=1e-8 4 epochs
+tags:
+  - transformer
+  - roberta
+  - roberta-base

keras_lstm/predict.py

@@ -57,8 +57,8 @@ 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_checkpoint = sorted(os.listdir(checkpoints))[-1]
-model.load_weights('checkpoints/saved-model-000018-0.03.hdf5','rb')
+model.load_weights('checkpoints/saved-model-000006-0.02.hdf5','rb')
 
 scaler = pickle.load(open('minmaxscaler.pickle', 'rb'))
 

keras_lstm/train.py

@@ -38,22 +38,22 @@ 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')]
 
-maxlen = 500
 
-#tokenizer = Tokenizer(num_words=vocab_size)
+tokenizer = Tokenizer(num_words=vocab_size)
-#tokenizer.fit_on_texts(train_text)
+tokenizer.fit_on_texts(train_text)
-#train_text_tokenized = tokenizer.texts_to_sequences(train_text)
+train_text_tokenized = tokenizer.texts_to_sequences(train_text)
-#train_text_tokenized = pad_sequences(train_text_tokenized, padding='post', maxlen=maxlen)
+maxlen = 500
-#pickle.dump(train_text_tokenized, open('train_text_30k_for_keras.pickle', 'wb'))
+train_text_tokenized = pad_sequences(train_text_tokenized, padding='post', maxlen=maxlen)
-#pickle.dump(tokenizer, open('tokenizer.pickle', 'wb'))
+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 = [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 = tokenizer.texts_to_sequences(eval_text_tokenized)
-#eval_text_tokenized = pad_sequences(eval_text_tokenized, padding='post', maxlen=maxlen)
+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'))
+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')]
@@ -84,5 +84,5 @@ 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=0.000001), loss='mse', metrics=['mse'])
+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=100, verbose=1, validation_data = (eval_text_tokenized, eval_year_scaled), callbacks = [es, checkpoint])
+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])