Update for the new version

gonito.yaml

@@ -0,0 +1,4 @@
+description: tfidf with linear regression
+tags:
+  - linear-regression
+  - tf-idf

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.load_weights('checkpoints/saved-model-000018-0.03.hdf5','rb')
+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'))
 

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.fit_on_texts(train_text)
-#train_text_tokenized = tokenizer.texts_to_sequences(train_text)
-#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'))
+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 = [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')]
@@ -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'])
-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])
+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])

roberta_regressor_head/train.py

@@ -65,7 +65,7 @@ class RegressorHead(torch.nn.Module):
 
 regressor_head = RegressorHead().to(device)
 
-optimizer = torch.optim.Adam(list(roberta.parameters()) + list(regressor_head.parameters()), lr=1e-8)
+optimizer = torch.optim.Adam(list(roberta.parameters()) + list(regressor_head.parameters()), lr=1e-6)
 criterion = torch.nn.MSELoss(reduction='sum').to(device)
 
 BATCH_SIZE = 1
@@ -118,8 +118,7 @@ def eval_short():
     loss = 0.0
     loss_clipped = 0.0
     loss_scaled = 0.0
-    eval_num = 10000
-    for batch, year in tqdm(get_train_batch(dev_in[:eval_num],dev_year_scaled[:eval_num])):
+    for batch, year in tqdm(get_train_batch(dev_in[:1000],dev_year_scaled[:1000])):
 
         x = regressor_head(batch.to(device)).squeeze()
         x_clipped = torch.clamp(x,0.0,1.0)
@@ -131,8 +130,8 @@ def eval_short():
         loss_scaled += criterion_eval(x, year).item()
         loss += criterion_eval(original_x, original_year).item()
         loss_clipped += criterion_eval(original_x_clipped, original_year).item()
-    print('valid loss scaled: ' + str(np.sqrt(loss_scaled/eval_num)))
-    print('valid loss: ' + str(np.sqrt(loss/eval_num)))
+    print('valid loss scaled: ' + str(np.sqrt(loss_scaled/1000)))
+    print('valid loss: ' + str(np.sqrt(loss/1000)))
     print('valid loss clipped: ' + str(np.sqrt(loss_clipped/len(dev_year))))
 
 

tfidflr/train_predict.py

@@ -10,18 +10,20 @@ import random
 import pickle
 import sys
 
+import lzma
+
 def tokenizer_space(text):
     return text.split(' ')
 
-type = sys.argv[1] # 1 or 2 
+type = sys.argv[1] # 1 or 2
 
 
 def run():
 
     # LOADING DATA
-    train_text = [a.rstrip('\n') for a in open('../train/in.tsv','r')]
-    dev_text = [a.rstrip('\n') for a in open('../dev-0/in.tsv','r')]
-    test_text = [a.rstrip('\n') for a in open('../test-A/in.tsv','r')]
+    train_text = [a.rstrip('\n') for a in lzma.open('../train/in.tsv.xz', 'rt')]
+    dev_text = [a.rstrip('\n') for a in lzma.open('../dev-0/in.tsv.xz', 'rt')]
+    test_text = [a.rstrip('\n') for a in lzma.open('../test-A/in.tsv.xz', 'rt')]
     global lowest
 
     train_year = [float(a.rstrip('\n')) for a in open(f'../train/expected{type}.tsv','r')]