roberta no year from scratch

roberta_no_year_from_scratch/01_create_datasets.py

@@ -0,0 +1,17 @@
+from config import LABELS_DICT
+
+with open('../test-A/in.tsv','r') as f_in, open(f'../test-A/huggingface_format.tsv', 'w') as f_hf:
+    f_hf.write('text\n')
+    for line_in in f_in:
+        _,_, text = line_in.split('\t')
+        f_hf.write(text)
+
+
+for dataset in 'train', 'dev-0':
+    with open(f'../{dataset}/in.tsv') as f_in, open(f'../{dataset}/expected.tsv') as f_exp, open(f'../{dataset}/huggingface_format.tsv','w') as f_hf:
+        f_hf.write('text\tlabel\n')
+        for line_in, line_exp in zip(f_in, f_exp):
+            label = LABELS_DICT[line_exp.rstrip('\n')]
+            _,_,text = line_in.rstrip('\n').split('\t')
+            f_hf.write(text +'\t'+ str(label) + '\n')
+

roberta_no_year_from_scratch/02_load_dataset.py

@@ -0,0 +1,32 @@
+import pickle
+from datasets import load_dataset
+from transformers import AutoTokenizer, RobertaTokenizer
+from config import MODEL
+
+dataset = load_dataset('csv', sep='\t', data_files={'train': ['../train/huggingface_format.tsv'], 'test': ['../dev-0/huggingface_format.tsv']})
+test_dataset = load_dataset('csv', sep='\t', data_files ='../test-A/huggingface_format.tsv')
+
+tokenizer = RobertaTokenizer.from_pretrained(MODEL)
+
+def tokenize_function(examples):
+    return tokenizer(examples["text"], padding="max_length", truncation=True)
+
+tokenized_datasets = dataset.map(tokenize_function, batched=True)
+test_tokenized_datasets = test_dataset.map(tokenize_function, batched=True)
+
+train_dataset = tokenized_datasets["train"].shuffle(seed=42)
+eval_dataset_full = tokenized_datasets["test"]
+eval_dataset_small = tokenized_datasets["test"].select(range(2000))
+test_dataset = test_tokenized_datasets["train"]
+
+with open('train_dataset.pickle','wb') as f_p:
+    pickle.dump(train_dataset, f_p)
+
+with open('eval_dataset_small.pickle','wb') as f_p:
+    pickle.dump(eval_dataset_small, f_p)
+
+with open('eval_dataset_full.pickle','wb') as f_p:
+    pickle.dump(eval_dataset_full, f_p)
+
+with open('test_dataset.pickle','wb') as f_p:
+    pickle.dump(test_dataset, f_p)

roberta_no_year_from_scratch/03_train.py

@@ -0,0 +1,78 @@
+import pickle
+from config import LABELS_LIST, MODEL
+
+with open('train_dataset.pickle','rb') as f_p:
+    train_dataset = pickle.load(f_p)
+
+with open('eval_dataset_small.pickle','rb') as f_p:
+    eval_dataset_small = pickle.load(f_p)
+
+with open('eval_dataset_full.pickle','rb') as f_p:
+    eval_dataset_full = pickle.load(f_p)
+
+with open('test_dataset.pickle','rb') as f_p:
+    test_dataset = pickle.load(f_p)
+
+
+from transformers import RobertaTokenizer, RobertaForSequenceClassification, RobertaConfig
+
+#model = RobertaForSequenceClassification(RobertaConfig(num_labels=7))
+model = RobertaForSequenceClassification.from_pretrained('roberta-base',num_labels=7)
+model = RobertaForSequenceClassification(model.config)
+
+from transformers import TrainingArguments
+
+
+training_args = TrainingArguments("test_trainer",
+        per_device_train_batch_size=4,
+        per_device_eval_batch_size=4,
+        evaluation_strategy='steps',
+        #eval_steps=2_000,
+        #save_steps=2_000,
+        eval_steps=20_000,
+        save_steps=20_000,
+        num_train_epochs=20,
+        gradient_accumulation_steps=2,
+        learning_rate = 1e-6,
+        #warmup_steps=4_000,
+        warmup_steps=4,
+        load_best_model_at_end=True,
+        )
+
+import numpy as np
+from datasets import load_metric
+
+metric = load_metric("accuracy")
+
+def compute_metrics(eval_pred):
+    logits, labels = eval_pred
+    predictions = np.argmax(logits, axis=-1)
+    return metric.compute(predictions=predictions, references=labels)
+
+
+from transformers import Trainer
+
+trainer = Trainer(
+            model=model,
+            args=training_args,
+            train_dataset=train_dataset,
+            eval_dataset=eval_dataset_small,
+            compute_metrics=compute_metrics,
+            )
+
+trainer.train(resume_from_checkpoint=True)
+#trainer.train()
+trainer.save_model("./roberta-retrained")
+trainer.evaluate()
+
+
+eval_predictions = trainer.predict(eval_dataset_full).predictions.argmax(1)
+
+with open('../dev-0/out.tsv', 'w') as f_out:
+    for pred in eval_predictions:
+        f_out.write(LABELS_LIST[pred] + '\n')
+
+test_predictions = trainer.predict(test_dataset).predictions.argmax(1)
+with open('../test-A/out.tsv', 'w') as f_out:
+    for pred in test_predictions:
+        f_out.write(LABELS_LIST[pred] + '\n')

roberta_no_year_from_scratch/04_predict.py

@@ -0,0 +1,24 @@
+import pickle
+from config import LABELS_LIST, MODEL
+from transformers import AutoTokenizer
+from tqdm import tqdm
+
+device = 'cpu'
+
+
+from transformers import AutoModelForSequenceClassification
+
+model = AutoModelForSequenceClassification.from_pretrained('test_trainer/checkpoint-620000/')
+tokenizer = AutoTokenizer.from_pretrained(MODEL)
+
+for dataset in ('dev-0', 'test-A'):
+    with open(f'../{dataset}/in.tsv') as f_in, open(f'../{dataset}/out.tsv','w') as f_out:
+        for line_in in tqdm(f_in, total=150_000):
+            _,_, text = line_in.split('\t')
+            text = text.rstrip('\n')
+            inputs = tokenizer(text, padding=True, truncation=True, return_tensors="pt").to(device)
+            outputs = model(**inputs)
+            probs = outputs[0].softmax(1)
+            prediction = LABELS_LIST[probs.argmax(1)]
+            f_out.write(prediction + '\n')
+

roberta_no_year_from_scratch/config.py

@@ -0,0 +1,18 @@
+LABELS_DICT = {'news':0,
+        'sport':1,
+        'business':2,
+        'opinion':3,
+        'culture':4,
+        'lifestyle':5,
+        'removed':6}
+
+
+LABELS_LIST = ['news',
+        'sport',
+        'business',
+        'opinion',
+        'culture',
+        'lifestyle',
+        'removed']
+
+MODEL = 'roberta-base'