hf_roberta_base

hf_roberta_base/01_create_datasets.py

@@ -0,0 +1,16 @@
+
+for split in 'train', 'dev-0':
+    with open(f'../{split}/in.tsv') as f_in, open(f'../{split}/expected.tsv') as f_exp, open(f'./{split}_huggingface_format.csv', 'w') as f_hf:
+        f_hf.write('year_start_float\tyear_end_float\tyear_middle_float\tyear_middle_int\ttext\n')
+        for line_in, line_exp in zip(f_in, f_exp):
+            year_start_float, year_end_float = line_exp.rstrip().split(',')
+            year_middle_float = (float(year_start_float) + float(year_end_float)) / 2
+            year_middle_int = round(year_middle_float)
+            f_hf.write(f'{year_start_float}\t{year_end_float}\t{year_middle_float}\t{year_middle_int}\t{line_in}')
+
+for split in ('test-A',):
+    with open(f'../{split}/in.tsv') as f_in,  open(f'./{split}_huggingface_format.csv', 'w') as f_hf:
+        f_hf.write('year_start_float\tyear_end_float\tyear_middle_float\tyear_middle_int\ttext\n')
+        for line_in in f_in:
+            expected = '0.0\t0.0\t0.0\t0'
+            f_hf.write(expected + '\t' + line_in)

hf_roberta_base/02_load_dataset.py

@@ -0,0 +1,70 @@
+from config import MODEL, TEST
+import pickle
+from datasets import load_dataset, Dataset
+from transformers import AutoTokenizer
+from tqdm import tqdm
+from sklearn.preprocessing import MinMaxScaler
+import numpy as np
+
+values = ('year_start_float', 'year_end_float', 'year_middle_float', 'year_middle_int', 'text')
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL)
+def tokenize_function(examples):
+    t = tokenizer(examples["text"], padding="max_length", truncation=True, max_length=512)
+    return t
+
+def get_dataset_dict(dataset):
+    with open(dataset) as f_in:
+        next(f_in)
+        d = dict()
+        for v in values:
+            d[v] = list()
+        for l in f_in:
+            args = l.rstrip().split('\t')
+            for v, a in zip(values, args):
+                d[v].append(a)
+    return d
+
+train_dataset = Dataset.from_dict(get_dataset_dict('train_huggingface_format.csv')).map(tokenize_function, batched=True).shuffle(seed=42)
+eval_dataset_full = Dataset.from_dict(get_dataset_dict('dev-0_huggingface_format.csv')).map(tokenize_function, batched=True)
+eval_dataset_small = eval_dataset_full.shuffle(seed=42).select(range(2000))
+test_dataset_A = Dataset.from_dict(get_dataset_dict('test-A_huggingface_format.csv')).map(tokenize_function, batched=True)
+
+if TEST:
+    train_dataset = train_dataset.select(range(500))
+    eval_dataset_full = eval_dataset_full.select(range(400))
+    eval_dataset_small = eval_dataset_small.select(range(50))
+    test_dataset_A = test_dataset_A.select(range(200))
+
+
+scalers = dict()
+values_to_scale = ('year_start_float', 'year_end_float', 'year_middle_float')
+for v in values_to_scale:
+    scalers[v] = MinMaxScaler().fit(np.array(train_dataset[v]).reshape(-1, 1))
+
+def add_scaled(example):
+    for factor in values_to_scale:
+        example[factor + '_scaled'] = scalers[factor].transform(np.array(example[factor]).reshape(-1,1)).reshape(1,-1)[0].item()
+    return example
+
+
+train_dataset = train_dataset.map(add_scaled)
+eval_dataset_full = eval_dataset_full.map(add_scaled)
+eval_dataset_small = eval_dataset_small.map(add_scaled)
+test_dataset_A = test_dataset_A.map(add_scaled)
+
+
+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_A.pickle', 'wb') as f_p:
+    pickle.dump(test_dataset_A, f_p)
+
+with open('scalers.pickle', 'wb') as f_p:
+    pickle.dump(scalers, f_p)

hf_roberta_base/03_train_pytorch_regression.py

@@ -0,0 +1,127 @@
+from config import MODEL, TEST
+import pickle
+from datasets import load_dataset
+from transformers import AutoTokenizer, RobertaModel, RobertaTokenizer
+from torch.utils.data import DataLoader
+from transformers import AutoModelForSequenceClassification
+from transformers import AdamW
+from transformers import get_scheduler
+import torch
+from tqdm.auto import tqdm
+
+BATCH_SIZE = 1
+EARLY_STOPPING = 3
+WARMUP_STEPS = 10_000
+
+STEPS_EVAL = 5_000
+
+if TEST:
+    STEPS_EVAL = 100
+    WARMUP_STEPS = 10
+
+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)
+
+
+
+train_dataloader = DataLoader(train_dataset, shuffle=True, batch_size=BATCH_SIZE)
+eval_dataloader_small = DataLoader(eval_dataset_small, batch_size=BATCH_SIZE)
+eval_dataloader_full = DataLoader(eval_dataset_full, batch_size=BATCH_SIZE)
+
+model = AutoModelForSequenceClassification.from_pretrained(MODEL, num_labels=1)
+optimizer = AdamW(model.parameters(), lr=1e-6)
+
+
+num_epochs = 1
+num_training_steps = num_epochs * len(train_dataloader)
+lr_scheduler = get_scheduler(
+    "linear",
+    optimizer=optimizer,
+    num_warmup_steps=WARMUP_STEPS,
+    num_training_steps=num_training_steps
+)
+
+
+device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+model.to(device)
+
+
+progress_bar = tqdm(range(num_training_steps))
+model.train()
+
+model.train()
+model.to(device)
+
+def transform_batch(batch):
+        batch['input_ids'] = torch.stack(batch['input_ids']).permute(1,0).to(device)
+        batch['attention_mask'] = torch.stack(batch['attention_mask']).permute(1,0).to(device)
+        batch['labels'] = batch['year_middle_float_scaled'].to(device).float()
+
+        batch['labels'].to(device)
+        batch['input_ids'].to(device)
+        batch['attention_mask'].to(device)
+
+        for c in set(batch.keys()) - {'input_ids', 'attention_mask', 'labels'}:
+            del batch[c]
+        return batch
+
+
+def eval(full=False):
+    model.eval()
+    eval_loss = 0.0
+    dataloader = eval_dataloader_full if full else eval_dataloader_small
+    for i, batch in enumerate(dataloader):
+        batch = transform_batch(batch)
+        outputs = model(**batch)
+        loss = outputs.loss
+        eval_loss += loss.item()
+    print(f'epoch {epoch} eval loss: {eval_loss /  i }')
+    model.train()
+    return eval_loss
+
+
+best_eval_loss = 9999
+epochs_without_progress = 0
+for epoch in range(num_epochs):
+    train_loss = 0.0
+    for i, batch in enumerate(train_dataloader):
+        batch = transform_batch(batch)
+        outputs = model(**batch)
+        loss = outputs.loss
+        loss.backward()
+        train_loss += loss.item()
+        progress_bar.update(1)
+
+        # DELAYED UPDATE
+        #if i % 16 == 1 and i > 1:
+        #    optimizer.step()
+        #    #lr_scheduler.step()
+        #    optimizer.zero_grad()
+
+        # DELAYED UPDATE
+        optimizer.step()
+        lr_scheduler.step()
+        optimizer.zero_grad()
+
+        if i % STEPS_EVAL == 0 and i > 1:
+            print(f' epoch {epoch} train loss: {train_loss /  STEPS_EVAL }', end='\t\t')
+            train_loss = 0.0
+            eval(full = False)
+
+    model.save_pretrained(f'roberta_year_prediction/epoch_{epoch}')
+    eval_loss = eval(full=True)
+
+    if eval_loss < best_eval_loss:
+        model.save_pretrained(f'roberta_year_prediction/epoch_best')
+        best_eval_loss = eval_loss
+    else:
+        epochs_without_progress += 1
+
+    if epochs_without_progress > EARLY_STOPPING:
+        break

hf_roberta_base/04_predict.py

@@ -0,0 +1,48 @@
+import pickle
+import torch
+from transformers import AutoModelForSequenceClassification
+from torch.utils.data import DataLoader
+from tqdm.auto import tqdm
+
+with open('eval_dataset_full.pickle','rb') as f_p:
+    eval_dataset_full = pickle.load(f_p)
+
+with open('test_dataset_A.pickle','rb') as f_p:
+    test_dataset = pickle.load(f_p)
+
+device = 'cuda'
+model = AutoModelForSequenceClassification.from_pretrained('./roberta_year_prediction/epoch_best')
+model.eval()
+model.to(device)
+
+with open('scalers.pickle', 'rb') as f_scaler:
+    scalers = pickle.load(f_scaler)
+
+def predict(dataset, out_f):
+    eval_dataloader = DataLoader(dataset, batch_size=1)
+    outputs = []
+
+    progress_bar = tqdm(range(len(eval_dataloader)))
+
+    for batch in eval_dataloader:
+        batch['input_ids'] = torch.stack(batch['input_ids']).permute(1,0).to(device)
+        batch['attention_mask'] = torch.stack(batch['attention_mask']).permute(1,0).to(device)
+        batch['labels'] = batch['year_middle_float_scaled'].to(device).float()
+
+        batch['labels'].to(device)
+        batch['input_ids'].to(device)
+        batch['attention_mask'].to(device)
+
+        for c in set(batch.keys()) - {'input_ids', 'attention_mask', 'labels'}:
+            del batch[c]
+        outputs.extend(model(**batch).logits.tolist())
+        progress_bar.update(1)
+    outputs_transformed = scalers['year_middle_float'].inverse_transform(outputs)
+
+    with open(out_f,'w') as f_out:
+
+        for o in outputs_transformed:
+            f_out.write(str(o[0]) + '\n')
+
+predict(eval_dataset_full, '../dev-0/out.tsv')
+predict(test_dataset, '../test-A/out.tsv')

hf_roberta_base/04_predict_from_file.py

@@ -0,0 +1,53 @@
+import pickle
+import torch
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+from torch.utils.data import DataLoader
+from tqdm.auto import tqdm
+
+#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_A.pickle','rb') as f_p:
+#    test_dataset_A = pickle.load(f_p)
+
+with open('dev-0_huggingface_format.csv','r') as f_p:
+    eval_dataset_full = f_p.readlines()
+
+with open('test-A_huggingface_format.csv','r') as f_p:
+    test_dataset = f_p.readlines()
+
+device = 'cuda'
+model = AutoModelForSequenceClassification.from_pretrained('./roberta_year_prediction/epoch_best')
+tokenizer = AutoTokenizer.from_pretrained('roberta-base')
+model.eval()
+model.to(device)
+
+with open('scalers.pickle', 'rb') as f_scaler:
+    scalers = pickle.load(f_scaler)
+
+tokenizer = AutoTokenizer.from_pretrained('roberta-base')
+def predict(dataset, out_f):
+    outputs = []
+
+    for sample in tqdm(dataset[1:]):
+        y, t = sample.split('\t')
+        t = t.rstrip()
+
+        t = tokenizer(t, padding="max_length", truncation=True, max_length=512, return_tensors='pt').to('cuda')
+
+        outputs.extend(model(**t).logits.tolist())
+        outputs_transformed = scalers['year'].inverse_transform(outputs)
+
+    with open(out_f,'w') as f_out:
+
+        for o in outputs_transformed:
+            f_out.write(str(o[0]) + '\n')
+
+predict(eval_dataset_full, '../dev-0/out.tsv')
+predict(test_dataset, '../test-A/out.tsv')

hf_roberta_base/config.py

@@ -0,0 +1,4 @@
+#MODEL = '/home/wmi/RoBERTa/without_date/checkpoint-1325000'
+MODEL = 'roberta-base'
+TEST=True
+