hf_roberta_base_as_in_ireland

hf_roberta_base_as_in_ireland/01_create_datasets.py

@@ -0,0 +1,62 @@
+import datetime
+import calendar
+
+def to_fractional_year(d: datetime.datetime) -> float:
+    """
+    Converts a date stamp to a fractional year (i.e. number like `1939.781`)
+    """
+    is_leap = calendar.isleap(d.year)
+    t = d.timetuple()
+    day_of_year = t.tm_yday
+    day_time = (60 * 60 * t.tm_hour + 60 * t.tm_min + t.tm_sec) / (24 * 60 * 60)
+
+    days_in_year = 366 if is_leap else 365
+
+    return d.year + ((day_of_year - 1 + day_time) / days_in_year)
+
+def fractional_to_date(fractional):
+    eps = 0.0001
+    year = int(fractional)
+    is_leap = calendar.isleap(year)
+    
+    modulus = fractional % 1
+
+    days_in_year = 366 if is_leap else 365
+
+    day_of_year =  int( days_in_year * modulus + eps )
+
+    d = datetime.datetime(year, 1,1) + datetime.timedelta(days = day_of_year )
+
+    return d
+
+dates = (datetime.datetime(1825,10,30),
+        datetime.datetime(1825,10,31),
+        datetime.datetime(1900,1,1),
+	datetime.datetime(1900,12,1),
+        datetime.datetime(1900,12,31),
+        datetime.datetime(1930,2,28),
+        datetime.datetime(1932,2,29),
+	)
+
+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_cont\tyear\tmonth\tday\tweekday\tday_of_year\ttext\n')
+        for line_in,line_exp in zip(f_in,f_exp):
+            line_in = line_in.split('\t')[-1]
+            year_cont = float(line_exp.rstrip())
+            date = fractional_to_date(year_cont)
+            year = date.year
+            month = date.month
+            day = date.day
+            weekday = date.weekday()
+            day_of_year = date.timetuple().tm_yday
+
+            #f_hf.write(line_exp.rstrip() + '\t' + line_in)
+            f_hf.write(f'{year_cont}\t{year}\t{month}\t{day}\t{weekday}\t{day_of_year}\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_cont\tyear\tmonth\tday\tweekday\tday_of_year\ttext\n')
+        for line_in in f_in:
+            line_in = line_in.split('\t')[-1]
+            f_hf.write(f'0\t0\t0\t0\t0\t0\t{line_in}')

hf_roberta_base_as_in_ireland/02_load_dataset.py

@@ -0,0 +1,74 @@
+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
+
+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()
+        d['year_cont'] = list()
+        d['year'] = list()
+        d['month'] = list()
+        d['day'] = list()
+        d['weekday'] = list()
+        d['day_of_year'] = list()
+        d['text'] = list()
+        for l in f_in:
+            yc,y,m,day,w,dy,t= l.rstrip().split('\t')
+            d['year_cont'].append(yc)
+            d['year'].append(int(y))
+            d['month'].append(int(m))
+            d['day'].append(int(day))
+            d['weekday'].append(int(w))
+            d['day_of_year'].append(int(dy))
+            d['text'].append(t)
+    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(25))
+    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()
+scalers['year'] = MinMaxScaler().fit(np.array(train_dataset['year']).reshape(-1,1))
+
+def add_scaled(example):
+    for factor in ('year',):
+        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_as_in_ireland/03_train_pytorch_regression.py

@@ -0,0 +1,141 @@
+from config import *
+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 torch.optim import Adam
+from transformers import get_scheduler
+import torch
+from tqdm.auto import tqdm
+
+
+if TEST:
+    STEPS_EVAL = 10
+    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 = Adam(model.parameters(), lr=LR)
+
+
+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_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()
+    with torch.no_grad():
+        eval_loss = 0.0
+        dataloader = eval_dataloader_full if full else eval_dataloader_small
+        items_passed = 0 
+        for i, batch in enumerate(dataloader):
+            items_passed += len(batch)
+            batch = transform_batch(batch)
+            labels = batch['labels']
+            del batch['labels']
+            outputs = model(**batch)
+            o = soft_clip(outputs['logits']).squeeze()
+            loss = criterion(o, labels)
+            eval_loss += loss.item()
+        eval_loss = (eval_loss / items_passed)
+        print(f'eval loss full={full}: {eval_loss:.5f}', end = '\n')
+    model.train()
+    return eval_loss
+
+criterion = torch.nn.MSELoss(reduction='sum').to(device)
+
+lrelu = torch.nn.LeakyReLU(0.1)
+def soft_clip(t):
+    t = lrelu(t)
+    t = -lrelu(-t + 1 ) + 1
+    return t
+
+best_eval_loss = 9999
+epochs_without_progress = 0
+for epoch in range(NUM_EPOCHS):
+    train_loss = 0.0
+    items_passed = 0 
+    for i, batch in enumerate(train_dataloader):
+        items_passed += len(batch)
+        batch = transform_batch(batch)
+        labels = batch['labels']
+        del batch['labels']
+        outputs = model(**batch)
+        o = soft_clip(outputs['logits']).squeeze()
+        loss = criterion(o, labels)
+        loss.backward()
+        train_loss += loss.item()
+        progress_bar.update(1)
+
+        optimizer.step()
+        #lr_scheduler.step()
+        optimizer.zero_grad()
+        model.zero_grad()
+
+        if i % STEPS_EVAL == 0 and i > 1 :
+            print(f' epoch {epoch} train loss: {(train_loss /  items_passed):.5f}', end = '\t')
+            items_passed = 0
+            train_loss = 0.0
+            eval(full = False)
+
+    eval_loss =  eval(full=True)
+    model.save_pretrained(f'roberta_year_prediction/epoch_{epoch}_loss{eval_loss:.5f}')
+    model.save_pretrained(f'roberta_year_prediction/epoch_last')
+
+    if eval_loss < best_eval_loss:
+        model.save_pretrained(f'roberta_year_prediction/epoch_best')
+        print('\nsaving best model')
+        best_eval_loss = eval_loss
+    else:
+        epochs_without_progress += 1
+        print(f'epochs_witohut_progress: {epochs_without_progress}')
+
+    if epochs_without_progress > EARLY_STOPPING:
+        print('early stopping')
+        break
+
+    print(f'best_eval_loss: {best_eval_loss:5f}', end = '\n')

hf_roberta_base_as_in_ireland/04_predict.py

@@ -0,0 +1,54 @@
+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)
+
+lrelu = torch.nn.LeakyReLU(0.0)
+def soft_clip(t):
+    t = lrelu(t)
+    t = -lrelu(-t + 1 ) + 1
+    return t
+
+with open('scalers.pickle', 'rb') as f_scaler:
+    scalers = pickle.load(f_scaler)
+
+def predict(dataset, out_f):
+    eval_dataloader = DataLoader(dataset, batch_size=50)
+    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_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(soft_clip(model(**batch).logits).tolist())
+        progress_bar.update(1)
+    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_as_in_ireland/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_as_in_ireland/config.py

@@ -0,0 +1,9 @@
+#MODEL = '../MODELS/without_date/checkpoint-395000'
+MODEL = 'roberta-base'
+BATCH_SIZE = 6
+EARLY_STOPPING = 3
+WARMUP_STEPS = 10_000
+LR=1e-6
+NUM_EPOCHS = 20
+STEPS_EVAL = 5_000
+TEST=False