train.py

# -*- coding: utf-8 -*
#!pip install transformers
#!pip install pandas
#!pip install numpy
#!pip install SentencePiece

import sys, argparse
import torch
from transformers import T5ForConditionalGeneration, T5Tokenizer
import pandas as pd
import os
import numpy as np
from tqdm.auto import tqdm, trange
import gc
from datetime import datetime
import time

st = time.time() #start time

parser = argparse.ArgumentParser()
parser.add_argument("-n","--model_name", type=str, default="d2t_model", required=False, help="Specify model name")
parser.add_argument("-e","--epochs", type=int, default=100, required=False, help="Specify training epochs")
args = parser.parse_args()

model_name = args.model_name
epochs = args.epochs

print("Model name: " + model_name + " Epochs: " + str(epochs))  


"""# Modelo T5
Importamos o modelo preadestrado
"""

model = T5ForConditionalGeneration.from_pretrained('google/mt5-base')
tokenizer = T5Tokenizer.from_pretrained('google/mt5-base')
model.cuda();
optimizer = torch.optim.Adam(params=[p for p in model.parameters() if p.requires_grad], lr=1e-5)


#Load dataset (dataset-gl.csv or dataset-es.csv)
all_data = pd.read_csv('./datasets/dataset-gl.csv', encoding="latin-1")

#seleccionamos 2733 registros para training (seria la particion 70-30 en dataset-es.csv)
#en dataset-gl.csv contamos con  mas registros, por lo que en test habria 500 en lugar de 300  casos

train_split = all_data.iloc[:2733, :]
test_split = all_data.iloc[2733:, :]

#Clean dataset rows
train_split=train_split.dropna()
train_split=train_split.dropna(axis=0)
train_split=train_split.reset_index()
print(torch.cuda.list_gpu_processes())

def split_batches(df, batch_size):
    batches = []
    for i in range(0, len(df), batch_size):
        if (i+batch_size) > len(df):
            batches.append(df[i:])
        else:
            batches.append(df[i: i+batch_size])
    return batches


def cleanup():
    gc.collect()
    torch.cuda.empty_cache()

cleanup()

optimizer.param_groups[0]['lr'] = 1e-5

"""# Adestramento"""

model.train();
batch_size = 8
max_len = 384
accumulation_steps = 1
save_steps = 1
epochs_tq = trange(epochs) #epochs

window = 4000
ewm = 0
errors = 0

cleanup()

batches = split_batches(train_split, batch_size)

for i in epochs_tq:
    print("Epoch:", i)
    batch_count = 0
    for batch in batches:
        batch_count += 1
        print("Batch:", batch_count)
        xx = batch.table.values.tolist()
        yy = batch.table.values.tolist()
        try:
          x = tokenizer(xx, return_tensors='pt', padding=True, truncation=True, max_length=max_len).to(model.device)
          y = tokenizer(yy, return_tensors='pt', padding=True, truncation=True, max_length=max_len).to(model.device)
          # do not force the model to predict pad tokens
          y.input_ids[y.input_ids==0] = -100

          loss = model(
              input_ids=x.input_ids,
              attention_mask=x.attention_mask,
              labels=y.input_ids,
              decoder_attention_mask=y.attention_mask,
              return_dict=True
          ).loss
          loss.backward()

        except RuntimeError as e:
            errors += 1
            print("ERROR")
            print(i, x.input_ids.shape[1], y.input_ids.shape[1], e)
            loss = None
            cleanup()
            continue

        w = 1 / min(i+1, window)
        ewm = ewm * (1-w) + loss.item() * w
        epochs_tq.set_description(f'loss: {ewm}')

        if i % accumulation_steps == 0:
            optimizer.step()
            optimizer.zero_grad()
            cleanup()

        if i % window == 0 and i > 0:
            print(ewm, errors)
            errors = 0
            cleanup()
            # optimizer.param_groups[0]['lr'] *= 0.999
        if i % save_steps == 0 and i > 0:
            model.save_pretrained(model_name + "_" + str(epochs))
            tokenizer.save_pretrained(model_name + "_" + str(epochs))
            print('saving...', i, optimizer.param_groups[0]['lr'])

model.save_pretrained(model_name + "_" + str(epochs))
tokenizer.save_pretrained(model_name + "_" + str(epochs))

total_time = time.time() - st
print("Training time:", time.strftime("%H:%M:%S", time.gmtime(total_time)))

"""# Test"""
model.eval();

def generate(text):
    x = tokenizer(text, return_tensors='pt', padding=True).to(model.device)
    out = model.generate(**x, do_sample=False, num_beams=10, max_length=100)
    return tokenizer.decode(out[0], skip_special_tokens=True)

with open(f"{model_name}_{epochs}_predictions_{datetime.now()}.txt", "w") as f:
    f.write("Training time:" + str(time.strftime("%H:%M:%S", time.gmtime(total_time))))
    for index, row in test_split.iterrows():
        text_id = str(row["id"])
        text1 = str(row["table"])
        text2 = str(row["caption"])

        f.write(text_id + "\n" + text1 + "\n")
        print(text_id + "\n" + text1)
        f.write("Prediction:\n")
        f.write(generate(text1) + "\n")
        print(generate(text1))
        f.write("Truth:\n")
        f.write(text2 + "\n\n")
        print(text2)