from locale import strcoll from datasets import load_dataset import numpy as np import torch from torch import optim from torch.nn import functional as F from transformers import AutoModelForSeq2SeqLM, AutoTokenizer from transformers.optimization import Adafactor from transformers import get_linear_schedule_with_warmup from tqdm.notebook import tqdm import random import sacrebleu import os import pandas as pd from sklearn.model_selection import train_test_split import torch.multiprocessing as mp from torch.multiprocessing import Process, Queue from joblib import Parallel, delayed,parallel_backend import sys from functools import partial import json import time import numpy as np from datetime import datetime class Config(): def __init__(self,args) -> None: self.homepath = args.homepath self.prediction_path = os.path.join(args.homepath,args.prediction_path) # Use 'google/mt5-small' for non-pro cloab users self.model_repo = 'google/mt5-base' self.model_path_dir = args.homepath self.model_name = f'{args.model_name}.pt' self.bt_data_dir = os.path.join(args.homepath,args.bt_data_dir) #Data part self.parallel_dir= os.path.join(args.homepath,args.parallel_dir) self.mono_dir= os.path.join(args.homepath,args.mono_dir) self.log = os.path.join(args.homepath,args.log) self.mono_data_limit = args.mono_data_limit self.mono_data_for_noise_limit=args.mono_data_for_noise_limit #Training params self.n_epochs = args.n_epochs self.n_bt_epochs=args.n_bt_epochs self.batch_size = args.batch_size self.max_seq_len = args.max_seq_len self.min_seq_len = args.min_seq_len self.checkpoint_freq = args.checkpoint_freq self.lr = 1e-4 self.print_freq = args.print_freq self.use_multiprocessing = args.use_multiprocessing self.num_cores = mp.cpu_count() self.NUM_PRETRAIN = args.num_pretrain_steps self.NUM_BACKTRANSLATION_TIMES =args.num_backtranslation_steps self.do_backtranslation=args.do_backtranslation self.now_on_bt=False self.bt_time=0 self.using_reconstruction= args.use_reconstruction self.num_return_sequences_bt=2 self.use_torch_data_parallel = args.use_torch_data_parallel self.gradient_accumulation_batch = args.gradient_accumulation_batch self.num_beams = args.num_beams self.best_loss = 1000 self.best_loss_delta = 0.00000001 self.patience=args.patience self.L2=0.0000001 self.dropout=args.dropout self.drop_prob=args.drop_probability self.num_swaps=args.num_swaps self.verbose=args.verbose self.now_on_test=False #Initialization of state dict which will be saved during training self.state_dict = {'batch_idx': 0,'epoch':0,'bt_time':self.bt_time,'best_loss':self.best_loss} self.state_dict_check = {'batch_idx': 0,'epoch':0,'bt_time':self.bt_time,'best_loss':self.best_loss} #this is for tracing training after abrupt end! self.device = torch.device('cuda' if True and torch.cuda.is_available() else 'cpu') #We will be leveraging parallel and monolingual data for each of these languages. #parallel data will be saved in a central 'parallel_data 'folder as 'src'_'tg'_parallel.tsv #monolingual data will be saved in another folder called 'monolingual_data' as 'lg'_mono.tsv #Each tsv file is of the form "input", "output" self.LANG_TOKEN_MAPPING = { 'ig': '', 'fon': '', 'en': '', 'fr': '', 'rw':'', 'yo':'', 'xh':'', 'sw':'' } self.truncation=True def beautify_time(time): hr = time//(3600) mins = (time-(hr*3600))//60 rest = time -(hr*3600) - (mins*60) #DARIA's implementation! sp = "" if hr >=1: sp += '{} hours'.format(hr) if mins >=1: sp += ' {} mins'.format(mins) if rest >=1: sp += ' {} seconds'.format(rest) return sp def word_delete(x,config): noise=[] words = x.split(' ') if len(words) == 1: return x for w in words: a= np.random.choice([0,1], 1, p=[config.drop_prob, 1-config.drop_prob]) if a[0]==1: #It means don't delete noise.append(w) #if you end up deleting all words, just return a random word if len(noise) == 0: rand_int = random.randint(0, len(words)-1) return [words[rand_int]] return ' '.join(noise) def swap_word(new_words): random_idx_1 = random.randint(0, len(new_words)-1) random_idx_2 = random_idx_1 counter = 0 while random_idx_2 == random_idx_1: random_idx_2 = random.randint(0, len(new_words)-1) counter += 1 if counter > 3: return new_words new_words[random_idx_1], new_words[random_idx_2] = new_words[random_idx_2], new_words[random_idx_1] return new_words def random_swap(words, n): words = words.split() new_words = words.copy() for _ in range(n): new_words = swap_word(new_words) sentence = ' '.join(new_words) return sentence def get_dict(input,target,src,tgt): inp = [i for i in input] target_ = [ i for i in target] s= [src for i in range(len(inp))] t = [tgt for i in range(len(target_))] return [{'inputs':inp_,'targets':target__,'src':s_,'tgt':t_} for inp_,target__,s_,t_ in zip(inp,target_,s,t)] def get_dict_mono(input,src,config): index = [i for i in range(len(input))] ids = random.sample(index,config.mono_data_limit) inp = [input[i] for i in ids] s= [src for i in range(len(inp))] data=[] for lang in config.LANG_TOKEN_MAPPING.keys(): if lang!=src and lang not in ['en','fr']: data.extend([{'inputs':inp_,'src':s_,'tgt':lang} for inp_,s_ in zip(inp,s)]) return data def get_dict_mono_noise(input,src,config): index = [i for i in range(len(input))] ids = random.sample(index,config.mono_data_for_noise_limit) inp = [input[i] for i in ids] noised = [word_delete(random_swap(str(x),config.num_swaps),config) for x in inp] s= [src for i in range(len(inp))] data=[] data.extend([{'inputs':noise_,'targets':inp_,'src':s_,'tgt':s_} for inp_,s_,noise_ in zip(inp,s,noised)]) return data def compress(input,target,src,tgt): return {'inputs':input,'targets':target,'src':src,'tgt':tgt} def make_dataset(config,mode): if mode!='eval' and mode!='train' and mode!='test': raise Exception('mode is either train or eval or test!') else: files = [f.name for f in os.scandir(config.parallel_dir) ] files = [f for f in files if f.split('.')[-1]=='tsv' and f.split('.tsv')[0].endswith(mode) and len(f.split('_'))>2 ] data = [(f_.split('_')[0],f_.split('_')[1],pd.read_csv(os.path.join(config.parallel_dir,f_), sep="\t")) for f_ in files] dict_ = [get_dict(df['input'],df['target'],src,tgt) for src,tgt,df in data] return [item for sublist in dict_ for item in sublist] def get_model_translation(config,model,tokenizer,sentence,tgt): if config.use_torch_data_parallel: max_seq_len_ = model.module.config.max_length else: max_seq_len_ = model.config.max_length input_ids = encode_input_str(config,text = sentence,target_lang = tgt,tokenizer = tokenizer,seq_len = max_seq_len_).unsqueeze(0).to(config.device) if config.use_torch_data_parallel: out = model.module.generate(input_ids,num_beams=3,do_sample=True, num_return_sequences=config.num_return_sequences_bt,max_length=config.max_seq_len,min_length=config.min_seq_len) else: out = model.generate(input_ids,num_beams=3, do_sample=True,num_return_sequences=config.num_return_sequences_bt,max_length=config.max_seq_len,min_length=config.min_seq_len) out_id = [i for i in range(config.num_return_sequences_bt)] id_ = random.sample(out_id,1) return tokenizer.decode(out[id_][0], skip_special_tokens=True) def do_job(t,id_,tokenizers): tokenizer = tokenizers[id_ % len(tokenizers)] #We flip the input as target and vice versa in order to have target-side backtranslation (where source side is synthetic). return {'inputs':get_model_translation(config,model,tokenizer,t['inputs'],t['tgt']),'targets':t['inputs'],'src':t['tgt'],'tgt':t['src']} #return {'inputs':t['inputs'],'targets':get_model_translation(config,model,tokenizer,t['inputs'],t['tgt']),'src':t['src'],'tgt':t['tgt']} def do_job_pmap(t): #tokenizer = tokenizers[id_ % len(tokenizers)] return {'inputs':t['inputs'],'targets':get_model_translation(config,model,tokenizer,t['inputs'],t['tgt']),'src':t['src'],'tgt':t['tgt']} def do_job_pool(bt_data,model,id_,tokenizers,config,mono_data): tokenizer = tokenizers[id_] if config.verbose: print(f"Mono data inside job pool: {mono_data}") sys.stdout.flush() res = [{'inputs':t['inputs'],'targets':get_model_translation(config,model,tokenizer,t['inputs'],t['tgt']),'src':t['src'],'tgt':t['tgt']} for t in mono_data] bt_data.put(res) return None def mono_data_(config): #Find and prepare all the mono data in the directory files_ = [f.name for f in os.scandir(config.mono_dir) ] files = [f for f in files_ if f.endswith('tsv') and f.split('.tsv')[0].endswith('mono')] if config.verbose: print("Generating data for back translation") print(f"Files found in mono dir: {files}") data = [(f_.split('_')[0],pd.read_csv(os.path.join(config.mono_dir,f_), sep="\t")) for f_ in files] dict_ = [get_dict_mono(df['input'],src,config) for src,df in data] mono_data = [item for sublist in dict_ for item in sublist] return mono_data def mono_data_noise(config): #Find and prepare all the mono data in the directory files_ = [f.name for f in os.scandir(config.mono_dir) ] files = [f for f in files_ if f.endswith('tsv') and f.split('.tsv')[0].endswith('mono')] if config.verbose: print("Generating data for back translation") print(f"Files found in mono dir: {files}") data = [(f_.split('_')[0],pd.read_csv(os.path.join(config.mono_dir,f_), sep="\t")) for f_ in files] dict_ = [get_dict_mono_noise(df['input'],src,config) for src,df in data] mono_data = [item for sublist in dict_ for item in sublist] return mono_data def get_mono_data(config,model): mono_data = mono_data_(config) if config.use_multiprocessing: if config.verbose: print(f"Using multiprocessing on {config.num_cores} processes") if __name__ == "__main__": ctx = mp.get_context('spawn') #mp.set_start_method("spawn",force=True) bt_data = ctx.Queue() model.share_memory() num_processes = config.num_cores NUM_TO_USE = len(mono_data)//num_processes mini_mono_data = [mono_data[i:i + NUM_TO_USE] for i in range(0, len(mono_data), NUM_TO_USE)] #print(f"Length of mini mono data {len(mini_mono_data)}. Length of processes: {num_processes}") assert len(mini_mono_data) == num_processes, "Length of mini mono data and number of processes do not match." num_processes_range = [i for i in range(num_processes)] processes = [] for rank,data_ in tqdm(zip(num_processes_range,mini_mono_data)): p = ctx.Process(target=do_job_pool, args=(bt_data,model,rank,tokenizers_for_parallel,config,data_)) p.start() if config.verbose: print(f"Bt data: {bt_data.get()}") sys.stdout.flush() processes.append(p) for p in processes: p.join() return bt_data #output = multiprocessing.Queue() #multiprocessing.set_start_method("spawn",force=True) #pool = mp.Pool(processes=config.num_cores) #bt_data = [pool.apply(do_job, args=(data_,i,tokenizers_for_parallel,)) for i,data_ in enumerate(mono_data)] ''' # Setup a list of processes that we want to run processes = [mp.Process(target=do_job, args=(5, output)) for x in range(config.num_cores)] if __name__ == "__main__": #pool = mp.Pool(processes=config.num_cores) with parallel_backend('loky'): bt_data = Parallel(n_jobs = config.num_cores, require='sharedmem')(delayed(do_job)(data_,i,tokenizers_for_parallel) for i,data_ in enumerate(mono_data)) ''' else: bt_data = [{'inputs':t['inputs'],'targets':get_model_translation(config,model,tokenizer,t['inputs'],t['tgt']),'src':t['src'],'tgt':t['tgt']} for t in tqdm(mono_data)] return bt_data def encode_input_str(config,text, target_lang, tokenizer, seq_len): target_lang_token = config.LANG_TOKEN_MAPPING[target_lang] # Tokenize and add special tokens input_ids = tokenizer.encode( text = str(target_lang_token) + str(text), return_tensors = 'pt', padding = 'max_length', truncation = config.truncation, max_length = seq_len) return input_ids[0] def encode_target_str(config,text, tokenizer, seq_len): token_ids = tokenizer.encode( text = str(text), return_tensors = 'pt', padding = 'max_length', truncation = config.truncation, max_length = seq_len) return token_ids[0] def format_translation_data(config,sample,tokenizer,seq_len): # sample is of the form {'inputs':input,'targets':target,'src':src,'tgt':tgt} # Get the translations for the batch input_lang = sample['src'] target_lang = sample['tgt'] input_text = sample['inputs'] target_text = sample['targets'] if input_text is None or target_text is None: return None input_token_ids = encode_input_str(config,input_text, target_lang, tokenizer, seq_len) target_token_ids = encode_target_str(config,target_text, tokenizer, seq_len) return input_token_ids, target_token_ids def transform_batch(config,batch,tokenizer,max_seq_len): inputs = [] targets = [] for sample in batch: formatted_data = format_translation_data(config,sample,tokenizer,max_seq_len) if formatted_data is None: continue input_ids, target_ids = formatted_data inputs.append(input_ids.unsqueeze(0)) targets.append(target_ids.unsqueeze(0)) batch_input_ids = torch.cat(inputs) batch_target_ids = torch.cat(targets) return batch_input_ids, batch_target_ids def get_data_generator(config,dataset,tokenizer,max_seq_len,batch_size): random.shuffle(dataset) for i in range(0, len(dataset), batch_size): raw_batch = dataset[i:i+batch_size] yield transform_batch(config,raw_batch, tokenizer,max_seq_len) def eval_model(config,tokenizer,model, gdataset, max_iters=8): test_generator = get_data_generator(config,gdataset,tokenizer,config.max_seq_len, config.batch_size) eval_losses = [] for i, (input_batch, label_batch) in enumerate(test_generator): input_batch, label_batch = input_batch.to(config.device), label_batch.to(config.device) model_out = model.forward( input_ids = input_batch, labels = label_batch) if config.use_torch_data_parallel: loss = torch.mean(model_out.loss) else: loss = model_out.loss eval_losses.append(loss.item()) return np.mean(eval_losses) def evaluate(config,tokenizer,model,test_dataset,src_lang=None,tgt_lang=None): if src_lang!=None and tgt_lang!=None: if config.verbose: with open(config.log,'a+') as fl: print(f"Getting evaluation set for source language -> {src_lang} and target language -> {tgt_lang}",file=fl) data = [t for t in test_dataset if t['src']==src_lang and t['tgt']==tgt_lang] else: data= [t for t in test_dataset] inp = [t['inputs'] for t in data] truth = [t['targets'] for t in data] tgt_lang_ = [t['tgt'] for t in data] seq_len__ = config.max_seq_len input_tokens = [encode_input_str(config,text = inp[i],target_lang = tgt_lang_[i],tokenizer = tokenizer,seq_len =seq_len__).unsqueeze(0).to(config.device) for i in range(len(inp))] if config.use_torch_data_parallel: output = [model.module.generate(input_ids, num_beams=config.num_beams, num_return_sequences=1,max_length=config.max_seq_len,min_length=config.min_seq_len) for input_ids in tqdm(input_tokens)] else: output = [model.generate(input_ids, num_beams=config.num_beams, num_return_sequences=1,max_length=config.max_seq_len,min_length=config.min_seq_len) for input_ids in tqdm(input_tokens)] output = [tokenizer.decode(out[0], skip_special_tokens=True) for out in tqdm(output)] df= pd.DataFrame({'predictions':output,'truth':truth,'inputs':inp}) if config.now_on_bt and config.using_reconstruction: filename = f'{src_lang}_{tgt_lang}_bt_{config.bt_time}_rec.tsv' elif config.now_on_bt: filename = f'{src_lang}_{tgt_lang}_bt_{config.bt_time}.tsv' elif config.now_on_test: filename = f'{src_lang}_{tgt_lang}_TEST.tsv' else: filename = f'{src_lang}_{tgt_lang}.tsv' df.to_csv(os.path.join(config.prediction_path,filename),sep='\t',index=False) try: spbleu = sacrebleu.corpus_bleu(output, [truth]) except Exception: raise Exception(f'There is a problem with {src_lang}_{tgt_lang}. Truth is {truth} \n Input is {inp} ') return spbleu.score def do_evaluation(config,tokenizer,model,test_dataset): LANGS = list(config.LANG_TOKEN_MAPPING.keys()) if config.now_on_bt and config.using_reconstruction: s=f'---------------------------AFTER BACKTRANSLATION {config.bt_time} with RECONSTRUCTION---------------------------'+'\n' elif config.now_on_bt: s=f'---------------------------AFTER BACKTRANSLATION {config.bt_time}---------------------------'+'\n' elif config.now_on_test: s=f'---------------------------TESTING EVALUATION---------------------------'+'\n' else: s=f'---------------------------EVALUATION ON DEV---------------------------'+'\n' for i in range(len(LANGS)): for j in range(len(LANGS)): if LANGS[j]!=LANGS[i]: eval_bleu = evaluate(config,tokenizer,model,test_dataset,src_lang=LANGS[i],tgt_lang=LANGS[j]) a = f'Bleu Score for {LANGS[i]} to {LANGS[j]} -> {eval_bleu} '+'\n' s+=a s+='------------------------------------------------------' with open(os.path.join(config.homepath,'bleu_log.txt'), 'a+') as fl: print(s,file=fl) def train(config,n_epochs,optimizer,tokenizer,train_dataset,dev_dataset,n_batches,model,save_with_bt=False): patience=0 losses = [] for epoch_idx in range(n_epochs): if epoch_idx>=config.state_dict_check['epoch']+1: st_time = time.time() avg_loss=0 # Randomize data order data_generator = get_data_generator(config,train_dataset,tokenizer,config.max_seq_len, config.batch_size) optimizer.zero_grad() for batch_idx, (input_batch, label_batch) in tqdm(enumerate(data_generator), total=n_batches): if batch_idx >= config.state_dict_check['batch_idx']: input_batch,label_batch = input_batch.to(config.device),label_batch.to(config.device) # Forward pass model_out = model.forward(input_ids = input_batch, labels = label_batch) # Calculate loss and update weights if config.use_torch_data_parallel: loss = torch.mean(model_out.loss) else: loss = model_out.loss losses.append(loss.item()) loss.backward() #Gradient accumulation if (batch_idx+1) % config.gradient_accumulation_batch == 0: optimizer.step() optimizer.zero_grad() # Print training update info if (batch_idx + 1) % config.print_freq == 0: avg_loss = np.mean(losses) losses=[] if config.verbose: with open(config.log,'a+') as fl: print('Epoch: {} | Step: {} | Avg. loss: {:.3f}'.format(epoch_idx+1, batch_idx+1, avg_loss),file=fl) if (batch_idx + 1) % config.checkpoint_freq == 0: test_loss = eval_model(config,tokenizer,model, dev_dataset) if config.best_loss-test_loss > config.best_loss_delta: config.best_loss = test_loss patience=0 if config.verbose: with open(config.log,'a+') as fl: print('Saving model with best test loss of {:.3f}'.format(test_loss),file=fl) if save_with_bt: model_name = config.model_name.split('.')[0]+'_bt.pt' else: model_name = config.model_name config.state_dict.update({'batch_idx': batch_idx,'epoch':epoch_idx,'bt_time':config.bt_time-1,'best_loss':config.best_loss}) if config.use_torch_data_parallel: config.state_dict['model_state_dict']=model.module.state_dict() torch.save(config.state_dict, os.path.join(config.model_path_dir,model_name)) else: config.state_dict['model_state_dict']=model.state_dict() torch.save(config.state_dict, os.path.join(config.model_path_dir,model_name)) else: if config.verbose: with open(config.log,'a+') as fl: print(f'No improvement in loss {test_loss} over best loss {config.best_loss}. Not saving model checkpoint',file=fl) patience+=1 if patience >= config.patience: with open(config.log,'a+') as fl: print("Stopping model training due to early stopping",file=fl) break with open(config.log,'a+') as fl: print('Epoch: {} | Step: {} | Avg. loss: {:.3f} | Time taken: {} | Time: {}'.format(epoch_idx+1, batch_idx+1, avg_loss, beautify_time(time.time()-st_time),datetime.now()),file=fl) # Do this after epochs to get status of model at end of training---- test_loss = eval_model(config,tokenizer,model, dev_dataset) if config.best_loss-test_loss > config.best_loss_delta: config.best_loss = test_loss patience=0 if config.verbose: with open(config.log,'a+') as fl: print('Saving model with best test loss of {:.3f}'.format(test_loss),file=fl) if save_with_bt: model_name = config.model_name.split('.')[0]+'_bt.pt' else: model_name = config.model_name config.state_dict.update({'batch_idx': n_batches-1,'epoch':n_epochs-1,'bt_time':config.bt_time-1,'best_loss':config.best_loss}) if config.use_torch_data_parallel: config.state_dict['model_state_dict']=model.module.state_dict() torch.save(config.state_dict, os.path.join(config.model_path_dir,model_name)) else: config.state_dict['model_state_dict']=model.state_dict() torch.save(config.state_dict, os.path.join(config.model_path_dir,model_name)) else: if config.verbose: with open(config.log,'a+') as fl: print(f'No improvement in loss {test_loss} over best loss {config.best_loss}. Not saving model checkpoint',file=fl) patience+=1 #--------------------------------------------- def main(args): if not os.path.exists(args.homepath): raise Exception(f'HOMEPATH {args.homepath} does not exist!') config = Config(args) if not os.path.exists(config.prediction_path): os.makedirs(config.prediction_path) if not os.path.exists(config.bt_data_dir): os.makedirs(config.bt_data_dir) """# Load Tokenizer & Model""" tokenizer = AutoTokenizer.from_pretrained(config.model_repo) if config.use_multiprocessing: tokenizers_for_parallel = [AutoTokenizer.from_pretrained(config.model_repo) for i in range(config.num_cores)] model = AutoModelForSeq2SeqLM.from_pretrained(config.model_repo) if not os.path.exists(config.parallel_dir): raise Exception(f'Directory `{config.parallel_dir}` cannot be empty! It must contain the parallel files') train_dataset = make_dataset(config,'train') with open(config.log,'a+') as fl: print(f"Length of train dataset: {len(train_dataset)}",file=fl) dev_dataset = make_dataset(config,'eval') with open(config.log,'a+') as fl: print(f"Length of dev dataset: {len(dev_dataset)}",file=fl) """## Update tokenizer""" special_tokens_dict = {'additional_special_tokens': list(config.LANG_TOKEN_MAPPING.values())} tokenizer.add_special_tokens(special_tokens_dict) if config.use_multiprocessing: for tk in tokenizers_for_parallel: tk.add_special_tokens(special_tokens_dict) model.resize_token_embeddings(len(tokenizer)) """# Train/Finetune MT5""" if os.path.exists(os.path.join(config.model_path_dir,config.model_name)): if config.verbose: with open(config.log,'a+') as fl: print("-----------Using model checkpoint-----------",file=fl) try: state_dict = torch.load(os.path.join(config.model_path_dir,config.model_name.split('.')[0]+'_bt.pt')) except Exception: with open(config.log,'a+') as fl: print('No mmt_translation_bt.pt present. Default to original mmt_translation.pt',file=fl) state_dict = torch.load(os.path.join(config.model_path_dir,config.model_name)) # Note to self: Make this beter. config.state_dict_check['epoch']=state_dict['epoch'] config.state_dict_check['bt_time']=state_dict['bt_time'] config.state_dict_check['best_loss']=state_dict['best_loss'] config.best_loss = config.state_dict_check['best_loss'] config.state_dict_check['batch_idx']=state_dict['batch_idx'] model.load_state_dict(state_dict['model_state_dict']) #Temp change config.state_dict_check['epoch']=-1 config.state_dict_check['batch_idx']=0 config.state_dict_check['bt_time']=-1 #Using DataParallel if config.use_torch_data_parallel: model = torch.nn.DataParallel(model, device_ids=list(range(torch.cuda.device_count()))) model = model.to(config.device) #----- # Optimizer optimizer = Adafactor(model.parameters(), scale_parameter=False, relative_step=False, warmup_init=False, lr=config.lr) #Normal training n_batches = int(np.ceil(len(train_dataset) / config.batch_size)) total_steps = config.n_epochs * n_batches n_warmup_steps = int(total_steps * 0.01) #scheduler = get_linear_schedule_with_warmup(optimizer, n_warmup_steps, total_steps) #scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=config.lr, max_lr=0.001,cycle_momentum=False) train(config,config.n_epochs,optimizer,tokenizer,train_dataset,dev_dataset,n_batches,model) if config.verbose: with open(config.log,'a+') as fl: print('Evaluaton...',file=fl) do_evaluation(config,tokenizer,model,dev_dataset) config.state_dict_check['epoch']=-1 config.state_dict_check['batch_idx']=0 if config.do_backtranslation: #Backtranslation time config.now_on_bt=True with open(config.log,'a+') as fl: print('---------------Start of Backtranslation---------------',file=fl) for n_bt in range(config.NUM_BACKTRANSLATION_TIMES): if n_bt>=config.state_dict_check['bt_time']+1: with open(config.log,'a+') as fl: print(f"Backtranslation {n_bt+1} of {config.NUM_BACKTRANSLATION_TIMES}--------------",file=fl) config.bt_time = n_bt+1 save_bt_file_path = os.path.join(config.bt_data_dir,'bt'+str(n_bt+1)+'.json') if not os.path.exists(save_bt_file_path): mono_data = mono_data_(config) start_time = time.time() if config.use_multiprocessing: if config.verbose: with open(config.log,'a+') as fl: print(f"Using multiprocessing on {config.num_cores} processes",file=fl) if __name__ == "__main__": model.share_memory() with parallel_backend('loky'): bt_data = Parallel(n_jobs = config.num_cores, require='sharedmem')(delayed(do_job)(data_,i,tokenizers_for_parallel) for i,data_ in tqdm(enumerate(mono_data))) else: bt_data = [{'inputs':get_model_translation(config,model,tokenizer,t['inputs'],t['tgt']),'targets':t['inputs'],'src':t['tgt'],'tgt':t['src']} for t in tqdm(mono_data)] with open(config.log,'a+') as fl: print(f'Time taken for backtranslation of data: {beautify_time(time.time()-start_time)}',file=fl) with open(save_bt_file_path,'w') as fp: json.dump(bt_data,fp) else: with open(save_bt_file_path,'r') as f: bt_data = json.load(f) with open(config.log,'a+') as fl: print('-'*15+'Printing 5 random BT Data'+'-'*15,file=fl) ids_print = random.sample([i for i in range(len(bt_data))],5) with open(config.log,'a+') as fl: for ids_print_ in ids_print: print(bt_data[ids_print_],file=fl) augmented_dataset = train_dataset + bt_data + mono_data_noise(config) #mono_data_noise adds denoising objective random.shuffle(augmented_dataset) with open(config.log,'a+') as fl: print(f'New length of dataset: {len(augmented_dataset)}',file=fl) n_batches = int(np.ceil(len(augmented_dataset) / config.batch_size)) total_steps = config.n_bt_epochs * n_batches n_warmup_steps = int(total_steps * 0.01) #scheduler = get_linear_schedule_with_warmup(optimizer, n_warmup_steps, total_steps) #scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=config.lr, max_lr=0.001,cycle_momentum=False) train(config,config.n_bt_epochs,optimizer,tokenizer,augmented_dataset,dev_dataset,n_batches,model,save_with_bt=True) if config.verbose: with open(config.log,'a+') as fl: print('Evaluaton...',file=fl) do_evaluation(config,tokenizer,model,dev_dataset) config.state_dict_check['epoch']=-1 config.state_dict_check['batch_idx']=0 with open(config.log,'a+') as fl: print('---------------End of Backtranslation---------------',file=fl) with open(config.log,'a+') as fl: print('---------------End of Training---------------',file=fl) config.now_on_bt=False config.now_on_test=True with open(config.log,'a+') as fl: print('Evaluating on test set',file=fl) test_dataset = make_dataset(config,'test') with open(config.log,'a+') as fl: print(f"Length of test dataset: {len(test_dataset)}",file=fl) do_evaluation(config,tokenizer,model,test_dataset) with open(config.log,'a+') as fl: print("ALL DONE",file=fl) def load_params(args: dict) -> dict: """ Load the parameters passed to `translate` """ #if not os.path.exists(args['checkpoint']): # raise Exception(f'Checkpoint file does not exist') params = {} model_repo = 'google/mt5-base' LANG_TOKEN_MAPPING = { 'ig': '', 'fon': '', 'en': '', 'fr': '', 'rw':'', 'yo':'', 'xh':'', 'sw':'' } tokenizer = AutoTokenizer.from_pretrained(model_repo) model = AutoModelForSeq2SeqLM.from_pretrained(model_repo) """## Update tokenizer""" special_tokens_dict = {'additional_special_tokens': list(LANG_TOKEN_MAPPING.values())} tokenizer.add_special_tokens(special_tokens_dict) model.resize_token_embeddings(len(tokenizer)) state_dict = torch.load(args['checkpoint'],map_location=args['device']) model.load_state_dict(state_dict['model_state_dict']) model = model.to(args['device']) #Load the model, load the tokenizer, max and min seq len params['model'] = model params['device'] = args['device'] params['max_seq_len'] = args['max_seq_len'] if 'max_seq_len' in args else 50 params['min_seq_len'] = args['min_seq_len'] if 'min_seq_len' in args else 2 params['tokenizer'] = tokenizer params['num_beams'] = args['num_beams'] if 'num_beams' in args else 4 params['lang_token'] = LANG_TOKEN_MAPPING params['truncation'] = args['truncation'] if 'truncation' in args else True return params def encode_input_str_translate(params,text, target_lang, tokenizer, seq_len): target_lang_token = params['lang_token'][target_lang] # Tokenize and add special tokens input_ids = tokenizer.encode( text = str(target_lang_token) + str(text), return_tensors = 'pt', padding = 'max_length', truncation = params['truncation'] , max_length = seq_len) return input_ids[0] def translate( params: dict, sentence: str, source_lang: str, target_lang: str ) -> str: """ Given a sentence and its source and target sentences, this translates the sentence to the given target sentence. """ if source_lang!='' and target_lang!='': inp = [sentence] input_tokens = [encode_input_str_translate(params,text = inp[i],target_lang = target_lang,tokenizer = params['tokenizer'],seq_len =params['max_seq_len']).unsqueeze(0).to(params['device']) for i in range(len(inp))] output = [params['model'].generate(input_ids, num_beams=params['num_beams'], num_return_sequences=1,max_length=params['max_seq_len'],min_length=params['min_seq_len']) for input_ids in input_tokens] output = [params['tokenizer'].decode(out[0], skip_special_tokens=True) for out in tqdm(output)] return output[0] else: return '' if __name__=="__main__": from argparse import ArgumentParser import json import os parser = ArgumentParser('MMTArica Experiments') parser.add_argument('-homepath', type=str, default=os.getcwd(), help="Homepath directory. Where all experiments are saved and all \ necessary files/folders are saved. (default: current working directory)") parser.add_argument('--prediction_path', type=str, default='./predictions', help='directory path to save predictions (default: %(default)s)') parser.add_argument('--model_name', type=str, default='mmt_translation', help='Name of model (default: %(default)s)') parser.add_argument('--bt_data_dir', type=str, default='btData', help='Directory to save back-translation files (default: %(default)s)') parser.add_argument('--parallel_dir', type=str, default='parallel', help='name of directory where parallel corpora is saved') parser.add_argument('--mono_dir', type=str, default='mono', help='name of directory where monolingual files are saved (default: %(default)s)') parser.add_argument('--log', type=str, default='train.log', help='name of file to log experiments (default: %(default)s)') parser.add_argument('--mono_data_limit', type=int, default=300, help='limit of monolingual sentences to use for training (default: %(default)s)') parser.add_argument('--mono_data_for_noise_limit', type=int, default=50, help='limit of monolingual sentences to use for noise (default: %(default)s)') parser.add_argument('--n_epochs', type=int, default=10, help='number of training epochs (default: %(default)s)') parser.add_argument('--n_bt_epochs', type=int, default=3, help='number of backtranslation epochs (default: %(default)s)') parser.add_argument('--batch_size', type=int, default=64, help='batch size (default: %(default)s)') parser.add_argument('--max_seq_len', type=int, default=50, help='maximum length of sentence. All sentences beyond this length will be skipped. (default: %(default)s)') parser.add_argument('--min_seq_len', type=int, default=2, help='mnimum length of sentence. All sentences beyond this length will be skipped. (default: %(default)s)') parser.add_argument('--checkpoint_freq', type=int, default=10_000, help='maximum length of sentence. All sentences beyond this length will be skipped. (default: %(default)s)') parser.add_argument('--lr', type=int, default=1e-4, help='learning rate. (default: %(default)s)') parser.add_argument('--print_freq', type=int, default=5_000, help='frequency at which to print to log. (default: %(default)s)') parser.add_argument('--use_multiprocessing', type=bool, default=False, help='whether or not to use multiprocessing. (default: %(default)s)') parser.add_argument('--num_pretrain_steps', type=int, default=20, help='number of pretrain steps. (default: %(default)s)') parser.add_argument('--num_backtranslation_steps', type=int, default=5, help='number of pretrain steps. (default: %(default)s)') parser.add_argument('--do_backtranslation', type=bool, default=True, help='whether or not to do backtranslation during training. (default: %(default)s)') parser.add_argument('--use_reconstruction', type=bool, default=True, help='whether or not to use reconstruction during training. (default: %(default)s)') parser.add_argument('--use_torch_data_parallel', type=bool, default=False, help='whether or not to use torch data parallelism. (default: %(default)s)') parser.add_argument('--gradient_accumulation_batch', type=int, default=4096//64, help='batch size for gradient accumulation. (default: %(default)s)') parser.add_argument('--num_beams', type=int, default=4, help='number of beams to use for inference. (default: %(default)s)') parser.add_argument('--patience', type=int, default=15_000_000, help='patience for early stopping. (default: %(default)s)') parser.add_argument('--drop_probability', type=float, default=0.2, help='drop probability for reconstruction. (default: %(default)s)') parser.add_argument('--dropout', type=float, default=0.1, help='dropout probability. (default: %(default)s)') parser.add_argument('--num_swaps', type=int, default=3, help='number of word swaps to perform during reconstruction. (default: %(default)s)') parser.add_argument('--verbose', type=bool, default=True, help='whether or not to print information during experiments. (default: %(default)s)') args = parser.parse_args() main(args)