Thinking-while-Observing / code /train4LXMT5_DDP.py
QingyiSi's picture
Upload 14 files
44ddffd
#!user/bin/env python
# -*- coding:utf-8 -*-
import argparse
import json
import os
import datetime
import pickle
import random
import torch
import torch.backends.cudnn as cudnn
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from bisect import bisect
from math import fabs
from torch.optim import lr_scheduler
from torch.utils.data import DataLoader
from tqdm import tqdm
from transformers import LxmertTokenizer
from dist_train import get_world_size, get_rank, get_local_rank, barrier, reduce_sum
import numpy as np
from transformers.tokenization_utils_base import ENCODE_KWARGS_DOCSTRING
from config4LXMT5_DDP import args
from dataset4LXMT5 import KgDataset, my_collate, my_val_gpt3_collate, my_val_collate
from dataset_val4LXMT5 import KgDatasetVal
if args.visualBERT:
from model_ViB2T5 import T5tokenizer, ViBT52T5, LXMtokenizer
else:
from model_LXM2T5 import T5tokenizer, LXMT52T5, LXMtokenizer
from transformers import get_linear_schedule_with_warmup
from transformers import LxmertConfig, LxmertTokenizer, LxmertModel,BertTokenizer
dist.init_process_group(backend='nccl',timeout=datetime.timedelta(seconds=5400))
torch.cuda.set_device(args.local_rank)
# LR = 1e-5
LR = args.learning_rate
LR_LXM = args.learning_rate_LXM
# LR = 1e-4
torch.multiprocessing.set_sharing_strategy('file_system')
torch.cuda.set_device(get_local_rank())
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
def reduce_tensor(tensor: torch.Tensor):
rt = tensor.clone().float()
dist.all_reduce(rt,op=dist.ReduceOp.SUM)
rt /= dist.get_world_size()#.float()
return rt
def set_seed(rank):
random.seed(args.seed+rank)
np.random.seed(args.seed+rank)
torch.manual_seed(args.seed+rank)
torch.cuda.manual_seed(args.seed+rank)
torch.cuda.manual_seed_all(args.seed+rank)
torch.backends.cudnn.deterministic = True
set_seed(get_rank())
def cal_acc_multi(ground_truth, preds, return_id = False):
all_num = len(ground_truth)
acc_num = 0
ids = []
temp = []
for i, answer_id in enumerate(ground_truth):
pred = preds[i]
# ids.append([i, int(pred)])
cnt = 0
for aid in answer_id:
if pred == aid:
cnt += 1
if cnt ==1:
acc_num += 1/3
elif cnt == 2:
acc_num += 2/3
elif cnt > 2:
acc_num += 1
if return_id:
return acc_num / all_num, ids
else:
return acc_num, all_num
def cal_acc(ground_truth, preds, return_id = False):
all_num = len(ground_truth)
acc_num = 0
ids = []
temp = []
for i, answer_id in enumerate(ground_truth):
pred = preds[i]
# ids.append([i, int(pred)])
cnt = 0
for aid in answer_id:
if pred == aid:
acc_num += 1
if return_id:
return acc_num / all_num, ids
else:
return acc_num, all_num
def train():
if not args.describe:
print('please set the description for the saved-model name! use --describe !')
assert 1==0
else:
model_name=args.describe
if not args.pretrain:
train_dataset = KgDataset(val=False)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
train_dataloader = DataLoader(dataset=train_dataset, batch_size=args.batch_size, sampler=train_sampler,#shuffle=True,
num_workers=0, collate_fn=my_collate)#, pin_memory=True)
if args.validate:
test_dataset = KgDatasetVal(val=False)
if args.gpt3:
test_dataloader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=0, collate_fn=my_val_gpt3_collate)
elif not args.gpt3:
test_dataloader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=0, collate_fn=my_val_collate)
else:
train_dataset = KgDataset(val=False)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
train_dataloader = DataLoader(dataset=train_dataset, batch_size=args.batch_size,#pin_memory=True,
num_workers=0, collate_fn=my_collate, sampler=train_sampler)
if args.validate:
test_dataset = KgDatasetVal(val=False)
if args.gpt3:
test_dataloader = DataLoader(dataset=test_dataset, batch_size=args.batch_size,
num_workers=0, collate_fn=my_val_gpt3_collate, shuffle=False)#sampler=test_sampler)
elif not args.gpt3:
test_dataloader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=0, collate_fn=my_val_collate)
if args.pretrain:
if get_rank() == 0:
print('pre-training!')
if args.visualBERT:
model= ViBT52T5()
else:
model = LXMT52T5()
else:
if get_rank() == 0:
print('fine-tuning!')
if args.visualBERT:
model = ViBT52T5()
else:
model = LXMT52T5()
model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
model = model.to(device)
if get_world_size() > 1:
if get_rank() == 0:
print("Let's use", get_world_size(), "GPUs!")
model = nn.parallel.DistributedDataParallel(model, device_ids=[get_local_rank()], output_device=get_local_rank(),find_unused_parameters=True)
print(model.named_modules)
if get_world_size() > 1:
if args.visualBERT:
optimizer = optim.AdamW([
{'params': model.module.T5model.parameters(), 'lr': LR},
{'params': model.module.ViBmodel.parameters(), 'lr': LR_LXM},
{'params': model.module.mapping.parameters(), 'lr': LR_LXM},
])
else:
optimizer = optim.AdamW([
{'params': model.module.T5model.parameters(), 'lr': LR},
{'params': model.module.LXMmodel.parameters(), 'lr': LR_LXM},
{'params': model.module.mapping.parameters(), 'lr': LR_LXM},
])
else:
if args.visualBERT:
optimizer = optim.AdamW([
{'params': model.T5model.parameters(), 'lr': LR},
{'params': model.ViBmodel.parameters(), 'lr': LR_LXM},
{'params': model.mapping.parameters(), 'lr': LR_LXM},
])
else:
optimizer = optim.AdamW([
{'params': model.T5model.parameters(), 'lr': LR},
{'params': model.LXMmodel.parameters(), 'lr': LR_LXM},
{'params': model.mapping.parameters(), 'lr': LR_LXM},
])
if args.pretrain:
steps_num = 100000 # batch_size should be set small
else:
steps_num = 4000
args.num_epochs = steps_num // (len(train_dataset) / (args.batch_size * get_world_size())) \
if len(train_dataset) % args.batch_size == 0 \
else (steps_num // (len(train_dataset) / (args.batch_size * get_world_size())) )+1
args.num_epochs = int(args.num_epochs)
if get_rank() == 0:
print('total_epoch', args.num_epochs)
print('total_steps', "we set steps=",steps_num)
print('warmup_steps', int(steps_num/10)) #0.05*total_steps)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=int(steps_num/10), #0.01 * total_steps,
num_training_steps=steps_num)
if args.load_pthpath == "":
start_epoch = 0
else:
if get_rank() == 0:
print('load model')
start_epoch = 0
if get_world_size() > 1:
model.module.load_state_dict(torch.load(args.load_pthpath))
else:
model.load_state_dict(torch.load(args.load_pthpath))
best_acc_t = 0
best_epoch_t = 0
best_acc_t3 = 0
step_ind = 0
for epoch in range(start_epoch, args.num_epochs):
train_preds_trip = []
train_sampler.set_epoch(epoch)
train_answers_trip = []
s=0
for batch_data in tqdm(train_dataloader):
step_ind+=1
if get_rank()==0:
print("step_ind",step_ind)
s=s+1
visual_faetures = torch.from_numpy(np.array(batch_data['img'], dtype=float)).float().to(device)
spatial_features = torch.tensor(np.array(batch_data['spatial'])).float().to(device)
if 1:
T5_input_id = torch.stack(batch_data['T5_input_ids']).to(device)
T5_input_mask = torch.stack(batch_data['T5_input_masks']).to(device)
LXM_input_id = torch.stack(batch_data['LXM_input_ids']).to(device)
LXM_input_mask = torch.stack(batch_data['LXM_input_masks']).to(device)
LXM_token_type_ids = torch.stack(batch_data['LXM_token_type_ids']).to(device)
T5_target_id = torch.stack(batch_data['T5_target_ids']).to(device)
neg100 = torch.ones_like(T5_target_id)*(-100)
T5_target_id = torch.where(T5_target_id==T5tokenizer.pad_token_id,neg100, T5_target_id)
model.zero_grad()
optimizer.zero_grad()
if args.pretrain:
outputs = model(train=True, LXM_source_ids=LXM_input_id, LXM_source_masks=LXM_input_mask,T5_source_ids=T5_input_id, T5_source_masks=T5_input_mask,token_type_ids=LXM_token_type_ids, visual_features=visual_faetures, spatial_features=spatial_features,T5_target_ids=T5_target_id)#,T5_target_masks=None
else:
outputs = model(train=True, LXM_source_ids=LXM_input_id, LXM_source_masks=LXM_input_mask,T5_source_ids=T5_input_id, T5_source_masks=T5_input_mask,token_type_ids=LXM_token_type_ids, visual_features=visual_faetures, spatial_features=spatial_features,T5_target_ids=T5_target_id)#,T5_target_masks=None
loss = outputs.loss
loss_stat = torch.mean(loss.detach()).item()
if get_rank() == 0:
print("loss on GPU0", loss_stat)
loss.sum().backward()
optimizer.step()
scheduler.step()
model.eval()
with torch.no_grad():
if args.pretrain:
eval_outputs = model(train=False, LXM_source_ids=LXM_input_id, LXM_source_masks=LXM_input_mask,T5_source_ids=T5_input_id, T5_source_masks=T5_input_mask,token_type_ids=LXM_token_type_ids, visual_features=visual_faetures, spatial_features=spatial_features,T5_target_ids=T5_target_id)
else:
eval_outputs = model(train=False, LXM_source_ids=LXM_input_id, LXM_source_masks=LXM_input_mask,T5_source_ids=T5_input_id, T5_source_masks=T5_input_mask,token_type_ids=LXM_token_type_ids, visual_features=visual_faetures, spatial_features=spatial_features,T5_target_ids=T5_target_id)
trip_predict = T5tokenizer.batch_decode(eval_outputs, skip_special_tokens=True)
if get_rank() == 0:
print('epoch', epoch, 'step', s, '>>>', '\tans:', batch_data['ans'][0], 'pred:', trip_predict[0])
for i, pre in enumerate(batch_data['ans']):
train_answers_trip.append(batch_data['ans'][i])
train_preds_trip.append(trip_predict[i])
model.train()
barrier()
barrier()
if 1:
train_acc_1_num, train_total_1_num = cal_acc_multi(train_answers_trip, train_preds_trip)
train_reduce_acc_num=reduce_tensor(torch.tensor(train_acc_1_num).cuda(args.local_rank)).item()
train_reduce_total_num=reduce_tensor(torch.tensor(train_total_1_num).cuda(args.local_rank)).item()
train_acc_1_trip = train_reduce_acc_num/train_reduce_total_num
if get_rank() == 0:
print('epoch %d train_loss of GPU0= %.1f, acc_trip on all GPUs= %.4f' % (epoch, loss_stat,
train_acc_1_trip))
if args.validate:
model.eval()
answers = [] # [batch_answers,...]
preds = [] # [batch_preds,...]
preds_trip = []
preds_trip_3 = []
answers_trip = []
id2pred_trip = {}
print(f"\nValidation after epoch {epoch}:")
for i, batch_data in enumerate(tqdm(test_dataloader)):
with torch.no_grad():
val_T5_input_id = torch.stack(batch_data['T5_input_ids']).to(device)
val_T5_input_mask = torch.stack(batch_data['T5_input_masks']).to(device)
val_visual_faetures = torch.tensor(np.array(batch_data['img'])).float().to(device)
val_spatial_features = torch.tensor(np.array(batch_data['spatial'])).float().to(device)
val_LXM_input_id = torch.stack(batch_data['LXM_input_ids']).to(device)
val_LXM_input_mask = torch.stack(batch_data['LXM_input_masks']).to(device)
val_LXM_token_type_ids = torch.stack(batch_data['LXM_token_type_ids']).to(device)
if args.pretrain:
val_outputs = model(train=False, LXM_source_ids=val_LXM_input_id, LXM_source_masks=val_LXM_input_mask,T5_source_ids=val_T5_input_id, T5_source_masks=val_T5_input_mask,token_type_ids=val_LXM_token_type_ids, visual_features=val_visual_faetures, spatial_features=val_spatial_features,T5_target_ids=None)
else:
val_outputs = model(train=False, LXM_source_ids=val_LXM_input_id, LXM_source_masks=val_LXM_input_mask,T5_source_ids=val_T5_input_id, T5_source_masks=val_T5_input_mask,token_type_ids=val_LXM_token_type_ids, visual_features=val_visual_faetures, spatial_features=val_spatial_features,T5_target_ids=None)
val_trip_predict = T5tokenizer.batch_decode(val_outputs, skip_special_tokens=True)
for i, pre in enumerate(batch_data['ans']):
preds_trip.append(val_trip_predict[i])
answers_trip.append(batch_data['ans'][i])
id2pred_trip[str(batch_data['id'][i])]=val_trip_predict[i]
if args.dataset == 'krvqa':
acc_1_num, total_1_num = cal_acc(answers_trip, preds_trip)
reduce_acc_num=reduce_tensor(torch.tensor(acc_1_num).cuda(args.local_rank)).item()
reduce_total_num=reduce_tensor(torch.tensor(total_1_num).cuda(args.local_rank)).item()
acc_1_trip = reduce_acc_num/reduce_total_num
if get_rank() == 0:
print('epoch %d , acc_trip on all GPUs= %.4f' % (epoch, acc_1_trip))
else:
acc_1_num, total_1_num = cal_acc_multi(answers_trip, preds_trip)
reduce_acc_num=reduce_tensor(torch.tensor(acc_1_num).cuda(args.local_rank)).item()
reduce_total_num=reduce_tensor(torch.tensor(total_1_num).cuda(args.local_rank)).item()
acc_1_trip = reduce_acc_num/reduce_total_num
if get_rank() == 0:
print('epoch %d , acc_trip on all GPUs= %.4f' % (epoch, acc_1_trip))
if acc_1_trip > best_acc_t:
best_acc_t = acc_1_trip
best_epoch_t = epoch
if not args.pretrain:
if get_rank() == 0:
f=open(args.model_dir+"/predictions.json", 'w')
json.dump(id2pred_trip, f)
f.close()
"""
# ablations on two encoders
# LXMERTenc-T5dec
if args.load_pthpath == "":
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/C1_LXMERTencOnly_noPre_predictions.json", 'w') #GPT-noPre
else:
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/C3_LXMERTencOnly_predictions.json", 'w') #GPT
json.dump(id2pred_trip, fx)
fx.close()
# T5enc-T5dec
if args.load_pthpath == "":
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/C2_T5encOnly_noPre_predictions.json", 'w') #GPT-noPre
else:
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/C4_T5encOnly_predictions.json", 'w') #GPT
json.dump(id2pred_trip, fx)
fx.close()
"""
"""
# ablations on Knowledge types
if args.gpt3:
if args.input_type==0 and args.load_pthpath == "":
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/A2_noPre_predictions.json", 'w') #GPT-noPre
json.dump(id2pred_trip, fx)
fx.close()
elif args.input_type==1 and (args.load_pthpath != ""):
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/A3_noWiki_predictions.json", 'w') #GPT
json.dump(id2pred_trip, fx)
fx.close()
elif args.input_type==2 and (args.load_pthpath != ""):
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/A4_noOFA_predictions.json", 'w') #GPT
json.dump(id2pred_trip, fx)
fx.close()
elif args.input_type==3 and (args.load_pthpath == ""):
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/B1_onlyGPT3_predictions.json", 'w') #GPT-noPre
json.dump(id2pred_trip, fx)
fx.close()
else:
if args.input_type==0 and args.load_pthpath != "":
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/A5_noGPT3_predictions.json", 'w') #noGPT
json.dump(id2pred_trip, fx)
fx.close()
elif args.input_type==0 and args.load_pthpath == "":
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/A6_noGPT3noPre_predictions.json", 'w') #noGPT
json.dump(id2pred_trip, fx)
fx.close()
elif args.input_type==1 and (args.load_pthpath == ""):
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/B2_onlyOFA_predictions.json", 'w') #noGPT-noPre
json.dump(id2pred_trip, fx)
fx.close()
elif args.input_type==2 and (args.load_pthpath == ""):
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/B3_onlyWiki_predictions.json", 'w') #noGPT-noPre
json.dump(id2pred_trip, fx)
fx.close()
elif args.input_type==3 and (args.load_pthpath == ""):
fx=open("/mnt/bn/qingyi-bn-lq/okvqa-output/B4_onlyVisualNoKnowledge_predictions.json", 'w') #noGPT-noPre
json.dump(id2pred_trip, fx)
fx.close()
"""
print('saving model at epoch', epoch, '!!')
if get_world_size() > 1:
torch.save(model.module.state_dict(), args.model_dir+'/best_finetuned_model_'+model_name+'.pth')
else:
torch.save(model.state_dict(), args.model_dir+'/best_finetuned_model_'+model_name+'.pth')
if get_rank() == 0:
print("best_acc@1t={:.2%}, epoch{}\n\n".format(best_acc_t, best_epoch_t))
model.train()
if args.pretrain:
if get_rank() == 0:
if get_world_size() > 1:
torch.save(model.module.state_dict(), args.model_dir+ '/model_for_epoch_%d.pth' % epoch)
else:
torch.save(model.state_dict(), args.model_dir+ '/model_for_epoch_%d.pth' % epoch)
barrier()
dist.destroy_process_group()
if __name__ == "__main__":
train()