Spaces:
Sleeping
Sleeping
File size: 7,790 Bytes
08d7644 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 |
# coding=utf-8
# Copyleft 2019 project LXRT.
import os
import collections
import torch
import torch.nn as nn
from torch.utils.data.dataloader import DataLoader
from tqdm import tqdm
from ..param import args
from ..pretrain.qa_answer_table import load_lxmert_qa
from .vqa_model import VQAModel
from .vqa_data import VQADataset, VQATorchDataset, VQAEvaluator
DataTuple = collections.namedtuple("DataTuple", 'dataset loader evaluator')
def get_data_tuple(splits: str, bs:int, shuffle=False, drop_last=False) -> DataTuple:
dset = VQADataset(splits)
tset = VQATorchDataset(dset)
evaluator = VQAEvaluator(dset)
data_loader = DataLoader(
tset, batch_size=bs,
shuffle=shuffle, num_workers=args.num_workers,
drop_last=drop_last, pin_memory=True
)
return DataTuple(dataset=dset, loader=data_loader, evaluator=evaluator)
class VQA:
def __init__(self):
# Datasets
self.train_tuple = get_data_tuple(
args.train, bs=args.batch_size, shuffle=True, drop_last=True
)
if args.valid != "":
self.valid_tuple = get_data_tuple(
args.valid, bs=1024,
shuffle=False, drop_last=False
)
else:
self.valid_tuple = None
# Model
self.model = VQAModel(self.train_tuple.dataset.num_answers)
# Load pre-trained weights
if args.load_lxmert is not None:
self.model.lxrt_encoder.load(args.load_lxmert)
if args.load_lxmert_qa is not None:
load_lxmert_qa(args.load_lxmert_qa, self.model,
label2ans=self.train_tuple.dataset.label2ans)
# GPU options
self.model = self.model.cuda()
if args.multiGPU:
self.model.lxrt_encoder.multi_gpu()
# Loss and Optimizer
self.bce_loss = nn.BCEWithLogitsLoss()
if 'bert' in args.optim:
batch_per_epoch = len(self.train_tuple.loader)
t_total = int(batch_per_epoch * args.epochs)
print("BertAdam Total Iters: %d" % t_total)
from ..lxrt.optimization import BertAdam
self.optim = BertAdam(list(self.model.parameters()),
lr=args.lr,
warmup=0.1,
t_total=t_total)
else:
self.optim = args.optimizer(self.model.parameters(), args.lr)
# Output Directory
self.output = args.output
os.makedirs(self.output, exist_ok=True)
def train(self, train_tuple, eval_tuple):
dset, loader, evaluator = train_tuple
iter_wrapper = (lambda x: tqdm(x, total=len(loader))) if args.tqdm else (lambda x: x)
best_valid = 0.
for epoch in range(args.epochs):
quesid2ans = {}
for i, (ques_id, feats, boxes, sent, target) in iter_wrapper(enumerate(loader)):
self.model.train()
self.optim.zero_grad()
feats, boxes, target = feats.cuda(), boxes.cuda(), target.cuda()
logit = self.model(feats, boxes, sent)
assert logit.dim() == target.dim() == 2
loss = self.bce_loss(logit, target)
loss = loss * logit.size(1)
loss.backward()
nn.utils.clip_grad_norm_(self.model.parameters(), 5.)
self.optim.step()
score, label = logit.max(1)
for qid, l in zip(ques_id, label.cpu().numpy()):
ans = dset.label2ans[l]
quesid2ans[qid.item()] = ans
log_str = "\nEpoch %d: Train %0.2f\n" % (epoch, evaluator.evaluate(quesid2ans) * 100.)
if self.valid_tuple is not None: # Do Validation
valid_score = self.evaluate(eval_tuple)
if valid_score > best_valid:
best_valid = valid_score
self.save("BEST")
log_str += "Epoch %d: Valid %0.2f\n" % (epoch, valid_score * 100.) + \
"Epoch %d: Best %0.2f\n" % (epoch, best_valid * 100.)
print(log_str, end='')
with open(self.output + "/log.log", 'a') as f:
f.write(log_str)
f.flush()
self.save("LAST")
def predict(self, eval_tuple: DataTuple, dump=None):
"""
Predict the answers to questions in a data split.
:param eval_tuple: The data tuple to be evaluated.
:param dump: The path of saved file to dump results.
:return: A dict of question_id to answer.
"""
self.model.eval()
dset, loader, evaluator = eval_tuple
quesid2ans = {}
for i, datum_tuple in enumerate(loader):
ques_id, feats, boxes, sent = datum_tuple[:4] # Avoid seeing ground truth
with torch.no_grad():
feats, boxes = feats.cuda(), boxes.cuda()
logit = self.model(feats, boxes, sent)
score, label = logit.max(1)
for qid, l in zip(ques_id, label.cpu().numpy()):
ans = dset.label2ans[l]
quesid2ans[qid.item()] = ans
if dump is not None:
evaluator.dump_result(quesid2ans, dump)
return quesid2ans
def evaluate(self, eval_tuple: DataTuple, dump=None):
"""Evaluate all data in data_tuple."""
quesid2ans = self.predict(eval_tuple, dump)
return eval_tuple.evaluator.evaluate(quesid2ans)
@staticmethod
def oracle_score(data_tuple):
dset, loader, evaluator = data_tuple
quesid2ans = {}
for i, (ques_id, feats, boxes, sent, target) in enumerate(loader):
_, label = target.max(1)
for qid, l in zip(ques_id, label.cpu().numpy()):
ans = dset.label2ans[l]
quesid2ans[qid.item()] = ans
return evaluator.evaluate(quesid2ans)
def save(self, name):
torch.save(self.model.state_dict(),
os.path.join(self.output, "%s.pth" % name))
def load(self, path):
print("Load model from %s" % path)
state_dict = torch.load("%s.pth" % path)
self.model.load_state_dict(state_dict)
if __name__ == "__main__":
# Build Class
vqa = VQA()
# Load VQA model weights
# Note: It is different from loading lxmert pre-trained weights.
if args.load is not None:
vqa.load(args.load)
# Test or Train
if args.test is not None:
args.fast = args.tiny = False # Always loading all data in test
if 'test' in args.test:
vqa.predict(
get_data_tuple(args.test, bs=950,
shuffle=False, drop_last=False),
dump=os.path.join(args.output, 'test_predict.json')
)
elif 'val' in args.test:
# Since part of valididation data are used in pre-training/fine-tuning,
# only validate on the minival set.
result = vqa.evaluate(
get_data_tuple('minival', bs=950,
shuffle=False, drop_last=False),
dump=os.path.join(args.output, 'minival_predict.json')
)
print(result)
else:
assert False, "No such test option for %s" % args.test
else:
print('Splits in Train data:', vqa.train_tuple.dataset.splits)
if vqa.valid_tuple is not None:
print('Splits in Valid data:', vqa.valid_tuple.dataset.splits)
print("Valid Oracle: %0.2f" % (vqa.oracle_score(vqa.valid_tuple) * 100))
else:
print("DO NOT USE VALIDATION")
vqa.train(vqa.train_tuple, vqa.valid_tuple)
|