dest='folder',
help='folder location',
type=str,
)
parser.add_argument(
'-no_tqdm',
dest='no_tqdm',
action='store_true',
)
args = parser.parse_args()
params = Params.from_file(args.params)
train, val, test = VCR.splits(
mode='rationale' if args.rationale else 'answer',
embs_to_load=params['dataset_reader'].get('embs', 'bert_da'),
only_use_relevant_dets=params['dataset_reader'].get(
'only_use_relevant_dets', True))
NUM_GPUS = torch.cuda.device_count()
NUM_CPUS = multiprocessing.cpu_count()
if NUM_GPUS == 0:
raise ValueError("you need gpus!")
def _to_gpu(td):
if NUM_GPUS > 1:
return td
for k in td:
td[k] = {k2: v.cuda(non_blocking=True)
for k2, v in td[k].items()} if isinstance(
td[k], dict) else td[k].cuda(non_blocking=True)
from dataloaders.vcr import VCR
from utils.extractionKeyword import extractionKeyword
from utils.extractionKnowledge import extractionKnowledge
from utils.topKknowledge import topKknowledge
import time # this is test
import torch
train_answer = VCR('train', 'answer')
train_rationale = VCR('train', 'rationale')
val_answer = VCR('val', 'answer')
val_rationale = VCR('val', 'rationale')
#test_answer = VCR('test','answer')
#test_rationale = VCR('test', 'rationale')
# definee keyword extractor
keywordExtractor = extractionKeyword()
# define knowledge extractor
knowledgeExtractor = extractionKnowledge(5, 10)
# define topK extractor
topKExtractor = topKknowledge(50, 10)
print('start!!!') # this is test
start = time.time() # this is test
k = 0
for t_answer, t_rationale in zip(train_answer, train_rationale):
answer_list = []
rationale_list = []
print('k : ', k)
k += 1
for i in range(4):
td[k] = {k2: torch.tensor(v).cuda()
for k2, v in td[k].items()} if isinstance(
td[k], dict) else td[k].cuda()
return td
num_workers = (4 * NUM_GPUS if NUM_CPUS == 32 else 2 * NUM_GPUS) - 1
print(f"Using {num_workers} workers out of {NUM_CPUS} possible", flush=True)
loader_params = {
'batch_size': 96 // NUM_GPUS,
'num_gpus': NUM_GPUS,
'num_workers': num_workers
}
vcr_modes = VCR.eval_splits(
embs_to_load=params['dataset_reader'].get('embs', 'bert_da'),
only_use_relevant_dets=params['dataset_reader'].get(
'only_use_relevant_dets', True))
probs_grp = []
ids_grp = []
for (vcr_dataset,
mode_long) in zip(vcr_modes,
['answer'] + [f'rationale_{i}' for i in range(4)]):
mode = mode_long.split('_')[0]
test_loader = VCRLoader.from_dataset(vcr_dataset, **loader_params)
# Load the params again because allennlp will delete them... ugh.
params = Params.from_file(args.params)
print("Loading {} for {}".format(params['model'].get('type', 'WTF?'),
mode),
flush=True)
import h5py
import numpy as np
import torch
from utils.detector_101 import SimpleDetector
from config import VCR_IMAGES_DIR, VCR_ANNOTS_DIR
from dataloaders.vcr import VCR, VCRLoader
################################################
#this is for vcr dataset
################################################
################################################
#data splits!
################################################
train, val = VCR.splits(mode='rationale',
embs_to_load='bert_da',
only_use_relevant_dets=False)
print('split is ok!')
################################################
#data loader!
################################################
loader_params = {'batch_size': 1, 'num_gpus': 1, 'num_workers': 1}
train_loader = VCRLoader.from_dataset(train, **loader_params)
val_loader = VCRLoader.from_dataset(val, **loader_params)
print('loader is ok!')
################################################
#define detector!