def main(args):
dataset_splitBy = args.dataset + '_' + args.splitBy
if not osp.isdir(osp.join('cache/feats/', dataset_splitBy)):
os.makedirs(osp.join('cache/feats/', dataset_splitBy))
# Image Directory
if 'coco' in dataset_splitBy or 'combined' in dataset_splitBy:
IMAGE_DIR = 'data/images/mscoco/images/train2014'
elif 'clef' in dataset_splitBy:
IMAGE_DIR = 'data/images/saiapr_tc-12'
elif 'sunspot' in dataset_splitBy:
IMAGE_DIR = 'data/images/SUNRGBD'
else:
print('No image directory prepared for ', args.dataset)
sys.exit(0)
# load dataset
data_json = osp.join('cache/prepro', dataset_splitBy, 'data.json')
data_h5 = osp.join('cache/prepro', dataset_splitBy, 'data.h5')
loader = Loader(data_json, data_h5)
images = loader.images
anns = loader.anns
num_anns = len(anns)
assert sum([len(image['ann_ids']) for image in images]) == num_anns
# load mrcn model
mrcn = inference_no_imdb.Inference(args)
# feats_h5
# feats_h5 = osp.join('cache/feats', dataset_splitBy, args.file_name)
file_name = '%s_%s_%s_ann_feats.h5' % (args.net_name, args.imdb_name,
args.tag)
feats_h5 = osp.join('cache/feats', dataset_splitBy, 'mrcn', file_name)
f = h5py.File(feats_h5, 'w')
pool5_set = f.create_dataset('pool5', (num_anns, 1024), dtype=np.float32)
fc7_set = f.create_dataset('fc7', (num_anns, 2048), dtype=np.float32)
# extract
feats_dir = '%s_%s_%s' % (args.net_name, args.imdb_name, args.tag)
head_feats_dir = osp.join('cache/feats/', dataset_splitBy, 'mrcn',
feats_dir)
for i, image in enumerate(images):
image_id = image['image_id']
net_conv, im_info = image_to_head(head_feats_dir, image_id)
ann_ids = image['ann_ids']
for ann_id in ann_ids:
ann = loader.Anns[ann_id]
ann_pool5, ann_fc7 = ann_to_pool5_fc7(mrcn, ann, net_conv, im_info)
ann_h5_id = ann['h5_id']
pool5_set[ann_h5_id] = ann_pool5.data.cpu().numpy()
fc7_set[ann_h5_id] = ann_fc7.data.cpu().numpy()
if i % 20 == 0:
print('%s/%s done.' % (i + 1, len(images)))
f.close()
print('%s written.' % feats_h5)
def main():
data_json = '/data/ryli/kcli/refer_seg/MAttNet/cache/prepro/refcoco_unc/data.json'
data_h5 = '/data/ryli/kcli/refer_seg/MAttNet/cache/prepro/refcoco_unc/data.h5'
loader = Loader(data_json, data_h5)
parser = SubjectParser(len(loader.word_to_ix), 512, 512, 90)
parser.cuda()
learning_rate = 1e-4
optimizer = torch.optim.Adam(parser.parameters(), lr=learning_rate)
def lossFun(loader, optimizer, parser, input_label, class_label):
parser.train()
optimizer.zero_grad()
class_pred, rnn_output, rnn_hidden = parser(input_label)
cls_error = F.cross_entropy(class_pred, class_label)
#loss = confidence * torch.exp(F.threshold(-cls_error, -1.0, -1.0)) + 0.1 * (1 - confidence) * cls_error
loss = cls_error
loss.backward()
optimizer.step()
return loss.data[0], cls_error, class_pred
sent_count = 0
avg_accuracy = 0
for ref_id in loader.Refs:
ref = loader.Refs[ref_id]
class_label = loader.Anns[ref['ann_id']]['category_id'] - 1
for sent_id in ref['sent_ids']:
sent = loader.sentences[sent_id]
if len(sent['tokens']) > 2: continue
sent_count += 1
input_sent = [t for t in sent['tokens'] if t in loader.word_to_ix else '<UNK>']
input_label = [[loader.word_to_ix[word] for word in input_sent]]
input_label = Variable(torch.cuda.LongTensor(input_label))
class_label = Variable(torch.cuda.LongTensor(class_label))
loss, cls_error, cls_pred = lossFun(loader, optimizer, parser,
input_label, class_label)
_, pred = torch.max(cls_pred, 1)
if pred == class_label:
avg_accuracy = avg_accuracy * 0.99 + 0.01
else:
avg_accuracy *= 0.999
if sent_count % 100 == 0:
print('Sentence %d: id(%d)' % (sent_count, sent_id))
print(' %-12s: loss = %f, cls_error = %f, avg_accuracy = %.4f, lr = %.2E' %
(word, loss, cls_error, avg_accuracy, learning_rate))
def Loader(*args, **kwargs):
from loaders.loader import Loader
return Loader(*args, **kwargs)