def test_image_hred_model(config):
db = vg(config, 'train')
loaddb = caption_loader(db)
loader = DataLoader(loaddb,
batch_size=3 * config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=caption_collate_fn)
net = ImageHREDModel(config)
net.train()
for name, param in net.named_parameters():
print(name, param.size())
for cnt, batched in enumerate(loader):
images = batched['images'].float()
sent_inds = batched['sent_inds'].long()
sent_msks = batched['sent_msks'].long()
img_feats, txt_feats = net(sent_inds, sent_msks, None, images)
print('images', images.size())
print('img_feats', img_feats.size())
print('txt_feats', txt_feats.size())
loss = net.forward_loss(img_feats, txt_feats)
print(loss)
metrics, caches = net.evaluate(img_feats, txt_feats)
print(metrics)
break
def dump_image_features(config):
output_dir = osp.join(config.data_dir, 'vg', 'global_features')
maybe_create(output_dir)
db = vg(config)
loaddb = caption_loader(db)
loader = DataLoader(loaddb,
batch_size=1,
shuffle=False,
num_workers=0,
collate_fn=caption_collate_fn)
net = ImageEncoder(config)
if config.cuda:
net = net.cuda()
net.eval()
for cnt, batched in enumerate(loader):
images = batched['images'].float()
if config.cuda:
images = images.cuda()
indices = batched['image_inds']
image_index = int(indices[0])
output_path = osp.join(output_dir, str(image_index).zfill(12) + '.npy')
features = net(images).squeeze().cpu().data.numpy()
assert (len(features) == 2048)
pickle_save(output_path, features)
print(cnt, image_index)
def test_region_grounding_model(config):
db = vg(config, 'test')
loaddb = region_loader(db)
loader = DataLoader(loaddb,
batch_size=3 * config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=region_collate_fn)
net = RegionGroundingModel(config)
if config.pretrained is not None:
pretrained_path = osp.join(config.data_dir,
'caches/region_grounding_ckpts',
config.pretrained + '.pkl')
states = torch.load(pretrained_path,
map_location=lambda storage, loc: storage)
net.load_state_dict(states['state_dict'], strict=False)
net.train()
for name, param in net.named_parameters():
print(name, param.size())
for cnt, batched in enumerate(loader):
scene_inds = batched['scene_inds'].long()
sent_inds = batched['sent_inds'].long()
sent_msks = batched['sent_msks'].long()
region_feats = batched['region_feats'].float()
region_clses = batched['region_clses'].long()
region_masks = batched['region_masks'].float()
img_feats, masked_feats, txt_feats, subspace_masks, sample_logits, sample_indices = \
net(scene_inds, sent_inds, sent_msks, None, None, None, region_feats, region_clses, region_masks, config.explore_mode)
if config.instance_dim > 1:
print(sample_indices[0])
# print('sample_logits', sample_logits.size())
# print('sample_indices', sample_indices.size())
txt_masks = txt_feats.new_ones(txt_feats.size(0), txt_feats.size(1))
losses = net.final_loss(img_feats, masked_feats, region_masks,
txt_feats, txt_masks, sample_logits,
sample_indices)
print('losses', losses.size(), torch.mean(losses))
if config.subspace_alignment_mode > 0:
metrics, cache_results = net.evaluate(masked_feats, region_masks,
txt_feats)
else:
metrics, cache_results = net.evaluate(img_feats, region_masks,
txt_feats)
print('metrics', metrics)
print('txt_feats', txt_feats.size())
print('img_feats', img_feats.size())
break
def create_text_reference_html(config):
config_html = HTML()
config_table = config_html.table(border='1')
testdb = vg(config, 'val')
for i in range(len(testdb.scenedb)):
scene = testdb.scenedb[i]
image_index = scene['image_index']
path = "http://www.cs.virginia.edu/~ft3ex/data/language_vision/val_image_htmls/%d.html"%image_index
r = config_table.tr
c = r.td()
c.a('%04d'%i, href='%s'%path)
html_file = open('reference.html', 'w')
print(config_table, file=html_file)
html_file.close()
def create_img_reference_html(config):
config_html = HTML()
config_table = config_html.table(border='1')
testdb = vg(config, 'test')
for i in range(len(testdb.scenedb)):
scene = testdb.scenedb[i]
image_index = scene['image_index']
img_path = "http://www.cs.virginia.edu/~ft3ex/data/vg/VG_100K/%d.jpg"%image_index
html_path = "http://www.cs.virginia.edu/~ft3ex/data/language_vision/test_image_htmls/%d.html"%image_index
c = config_table.tr
a = c.a(href='%s'%html_path)
a.img(src='%s'%img_path, height='150')
html_file = open('img_reference.html', 'w')
print(config_table, file=html_file)
html_file.close()
def test_image_encoder(config):
db = vg(config, 'test')
loaddb = caption_loader(db)
loader = DataLoader(loaddb,
batch_size=3 * config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=caption_collate_fn)
net = ImageEncoder(config)
for cnt, batched in enumerate(loader):
images = batched['images'].float()
print('images', images.size())
feats = net(images)
print('features', feats.size())
break
def test_caption_loader(config):
db = vg(config, 'train')
# db = coco(config, 'train')
loaddb = caption_loader(db)
output_dir = osp.join(config.model_dir, 'test_caption_dataloader')
maybe_create(output_dir)
loader = DataLoader(loaddb,
batch_size=config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=caption_collate_fn)
start = time()
plt.switch_backend('agg')
for cnt, batched in enumerate(loader):
sent_inds = batched['sent_inds'].long()
sent_msks = batched['sent_msks'].long()
images = batched['images'].float()
captions = batched['captions']
print('sent_inds', sent_inds.size())
print('sent_msks', sent_msks.size())
print('images', images.size())
for i in range(config.batch_size):
color = cv2.imread(
db.color_path_from_index(batched['image_inds'][i]),
cv2.IMREAD_COLOR)
out_path = osp.join(output_dir,
'%d.png' % batched['image_inds'][i])
fig = plt.figure(figsize=(32, 16))
for j in range(min(config.max_turns, 10)):
plt.subplot(2, 5, j + 1)
plt.title(
captions[i][j] + '\n' +
' '.join([str(x.data.item())
for x in sent_inds[i, j]]) + '\n' +
' '.join([str(x.data.item()) for x in sent_msks[i, j]]))
plt.imshow(color[:, :, ::-1])
plt.axis('off')
fig.savefig(out_path, bbox_inches='tight')
plt.close(fig)
print('------------------')
if cnt == 2:
break
print("Time", time() - start)
def test_region_model(config):
db = vg(config, 'test')
loaddb = region_loader(db)
loader = DataLoader(loaddb,
batch_size=3 * config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=region_collate_fn)
net = RegionModel(config)
net.train()
for name, param in net.named_parameters():
print(name, param.size())
for cnt, batched in enumerate(loader):
start = time()
scene_inds = batched['scene_inds'].long()[:config.batch_size]
sent_inds = batched['sent_inds'].long()[:config.batch_size]
sent_msks = batched['sent_msks'].long()[:config.batch_size]
region_feats = batched['region_feats'].float()[:config.batch_size]
region_clses = batched['region_clses'].long()[:config.batch_size]
region_masks = batched['region_masks'].float()[:config.batch_size]
src_region_feats = batched['region_feats'].float(
)[config.batch_size:2 * config.batch_size]
src_region_clses = batched['region_clses'].long()[config.batch_size:2 *
config.batch_size]
src_region_masks = batched['region_masks'].float(
)[config.batch_size:2 * config.batch_size]
img_feats, masked_feats, txt_feats, subspace_masks, sample_logits, sample_indices = \
net(scene_inds, sent_inds, sent_msks,
src_region_feats, src_region_clses, src_region_masks,
region_feats, region_clses, region_masks,
config.explore_mode)
print('img_feats', img_feats.size())
print('txt_feats', txt_feats.size())
if config.subspace_alignment_mode > 0:
print('masked_feats', masked_feats.size())
print('subspace_masks', subspace_masks.size())
if config.instance_dim > 1:
print('sample_logits', sample_logits.size())
print('sample_indices', sample_indices.size())
print('time:', time() - start)
break
def main_rnn(config):
testdb = vg(config, 'test')
trainer = RegionGroundingTrainer(config)
with open(
osp.join(testdb.cache_dir,
'img_features/vg_rnn_1280_img_features.pkl'),
'rb') as fid:
data_ = pickle.load(fid)
all_img_feats = data_['feats']
all_img_masks = data_['masks']
all_img_feats = torch.from_numpy(all_img_feats).float()
all_img_masks = torch.from_numpy(all_img_masks).float()
if config.cuda:
all_img_feats = all_img_feats.cuda()
all_img_masks = all_img_masks.cuda()
print('all_img_feats', all_img_feats.size())
print('all_img_masks', all_img_masks.size())
count = 0
all_captions = []
while count < 10:
print('Please input the query:\n')
query = input()
r, top5_img_inds = trainer.net.demo_step(query, all_captions,
all_img_feats, all_img_masks,
testdb)
top5_imgs = []
for x in top5_img_inds:
cur_img = cv2.imread(testdb.color_path_from_index(x),
cv2.IMREAD_COLOR)
cur_img, _, _ = create_squared_image(cur_img)
cur_img = cv2.resize(cur_img, (500, 500))
top5_imgs.append(cur_img)
fig = plt.figure(figsize=(32, 8))
plt.suptitle(query, fontsize=20)
for i in range(len(top5_imgs)):
cur_img = top5_imgs[i]
plt.subplot(1, 5, i + 1)
plt.imshow(cur_img[:, :, ::-1].astype(np.uint8))
plt.axis('off')
plt.show()
count += 1
print('turn:', count)
def check_region_clses(config):
db = vg(config, 'train')
loaddb = region_loader(db)
loader = DataLoader(loaddb,
batch_size=config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=region_collate_fn)
min_index = 1000000
max_index = -1
for cnt, batched in enumerate(loader):
region_clses = batched['region_clses'].long()
min_index = min(min_index, torch.min(region_clses).item())
max_index = max(max_index, torch.max(region_clses).item())
if cnt % 1000:
print('iter:', cnt)
print('min_index', min_index)
print('max_index', max_index)
def test_image_model(config):
db = vg(config, 'test')
loaddb = caption_loader(db)
loader = DataLoader(loaddb,
batch_size=config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=caption_collate_fn)
net = ImageModel(config)
for cnt, batched in enumerate(loader):
images = batched['images'].float()
sent_inds = batched['sent_inds'].long()
sent_msks = batched['sent_msks'].long()
img_feats, txt_feats = net(sent_inds, sent_msks, None, images)
print('images', images.size())
print('img_feats', img_feats.size())
print('txt_feats', txt_feats.size())
break
def test_grounding_loss(config):
db = vg(config, 'test')
loaddb = region_loader(db)
loader = DataLoader(loaddb,
batch_size=3 * config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=region_collate_fn)
net = RegionModel(config)
criterion = GroundingLoss(config)
for cnt, batched in enumerate(loader):
scene_inds = batched['scene_inds'].long()[:config.batch_size]
sent_inds = batched['sent_inds'].long()[:config.batch_size]
sent_msks = batched['sent_msks'].long()[:config.batch_size]
region_feats = batched['region_feats'].float()[:config.batch_size]
region_clses = batched['region_clses'].long()[:config.batch_size]
region_masks = batched['region_masks'].float()[:config.batch_size]
src_region_feats = batched['region_feats'].float(
)[config.batch_size:2 * config.batch_size]
src_region_clses = batched['region_clses'].long()[config.batch_size:2 *
config.batch_size]
src_region_masks = batched['region_masks'].float(
)[config.batch_size:2 * config.batch_size]
img_feats, masked_feats, txt_feats, subspace_masks, sample_logits, sample_indices = \
net(scene_inds, sent_inds, sent_msks,
src_region_feats, src_region_clses, src_region_masks,
region_feats, region_clses, region_masks,
config.explore_mode)
masked_feats = img_feats
sim1 = criterion.compute_batch_mutual_similarity(
masked_feats, region_masks, txt_feats)
sim2 = criterion.debug_compute_batch_mutual_similarity(
masked_feats, region_masks, txt_feats)
print('sim1', sim1.size())
print('sim2', sim2.size())
print('diff', torch.sum(torch.abs(sim1 - sim2)))
txt_masks = txt_feats.new_ones(txt_feats.size(0), txt_feats.size(1))
losses = criterion.forward_loss(masked_feats, region_masks, txt_feats,
txt_masks, config.loss_reduction_mode)
print('losses', losses.size())
break
def test_paragraph_model(config):
db = vg(config, 'test')
loaddb = paragraph_loader(db)
loader = DataLoader(loaddb,
batch_size=3 * config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=paragraph_collate_fn)
net = ParagraphModel(config)
net.train()
for name, param in net.named_parameters():
print(name, param.size())
for cnt, batched in enumerate(loader):
start = time()
scene_inds = batched['scene_inds'].long()[:config.batch_size]
sent_inds = batched['sent_inds'].long()[:config.batch_size]
sent_msks = batched['sent_msks'].long()[:config.batch_size]
region_feats = batched['region_feats'].float()[:config.batch_size]
region_clses = batched['region_clses'].long()[:config.batch_size]
region_masks = batched['region_masks'].float()[:config.batch_size]
img_feats, txt_feats = net(sent_inds, sent_msks, region_feats,
region_clses, region_masks)
losses = net.loss(img_feats, region_masks, txt_feats.unsqueeze(1))
print('losses', losses.size(), torch.mean(losses))
metrics, cache_results = net.evaluate(img_feats, region_masks,
txt_feats.unsqueeze(1))
print('metrics', metrics)
print('sent_inds', sent_inds.size())
print('sent_msks', sent_msks.size())
print('region_feats', region_feats.size())
print('region_clses', region_clses.size())
print('region_masks', region_masks.size())
print('img_feats', img_feats.size())
print('txt_feats', txt_feats.size())
print('time:', time() - start)
break
def test_text_encoder(config):
db = vg(config, 'test')
loaddb = region_loader(db)
loader = DataLoader(loaddb,
batch_size=3 * config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=region_collate_fn)
net = TextEncoder(config)
for cnt, batched in enumerate(loader):
sent_inds = batched['sent_inds'].long()
sent_msks = batched['sent_msks'].float()
bsize, slen, fsize = sent_inds.size()
print('sent_inds', sent_inds.size())
print('sent_msks', sent_msks.size())
f1, f2, h = net(sent_inds.view(bsize * slen, fsize),
sent_msks.view(bsize * slen, fsize))
print(f1.size(), f2.size(), h.size())
break
def test_region_encoder(config):
db = vg(config, 'test')
loaddb = region_loader(db)
loader = DataLoader(loaddb,
batch_size=3 * config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=region_collate_fn)
net = RegionEncoder(config)
for cnt, batched in enumerate(loader):
region_feats = batched['region_feats'].float()
region_clses = batched['region_clses'].long()
print('region_feats', region_feats.size())
print('region_clses', region_clses.size())
img_feats, masked_feats, mm = net(region_feats, region_clses)
print('img_feats', img_feats.size())
if config.subspace_alignment_mode > 0:
print('masked_feats', masked_feats.size())
print('mm', mm.size())
break
def create_html_per_image(config):
testdb = vg(config, 'test')
image_folder_name = 'test_image_htmls'
maybe_create(image_folder_name)
for i in range(len(testdb.scenedb)):
scene = testdb.scenedb[i]
all_meta_regions = [scene['regions'][x] for x in sorted(list(scene['regions'].keys()))]
captions = [x['caption'] for x in all_meta_regions[:config.max_turns]]
image_index = scene['image_index']
text = '\n'.join(captions)
path = "http://www.cs.virginia.edu/~ft3ex/data/vg/VG_100K/%d.jpg"%image_index
config_html = HTML()
config_table = config_html.table(border='1')
r = config_table.tr
c1 = r.td()
c1.img(src='%s'%path, height='700')
c2 = r.td()
for j in range(len(captions)):
c2.p(captions[j])
html_file = open(osp.join(image_folder_name, '%d.html'%image_index), 'w')
print(config_table, file=html_file)
html_file.close()
print(i)
def test_model(config):
testdb = vg(config, 'test')
trainer = RegionGroundingTrainer(config)
trainer.test(testdb)
def test_response_gen(config):
s = time()
db = vg(config)
html_folder_name = 'template_htmls'
maybe_create(html_folder_name)
with open('candidates.json', 'r') as fp:
candidates = json.load(fp)
captioner = RelativeCaptioner(db)
for k, v in candidates.items():
target_scene = db.scenedb[v['src']]
decoy_scenes = [db.scenedb[x] for x in v['top5']]
unmention_candidates = captioner.collect_unmentioned_candidates(
target_scene, decoy_scenes)
captions_1 = []
if len(unmention_candidates) > 0:
cap1 = caption_1(
unmention_candidates[np.random.randint(
0, len(unmention_candidates))], target_scene)
if cap1 is not None:
captions_1.append(cap1)
cap2 = caption_2(
unmention_candidates[np.random.randint(
0, len(unmention_candidates))], target_scene)
if cap2 is not None:
captions_1.append(cap2)
cap3 = caption_3(
unmention_candidates[np.random.randint(
0, len(unmention_candidates))], target_scene)
if cap3 is not None:
captions_1.append(cap3)
# print(cap3)
captions_2 = []
mention_candidates = captioner.collect_mentioned_candidates(
target_scene, decoy_scenes)
if len(mention_candidates) > 0:
cap1 = caption_1(
mention_candidates[np.random.randint(0,
len(mention_candidates))],
target_scene)
if cap1 is not None:
captions_2.append(cap1)
cap2 = caption_2(
mention_candidates[np.random.randint(0,
len(mention_candidates))],
target_scene)
if cap2 is not None:
captions_2.append(cap2)
cap3 = caption_3(
mention_candidates[np.random.randint(0,
len(mention_candidates))],
target_scene)
if cap3 is not None:
captions_2.append(cap3)
# print(cap3)
cap4 = caption_4(
mention_candidates[np.random.randint(0,
len(mention_candidates))],
target_scene)
if cap4 is not None:
captions_2.append(cap4)
cap5 = caption_5(
mention_candidates[np.random.randint(0,
len(mention_candidates))],
target_scene)
if cap5 is not None:
captions_2.append(cap5)
# print(cap5)
# query_path = "http://www.cs.virginia.edu/~ft3ex/data/vg/VG_100K/%d.jpg"%v['src']
# top5_paths = ["http://www.cs.virginia.edu/~ft3ex/data/vg/VG_100K/%d.jpg"%x for x in v['top5']]
query_path = "file:///Users/fuwentan/datasets/vg/VG_100K/%d.jpg" % v[
'src']
top5_paths = [
"file:///Users/fuwentan/datasets/vg/VG_100K/%d.jpg" % x
for x in v['top5']
]
config_html = HTML()
config_table = config_html.table(border='1')
r1 = config_table.tr
c1 = r1.td(colspan="2")
for j in range(len(captions_1)):
c1.p(captions_1[j])
c2 = r1.td()
c2.img(src='%s' % query_path, height='200')
c3 = r1.td(colspan="2")
for j in range(len(captions_2)):
c3.p(captions_2[j])
r2 = config_table.tr
for j in range(5):
c2_r2_c = r2.td()
c2_r2_c.img(src='%s' % top5_paths[j], height='200')
html_file = open(
osp.join(html_folder_name, '%d_%d.html' % (v['src'], v['turn'])),
'w')
print(config_table, file=html_file)
html_file.close()
print(k)
elif args.vg_dataset == "vg6":
vg_version = "1600-400-400"
vg_split = imdb_vg_name
set_cfgs = [
'ANCHOR_SCALES', '[4, 8, 16, 32]', 'ANCHOR_RATIOS',
'[0.25, 0.5, 1, 2, 4]'
]
elif args.vg_dataset == "vg_bm":
vg_version = "150-50-50"
vg_split = imdb_vg_name
set_cfgs = [
'ANCHOR_SCALES', '[4, 8, 16, 32]', 'ANCHOR_RATIOS',
'[0.25, 0.5, 1, 2, 4]'
]
imdb_vg = vg(vg_version, vg_split)
def bbox_proposal_fast(obj_prob, att_prob, rois):
batch_size = obj_prob.size(0)
# get the top obj cls, excluded the background class.
max_obj_prob, max_obj_clss = obj_prob[:, :, 1:].max(2)
# get the top att cls, exclude the background class.
max_att_prob, max_att_clss = att_prob[:, :, 1:].max(2)
# get the top rel cls, exlude the background class
# max_rel_scores, max_rel_ind = rel_prob[:, :, 1:].max(2)
# compute the final score, B x N
obj_att_scores = max_obj_prob * max_att_prob
def test_region_loader(config):
db = vg(config, 'train')
# db = coco(config, 'train')
loaddb = region_loader(db)
loader = DataLoader(loaddb,
batch_size=config.batch_size,
shuffle=True,
num_workers=config.num_workers,
collate_fn=region_collate_fn)
output_dir = osp.join(config.model_dir, 'test_region_loader')
maybe_create(output_dir)
start = time()
plt.switch_backend('agg')
for cnt, batched in enumerate(loader):
print('scene_inds', batched['scene_inds'])
sent_inds = batched['sent_inds'].long()
sent_msks = batched['sent_msks'].long()
widths = batched['widths']
heights = batched['heights']
captions = batched['captions']
region_boxes = batched['region_boxes'].float()
region_feats = batched['region_feats'].float()
region_clses = batched['region_clses'].long()
region_masks = batched['region_masks'].long()
print('sent_inds', sent_inds.size())
print('sent_msks', sent_msks.size())
print('region_boxes', region_boxes.size())
print('region_feats', region_feats.size())
print('region_clses', region_clses.size())
print('region_masks', region_masks.size())
print('clses', torch.min(region_clses), torch.max(region_clses))
print('widths', widths)
print('heights', heights)
for i in range(len(sent_inds)):
# print('####')
# print(len(captions), len(captions[0]))
entry = {}
image_index = batched['image_inds'][i]
entry['width'] = widths[i]
entry['height'] = heights[i]
nr = torch.sum(region_masks[i])
entry['region_boxes'] = xyxys_to_xywhs(
region_boxes[i, :nr].cpu().data.numpy())
color = cv2.imread(db.color_path_from_index(image_index),
cv2.IMREAD_COLOR)
color, _, _ = create_squared_image(color)
out_path = osp.join(output_dir, '%d.png' % image_index)
layouts = db.render_regions_as_output(
entry,
bg=cv2.resize(
color,
(config.visu_size[0], config.visu_size[0]))[:, :, ::-1])
fig = plt.figure(figsize=(32, 16))
for j in range(min(14, len(layouts))):
plt.subplot(3, 5, j + 1)
if j < config.max_turns:
plt.title(
captions[i][j] + '\n' +
' '.join([str(x.data.item())
for x in sent_inds[i, j]]) + '\n' +
' '.join([str(x.data.item())
for x in sent_msks[i, j]]))
plt.imshow(layouts[j].astype(np.uint8))
plt.axis('off')
plt.subplot(3, 5, 15)
plt.imshow(color[:, :, ::-1])
plt.axis('off')
fig.savefig(out_path, bbox_inches='tight')
plt.close(fig)
print('------------------')
if cnt == 2:
break
print("Time", time() - start)
def overfit_model(config):
valdb = vg(config, 'val')
valdb.scenedb = valdb.scenedb[:31]
trainer = ImageHREDTrainer(config)
trainer.train(valdb, valdb, valdb)
def train_model(config):
traindb = vg(config, 'train')
valdb = vg(config, 'val')
trainer = ImageHREDTrainer(config)
trainer.train(traindb, valdb, valdb)
def overfit_model(config):
valdb = vg(config, 'val')
valdb.scenedb = valdb.scenedb[:31]
trainer = RegionGroundingTrainer(config)
trainer.train(valdb, valdb, valdb)
# Set up coco_2015_<split>
for year in ['2015']:
for split in ['test', 'test-dev']:
name = 'coco_{}_{}'.format(year, split)
__sets[name] = (lambda split=split, year=year: coco(split, year))
# Set up vg_<split>
# for version in ['1600-400-20']:
# for split in ['minitrain', 'train', 'minival', 'val', 'test']:
# name = 'vg_{}_{}'.format(version,split)
# __sets[name] = (lambda split=split, version=version: vg(version, split))
for version in ['150-50-20', '150-50-50', '500-150-80', '750-250-150', '1750-700-450', '1600-400-20']:
for split in ['minitrain', 'smalltrain', 'train', 'minival', 'smallval', 'val', 'test']:
name = 'vg_{}_{}'.format(version,split)
__sets[name] = (lambda split=split, version=version: vg(version, split))
# set up image net.
for split in ['train', 'val', 'val1', 'val2', 'test']:
name = 'imagenet_{}'.format(split)
devkit_path = 'data/imagenet/ILSVRC/devkit'
data_path = 'data/imagenet/ILSVRC'
__sets[name] = (lambda split=split, devkit_path=devkit_path, data_path=data_path: imagenet(split,devkit_path,data_path))
for split in ['train', 'val', 'test']:
name = 'wider_face_{}'.format(split)
__sets[name] = (lambda split=split: wider_face(split))
for split in ['train', 'val', 'test']:
name = 'MI3_{}'.format(split)
__sets[name] = (lambda split=split: mi3(split))
def test_vg_dataset(config):
s = time()
db = vg(config, 'train')
def dump_trained_features(config):
traindb = vg(config, 'train')
trainer = RegionGroundingTrainer(config)
trainer.test(traindb)
def test_model(config):
testdb = vg(config, 'test')
trainer = ImageHREDTrainer(config)
trainer.test(testdb)
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
__sets = {}
from datasets.vg import vg
from datasets.clevr import clevr
from datasets.visualgenome import visualgenome
from datasets.vrd import vrd
import numpy as np
# Set up vg_<split>
for split in ['train', 'validation', 'test']:
name = 'visual_genome_{}'.format(split)
__sets[name] = (lambda split=split: vg(split))
# Set up clevr_<split>
for split in ['train', 'val', 'test']:
name = 'clevr_{}'.format(split)
__sets[name] = (lambda split=split: clevr(split))
# Set up vg_<split>
for split in ['train', 'validation', 'test']:
name = 'visualgenome_{}'.format(split)
__sets[name] = (lambda split=split: visualgenome(split))
# Set up vrd_<split>
for split in ['train', 'validation', 'test']:
name = 'vrd_{}'.format(split)
__sets[name] = (lambda split=split: vrd(split))
def dump_trained_features(config):
traindb = vg(config, 'train')
trainer = ImageHREDTrainer(config)
trainer.test(traindb)
# Set up vg_<split>
# for version in ['1600-400-20']:
# for split in ['minitrain', 'train', 'minival', 'val', 'test']:
# name = 'vg_{}_{}'.format(version,split)
# __sets[name] = (lambda split=split, version=version: vg(version, split))
for version in [
'150-50-20', '150-50-50', '500-150-80', '750-250-150', '1750-700-450',
'1600-400-20'
]:
for split in [
'minitrain', 'smalltrain', 'train', 'minival', 'smallval', 'val',
'test'
]:
name = 'vg_{}_{}'.format(version, split)
__sets[name] = (
lambda split=split, version=version: vg(version, split))
# set up image net.
for split in ['train', 'val', 'val1', 'val2', 'test']:
name = 'imagenet_{}'.format(split)
devkit_path = 'data/imagenet/ILSVRC/devkit'
data_path = 'data/imagenet/ILSVRC'
__sets[name] = (lambda split=split, devkit_path=devkit_path, data_path=
data_path: imagenet(split, devkit_path, data_path))
def get_imdb(name):
"""Get an imdb (image database) by name."""
if name not in __sets:
raise KeyError('Unknown dataset: {}'.format(name))
return __sets[name]()
# Set up coco_2015_<split>
for year in ['2015']:
for split in ['test', 'test-dev']:
name = 'coco_{}_{}'.format(year, split)
__sets[name] = (lambda split=split, year=year: coco(split, year))
# Set up vg_<split>
# for version in ['1600-400-20']:
# for split in ['minitrain', 'train', 'minival', 'val', 'test']:
# name = 'vg_{}_{}'.format(version,split)
# __sets[name] = (lambda split=split, version=version: vg(version, split))
for version in ['150-50-20', '150-50-50', '500-150-80', '750-250-150', '1750-700-450', '1600-400-20']:
for split in ['minitrain', 'smalltrain', 'train', 'minival', 'smallval', 'val', 'test']:
name = 'vg_{}_{}'.format(version,split)
__sets[name] = (lambda split=split, version=version: vg(version, split))
# set up image net.
for split in ['train', 'val', 'val1', 'val2', 'test']:
name = 'imagenet_{}'.format(split)
devkit_path = 'data/imagenet/ILSVRC/devkit'
data_path = 'data/imagenet/ILSVRC'
__sets[name] = (lambda split=split, devkit_path=devkit_path, data_path=data_path: imagenet(split,devkit_path,data_path))
def get_imdb(name):
"""Get an imdb (image database) by name."""
if name not in __sets:
raise KeyError('Unknown dataset: {}'.format(name))
return __sets[name]()