def train_model(config):
transformer = image_normalize('background')
train_db = abstract_scene(config, split='train', transform=transformer)
val_db = abstract_scene(config, split='val', transform=transformer)
test_db = abstract_scene(config, split='test', transform=transformer)
trainer = SupervisedTrainer(train_db)
trainer.train(train_db, val_db, test_db)
def test_loader(config):
from torch.utils.data import DataLoader
transformer = image_normalize('background')
# traindb = abstract_scene(config, 'train', transform=transformer)
# print('traindb', len(traindb))
# valdb = abstract_scene(config, 'val', transform=transformer)
# print('valdb', len(valdb))
# testdb = abstract_scene(config, 'test', transform=transformer)
# print('testdb', len(testdb))
# print(testdb.scenedb[0]['scene_idx'], testdb.scenedb[-1]['scene_idx'])
db = abstract_scene(config, 'val', transform=transformer)
loader = DataLoader(db,
batch_size=config.batch_size,
shuffle=True,
num_workers=config.num_workers)
for cnt, batched in enumerate(loader):
print(batched['scene_idx'].size())
print(batched['word_inds'].size())
print(batched['word_lens'].size())
print(batched['word_inds'][0])
print(batched['word_lens'][0])
print(batched['background'].size())
print(batched['out_inds'].size())
print(batched['out_msks'].size())
print(batched['fg_inds'].size())
print(batched['hmaps'].size())
print(batched['out_inds'][0,:,0])
print(batched['fg_inds'][0])
break
def test_dataset(config):
db = abstract_scene(config, 'train')
plt.switch_backend('agg')
output_dir = osp.join(config.model_dir, 'test_abstract_scene')
maybe_create(output_dir)
indices = np.random.permutation(range(len(db)))
indices = indices[:config.n_samples]
for i in indices:
entry = db[i]
scene = db.scenedb[i]
# print('cls_inds: ', scene['cls_inds'])
imgs = db.render_scene_as_input(scene, True, True)
name = osp.splitext(osp.basename(entry['color_path']))[0]
out_path = osp.join(output_dir, name+'.png')
fig = plt.figure(figsize=(16, 8))
plt.suptitle(entry['sentence'])
for j in range(min(len(imgs), 11)):
plt.subplot(4, 3, j+1)
plt.imshow(imgs[j,:,:,::-1].astype(np.uint8))
plt.axis('off')
target = cv2.imread(entry['color_path'], cv2.IMREAD_COLOR)
plt.subplot(4, 3, 12)
plt.imshow(target[:,:,::-1])
plt.axis('off')
fig.savefig(out_path, bbox_inches='tight')
plt.close(fig)
def test_lang_vocab(config):
db = abstract_scene(config, 'train')
scenedb = db.scenedb
lang_vocab = db.lang_vocab
sent_lens = []
for i in range(len(scenedb)):
group_sents = scenedb[i]['scene_sentences']
for j in range(len(group_sents)):
triplet = group_sents[j]
for k in range(len(triplet)):
sentence = triplet[k]
tokens = word_tokenize(sentence.lower())
tokens = further_token_process(tokens)
word_inds = [lang_vocab.word_to_index(w) for w in tokens]
# word_inds = [wi for wi in word_inds if wi > config.EOS_idx]
sent_lens.append(len(word_inds))
print('sent len: ', np.median(sent_lens), np.amax(sent_lens), np.amin(sent_lens))
ps = [80.0, 90.0, 95.0, 99.0]
for p in ps:
print('p %d/100: '%(int(p)), np.percentile(sent_lens, p))
# 10.0 50 6
print("vocab size: ", len(lang_vocab.index2word))
print("vocab: ", lang_vocab.index2word[:10])
obj_lens = []
for i in range(len(scenedb)):
clses = scenedb[i]['clips']
obj_lens.append(len(clses))
print('obj len: ', np.median(obj_lens), np.amax(obj_lens), np.amin(obj_lens))
for p in ps:
print('p %d/100: '%(int(p)), np.percentile(obj_lens, p))
def test_clip_and_triplet(config):
db = abstract_scene(config)
clip_vocab = db.clip_vocab
for i in range(len(clip_vocab.index2word)):
o, p, e = db.clip_to_triplet(i)
w = clip_vocab.index2word[i]
print(i, o, p, e, w)
j = db.triplet_to_clip([o, p, e])
assert(i == j)
def test_decoder(config):
transformer = image_normalize('background')
db = abstract_scene(config, 'train', transform=transformer)
text_encoder = TextEncoder(db)
img_encoder = ImageEncoder(config)
what_decoder = WhatDecoder(config)
where_decoder = WhereDecoder(config)
# print(where_decoder)
loader = DataLoader(db,
batch_size=4,
shuffle=False,
num_workers=config.num_workers)
for cnt, batched in enumerate(loader):
word_inds = batched['word_inds'].long()
word_lens = batched['word_lens'].long()
bg_imgs = batched['background'].float()
encoder_states = text_encoder(word_inds, word_lens)
bg_feats = img_encoder(bg_imgs)
prev_bgfs = bg_feats[:, 0].unsqueeze(1)
prev_states = {}
prev_states['bgfs'] = prev_bgfs
what_outs = what_decoder(prev_states, encoder_states)
print('------------------------------------------')
print('obj_logits', what_outs['obj_logits'].size())
print('rnn_outs', what_outs['rnn_outs'][0].size())
print('hids', what_outs['hids'][0].size())
print('attn_ctx', what_outs['attn_ctx'].size())
print('attn_wei', what_outs['attn_wei'].size())
obj_logits = what_outs['obj_logits']
# print('obj_logits ', obj_logits.size())
_, obj_inds = torch.max(obj_logits + 1.0, dim=-1)
curr_fgfs = indices2onehots(obj_inds.cpu().data,
config.output_cls_size)
# curr_fgfs = curr_fgfs.unsqueeze(1)
if config.cuda:
curr_fgfs = curr_fgfs.cuda()
curr_fgfs = curr_fgfs.float()
what_outs['fgfs'] = curr_fgfs
where_outs = where_decoder(what_outs, encoder_states)
print('coord_logits ', where_outs['coord_logits'].size())
print('attri_logits ', where_outs['attri_logits'].size())
print('attn_ctx', where_outs['attn_ctx'].size())
print('attn_wei', where_outs['attn_wei'].size())
break
def abstract_demo(config, input_app):
transformer = image_normalize('background')
train_db = abstract_scene(config, split='train', transform=transformer)
trainer = SupervisedTrainer(train_db)
input_sentences = json_load('examples/abstract_samples.json')
#print(type(input_sentences))
#print(input_sentences[0])
#input_sentences_2 = prepare_data('Mike talks to the dog.,Jenny kicks the soccer ball.,The duck wants to play.')
#print(input_sentences_2)
input_app = prepare_data(input_app)
#print(type(input_sentences_2))
#print(input_app)
#print(type(input_app))
trainer.sample_demo(input_app)
def test_img_encoder(config):
transformer = image_normalize('background')
db = abstract_scene(config, 'train', transform=transformer)
img_encoder = ImageEncoder(config)
print(get_n_params(img_encoder))
loader = DataLoader(db,
batch_size=config.batch_size,
shuffle=False,
num_workers=config.num_workers)
for cnt, batched in enumerate(loader):
bg_imgs = batched['background'].float()
y = img_encoder(bg_imgs)
print(y.size())
break
def test_txt_encoder_abstract(config):
transformer = image_normalize('background')
db = abstract_scene(config, 'train', transform=transformer)
net = TextEncoder(db)
loader = DataLoader(db,
batch_size=config.batch_size,
shuffle=False,
num_workers=config.num_workers)
for cnt, batched in enumerate(loader):
input_inds = batched['word_inds'].long()
input_lens = batched['word_lens'].long()
print('Checking the output shapes')
out = net(input_inds, input_lens)
out_rfts = out['rfts']
out_embs = out['embs']
out_msks = out['msks']
out_hids = out['hids']
print(out_rfts.size(), out_embs.size(), out_msks.size())
if isinstance(out_hids[0], tuple):
print(out_hids[0][0].size())
else:
print(out_hids[0].size())
print('m: ', out_msks[-1])
print('Checking the embedding')
embeded = net.embedding(input_inds[:, 0, :])
v1 = embeded[0, 0]
idx = input_inds[0, 0, 0].data.item()
v2 = db.lang_vocab.vectors[idx]
diff = v2 - v1
print('Diff: (should be zero)', torch.sum(diff.abs_()))
break
def test_model(config):
transformer = image_normalize('background')
db = abstract_scene(config, 'val', transform=transformer)
net = DrawModel(db)
plt.switch_backend('agg')
output_dir = osp.join(config.model_dir, 'test_model')
maybe_create(output_dir)
pretrained_path = osp.join(
'../data/caches/abstract_ckpts/supervised_abstract_top1.pkl')
assert osp.exists(pretrained_path)
if config.cuda:
states = torch.load(pretrained_path)
else:
states = torch.load(pretrained_path,
map_location=lambda storage, loc: storage)
net.load_state_dict(states)
loader = DataLoader(db,
batch_size=config.batch_size,
shuffle=False,
num_workers=config.num_workers)
net.eval()
for cnt, batched in enumerate(loader):
word_inds = batched['word_inds'].long()
word_lens = batched['word_lens'].long()
bg_images = batched['background'].float()
fg_inds = batched['fg_inds'].long()
gt_inds = batched['out_inds'].long()
gt_msks = batched['out_msks'].float()
hmaps = batched['hmaps'].float()
fg_onehots = indices2onehots(fg_inds, config.output_cls_size)
fg_onehots = fg_onehots
inf_outs = net.teacher_forcing(word_inds, word_lens, bg_images,
fg_onehots, hmaps)
print('teacher forcing')
print('obj_logits ', inf_outs['obj_logits'].size())
print('coord_logits ', inf_outs['coord_logits'].size())
print('attri_logits ', inf_outs['attri_logits'].size())
if config.what_attn:
print('what_att_logits ', inf_outs['what_att_logits'].size())
if config.where_attn > 0:
print('where_att_logits ', inf_outs['where_att_logits'].size())
print('----------------------')
# inf_outs, env = net(word_inds, word_lens, -1, 0, 0, gt_inds)
inf_outs, env = net(word_inds, word_lens, 0, 1, 0, None)
print('scheduled sampling')
print('obj_logits ', inf_outs['obj_logits'].size())
print('coord_logits ', inf_outs['coord_logits'].size())
print('attri_logits ', inf_outs['attri_logits'].size())
if config.what_attn:
print('what_att_logits ', inf_outs['what_att_logits'].size())
if config.where_attn > 0:
print('where_att_logits ', inf_outs['where_att_logits'].size())
print('----------------------')
pred_out_inds, pred_out_msks = env.get_batch_inds_and_masks()
print('pred_out_inds', pred_out_inds[0, 1], pred_out_inds.shape)
print('gt_inds', gt_inds[0, 1], gt_inds.size())
print('pred_out_msks', pred_out_msks[0, 1], pred_out_msks.shape)
print('gt_msks', gt_msks[0, 1], gt_msks.size())
batch_frames = env.batch_redraw(True)
scene_inds = batched['scene_idx']
for i in range(len(scene_inds)):
sid = scene_inds[i]
entry = db[sid]
name = osp.splitext(osp.basename(entry['color_path']))[0]
imgs = batch_frames[i]
out_path = osp.join(output_dir, name + '.png')
fig = plt.figure(figsize=(16, 8))
plt.suptitle(entry['sentence'])
for j in range(len(imgs)):
plt.subplot(4, 3, j + 1)
plt.imshow(imgs[j, :, :, ::-1].astype(np.uint8))
plt.axis('off')
target = cv2.imread(entry['color_path'], cv2.IMREAD_COLOR)
plt.subplot(4, 3, 12)
plt.imshow(target[:, :, ::-1])
plt.axis('off')
fig.savefig(out_path, bbox_inches='tight')
plt.close(fig)
break
def test_simulator(config):
plt.switch_backend('agg')
output_dir = osp.join(config.model_dir, 'simulator')
maybe_create(output_dir)
transformer = image_normalize('background')
db = abstract_scene(config, 'val', transform=transformer)
loader = DataLoader(db,
batch_size=config.batch_size,
shuffle=False,
num_workers=config.num_workers)
env = simulator(db, config.batch_size)
env.reset()
for cnt, batched in enumerate(loader):
out_inds = batched['out_inds'].long().numpy()
gt_paths = batched['color_path']
img_inds = batched['image_idx']
sents = batched['sentence']
sequences = []
masks = []
for i in range(out_inds.shape[1]):
frames = env.batch_render_to_pytorch(out_inds[:, i])
frames = tensor_to_img(frames)
msks = env.batch_location_maps(out_inds[:, i, 3])
for j in range(len(frames)):
frame = frames[j]
msk = cv2.resize(msks[j], (frame.shape[0], frame.shape[1]),
cv2.INTER_NEAREST)
frames[j] = frame * (1.0 - msk[..., None])
sequences.append(frames)
seqs1 = np.stack(sequences, 1)
print('seqs1', seqs1.shape)
seqs2 = env.batch_redraw(return_sequence=True)
seqs = seqs2
for i in range(len(seqs)):
imgs = seqs[i]
image_idx = img_inds[i]
name = '%03d_' % i + str(image_idx).zfill(9)
out_path = osp.join(output_dir, name + '.png')
color = cv2.imread(gt_paths[i], cv2.IMREAD_COLOR)
# color, _, _ = create_squared_image(color)
fig = plt.figure(figsize=(32, 16))
plt.suptitle(sents[i])
for j in range(len(imgs)):
plt.subplot(3, 5, j + 1)
plt.imshow(imgs[j, :, :, ::-1])
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)
break
def test_evaluator(config):
transformer = image_normalize('background')
db = abstract_scene(config, 'train', transform=transformer)
output_dir = osp.join(db.cfg.model_dir, 'test_evaluator')
maybe_create(output_dir)
ev = evaluator(db)
for i in range(0, len(db), 2):
# print('--------------------------------------')
entry_1 = db[i]
entry_2 = db[i + 1]
scene_1 = db.scenedb[i]
scene_2 = db.scenedb[i + 1]
name_1 = osp.splitext(osp.basename(entry_1['color_path']))[0]
name_2 = osp.splitext(osp.basename(entry_2['color_path']))[0]
graph_1 = scene_graph(db, scene_1, entry_1['out_inds'], True)
graph_2 = scene_graph(db, scene_2, entry_2['out_inds'], False)
color_1 = cv2.imread(entry_1['color_path'], cv2.IMREAD_COLOR)
color_2 = cv2.imread(entry_2['color_path'], cv2.IMREAD_COLOR)
color_1 = visualize_unigram(config, color_1, graph_1.unigrams,
(225, 0, 0))
color_2 = visualize_unigram(config, color_2, graph_2.unigrams,
(225, 0, 0))
color_1 = visualize_bigram(config, color_1, graph_1.bigrams,
(0, 255, 255))
color_2 = visualize_bigram(config, color_2, graph_2.bigrams,
(0, 255, 255))
scores = ev.evaluate_graph(graph_1, graph_2)
color_1 = visualize_unigram(config, color_1, ev.common_pred_unigrams,
(0, 225, 0))
color_2 = visualize_unigram(config, color_2, ev.common_gt_unigrams,
(0, 225, 0))
color_1 = visualize_bigram(config, color_1, ev.common_pred_bigrams,
(0, 0, 255))
color_2 = visualize_bigram(config, color_2, ev.common_gt_bigrams,
(0, 0, 255))
info = eval_info(config, scores[None, ...])
plt.switch_backend('agg')
fig = plt.figure(figsize=(16, 10))
title = entry_1['sentence'] + '\n' + entry_2['sentence'] + '\n'
title += 'unigram f3: %f, bigram f3: %f, bigram sim: %f\n' % (
info.unigram_F3()[0], info.bigram_F3()[0], info.bigram_coord()[0])
title += 'pose: %f, expr: %f, scale: %f, flip: %f, coord: %f \n' % (
info.pose()[0], info.expr()[0], info.scale()[0], info.flip()[0],
info.unigram_coord()[0])
plt.suptitle(title)
plt.subplot(1, 2, 1)
plt.imshow(color_1[:, :, ::-1])
plt.axis('off')
plt.subplot(1, 2, 2)
plt.imshow(color_2[:, :, ::-1])
plt.axis('off')
out_path = osp.join(output_dir, name_1 + '_' + name_2 + '.png')
fig.savefig(out_path, bbox_inches='tight')
plt.close(fig)
if i > 40:
break
def abstract_demo(config):
transformer = image_normalize('background')
train_db = abstract_scene(config, split='val', transform=transformer)
trainer = SupervisedTrainer(train_db)
input_sentences = json_load('examples/abstract_samples.json')
trainer.sample_demo(input_sentences)
