def extract_feature(params):
if params['dataset'] in ['refcoco', 'refcoco+', 'refcocog']:
image_root = params['coco_image_root']
elif params['dataset'] == 'refgta':
image_root = params['gta_image_root']
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
if params['old']:
params['data_json'] = 'old' + params['data_json']
params['data_h5'] = 'old' + params['data_h5']
params['ann_feats'] = 'old' + params['ann_feats']
loader = DataLoader(params)
# model setting
batch_size = params['batch_size']
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
res = L.ResNet152Layers()
res.to_gpu(gpu_id)
chainer.config.train = False
chainer.config.enable_backprop = False
anns = loader.anns
images = loader.Images
perm = np.arange(len(anns))
ann_feats = []
for bs in tqdm(range(0, len(anns), batch_size)):
batch = []
for ix in perm[bs:bs + batch_size]:
ann = anns[ix]
h5_id = ann['h5_id']
assert h5_id == ix, 'h5_id not match'
img = images[ann['image_id']]
x1, y1, w, h = ann['box']
image = Image.open(os.path.join(image_root,
img['file_name'])).convert('RGB')
if h <= w:
nh, nw = int(224 / w * h), 224
else:
nh, nw = 224, int(224 / h * w)
image = image.crop((x1, y1, x1 + w, y1 + h)).resize(
(nw, nh), Image.ANTIALIAS)
image = np.array(image).astype(np.float32)[:, :, ::-1]
image -= np.array([103.939, 116.779, 123.68], dtype=np.float32)
image = image.transpose((2, 0, 1))
pad_image = np.zeros((3, 224, 224), dtype=np.float32)
if nh <= nw:
pad_image[:, (224 - nh) // 2:(224 - nh) // 2 + nh, :] = image
else:
pad_image[:, :, (224 - nw) // 2:(224 - nw) // 2 + nw] = image
batch.append(pad_image)
batch = Variable(xp.array(batch, dtype=xp.float32))
feature = res(batch, layers=['pool5'])
feature = cuda.to_cpu(feature['pool5'].data)
ann_feats.extend(feature)
np.save(os.path.join(target_save_dir, params['ann_feats']), ann_feats)
def __init__(self, n_out, init_param=None, pretrained_model='auto'):
super(ResNet152, self).__init__()
self.n_out = n_out
with self.init_scope():
self.base = L.ResNet152Layers(pretrained_model=pretrained_model)
self.fc = L.Linear(None, n_out, initialW=init_param)
def get_resnet152(batchsize):
model = L.ResNet152Layers(pretrained_model=None)
model = Wrapper(model, 'fc6')
x = np.random.uniform(size=(batchsize, 3, 224, 224)).astype('f')
x = chainer.as_variable(x)
t = np.random.randint(size=(batchsize, ), low=0,
high=1000).astype(np.int32)
t = chainer.as_variable(t)
return [x, t], model
def extract_feature(params):
if params['dataset'] in ['refcoco', 'refcoco+', 'refcocog']:
image_root = params['coco_image_root']
elif params['dataset'] == 'refgta':
image_root = params['gta_image_root']
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
if params['old']:
params['data_json'] = 'old' + params['data_json']
params['data_h5'] = 'old' + params['data_h5']
params['ann_feats'] = 'old' + params['ann_feats']
loader = DataLoader(params)
# model setting
batch_size = params['batch_size']
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
res = L.ResNet152Layers()
res.to_gpu(gpu_id)
chainer.config.train = False
chainer.config.enable_backprop = False
anns = loader.anns
images = loader.Images
perm = np.arange(len(anns))
ann_feats = []
shapes = []
for bs in tqdm(range(0, len(anns), batch_size)):
batch = []
for ix in perm[bs:bs + batch_size]:
ann = anns[ix]
h5_id = ann['h5_id']
assert h5_id == ix, 'h5_id not match'
img = images[ann['image_id']]
x1, y1, w, h = ann['box']
image = Image.open(os.path.join(
image_root, img['file_name'])).convert('RGB').crop(
(x1, y1, x1 + w, y1 + h))
image, resize_shape = keep_asR_resize(image)
shapes.append(resize_shape)
image = np.array(image).astype(np.float32)[:, :, ::-1]
image -= np.array([103.939, 116.779, 123.68], dtype=np.float32)
image = image.transpose((2, 0, 1))
batch.append(image)
batch = Variable(xp.array(batch, dtype=xp.float32))
feature = res(batch, layers=['res5'])
feature = cuda.to_cpu(feature['res5'].data)
ann_feats.extend(
np.transpose(feature, (0, 2, 3, 1)).reshape(-1, 36, 2048))
np.save(os.path.join(target_save_dir, params['sp_ann_feats']), ann_feats)
np.save(os.path.join(target_save_dir, params['ann_shapes']), shapes)
def extract_feature(params):
if params['dataset'] in ['refcoco', 'refcoco+', 'refcocog']:
image_root = params['coco_image_root']
elif params['dataset'] == 'refgta':
image_root = params['gta_image_root']
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
if params['old']:
params['data_json'] = 'old' + params['data_json']
params['data_h5'] = 'old' + params['data_h5']
params['image_feats'] = 'old' + params['image_feats']
loader = DataLoader(params)
# model setting
batch_size = params['batch_size']
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
res = L.ResNet152Layers()
res.to_gpu(gpu_id)
chainer.config.train = False
chainer.config.enable_backprop = False
images = loader.images
perm = np.arange(len(images))
image_feats = []
for bs in tqdm(range(0, len(images), batch_size)):
batch = []
for ix in perm[bs:bs + batch_size]:
image = Image.open(
os.path.join(image_root,
images[ix]['file_name'])).convert('RGB')
if params['dataset'] in ['refcoco', 'refcoco+', 'refcocog']:
image = image.resize((224, 224), Image.ANTIALIAS)
else:
image = image.resize((480, 288), Image.ANTIALIAS)
image = np.array(image).astype(np.float32)[:, :, ::-1]
image -= np.array([103.939, 116.779, 123.68], dtype=np.float32)
image = image.transpose((2, 0, 1))
batch.append(image)
batch = Variable(xp.array(batch, dtype=xp.float32))
feature = res(batch, layers=['res5'])
feature = cuda.to_cpu(feature['res5'].data)
if params['dataset'] in ['refcoco', 'refcoco+', 'refcocog']:
image_feats.extend(
np.transpose(feature, (0, 2, 3, 1)).reshape(-1, 49, 2048))
else:
image_feats.extend(
np.transpose(feature, (0, 2, 3, 1)).reshape(-1, 135, 2048))
np.save(os.path.join(target_save_dir, params['image_feats']), image_feats)
def setup_extractor(extractor_name):
if extractor_name == 'resnet50':
extractor = L.ResNet50Layers()
elif extractor_name == 'resnet101':
extractor = L.ResNet101Layers()
elif extractor_name == 'resnet152':
extractor = L.ResNet152Layers()
else:
raise ValueError('Unknown extractor name: {}'.format(extractor_name))
return extractor
def __init__(self, args):
self.dropout = args.dropout
self.layer = args.layer
self.fch = args.fch
#w = chainer.initializers.HeNormal()
#bias = chainer.initializers.Zero()
super(Resnet, self).__init__(base=L.ResNet152Layers(), )
## add fc layers for finetuning
with self.init_scope():
# pointwise = L.Convolution2D(None, len(args.cols), 1, 1, 0, initialW=w, initial_bias=bias),
for i in range(len(args.fch) - 1):
setattr(self, 'fc' + str(i), L.Linear(None, args.fch[i]))
self.fcl = L.Linear(None, args.chs)
def set_model(cls, model_name, uses_device=0):
"""
Set model and device.
uses_device = -1 : CPU
uses_device >= 0 : GPU (default 0)
"""
# use gpu or cpu
cls.uses_device = uses_device
if uses_device >= 0:
chainer.cuda.get_device_from_id(uses_device).use()
chainer.cuda.check_cuda_available()
import cupy as xp
else:
xp = np
cls.xp = xp
# set model
cls.model_name = model_name
if model_name == "VGG16":
cls.model = L.VGG16Layers()
cls.last_layer = 'fc8'
cls.size = (224, 224)
cls.mean = [103.939, 116.779, 123.68]
elif model_name == "GoogLeNet":
cls.model = L.GoogLeNet()
cls.last_layer = 'loss3_fc'
cls.size = (224, 224)
cls.mean = [104.0, 117.0, 123.0]
elif model_name == "ResNet152":
cls.model = L.ResNet152Layers()
cls.last_layer = 'fc6'
cls.size = (224, 224)
cls.mean = [103.063, 115.903, 123.152]
else:
raise Exception("Invalid model")
if uses_device >= 0:
cls.model.to_gpu()
#for memory saving
for param in cls.model.params():
param._requires_grad = False
def train_all(params):
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
graph_dir = osp.join('log_graph',
params['dataset'] + '_' + params['splitBy'])
model_dir = osp.join(params['save_dir'], 'model',
params['dataset'] + '_' + params['splitBy'])
if params['old']:
params['data_json'] = 'old' + params['data_json']
params['data_h5'] = 'old' + params['data_h5']
params['image_feats'] = 'old' + params['image_feats']
params['sp_ann_feats'] = 'old' + params['sp_ann_feats']
params['ann_feats'] = 'old' + params['ann_feats']
params['ann_shapes'] = 'old' + params['ann_shapes']
params['id'] = 'old' + params['id']
params['word_emb_path'] = 'old' + params['word_emb_path']
if params['dataset'] in ['refcoco', 'refcoco+', 'refcocog']:
global_shapes = (224, 224)
elif params['dataset'] == 'refgta':
global_shapes = (480, 288)
loader = DataLoader(params)
# model setting
batch_size = params['batch_size']
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
featsOpt = {
'sp_ann': osp.join(target_save_dir, params['sp_ann_feats']),
'ann_input': osp.join(target_save_dir, params['ann_feats']),
'img': osp.join(target_save_dir, params['image_feats']),
'shapes': osp.join(target_save_dir, params['ann_shapes'])
}
loader.loadFeats(featsOpt, mmap_mode=False)
loader.shuffle('train')
ve = VisualEncoder(res6=L.ResNet152Layers().fc6,
global_shapes=global_shapes).to_gpu(gpu_id)
rl_crit = ListenerReward(len(loader.ix_to_word),
global_shapes=global_shapes).to_gpu(gpu_id)
lm = LanguageModel(len(loader.ix_to_word),
loader.seq_length,
global_shapes,
res6=L.ResNet152Layers().fc6).to_gpu(gpu_id)
serializers.load_hdf5(osp.join(model_dir, params['id'] + ".h5"), rl_crit)
ve_optim = optimizers.Adam(alpha=4e-5, beta1=0.8)
lm_optim = optimizers.Adam(alpha=4e-4, beta1=0.8)
ve_optim.setup(ve)
lm_optim.setup(lm)
ve_optim.add_hook(chainer.optimizer.GradientClipping(params['grad_clip']))
lm_optim.add_hook(chainer.optimizer.GradientClipping(params['grad_clip']))
## non-finetune layer
ve.joint_enc.W.update_rule.hyperparam.alpha = 4e-4
ve.joint_enc.b.update_rule.hyperparam.alpha = 4e-4
lm.gaussian_p.x_var.update_rule.hyperparam.alpha = 1e-2
lm.gaussian_p.y_var.update_rule.hyperparam.alpha = 1e-2
ve.gaussian_p.x_var.update_rule.hyperparam.alpha = 1e-2
ve.gaussian_p.y_var.update_rule.hyperparam.alpha = 1e-2
iteration = 0
epoch = 0
lam = params['rank_lam']
val_loss_history = []
val_loss_lm_s_history = []
val_loss_lm_l_history = []
val_loss_l_history = []
val_acc_history = []
val_rank_acc_history = []
min_val_loss = 100
while True:
chainer.config.train = True
chainer.config.enable_backprop = True
ve.zerograds()
lm.zerograds()
rl_crit.zerograds()
start = time.time()
data = loader.getBatch('train', params)
ref_ann_ids = data['ref_ann_ids']
pos_feats = Variable(xp.array(data['feats'], dtype=xp.float32))
pos_sp_cxt_feats = Variable(
xp.array(data['sp_cxt_feats'], dtype=xp.float32))
pos_sp_ann_feats = Variable(
xp.array(data['sp_ann_feats'], dtype=xp.float32))
neg_feats = Variable(xp.array(data['neg_feats'], dtype=xp.float32))
neg_pos_sp_cxt_feats = Variable(
xp.array(data['neg_sp_cxt_feats'], dtype=xp.float32))
neg_pos_sp_ann_feats = Variable(
xp.array(data['neg_sp_ann_feats'], dtype=xp.float32))
local_shapes = np.concatenate([
data['local_shapes'], data['neg_local_shapes'],
data['local_shapes']
],
axis=0)
feats = F.concat([pos_feats, neg_feats, pos_feats], axis=0)
sp_cxt_feats = F.concat(
[pos_sp_cxt_feats, neg_pos_sp_cxt_feats, pos_sp_cxt_feats], axis=0)
sp_ann_feats = F.concat(
[pos_sp_ann_feats, neg_pos_sp_ann_feats, pos_sp_ann_feats], axis=0)
seqz = np.concatenate([data['seqz'], data['seqz'], data['neg_seqz']],
axis=0)
lang_last_ind = calc_max_ind(seqz)
seqz = Variable(xp.array(seqz, dtype=xp.int32))
coord = cuda.to_cpu(
feats[:, sum(ve.feat_ind[:1]):sum(ve.feat_ind[:2])].data)
local_sp_coord, global_sp_coord = calc_coordinate_feature(
coord, local_shapes, global_shapes=global_shapes)
local_sp_coord, global_sp_coord = xp.array(local_sp_coord,
dtype=xp.float32), xp.array(
global_sp_coord,
dtype=xp.float32)
# encode vis feature
vis_feats = ve(feats, sp_cxt_feats, coord)
sp_feats, sp_feats_emb = lm.calc_spatial_features(
sp_cxt_feats, sp_ann_feats, local_sp_coord, global_sp_coord)
logprobs = lm(vis_feats, sp_feats, sp_feats_emb, coord, seqz,
lang_last_ind)
# lang loss
pairP, vis_unpairP, lang_unpairP = F.split_axis(logprobs, 3, axis=1)
pair_num, _, lang_unpair_num = np.split(lang_last_ind, 3)
num_labels = {'T': xp.array(pair_num), 'F': xp.array(lang_unpair_num)}
lm_flows = {
'T': pairP,
'visF': [pairP, vis_unpairP],
'langF': [pairP, lang_unpairP]
}
lm_loss = lm_crits(lm_flows,
num_labels,
params['lm_margin'],
vlamda=params['vis_rank_weight'],
llamda=params['lang_rank_weight'])
# RL loss (pos,pos)
rl_vis_feats = F.split_axis(vis_feats, 3, axis=0)[0]
rl_coord = np.split(coord, 3, axis=0)[0]
rl_sp_vis_feats = F.split_axis(sp_feats, 3, axis=0)[0]
rl_sp_vis_emb = F.split_axis(sp_feats_emb, 3, axis=0)[0]
sampled_seq, sample_log_probs = lm.sample(rl_vis_feats,
rl_sp_vis_feats,
rl_sp_vis_emb, rl_coord)
sampled_lang_last_ind = calc_max_ind(sampled_seq)
rl_loss = rl_crit(pos_feats, pos_sp_cxt_feats, rl_coord, sampled_seq,
sample_log_probs, sampled_lang_last_ind)
loss = lm_loss + rl_loss
print(lm_loss, rl_loss)
if params['dataset'] == 'refgta' and params[
'ranking'] and iteration > 8000:
lam += 0.4 / 8000
score = F.sum(pairP, axis=0) / (xp.array(pair_num + 1))
rank_loss = calc_rank_loss(score, data['rank'], margin=0.01) * lam
loss += rank_loss
loss.backward()
ve_optim.update()
lm_optim.update()
if data['bounds']['wrapped']:
print('one epoch finished!')
loader.shuffle('train')
if iteration % params['losses_log_every'] == 0:
acc = xp.where(rl_crit.reward > 0.5, 1, 0).mean()
print('{} iter : train loss {}, acc : {} reward_mean : {}'.format(
iteration, loss.data, acc, rl_crit.reward.mean()))
if (iteration % params['save_checkpoint_every'] == 0
and iteration > 0):
chainer.config.train = False
chainer.config.enable_backprop = False
loader.resetImageIterator('val')
loss_sum = 0
loss_generation = 0
loss_lm_margin = 0
loss_evals = 0
accuracy = 0
rank_acc = 0
rank_num = 0
while True:
data = loader.getImageBatch('val', params)
image_id = data['image_id']
img_ann_ids = data['img_ann_ids']
sent_ids = data['sent_ids']
gd_ixs = data['gd_ixs']
feats = Variable(xp.array(data['feats'], dtype=xp.float32))
sp_cxt_feats = Variable(
xp.array(data['sp_cxt_feats'], dtype=xp.float32))
sp_ann_feats = Variable(
xp.array(data['sp_ann_feats'], dtype=xp.float32))
local_shapes = data['local_shapes']
seqz = data['seqz']
lang_last_ind = calc_max_ind(seqz)
scores = []
for i, sent_id in enumerate(sent_ids):
gd_ix = gd_ixs[i]
labels = xp.zeros(len(img_ann_ids), dtype=xp.int32)
labels[gd_ix] = 1
labels = Variable(labels)
sent_seqz = np.concatenate(
[[seqz[i]] for _ in range(len(img_ann_ids))], axis=0)
one_last_ind = np.array([lang_last_ind[i]] *
len(img_ann_ids))
sent_seqz = Variable(xp.array(sent_seqz, dtype=xp.int32))
coord = cuda.to_cpu(
feats[:,
sum(ve.feat_ind[:1]):sum(ve.feat_ind[:2])].data)
local_sp_coord, global_sp_coord = calc_coordinate_feature(
coord, local_shapes, global_shapes=global_shapes)
local_sp_coord, global_sp_coord = xp.array(
local_sp_coord,
dtype=xp.float32), xp.array(global_sp_coord,
dtype=xp.float32)
vis_enc_feats = ve(feats, sp_cxt_feats, coord)
sp_feats, sp_feats_emb = lm.calc_spatial_features(
sp_cxt_feats, sp_ann_feats, local_sp_coord,
global_sp_coord)
vis_feats = vis_enc_feats
logprobs = lm(vis_feats, sp_feats, sp_feats_emb, coord,
sent_seqz, one_last_ind).data
gd_ix = gd_ixs[i]
lm_generation_loss = lm_crits(
{
'T': logprobs[:, gd_ix, xp.newaxis]
}, {
'T': one_last_ind[gd_ix, np.newaxis]
},
params['lm_margin'],
vlamda=0,
llamda=0).data
lm_scores = -computeLosses(logprobs, one_last_ind)
lm_margin_loss, pos_sc, max_neg_sc = compute_margin_loss(
lm_scores, gd_ix, params['lm_margin'])
scores.append(lm_scores[gd_ix])
loss_generation += lm_generation_loss
loss_lm_margin += lm_margin_loss
loss_sum += lm_generation_loss + lm_margin_loss
loss_evals += 1
if pos_sc > max_neg_sc:
accuracy += 1
if params['dataset'] == 'refgta':
rank_a, rank_n = calc_rank_acc(scores, data['rank'])
rank_acc += rank_a
rank_num += rank_n
print('{} iter | {}/{} validating acc : {}'.format(
iteration, data['bounds']['it_pos_now'],
data['bounds']['it_max'], accuracy / loss_evals))
if data['bounds']['wrapped']:
print('validation finished!')
fin_val_loss = cuda.to_cpu(loss_sum / loss_evals)
loss_generation = cuda.to_cpu(loss_generation / loss_evals)
loss_lm_margin = cuda.to_cpu(loss_lm_margin / loss_evals)
fin_val_acc = accuracy / loss_evals
break
val_loss_history.append(fin_val_loss)
val_loss_lm_s_history.append(loss_generation)
val_loss_lm_l_history.append(loss_lm_margin)
val_acc_history.append(fin_val_acc)
if min_val_loss > fin_val_loss:
print('val loss {} -> {} improved!'.format(
min_val_loss, val_loss_history[-1]))
min_val_loss = fin_val_loss
serializers.save_hdf5(
osp.join(model_dir,
params['id'] + params['id2'] + "ve.h5"), ve)
serializers.save_hdf5(
osp.join(model_dir,
params['id'] + params['id2'] + "lm.h5"), lm)
## graph
plt.title("accuracy")
plt.plot(np.arange(len(val_acc_history)),
val_acc_history,
label="val_accuracy")
plt.legend()
plt.savefig(
os.path.join(graph_dir,
params['id'] + params['id2'] + "_joint_acc.png"))
plt.close()
plt.title("loss")
plt.plot(np.arange(len(val_loss_history)),
val_loss_history,
label="all_loss")
plt.plot(np.arange(len(val_loss_history)),
val_loss_lm_s_history,
label="generation_loss")
plt.legend()
plt.savefig(
os.path.join(graph_dir,
params['id'] + params['id2'] + "_joint_loss.png"))
plt.close()
plt.title("loss")
plt.plot(np.arange(len(val_loss_history)),
val_loss_lm_l_history,
label="lm_comp_loss")
plt.legend()
plt.savefig(
os.path.join(
graph_dir,
params['id'] + params['id2'] + "_joint_comp_loss.png"))
plt.close()
if params['dataset'] == 'refgta':
val_rank_acc_history.append(rank_acc / rank_num)
plt.title("rank loss")
plt.plot(np.arange(len(val_rank_acc_history)),
val_rank_acc_history,
label="rank_acc")
plt.legend()
plt.savefig(
os.path.join(
graph_dir,
params['id'] + params['id2'] + "_joint_rank_acc.png"))
plt.close()
if iteration > params['learning_rate_decay_start'] and params[
'learning_rate_decay_start'] >= 0:
frac = (iteration - params['learning_rate_decay_start']
) / params['learning_rate_decay_every']
decay_factor = math.pow(0.1, frac)
ve_optim.alpha *= decay_factor
lm_optim.alpha *= decay_factor
iteration += 1
def train_all(params):
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
graph_dir = osp.join('log_graph',
params['dataset'] + '_' + params['splitBy'])
model_dir = osp.join(params['save_dir'], 'model',
params['dataset'] + '_' + params['splitBy'])
if params['old']:
params['data_json'] = 'old' + params['data_json']
params['data_h5'] = 'old' + params['data_h5']
params['image_feats'] = 'old' + params['image_feats']
params['ann_feats'] = 'old' + params['ann_feats']
params['id'] = 'old' + params['id']
params['word_emb_path'] = 'old' + params['word_emb_path']
with open('setting.json', 'w') as f:
json.dump(params, f)
if not osp.isdir(graph_dir):
os.mkdir(graph_dir)
loader = DataLoader(params)
# model setting
batch_size = params['batch_size']
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
featsOpt = {
'ann': osp.join(target_save_dir, params['ann_feats']),
'img': osp.join(target_save_dir, params['image_feats'])
}
loader.loadFeats(featsOpt)
loader.shuffle('train')
ve = VisualEncoder(res6=L.ResNet152Layers().fc6).to_gpu(gpu_id)
if 'attention' in params['id']:
print('attention language encoder')
le = LanguageEncoderAttn(len(loader.ix_to_word))
rl_crit = ListenerReward(len(loader.ix_to_word),
attention=True).to_gpu(gpu_id)
else:
le = LanguageEncoder(len(loader.ix_to_word))
rl_crit = ListenerReward(len(loader.ix_to_word),
attention=False).to_gpu(gpu_id)
cca = CcaEmbedding().to_gpu(gpu_id)
lm = LanguageModel(len(loader.ix_to_word), loader.seq_length)
if params['pretrained_w']:
print('pretrained word embedding...')
word_emb = load_vcab_init(
loader.word_to_ix,
osp.join(target_save_dir, params['word_emb_path']))
le.word_emb.W.data = word_emb
lm.word_emb = le.word_emb
le.to_gpu(gpu_id)
lm.to_gpu(gpu_id)
serializers.load_hdf5(osp.join(model_dir, params['id'] + ".h5"), rl_crit)
ve_optim = optimizers.Adam(alpha=4e-5, beta1=0.8)
le_optim = optimizers.Adam(alpha=4e-4, beta1=0.8)
cca_optim = optimizers.Adam(alpha=4e-4, beta1=0.8)
lm_optim = optimizers.Adam(alpha=4e-4, beta1=0.8)
ve_optim.setup(ve)
le_optim.setup(le)
cca_optim.setup(cca)
lm_optim.setup(lm)
ve_optim.add_hook(chainer.optimizer.GradientClipping(0.1))
le_optim.add_hook(chainer.optimizer.GradientClipping(0.1))
cca_optim.add_hook(chainer.optimizer.GradientClipping(0.1))
lm_optim.add_hook(chainer.optimizer.GradientClipping(0.1))
ve.joint_enc.W.update_rule.hyperparam.alpha = 4e-4
ve.joint_enc.b.update_rule.hyperparam.alpha = 4e-4
iteration = 0
epoch = 0
val_loss_history = []
val_loss_lm_s_history = []
val_loss_lm_l_history = []
val_loss_l_history = []
val_acc_history = []
val_rank_acc_history = []
min_val_loss = 100
while True:
chainer.config.train = True
chainer.config.enable_backprop = True
ve.zerograds()
le.zerograds()
cca.zerograds()
lm.zerograds()
rl_crit.zerograds()
data = loader.getBatch('train', params)
ref_ann_ids = data['ref_ann_ids']
pos_feats = Variable(xp.array(data['feats'], dtype=xp.float32))
neg_feats = Variable(xp.array(data['neg_feats'], dtype=xp.float32))
feats = F.concat([pos_feats, neg_feats, pos_feats], axis=0)
seqz = np.concatenate([data['seqz'], data['seqz'], data['neg_seqz']],
axis=0)
lang_last_ind = calc_max_ind(seqz)
seqz = Variable(xp.array(seqz, dtype=xp.int32))
vis_enc_feats = ve(feats)
lang_enc_feats = le(seqz, lang_last_ind)
cossim, vis_emb_feats = cca(vis_enc_feats, lang_enc_feats)
vis_feats = vis_combine(vis_enc_feats, vis_emb_feats)
logprobs = lm(vis_feats, seqz, lang_last_ind)
# emb loss
pairSim, vis_unpairSim, lang_unpairSim = F.split_axis(cossim,
3,
axis=0)
emb_flows = {
'vis': [pairSim, vis_unpairSim],
'lang': [pairSim, lang_unpairSim]
}
emb_loss = emb_crits(emb_flows, params['emb_margin'])
# lang loss
pairP, vis_unpairP, lang_unpairP = F.split_axis(logprobs, 3, axis=1)
pair_num, _, lang_unpair_num = np.split(lang_last_ind, 3)
num_labels = {'T': pair_num, 'F': lang_unpair_num}
lm_flows = {
'T': pairP,
'visF': [pairP, vis_unpairP],
'langF': [pairP, lang_unpairP]
}
lm_loss = lm_crits(lm_flows,
num_labels,
params['lm_margin'],
vlamda=params['vis_rank_weight'],
llamda=params['lang_rank_weight'])
# RL loss (pos,pos)のみ
rl_vis_feats = F.split_axis(vis_feats, 3, axis=0)[0]
sampled_seq, sample_log_probs = lm.sample(rl_vis_feats)
sampled_lang_last_ind = calc_max_ind(sampled_seq)
rl_loss = rl_crit(pos_feats, sampled_seq, sample_log_probs,
sampled_lang_last_ind) #, lm.baseline)
loss = emb_loss + lm_loss + rl_loss
print(emb_loss, lm_loss, rl_loss)
loss.backward()
ve_optim.update()
le_optim.update()
cca_optim.update()
lm_optim.update()
if data['bounds']['wrapped']:
print('one epoch finished!')
loader.shuffle('train')
if params['check_sent']:
sampled_sents = loader.decode_sequence(cuda.to_cpu(sampled_seq),
sampled_lang_last_ind)
for i in range(len(sampled_sents)):
print('sampled sentence : ', ' '.join(sampled_sents[i]))
print('reward : ', rl_crit.reward[i])
if iteration % params['losses_log_every'] == 0:
acc = xp.where(rl_crit.reward > 0.5, 1, 0).mean()
print('{} iter : train loss {}, acc : {}, reward_mean : {}'.format(
iteration, loss.data, acc, rl_crit.reward.mean()))
if iteration % params[
'mine_hard_every'] == 0 and iteration > 0 and params[
'mine_hard']:
make_graph(ve, cca, loader, 'train', params, xp)
if (iteration % params['save_checkpoint_every'] == 0
and iteration > 0):
chainer.config.train = False
chainer.config.enable_backprop = False
loader.resetImageIterator('val')
loss_sum = 0
loss_generation = 0
loss_lm_margin = 0
loss_emb_margin = 0
loss_evals = 0
accuracy = 0
rank_acc = 0
rank_num = 0
while True:
data = loader.getImageBatch('val', params)
image_id = data['image_id']
img_ann_ids = data['img_ann_ids']
sent_ids = data['sent_ids']
gd_ixs = data['gd_ixs']
feats = Variable(xp.array(data['feats'], dtype=xp.float32))
seqz = data['seqz']
lang_last_ind = calc_max_ind(seqz)
scores = []
for i, sent_id in enumerate(sent_ids):
gd_ix = gd_ixs[i]
labels = xp.zeros(len(img_ann_ids), dtype=xp.int32)
labels[gd_ix] = 1
labels = Variable(labels)
sent_seqz = np.concatenate(
[[seqz[i]] for _ in range(len(img_ann_ids))], axis=0)
one_last_ind = np.array([lang_last_ind[i]] *
len(img_ann_ids))
sent_seqz = Variable(xp.array(sent_seqz, dtype=xp.int32))
vis_enc_feats = ve(feats)
lang_enc_feats = le(sent_seqz, one_last_ind)
cossim, vis_emb_feats = cca(vis_enc_feats, lang_enc_feats)
vis_feats = vis_combine(vis_enc_feats, vis_emb_feats)
logprobs = lm(vis_feats, sent_seqz, one_last_ind).data
gd_ix = gd_ixs[i]
lm_generation_loss = lm_crits(
{
'T': logprobs[:, gd_ix, xp.newaxis]
}, {
'T': one_last_ind[gd_ix, np.newaxis]
},
params['lm_margin'],
vlamda=0,
llamda=0).data
lm_scores = -computeLosses(logprobs, one_last_ind)
lm_margin_loss, _, _ = compute_margin_loss(
lm_scores, gd_ix, params['lm_margin'])
scores.append(lm_scores[gd_ix])
emb_margin_loss, pos_sc, max_neg_sc = compute_margin_loss(
cossim.data, gd_ix, params['emb_margin'])
loss_generation += lm_generation_loss
loss_lm_margin += lm_margin_loss
loss_emb_margin += emb_margin_loss
loss_sum += lm_generation_loss + lm_margin_loss + emb_margin_loss
loss_evals += 1
if pos_sc > max_neg_sc:
accuracy += 1
if params['dataset'] == 'refgta':
rank_a, rank_n = calc_rank_acc(scores, data['rank'])
rank_acc += rank_a
rank_num += rank_n
print('{} iter | {}/{} validating acc : {}'.format(
iteration, data['bounds']['it_pos_now'],
data['bounds']['it_max'], accuracy / loss_evals))
if data['bounds']['wrapped']:
print('validation finished!')
fin_val_loss = cuda.to_cpu(loss_sum / loss_evals)
loss_generation = cuda.to_cpu(loss_generation / loss_evals)
loss_lm_margin = cuda.to_cpu(loss_lm_margin / loss_evals)
loss_emb_margin = cuda.to_cpu(loss_emb_margin / loss_evals)
fin_val_acc = accuracy / loss_evals
break
val_loss_history.append(fin_val_loss)
val_loss_lm_s_history.append(loss_generation)
val_loss_lm_l_history.append(loss_lm_margin)
val_loss_l_history.append(loss_emb_margin)
val_acc_history.append(fin_val_acc)
if min_val_loss > fin_val_loss:
print('val loss {} -> {} improved!'.format(
min_val_loss, val_loss_history[-1]))
min_val_loss = fin_val_loss
serializers.save_hdf5(
osp.join(model_dir,
params['id'] + params['id2'] + "ve.h5"), ve)
serializers.save_hdf5(
osp.join(model_dir,
params['id'] + params['id2'] + "le.h5"), le)
serializers.save_hdf5(
osp.join(model_dir,
params['id'] + params['id2'] + "cca.h5"), cca)
serializers.save_hdf5(
osp.join(model_dir,
params['id'] + params['id2'] + "lm.h5"), lm)
## graph
plt.title("accuracy")
plt.plot(np.arange(len(val_acc_history)),
val_acc_history,
label="val_accuracy")
plt.legend()
plt.savefig(
os.path.join(graph_dir,
params['id'] + params['id2'] + "_joint_acc.png"))
plt.close()
plt.title("loss")
plt.plot(np.arange(len(val_loss_history)),
val_loss_history,
label="all_loss")
plt.plot(np.arange(len(val_loss_history)),
val_loss_lm_s_history,
label="generation_loss")
plt.legend()
plt.savefig(
os.path.join(graph_dir,
params['id'] + params['id2'] + "_joint_loss.png"))
plt.close()
plt.title("loss")
plt.plot(np.arange(len(val_loss_history)),
val_loss_lm_l_history,
label="lm_comp_loss")
plt.plot(np.arange(len(val_loss_history)),
val_loss_l_history,
label="comp_loss")
plt.legend()
plt.savefig(
os.path.join(
graph_dir,
params['id'] + params['id2'] + "_joint_comp_loss.png"))
plt.close()
if params['dataset'] == 'refgta':
print(rank_num)
val_rank_acc_history.append(rank_acc / rank_num)
plt.title("rank loss")
plt.plot(np.arange(len(val_rank_acc_history)),
val_rank_acc_history,
label="rank_acc")
plt.legend()
plt.savefig(
os.path.join(
graph_dir,
params['id'] + params['id2'] + "_rank_acc.png"))
plt.close()
if iteration > params['learning_rate_decay_start'] and params[
'learning_rate_decay_start'] >= 0:
frac = (iteration - params['learning_rate_decay_start']
) / params['learning_rate_decay_every']
decay_factor = math.pow(0.1, frac)
ve_optim.alpha *= decay_factor
le_optim.alpha *= decay_factor
cca_optim.alpha *= decay_factor
lm_optim.alpha *= decay_factor
iteration += 1
def train_vl(params):
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
graph_dir = osp.join('log_graph',
params['dataset'] + '_' + params['splitBy'])
model_dir = osp.join(params['save_dir'], 'model',
params['dataset'] + '_' + params['splitBy'])
if not osp.isdir(graph_dir):
os.makedirs(graph_dir)
if not osp.isdir(model_dir):
os.makedirs(model_dir)
if params['old']:
params['data_json'] = 'old' + params['data_json']
params['data_h5'] = 'old' + params['data_h5']
params['image_feats'] = 'old' + params['image_feats']
params['ann_feats'] = 'old' + params['ann_feats']
params['id'] = 'old' + params['id']
params['word_emb_path'] = 'old' + params['word_emb_path']
loader = DataLoader(params)
featsOpt = {
'ann': osp.join(target_save_dir, params['ann_feats']),
'img': osp.join(target_save_dir, params['image_feats'])
}
loader.loadFeats(featsOpt)
loader.shuffle('train')
# model setting
batch_size = params['batch_size']
gpu_id = params['gpu_id']
seq_per_ref = params['seq_per_ref']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
ve = VisualEncoder(res6=L.ResNet152Layers().fc6).to_gpu(gpu_id)
if 'attention' in params['id']:
print('attention language encoder')
le = LanguageEncoderAttn(len(loader.ix_to_word))
save_model = ListenerReward(len(loader.ix_to_word), attention=True)
else:
le = LanguageEncoder(len(loader.ix_to_word))
save_model = ListenerReward(len(loader.ix_to_word), attention=False)
if params['pretrained_w']:
print('pretrained word embedding...')
word_emb = load_vcab_init(
loader.word_to_ix,
osp.join(target_save_dir, params['word_emb_path']))
le.word_emb.W.data = word_emb
le.to_gpu(gpu_id)
me = MetricNet().to_gpu(gpu_id)
ve_optim = optimizers.Adam(alpha=4e-5, beta1=0.8)
le_optim = optimizers.Adam(alpha=4e-4, beta1=0.8)
me_optim = optimizers.Adam(alpha=4e-4, beta1=0.8)
ve_optim.setup(ve)
le_optim.setup(le)
me_optim.setup(me)
ve_optim.add_hook(chainer.optimizer.GradientClipping(0.1))
le_optim.add_hook(chainer.optimizer.GradientClipping(0.1))
me_optim.add_hook(chainer.optimizer.GradientClipping(0.1))
ve.joint_enc.W.update_rule.hyperparam.alpha = 4e-4
ve.joint_enc.b.update_rule.hyperparam.alpha = 4e-4
iteration = 0
epoch = 0
val_loss_history = []
val_acc_history = []
val_rank_acc_history = []
min_val_loss = 100
max_acc = 0
while True:
chainer.config.train = True
chainer.config.enable_backprop = True
ve.zerograds()
le.zerograds()
me.zerograds()
data = loader.getBatch('train', params)
ref_ann_ids = data['ref_ann_ids']
pos_feats = Variable(xp.array(data['feats'], dtype=xp.float32))
neg_feats = Variable(xp.array(data['neg_feats'], dtype=xp.float32))
feats = F.concat([pos_feats, neg_feats, pos_feats], axis=0)
seqz = np.concatenate([data['seqz'], data['seqz'], data['neg_seqz']],
axis=0)
lang_last_ind = calc_max_ind(seqz)
seqz = Variable(xp.array(seqz, dtype=xp.int32))
labels = Variable(
xp.concatenate([
xp.ones((batch_size * seq_per_ref)),
xp.zeros((batch_size * seq_per_ref)),
xp.zeros((batch_size * seq_per_ref))
]).astype(xp.int32))
vis_enc_feats = ve(feats)
lang_enc_feats = le(seqz, lang_last_ind)
score = me(vis_enc_feats, lang_enc_feats).reshape(labels.shape)
loss = F.sigmoid_cross_entropy(score, labels)
loss.backward()
ve_optim.update()
le_optim.update()
me_optim.update()
if data['bounds']['wrapped']:
print('{} epoch finished!'.format(epoch))
loader.shuffle('train')
epoch += 1
if iteration % params['losses_log_every'] == 0:
print('{} iter ({} epoch): train loss {}'.format(
iteration, epoch, loss.data))
## validation
if (iteration % params['save_checkpoint_every'] == 0
and iteration > 0):
chainer.config.train = False
chainer.config.enable_backprop = False
loader.resetImageIterator('val')
loss_sum = 0
loss_evals = 0
accuracy = 0
rank_acc = 0
rank_num = 0
while True:
data = loader.getImageBatch('val', params)
image_id = data['image_id']
img_ann_ids = data['img_ann_ids']
sent_ids = data['sent_ids']
gd_ixs = data['gd_ixs']
feats = Variable(xp.array(data['feats'], dtype=xp.float32))
seqz = data['seqz']
scores = []
for i, sent_id in enumerate(sent_ids):
## image内の全ての候補領域とscoreを算出する
gd_ix = gd_ixs[i]
labels = xp.zeros(len(img_ann_ids), dtype=xp.int32)
labels[gd_ix] = 1
labels = Variable(labels)
sent_seqz = np.concatenate(
[[seqz[i]] for _ in range(len(img_ann_ids))], axis=0)
lang_last_ind = calc_max_ind(sent_seqz)
sent_seqz = Variable(xp.array(sent_seqz, dtype=xp.int32))
vis_enc_feats = ve(feats)
lang_enc_feats = le(sent_seqz, lang_last_ind)
score = me(vis_enc_feats,
lang_enc_feats).reshape(labels.shape)
loss = F.sigmoid_cross_entropy(score, labels)
scores.append(score[gd_ix].data)
loss_sum += loss.data
loss_evals += 1
_, pos_sc, max_neg_sc = compute_margin_loss(
score.data, gd_ix, 0)
if pos_sc > max_neg_sc:
accuracy += 1
if params['dataset'] == 'refgta':
rank_a, rank_n = calc_rank_acc(scores, data['rank'])
rank_acc += rank_a
rank_num += rank_n
print('{} iter | {}/{} validating acc : {}'.format(
iteration, data['bounds']['it_pos_now'],
data['bounds']['it_max'], accuracy / loss_evals))
if data['bounds']['wrapped']:
print('validation finished!')
fin_val_loss = cuda.to_cpu(loss_sum / loss_evals)
fin_val_acc = accuracy / loss_evals
break
val_loss_history.append(fin_val_loss)
val_acc_history.append(fin_val_acc)
if min_val_loss > fin_val_loss:
print('val loss {} -> {} improved!'.format(
min_val_loss, val_loss_history[-1]))
min_val_loss = fin_val_loss
if max_acc < fin_val_acc:
max_acc = fin_val_acc
save_model.ve = ve
save_model.le = le
save_model.me = me
serializers.save_hdf5(
osp.join(model_dir, params['id'] + ".h5"), save_model)
## graph
plt.title("accuracy")
plt.plot(np.arange(len(val_acc_history)),
val_acc_history,
label="val_accuracy")
plt.legend()
plt.savefig(os.path.join(graph_dir, params['id'] + "_acc.png"))
plt.close()
plt.title("loss")
plt.plot(np.arange(len(val_loss_history)),
val_loss_history,
label="val_loss")
plt.legend()
plt.savefig(os.path.join(graph_dir, params['id'] + "_loss.png"))
plt.close()
if params['dataset'] == 'refgta':
print(rank_num)
val_rank_acc_history.append(rank_acc / rank_num)
plt.title("rank loss")
plt.plot(np.arange(len(val_rank_acc_history)),
val_rank_acc_history,
label="rank_acc")
plt.legend()
plt.savefig(
os.path.join(graph_dir, params['id'] + "_rank_acc.png"))
plt.close()
# learning rate decay
if iteration > params['learning_rate_decay_start'] and params[
'learning_rate_decay_start'] >= 0:
frac = (iteration - params['learning_rate_decay_start']
) / params['learning_rate_decay_every']
decay_factor = math.pow(0.1, frac)
ve_optim.alpha *= decay_factor
le_optim.alpha *= decay_factor
me_optim.alpha *= decay_factor
iteration += 1
def __init__(self, class_labels):
super(ResNet152, self).__init__()
with self.init_scope():
self.base = L.ResNet152Layers()
self.fc6 = L.Linear(2048, class_labels)
def __init__(self, out_size):
super(Model_ResNet, self).__init__(base=L.ResNet152Layers(),
fc=L.Linear(None, out_size))