def test_create_training_rnn(self):
tf.reset_default_graph()
with tf.Session():
model = LanguageModel(self.num_layers, self.hidden_size, self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim)
model.create_training_rnn(self.input_keep_prob, self.output_keep_prob, self.grad_clip,
self.learning_rate, self.lr_decay_factor)
def test_build_dataset(self):
tf.reset_default_graph()
with tf.Session() as sess:
model = LanguageModel(self.num_layers, self.hidden_size, self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim)
# Create a Dataset from the train_set and the test_set
dataset = model.build_dataset(["the brown lazy fox", "the red quick fox"], self.batch_size,
self.max_input_seq_length, ENGLISH_CHAR_MAP)
iterator = dataset.make_initializable_iterator()
sess.run(iterator.initializer)
iterator_get_next_op = iterator.get_next()
input_dataset, input_length_dataset, label_dataset = sess.run(iterator_get_next_op)
# Rebuild the expected result for comparison
expected_result = []
one_hot = dataprocessor.DataProcessor.get_str_to_one_hot_encoded(ENGLISH_CHAR_MAP, "the brown lazy fox")
expected_result.append(one_hot)
one_hot = dataprocessor.DataProcessor.get_str_to_one_hot_encoded(ENGLISH_CHAR_MAP, "the red quick fox")
# Append the padding
one_hot.append(np.zeros(len(ENGLISH_CHAR_MAP)))
expected_result.append(one_hot)
# Check values
np.testing.assert_array_equal(input_dataset, expected_result)
np.testing.assert_array_equal(input_length_dataset, [18, 17])
np.testing.assert_array_equal(label_dataset[0],
[[0, 0], [0, 1], [0, 2], [0, 3], [0, 4], [0, 5], [0, 6], [0, 7],
[0, 8], [0, 9], [0, 10], [0, 11], [0, 12], [0, 13], [0, 14], [0, 15],
[1, 0], [1, 1], [1, 2], [1, 3], [1, 4], [1, 5], [1, 6], [1, 7],
[1, 8], [1, 9], [1, 10], [1, 11], [1, 12], [1, 13], [1, 14]
])
np.testing.assert_array_equal(label_dataset[1],
[33, 30, 53, 43, 40, 48, 39, 63, 26, 51, 50, 57, 40, 49, 79, 79,
33, 30, 69, 30, 29, 68, 46, 34, 28, 36, 57, 40, 49, 79, 79])
np.testing.assert_array_equal(label_dataset[2], [2, 1800])
def test_create_forward_rnn(self):
tf.reset_default_graph()
with tf.Session():
model = LanguageModel(self.num_layers, self.hidden_size,
self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim)
model.create_forward_rnn()
def test_create_training_rnn(self):
tf.reset_default_graph()
with tf.Session():
model = LanguageModel(self.num_layers, self.hidden_size,
self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim)
model.create_training_rnn(self.input_keep_prob,
self.output_keep_prob, self.grad_clip,
self.learning_rate, self.lr_decay_factor)
def generate_text(hyper_params):
with tf.Session() as sess:
# Create model
model = LanguageModel(hyper_params["num_layers"],
hyper_params["hidden_size"], 1, 1,
hyper_params["max_target_seq_length"],
hyper_params["char_map_length"])
model.create_forward_rnn()
model.initialize(sess)
model.restore(sess, hyper_params["checkpoint_dir"] + "/language/")
# Start with a letter
text = "O"
for _ in range(10):
print(text, end="")
# Convert to an one-hot encoded vector
input_vec = dataprocessor.DataProcessor.get_str_to_one_hot_encoded(
hyper_params["char_map"], text, add_eos=False)
feat_vec = np.array(input_vec)
(a, b) = feat_vec.shape
feat_vec = feat_vec.reshape((a, 1, b))
prediction = model.process_input(sess, feat_vec, [1])
text = dataprocessor.DataProcessor.get_labels_str(
hyper_params["char_map"], prediction[0])
print(text)
return
def test_build_dataset(self):
tf.reset_default_graph()
with tf.Session() as sess:
model = LanguageModel(self.num_layers, self.hidden_size,
self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim)
# Create a Dataset from the train_set and the test_set
dataset = model.build_dataset(
["the brown lazy fox", "the red quick fox"], self.batch_size,
self.max_input_seq_length, ENGLISH_CHAR_MAP)
iterator = dataset.make_initializable_iterator()
sess.run(iterator.initializer)
iterator_get_next_op = iterator.get_next()
input_dataset, input_length_dataset, label_dataset = sess.run(
iterator_get_next_op)
# Rebuild the expected result for comparison
expected_result = []
one_hot = dataprocessor.DataProcessor.get_str_to_one_hot_encoded(
ENGLISH_CHAR_MAP, "the brown lazy fox")
expected_result.append(one_hot)
one_hot = dataprocessor.DataProcessor.get_str_to_one_hot_encoded(
ENGLISH_CHAR_MAP, "the red quick fox")
# Append the padding
one_hot.append(np.zeros(len(ENGLISH_CHAR_MAP)))
expected_result.append(one_hot)
# Check values
np.testing.assert_array_equal(input_dataset, expected_result)
np.testing.assert_array_equal(input_length_dataset, [18, 17])
np.testing.assert_array_equal(
label_dataset[0],
[[0, 0], [0, 1], [0, 2], [0, 3], [0, 4], [0, 5], [0, 6],
[0, 7], [0, 8], [0, 9], [0, 10], [0, 11], [0, 12], [0, 13],
[0, 14], [0, 15], [1, 0], [1, 1], [1, 2], [1, 3], [1, 4],
[1, 5], [1, 6], [1, 7], [1, 8], [1, 9], [1, 10], [1, 11],
[1, 12], [1, 13], [1, 14]])
np.testing.assert_array_equal(label_dataset[1], [
33, 30, 53, 43, 40, 48, 39, 63, 26, 51, 50, 57, 40, 49, 79, 79,
33, 30, 69, 30, 29, 68, 46, 34, 28, 36, 57, 40, 49, 79, 79
])
np.testing.assert_array_equal(label_dataset[2], [2, 1800])
def test_create_training_rnn_with_iterators(self):
tf.reset_default_graph()
with tf.Session():
model = LanguageModel(self.num_layers, self.hidden_size, self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim)
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset(["the brown lazy fox", "the red quick fox"], self.batch_size,
self.max_input_seq_length, ENGLISH_CHAR_MAP)
model.add_dataset_input(train_dataset)
model.create_training_rnn(self.input_keep_prob, self.output_keep_prob, self.grad_clip,
self.learning_rate, self.lr_decay_factor, use_iterator=True)
def test_create_training_rnn_with_iterators(self):
tf.reset_default_graph()
with tf.Session():
model = LanguageModel(self.num_layers, self.hidden_size,
self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim)
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset(
["the brown lazy fox", "the red quick fox"], self.batch_size,
self.max_input_seq_length, ENGLISH_CHAR_MAP)
model.add_dataset_input(train_dataset)
model.create_training_rnn(self.input_keep_prob,
self.output_keep_prob,
self.grad_clip,
self.learning_rate,
self.lr_decay_factor,
use_iterator=True)
def generate_text(hyper_params):
with tf.Session() as sess:
# Create model
model = LanguageModel(hyper_params["num_layers"], hyper_params["hidden_size"], 1, 1,
hyper_params["max_target_seq_length"], hyper_params["char_map_length"])
model.create_forward_rnn()
model.initialize(sess)
model.restore(sess, hyper_params["checkpoint_dir"] + "/language/")
# Start with a letter
text = "O"
for _ in range(10):
print(text, end="")
# Convert to an one-hot encoded vector
input_vec = dataprocessor.DataProcessor.get_str_to_one_hot_encoded(hyper_params["char_map"], text,
add_eos=False)
feat_vec = np.array(input_vec)
(a, b) = feat_vec.shape
feat_vec = feat_vec.reshape((a, 1, b))
prediction = model.process_input(sess, feat_vec, [1])
text = dataprocessor.DataProcessor.get_labels_str(hyper_params["char_map"], prediction[0])
print(text)
return
def eval_all(params):
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
model_dir = osp.join(params['save_dir'], 'model',
params['dataset'] + '_' + params['splitBy'])
result_dir = osp.join('result',
params['dataset'] + '_' + params['splitBy'])
if not osp.isdir(result_dir):
os.makedirs(result_dir)
if params['old']:
params['data_json'] = 'old' + params['data_json']
params['data_h5'] = 'old' + params['data_h5']
params['image_feats_h5'] = 'old' + params['image_feats']
params['ann_feats_h5'] = 'old' + params['ann_feats']
params['ann_feats_input'] = 'old' + params['ann_feats_input']
params['shapes'] = 'old' + params['shapes']
params['id'] = 'old' + params['id']
if params['dataset'] in ['refcoco', 'refcoco+', 'refcocog']:
global_shapes = (224, 224)
elif params['dataset'] == 'refgta':
global_shapes = (480, 288)
loader = DataLoader(params)
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)
chainer.config.train = False
chainer.config.enable_backprop = False
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
ve = VisualEncoder(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).to_gpu(gpu_id)
serializers.load_hdf5(osp.join(model_dir, params['id'] + ".h5"), rl_crit)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "ve.h5"), ve)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "lm.h5"), lm)
accuracy = 0
loss_evals = 0
while True:
data = loader.getImageBatch(params['split'], 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)
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
lm_scores = -cuda.to_cpu(computeLosses(logprobs, one_last_ind))
score = cuda.to_cpu(
F.sigmoid(
rl_crit.calc_score(feats, sp_cxt_feats, coord, sent_seqz,
one_last_ind)).data)[:, 0]
if params['mode'] == 0:
_, pos_sc, max_neg_sc = compute_margin_loss(
lm_scores, gd_ix, 0)
elif params['mode'] == 1:
_, pos_sc, max_neg_sc = compute_margin_loss(score, gd_ix, 0)
elif params['mode'] == 2:
scores = score + params['lamda'] * lm_scores
_, pos_sc, max_neg_sc = compute_margin_loss(scores, gd_ix, 0)
if pos_sc > max_neg_sc:
accuracy += 1
loss_evals += 1
print('{}-th: evaluating [{}] ... image[{}/{}] sent[{}], acc={}'.
format(loss_evals, params['split'],
data['bounds']['it_pos_now'],
data['bounds']['it_max'], i,
accuracy * 100.0 / loss_evals))
if data['bounds']['wrapped']:
print('validation finished!')
f = open(
result_dir + '/' + params['id'] + params['id2'] +
str(params['mode']) + str(params['lamda']) + 'comp.txt', 'w')
f.write(str(accuracy * 100.0 / loss_evals))
f.close()
break
def eval_all(params):
target_save_dir = osp.join(params['save_dir'], 'prepro',
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']
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)
chainer.config.train = False
chainer.config.enable_backprop = False
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
if 'attention' in params['id']:
print('attn')
le = LanguageEncoderAttn(len(loader.ix_to_word)).to_gpu(gpu_id)
else:
le = LanguageEncoder(len(loader.ix_to_word)).to_gpu(gpu_id)
ve = VisualEncoder().to_gpu(gpu_id)
cca = CcaEmbedding().to_gpu(gpu_id)
lm = LanguageModel(len(loader.ix_to_word),
loader.seq_length).to_gpu(gpu_id)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "ve.h5"), ve)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "le.h5"), le)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "cca.h5"), cca)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "lm.h5"), lm)
predictions = []
beam_all_results = []
while True:
data = loader.getTestBatch(params['split'], params)
ref_ids = data['ref_ids']
image_id = data['image_id']
lang_last_ind = calc_max_ind(data['seqz'])
feats = Variable(xp.array(data['feats'], dtype=xp.float32))
vis_enc_feats = ve(feats)
lang_enc_feats = vis_enc_feats ##fake
_, vis_emb_feats = cca(vis_enc_feats, lang_enc_feats)
vis_feats = vis_combine(vis_enc_feats, vis_emb_feats)
if params['beam_width'] == 1:
results = lm.max_sample(vis_feats)
else:
beam_results = beam_search(lm, vis_feats, params['beam_width'])
results = [result[0]['sent'] for result in beam_results]
ppls = [result[0]['ppl'] for result in beam_results]
for i, result in enumerate(results):
gen_sentence = ' '.join(
[loader.ix_to_word[str(w)] for w in result])
if params['beam_width'] == 1:
print(gen_sentence)
else:
print(gen_sentence, 'image_id : ', image_id)
entry = {'ref_id': ref_ids[i], 'sent': gen_sentence}
predictions.append(entry)
if params['beam_width'] > 1:
beam_all_results.append({
'ref_id': ref_ids[i],
'beam': beam_results[i]
})
print('evaluating validation performance... {}/{}'.format(
data['bounds']['it_pos_now'], data['bounds']['it_max']))
if data['bounds']['wrapped']:
print('validation finished!')
break
lang_stats = language_eval(predictions, params['split'], params)
print(lang_stats)
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 build_language_training_rnn(sess, hyper_params, prog_params, train_set,
test_set):
model = LanguageModel(hyper_params["num_layers"],
hyper_params["hidden_size"],
hyper_params["batch_size"],
hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"],
hyper_params["char_map_length"])
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset(train_set, hyper_params["batch_size"],
hyper_params["max_input_seq_length"],
hyper_params["char_map"])
v_iterator = None
if test_set is []:
t_iterator = model.add_dataset_input(train_dataset)
sess.run(t_iterator.initializer)
else:
test_dataset = model.build_dataset(
test_set, hyper_params["batch_size"],
hyper_params["max_input_seq_length"], hyper_params["char_map"])
# Build the input stream from the different datasets
t_iterator, v_iterator = model.add_datasets_input(
train_dataset, test_dataset)
sess.run(t_iterator.initializer)
sess.run(v_iterator.initializer)
# Create the model
model.create_training_rnn(hyper_params["dropout_input_keep_prob"],
hyper_params["dropout_output_keep_prob"],
hyper_params["grad_clip"],
hyper_params["learning_rate"],
hyper_params["lr_decay_factor"],
use_iterator=True)
model.add_tensorboard(sess, hyper_params["tensorboard_dir"],
prog_params["tb_name"], prog_params["timeline"])
model.initialize(sess)
model.restore(sess, hyper_params["checkpoint_dir"] + "/language/")
# Override the learning rate if given on the command line
if prog_params["learn_rate"] is not None:
model.set_learning_rate(sess, prog_params["learn_rate"])
return model, t_iterator, v_iterator
def eval_all(params):
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
model_dir = osp.join(params['save_dir'], 'model',
params['dataset'] + '_' + params['splitBy'])
result_dir = osp.join('result',
params['dataset'] + '_' + params['splitBy'])
if not osp.isdir(result_dir):
os.makedirs(result_dir)
if params['old']:
params['data_json'] = 'old' + params['data_json']
params['data_h5'] = 'old' + params['data_h5']
params['image_feats_h5'] = 'old' + params['image_feats_h5']
params['ann_feats_h5'] = 'old' + params['ann_feats_h5']
params['id'] = 'old' + params['id']
if params['dataset'] in ['refcoco', 'refcoco+', 'refcocog']:
global_shapes = (224, 224)
elif params['dataset'] == 'refgta':
global_shapes = (480, 288)
loader = DataLoader(params)
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)
chainer.config.train = False
chainer.config.enable_backprop = False
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
ve = VisualEncoder(global_shapes=global_shapes).to_gpu(gpu_id)
lm = LanguageModel(len(loader.ix_to_word), loader.seq_length,
global_shapes).to_gpu(gpu_id)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "ve.h5"), ve)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "lm.h5"), lm)
predictions = []
beam_all_results = []
while True:
data = loader.getTestBatch(params['split'], params)
ref_ids = data['ref_ids']
lang_last_ind = calc_max_ind(data['seqz'])
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']
coord = data['feats'][:, sum(ve.feat_ind[:1]):sum(ve.feat_ind[:2])]
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)
vis_feats = vis_enc_feats
sp_feats, sp_feats_emb = lm.calc_spatial_features(
sp_cxt_feats, sp_ann_feats, local_sp_coord, global_sp_coord)
if params['beam_width'] == 1:
results = lm.max_sample(vis_feats)
else:
beam_results, _ = beam_search(lm, vis_feats, sp_feats,
sp_feats_emb, coord,
params['beam_width'])
results = [result[0]['sent'] for result in beam_results]
ppls = [result[0]['ppl'] for result in beam_results]
for i, result in enumerate(results):
gen_sentence = ' '.join(
[loader.ix_to_word[str(w)] for w in result])
if params['beam_width'] == 1:
print(gen_sentence)
else:
print(gen_sentence, ', ppl : ', ppls[i])
entry = {'ref_id': ref_ids[i], 'sent': gen_sentence}
predictions.append(entry)
if params['beam_width'] > 1:
beam_all_results.append({
'ref_id': ref_ids[i],
'beam': beam_results[i]
})
print('evaluating validation performance... {}/{}'.format(
data['bounds']['it_pos_now'], data['bounds']['it_max']))
if data['bounds']['wrapped']:
print('validation finished!')
break
lang_stats = language_eval(predictions, params['split'], params)
print(lang_stats)
print('sentence mean length: ',
np.mean([len(pred['sent'].split()) for pred in predictions]))
with open(
result_dir + '/' + params['id'] + params['id2'] +
str(params['beam_width']) + params['split'] + 'raw.json',
'w') as f:
json.dump(predictions, f)
with open(
result_dir + '/' + params['id'] + params['id2'] +
str(params['beam_width']) + params['split'] + '.json', 'w') as f:
json.dump(lang_stats, f)
with open(
result_dir + '/' + params['id'] + params['id2'] +
str(params['beam_width']) + params['split'] + 'all_beam.json',
'w') as f:
json.dump(beam_all_results, f)
def eval_all(params):
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
model_dir = osp.join(params['save_dir'], 'model',
params['dataset'] + '_' + params['splitBy'])
batch_size = params['batch_size']
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']
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)
chainer.config.train = False
chainer.config.enable_backprop = False
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
if 'attention' in params['id']:
print('attn')
le = LanguageEncoderAttn(len(loader.ix_to_word)).to_gpu(gpu_id)
else:
le = LanguageEncoder(len(loader.ix_to_word)).to_gpu(gpu_id)
ve = VisualEncoder().to_gpu(gpu_id)
cca = CcaEmbedding().to_gpu(gpu_id)
lm = LanguageModel(len(loader.ix_to_word),
loader.seq_length).to_gpu(gpu_id)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "ve.h5"), ve)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "le.h5"), le)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "cca.h5"), cca)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "lm.h5"), lm)
for num, entry in enumerate(train_entries):
print("{}/{}".format(num, len(train_entries)))
image_id = entry['image_id']
idx = train_iid2id[image_id]
features = all_image_feats[idx]
boxes = all_boxes[idx]
referring_expression = []
for m in range(36):
bbox = boxes[m]
bbox[2] -= bbox[0]
bbox[3] -= bbox[1]
boxes[m] = bbox
for k in range(36):
feats = fetch_feats(entry, features, boxes, loader, k, params)
feats = Variable(xp.array(feats, dtype=xp.float32))
vis_enc_feats = ve(feats)
lang_enc_feats = vis_enc_feats
_, vis_emb_feats = cca(vis_enc_feats, lang_enc_feats)
vis_feats = vis_combine(vis_enc_feats, vis_emb_feats)
beam_results = beam_search(lm, vis_feats, params['beam_width'])
results = [result['sent'] for result in beam_results[0]]
results = results[:3]
gen_sentence = []
for i, result in enumerate(results):
gen_sentence.append(' '.join(
[loader.ix_to_word[str(w)] for w in result]))
referring_expression.append(gen_sentence)
entry['object_captions'] = referring_expression
pickle.dump(train_entries, open('VQA_ref_testdataset_v3.pkl', 'wb'))
def build_language_training_rnn(sess, hyper_params, prog_params, train_set, test_set):
model = LanguageModel(hyper_params["num_layers"], hyper_params["hidden_size"], hyper_params["batch_size"],
hyper_params["max_input_seq_length"], hyper_params["max_target_seq_length"],
hyper_params["char_map_length"])
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset(train_set, hyper_params["batch_size"], hyper_params["max_input_seq_length"],
hyper_params["char_map"])
v_iterator = None
if test_set is []:
t_iterator = model.add_dataset_input(train_dataset)
sess.run(t_iterator.initializer)
else:
test_dataset = model.build_dataset(test_set, hyper_params["batch_size"], hyper_params["max_input_seq_length"],
hyper_params["char_map"])
# Build the input stream from the different datasets
t_iterator, v_iterator = model.add_datasets_input(train_dataset, test_dataset)
sess.run(t_iterator.initializer)
sess.run(v_iterator.initializer)
# Create the model
model.create_training_rnn(hyper_params["dropout_input_keep_prob"], hyper_params["dropout_output_keep_prob"],
hyper_params["grad_clip"], hyper_params["learning_rate"],
hyper_params["lr_decay_factor"], use_iterator=True)
model.add_tensorboard(sess, hyper_params["tensorboard_dir"], prog_params["tb_name"], prog_params["timeline"])
model.initialize(sess)
model.restore(sess, hyper_params["checkpoint_dir"] + "/language/")
# Override the learning rate if given on the command line
if prog_params["learn_rate"] is not None:
model.set_learning_rate(sess, prog_params["learn_rate"])
return model, t_iterator, v_iterator
def eval_all(params):
target_save_dir = osp.join(params['save_dir'], 'prepro',
params['dataset'] + '_' + params['splitBy'])
model_dir = osp.join(params['save_dir'], 'model',
params['dataset'] + '_' + params['splitBy'])
batch_size = params['batch_size']
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']
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)
chainer.config.train = False
chainer.config.enable_backprop = False
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
if 'attention' in params['id']:
print('attn')
le = LanguageEncoderAttn(len(loader.ix_to_word)).to_gpu(gpu_id)
else:
le = LanguageEncoder(len(loader.ix_to_word)).to_gpu(gpu_id)
ve = VisualEncoder().to_gpu(gpu_id)
cca = CcaEmbedding().to_gpu(gpu_id)
lm = LanguageModel(len(loader.ix_to_word),
loader.seq_length).to_gpu(gpu_id)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "ve.h5"), ve)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "le.h5"), le)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "cca.h5"), cca)
serializers.load_hdf5(
osp.join(model_dir, params['id'] + params['id2'] + "lm.h5"), lm)
# train: 82783 , val: 40504, test: 81434
add_feats = np.zeros([81434, 36, 1024])
for num, entry in enumerate(train_entries):
print("{}/{}".format(num, len(train_entries)))
image_id = entry['image_id']
idx = train_iid2id[image_id]
features = all_image_feats[idx]
boxes = all_boxes[idx]
for m in range(36):
bbox = boxes[m]
bbox[2] -= bbox[0]
bbox[3] -= bbox[1]
boxes[m] = bbox
for k in range(36):
# feat 's shape: [1, 6249]
feats = fetch_feats(entry, features, boxes, loader, k, params)
feats = Variable(xp.array(feats, dtype=xp.float32))
# vis_enc_featus 's shape: [1, 512]
vis_enc_feats = ve(feats)
# lang_enc_featus 's shape: [1, 512]
lang_enc_feats = vis_enc_feats
# vis_emb_feats 's shape: [1, 512]
_, vis_emb_feats = cca(vis_enc_feats, lang_enc_feats)
# vis_feats 's shape: [1, 1024]
vis_feats = vis_combine(vis_enc_feats, vis_emb_feats)
# add_feats = np.zeros([82783, 36, 1024])
vis_feats = chainer.cuda.to_cpu(vis_feats.data)
add_feats[idx][k] = np.array(vis_feats)
# all_image_feats = cxt_features.get('image_features')
# all_sp_feats = cxt_features.get('spatial_features')
# all_boxes = cxt_features.get('image_bb')
with h5py.File('new_hdf5/add_test36.hdf5', 'w') as hf:
hf.create_dataset('image_features',
data=all_image_feats,
maxshape=(82783, 36, 2048))
hf.create_dataset('spatial_features',
data=all_sp_feats,
maxshape=(82783, 36, 6))
hf.create_dataset('image_bb', data=all_boxes, maxshape=(82783, 36, 4))
hf.create_dataset('additional_feats',
data=add_feats,
maxshape=(82783, 36, 1024))
def eval_all(params):
target_save_dir = osp.join(params['save_dir'],'prepro', params['dataset']+'_'+params['splitBy'])
model_dir = osp.join(params['save_dir'],'model', params['dataset']+'_'+params['splitBy'])
if params['old'] and params['dataset'] in ['refcoco','refcoco+','refcocog']:
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']
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)
chainer.config.train = False
chainer.config.enable_backprop = False
gpu_id = params['gpu_id']
cuda.get_device(gpu_id).use()
xp = cuda.cupy
ve = VisualEncoder().to_gpu(gpu_id)
if 'attention' in params['id']:
print('attn')
le = LanguageEncoderAttn(len(loader.ix_to_word)).to_gpu(gpu_id)
else:
le = LanguageEncoder(len(loader.ix_to_word)).to_gpu(gpu_id)
cca = CcaEmbedding().to_gpu(gpu_id)
lm = LanguageModel(len(loader.ix_to_word), loader.seq_length).to_gpu(gpu_id)
serializers.load_hdf5(osp.join(model_dir, params['id']+params['id2']+"ve.h5"), ve)
serializers.load_hdf5(osp.join(model_dir, params['id']+params['id2']+"le.h5"), le)
serializers.load_hdf5(osp.join(model_dir, params['id']+params['id2']+"cca.h5"), cca)
serializers.load_hdf5(osp.join(model_dir, params['id']+params['id2']+"lm.h5"), lm)
accuracy = 0
loss_evals = 0
while True:
data = loader.getImageBatch(params['split'], 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)
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
lm_scores = -computeLosses(logprobs, one_last_ind)
if params['mode']==0:
_, pos_sc, max_neg_sc = compute_margin_loss(lm_scores, gd_ix, 0)
elif params['mode']==1:
_, pos_sc, max_neg_sc = compute_margin_loss(cossim.data, gd_ix, 0)
elif params['mode']==2:
scores = cossim.data + params['lamda'] * lm_scores
_, pos_sc, max_neg_sc = compute_margin_loss(scores, gd_ix, 0)
if pos_sc > max_neg_sc:
accuracy += 1
loss_evals += 1
print('{}-th: evaluating [{}] ... image[{}/{}] sent[{}], acc={}'.format(loss_evals, params['split'], data['bounds']['it_pos_now'], data['bounds']['it_max'], i, accuracy*100.0/loss_evals))
if data['bounds']['wrapped']:
print('validation finished!')
f = open('result/'+params['dataset']+params['split']+params['id']+str(params['mode'])+str(params['lamda'])+'comp.txt', 'w') # 書き込みモードで開く
f.write(str(accuracy*100.0/loss_evals)) # 引数の文字列をファイルに書き込む
f.close()
break
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 test_create_forward_rnn(self):
tf.reset_default_graph()
with tf.Session():
model = LanguageModel(self.num_layers, self.hidden_size, self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim)
model.create_forward_rnn()