def main():
g = ModelGraph()
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
hits = 0
num_imgs = 0
with tf.sg_queue_context(sess):
# loop end-of-queue
while True:
try:
logits, y = sess.run([g.logits, g.y]) # (16, 28)
preds = np.squeeze(np.argmax(logits, -1)) # (16,)
hits += np.equal(preds, y).astype(np.int32).sum()
num_imgs += len(y)
print "%d/%d = %.02f" % (hits, num_imgs, float(hits) / num_imgs)
except:
break
print "\nFinal result is\n%d/%d = %.02f" % (hits, num_imgs, float(hits) / num_imgs)
def run_generator(fig_name, cval):
with tf.Session() as sess:
# init session
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
feed_dic = {}
for t, c in zip(target_cval, cval):
feed_dic[t] = c
# run generator
imgs = sess.run(gen, feed_dic)
# plot result
_, ax = plt.subplots(10, 10, sharex=True, sharey=True)
for i in range(10):
for j in range(10):
ax[i][j].plot(imgs[i * 10 + j])
ax[i][j].set_axis_off()
plt.savefig('asset/train/' + fig_name, dpi=600)
tf.sg_info('Sample image saved to "asset/train/%s"' % fig_name)
plt.close()
def generate(sample_image):
start_time = time.time()
g = ModelGraph()
with tf.Session() as sess:
# We need to initialize variables in this case because the Variable `generator/x` will not restored.
tf.sg_init(sess)
vars = [v for v in tf.global_variables() if "generator" not in v.name]
saver = tf.train.Saver(vars)
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
i = 0
while True:
mse, _ = sess.run([g.mse, g.train_gen],
{g.y: transform_image(sample_image)}) # (16, 28)
if time.time() - start_time > 60: # Save every 60 seconds
gen_image = sess.run(g.x)
gen_image = np.squeeze(gen_image)
misc.imsave('gen_images/%s/gen_%.2f.jpg' % (label, mse),
gen_image)
start_time = time.time()
i += 1
if i == 60: break # Finish after 1 hour
def run_generator(num, x1, x2, fig_name='sample.png'):
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
tf.sg_restore(sess,
tf.train.latest_checkpoint('asset/train/infogan'),
category='generator')
# run generator
imgs = sess.run(gen, {
target_num: num,
target_cval_1: x1,
target_cval_2: x2
})
# plot result
_, ax = plt.subplots(10, 10, sharex=True, sharey=True)
for i in range(10):
for j in range(10):
ax[i][j].imshow(imgs[i * 10 + j], 'gray')
ax[i][j].set_axis_off()
plt.savefig('asset/train/infogan/' + fig_name, dpi=600)
tf.sg_info('Sample image saved to "asset/train/infogan/%s"' % fig_name)
plt.close()
def __init__(self):
# set log level to debug
tf.sg_verbosity(10)
# batch size
self.batch_size = 1
# vocabulary size
self.voca_size = sttwdata.voca_size
# mfcc feature of audio
self.x = tf.placeholder(dtype=tf.sg_floatx,
shape=(self.batch_size, None, 20))
# encode audio feature
self.logit = get_logit(self.x, voca_size=self.voca_size)
# sequence length except zero-padding
self.seq_len = tf.not_equal(self.x.sg_sum(axis=2),
0.).sg_int().sg_sum(axis=1)
# run network
self.session = tf.Session()
tf.sg_init(self.session)
self.saver = tf.train.Saver()
self.saver.restore(self.session,
tf.train.latest_checkpoint('asset/train'))
def train(self):
predict = self.forward(Mnist.train.image)
#######GP
sess = tf.Session()
with tf.sg_queue_context(sess):
tf.sg_init(sess)
trainf = sess.run([Mnist.train.image])[0]
n, w, h, c = trainf.shape
print trainf.shape
np.savetxt('./image.txt', trainf[1, :, :, 0])
#print trainf[1, :, :, 0]
#plt.imshow(trainf[1, :, :, 0])
#plt.axis('off')
#plt.show()
#print type(trainf[1, :, :, 0])
transfer = np.zeros((n, w, h, c))
for i in range(n):
candi = random.randint(0, n - 1)
#print GP(trainf[i, :, :, 0], trainf[candi, :, :, 0])
#transfer[i, :, :, :] = GP(trainf[i, :, :, :], trainf[candi, :, :, :])
#print trainsfer[i, :, :, :]
t = tf.convert_to_tensor(transfer, dtype=tf.float32)
gp_predict = predict.sg_reuse(input=t)
#print trainf.shape
sess.close()
def train(self): # train baseline model
input_ph = tf.placeholder(shape=[batch_size, 28, 28, 1],
dtype=tf.float32)
label_ph = tf.placeholder(shape=[
batch_size,
], dtype=tf.int32)
predict = self.forward(input_ph)
loss_tensor = tf.reduce_mean(predict.sg_ce(target=label_ph))
# use to update network parameters
optim = tf.sg_optim(loss_tensor, optim='Adam', lr=1e-3)
# use saver to save a new model
saver = tf.train.Saver()
sess = tf.Session()
with tf.sg_queue_context(sess):
# inital
tf.sg_init(sess)
# validation
acc = (predict.sg_reuse(
input=Mnist.valid.image).sg_softmax().sg_accuracy(
target=Mnist.valid.label, name='validation'))
tf.sg_train(loss=loss,
eval_metric=[acc],
max_ep=max_ep,
save_dir=save_dir,
ep_size=Mnist.train.num_batch,
log_interval=10)
def run_generator(num, x1, x2, fig_name='sample.png'):
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
# run generator
imgs = sess.run(gen, {target_num: num,
target_cval_1: x1,
target_cval_2: x2})
# plot result
_, ax = plt.subplots(10, 10, sharex=True, sharey=True)
for i in range(10):
for j in range(10):
ax[i][j].plot(imgs[i * 10 + j, :, 0], color='b', linewidth=0.25)
# Turn off tick labels only
# ax[i][j].set_axis_off()
ax[i][j].set_xticks([])
ax[i][j].set_yticks([])
plt.savefig('asset/train/' + fig_name, dpi=600)
tf.sg_info('Sample image saved to "asset/train/%s"' % fig_name)
plt.close()
def eval():
# Load graph
g = Graph(mode="inference")
print("Graph Loaded")
with tf.Session() as sess:
# Initialize variables
tf.sg_init(sess)
# Restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
print("Restored!")
mname = open('asset/train/checkpoint',
'r').read().split('"')[1] # model name
# Load data
X, Sources, Targets = load_test_data(input_reverse=Hp.reverse_inputs)
char2idx, idx2char = load_vocab()
with codecs.open(mname, "w", "utf-8") as fout:
list_of_refs, hypotheses = [], []
for i in range(len(X) // Hp.batch_size):
# Get mini-batches
x = X[i * Hp.batch_size:(i + 1) * Hp.batch_size] # mini-batch
sources = Sources[i * Hp.batch_size:(i + 1) * Hp.batch_size]
targets = Targets[i * Hp.batch_size:(i + 1) * Hp.batch_size]
preds_prev = np.zeros((Hp.batch_size, Hp.maxlen), np.int32)
preds = np.zeros((Hp.batch_size, Hp.maxlen), np.int32)
for j in range(Hp.maxlen):
# predict next character
outs = sess.run(g.preds, {g.x: x, g.y_src: preds_prev})
# update character sequence
if j < Hp.maxlen - 1:
preds_prev[:, j + 1] = outs[:, j]
preds[:, j] = outs[:, j]
# Write to file
for source, target, pred in zip(sources, targets,
preds): # sentence-wise
got = "".join(idx2char[idx] for idx in pred).split(u"␃")[0]
fout.write("- source: " + source + "\n")
fout.write("- expected: " + target + "\n")
fout.write("- got: " + got + "\n\n")
fout.flush()
# For bleu score
ref = target.split()
hypothesis = got.split()
if len(ref) > 2:
list_of_refs.append([ref])
hypotheses.append(hypothesis)
# Get bleu score
score = corpus_bleu(list_of_refs, hypotheses)
fout.write("Bleu Score = " + str(100 * score))
def generate(self, prev_midi):
with tf.Session() as sess:
tf.sg_init(sess)
# saver = tf.train.Saver()
# saver.restore(sess, tf.train.latest_checkpoint('save/train/small'))
# KDK: choose self.next_token or self.preds
# out = sess.run(self.next_token, {self.x: prev_midi})
tf.sg_restore(sess, tf.train.latest_checkpoint('save/train/small'))
out = sess.run(self.next_token, {self.x: prev_midi})
return out
def test(self):
# predict = self.forward(Mnist.test.image)
# acc = (predict.sg_softmax()
# .sg_accuracy(target=Mnist.test.label, name='test'))
sess = tf.Session()
with tf.sg_queue_context(sess):
tf.sg_init(sess)
testf = sess.run([Mnist.test.image])[0]
# print testf.shape
n, w, h, c = testf.shape
tmp0 = np.zeros((n * w, h))
tmp02 = np.zeros((n * w, h))
tmp05 = np.zeros((n * w, h))
tmp08 = np.zeros((n * w, h))
tmp90 = np.zeros((n * w, h))
tmp_90 = np.zeros((n * w, h))
for i in range(n):
tmp0[i * w : (i + 1) * w, 0 : h] = testf[i, :, :, 0]
tmp02[i * w : (i + 1) * w, 0 : h] = addnoisy(testf[i, :, :, 0], 0.2)
tmp05[i * w : (i + 1) * w, 0 : h] = addnoisy(testf[i, :, :, 0], 0.5)
tmp08[i * w : (i + 1) * w, 0 : h] = addnoisy(testf[i, :, :, 0], 0.8)
tmp90[i * w : (i + 1) * w, 0 : h] = rotate90(testf[i, :, :, 0])
tmp_90[i * w : (i + 1) * w, 0 : h] = rotate_90(testf[i, :, :, 0])# addnoisy(testf[i, :, :, 0], 0.8)
#testf[i, :, :, 0] = addnoisy(testf[i, :, :, 0], 0.0)
#testf[i, :, :, 0] = rotate90(testf[i, :, :, 0])
#testf[i, :, :, 0] = rotate_90(testf[i, :, :, 0])
#print testf[i, :, :, 0]
np.savetxt('./image0.txt', tmp0)
np.savetxt('./image02.txt', tmp02)
np.savetxt('./image05.txt', tmp05)
np.savetxt('./image08.txt', tmp08)
np.savetxt('./image90.txt', tmp90)
np.savetxt('./image_90.txt', tmp_90)
testf_tensor = tf.convert_to_tensor(testf, dtype=tf.float32)
predict = self.forward(testf_tensor)
acc = (predict.sg_softmax()
.sg_accuracy(target=Mnist.test.label, name='test'))
saver=tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(save_dir))
total_accuracy = 0
for i in range(Mnist.test.num_batch):
total_accuracy += np.sum(sess.run([acc])[0])
print('Evaluation accuracy: {}'.format(float(total_accuracy)/(Mnist.test.num_batch*batch_size)))
# close session
sess.close()
def main():
g = ModelGraph(mode="test")
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
if Hyperparams.isqwerty:
save_path = "qwerty/asset/train/ckpt"
else:
save_path = "nine/asset/train/ckpt"
saver.restore(sess, tf.train.latest_checkpoint(save_path))
mname = open(save_path + "/checkpoint", 'r').read().split('"')[1]
nums, X, expected_list = load_test_data()
pnyn2idx, idx2pnyn, hanzi2idx, idx2hanzi = load_vocab()
with codecs.open('data/output_{}.txt'.format(mname), 'w',
'utf-8') as fout:
cum_score = 0
full_score = 0
for step in range(len(X) // 64 + 1):
n = nums[step * 64:(step + 1) * 64] #number batch
x = X[step * 64:(step + 1) * 64] # input batch
e = expected_list[step * 64:(step + 1) *
64] # batch of ground truth strings
# predict characters
logits = sess.run(g.logits, {g.x: x})
preds = np.squeeze(np.argmax(logits, -1))
for nn, xx, pp, ee in zip(n, x, preds, e): # sentence-wise
got = ''
for xxx, ppp in zip(xx, pp): # character-wise
if xxx == 0: break
if xxx == 1 or ppp == 1:
got += "*"
else:
got += idx2hanzi.get(ppp, "*")
got = got.replace("_", "") # Remove blanks
error = distance.levenshtein(ee, got)
score = len(ee) - error
cum_score += score
full_score += len(ee)
fout.write(u"{}\t{}\t{}\t{}\n".format(nn, ee, got, score))
fout.write(u"Total acc.: {}/{}={}\n".format(
cum_score, full_score, round(float(cum_score) / full_score,
2)))
def predict_wavenet(mfcc):
with tf.Session().as_default() as sess:
# init variables
tf.sg_init(sess)
# # restore parameters
# saver = tf.train.Saver()
saver.restore(sess, train_model)
# run session
label = sess.run(y, feed_dict={x: mfcc})
# print label
return data.print_index(label)
def load_model():
# load the pre-trained Keras model (here we are using a model
# pre-trained on ImageNet and provided by Keras, but you can
# substitute in your own networks just as easily)
global model, sess
init_model()
# run network
sess = tf.Session()
# init variables
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
model = saver.restore(sess, tf.train.latest_checkpoint('wavenet_train'))
def main():
graph = ModelGraph("test")
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
X, Y = load_data("test")
idx2chr = load_charmaps()[0]
with codecs.open('results.txt', 'w', 'utf-8') as fout:
results = []
for step in range(len(X) // Hyperparams.batch_size - 1):
X_batch = X[step * Hyperparams.batch_size:(step + 1) *
Hyperparams.batch_size]
Y_batch = Y[step * Hyperparams.batch_size:(step + 1) *
Hyperparams.batch_size]
# predict characters
logits = sess.run(graph.logits, {graph.X_batch: X_batch})
preds = np.squeeze(np.argmax(logits, -1))
for x, y, p in zip(X_batch, Y_batch, preds): # sentence-wise
ground_truth = ''
predicted = ''
for xx, yy, pp in zip(x, y, p): # character-wise
if xx == 0: break
else:
predicted += idx2chr.get(xx, "*")
ground_truth += idx2chr.get(xx, "*")
if pp == 1: predicted += " "
if yy == 1: ground_truth += " "
if pp == yy: results.append(1)
else: results.append(0)
fout.write(u"▌Expected: " + ground_truth + "\n")
fout.write(u"▌Got: " + predicted + "\n\n")
fout.write(u"Final Accuracy = %d/%d=%.2f" %
(sum(results), len(results),
float(sum(results)) / len(results)))
def test(self):
print 'Testing model {}: addnoise={}, rotate={}, var={}'.format(
save_dir, addnoise, rotate, var)
input_ph = tf.placeholder(shape=[batch_size, 28, 28, 1],
dtype=tf.float32)
label_ph = tf.placeholder(shape=[
batch_size,
], dtype=tf.int32)
predict = self.forward(input_ph)
acc = (predict.sg_softmax().sg_accuracy(target=label_ph, name='test'))
sess = tf.Session()
with tf.sg_queue_context(sess):
tf.sg_init(sess)
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint(save_dir))
total_accuracy = 0
for i in range(Mnist.test.num_batch):
[image_array,
label_array] = sess.run([Mnist.test.image, Mnist.test.label])
if addnoise:
image_array[0, :, :, 0] = addnoisy(image_array[0, :, :, 0],
var)
if rotate:
image_array[0, :, :, 0] = rotate_90(image_array[0, :, :,
0])
acc_value = sess.run([acc],
feed_dict={
input_ph: image_array,
label_ph: label_array
})[0]
total_accuracy += np.sum(acc_value)
print 'Evaluation accuracy: {}'.format(
float(total_accuracy) / (Mnist.test.num_batch * batch_size))
# close session
sess.close()
def main():
g = ModelGraph(mode="test")
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
print("Restored!")
mname = open('asset/train/checkpoint',
'r').read().split('"')[1] # model name
char2idx, idx2char = load_char_vocab()
word2idx, idx2word = load_word_vocab()
previous = [0] * 50 # a stack for previous words
para = "EEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEE"
ctx = [0] * 50
while True:
key = readchar.readkey().lower()
if key == readchar.key.BACKSPACE:
ctx.insert(0, previous.pop())
ctx.pop()
previous.insert(0, 0)
elif key == readchar.key.ESC:
break
else:
key_idx = char2idx[key]
ctx.append(key_idx)
ctx.pop(0)
logits = sess.run(g.logits, {g.x: np.expand_dims(ctx, 0)})
preds = logits.argsort()[0][-3:]
# pred = np.argmax(logits, -1)[0]
predword1, predword2, predword3 = [
idx2word.get(pred) for pred in preds
]
print(predword1, ' ', predword2, ' ', predword3)
def test1():
'''
Predicts all at once.
'''
X, Y = preprocess()
g = Graph(is_train=False)
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
total_blanks, total_hits = 0, 0
for x_3d, y_2d in zip(X, Y): # problem-wise x: (9, 9, 1), y: (9, 9)
x_2d = np.squeeze(x_3d, -1) #(9, 9)
x_4d = np.expand_dims(x_3d, 0) # (1, 9, 9, 1)
while 1:
logits = sess.run(g.logits,
{g.X: x_4d}) # (1, 9, 9, 10) float32
preds = np.squeeze(np.argmax(logits, axis=-1),
0) # (9, 9) # most probable numbers
expected = y_2d[x_2d == 0]
got = preds[x_2d == 0]
hits = np.equal(expected, got).sum()
result = np.where(x_2d == 0, preds, y_2d).astype(int)
print result
print "Acc.=%d/%d=%.2f" % (hits, len(expected),
float(hits) / len(expected))
total_blanks += len(expected)
total_hits += hits
break
print "Total Accuracy = %d/%d=%.2f" % (
total_hits, total_blanks, float(total_hits) / total_blanks)
def run_generator(num, x1, x2, fig_name='sample.png'):
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
# run generator
imgs = sess.run(gen, {target_num: num,
target_cval_1: x1,
target_cval_2: x2})
# plot result
_, ax = plt.subplots(10, 10, sharex=True, sharey=True)
for i in range(10):
for j in range(10):
ax[i][j].plot(imgs[i * 10 + j, :, 0])
ax[i][j].plot(imgs[i * 10 + j, :, 1])
ax[i][j].set_axis_off()
plt.savefig('asset/train/' + fig_name, dpi=600)
tf.sg_info('Sample image saved to "asset/train/%s"' % fig_name)
plt.close()
def main():
g = ModelGraph(is_train=False)
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
# Or you could use pretrained model which can be downloaded from here
# https://drive.google.com/open?id=0B5M-ed49qMsDQ1dEYXF3cTVNM1E
# saver.restore(sess, 'model-019-1239684')
sents, X = vectorize_input()
idx2hanja = load_charmaps()[-1]
with codecs.open('data/output.txt', 'w', 'utf-8') as fout:
for step in range(len(X) // Hyperparams.batch_size + 1):
inputs = sents[step * Hyperparams.batch_size:(step + 1) *
Hyperparams.batch_size] # batch
x = X[step * Hyperparams.batch_size:(step + 1) *
Hyperparams.batch_size] # batch
# predict characters
logits = sess.run(g.logits, {g.x: x})
preds = np.squeeze(np.argmax(logits, -1))
for ii, xx, pp in zip(inputs, x, preds): # sentence-wise
got = ''
for ii, xxx, ppp in zip(ii, xx, pp): # character-wise
if xxx == 0: break
elif xxx == 1 or ppp == 1:
got += ii
else:
got += idx2hanja.get(ppp, "*")
fout.write(got + "\n")
def train_with_GP(self):
input_ph = tf.placeholder(shape=[batch_size, 28, 28, 1],
dtype=tf.float32)
label_ph = tf.placeholder(shape=[
batch_size,
], dtype=tf.int32)
predict = self.forward(input_ph)
loss_tensor = tf.reduce_mean(predict.sg_ce(target=label_ph))
# use to update network parameters
optim = tf.sg_optim(loss_tensor, optim='Adam', lr=1e-3)
# use saver to save a new model
saver = tf.train.Saver()
sess = tf.Session()
with tf.sg_queue_context(sess):
# inital
tf.sg_init(sess)
# train by GP guilding
for e in range(max_ep):
previous_loss = None
for i in range(Mnist.train.num_batch):
[image_array, label_array
] = sess.run([Mnist.train.image, Mnist.train.label])
if (e == 0 or e == 1
): # first and second epoch train no noisy image
loss = sess.run([loss_tensor, optim],
feed_dict={
input_ph: image_array,
label_ph: label_array
})[0]
print 'Baseline loss = ', loss
elif (
e == 2
): # third epoch train with gp image and original image
gpIn1 = np.squeeze(image_array)
gpIn2 = np.zeros((28, 28))
image_gp = GP(gpIn1, gpIn2, seed=0.8)
image_gp2 = image_gp[np.newaxis, ...]
image_gp2 = image_gp2[..., np.newaxis]
loss = sess.run([loss_tensor, optim],
feed_dict={
input_ph: image_array,
label_ph: label_array
})[0]
print 'GP without nosiy loss = ', loss
loss = sess.run([loss_tensor, optim],
feed_dict={
input_ph: image_gp2,
label_ph: label_array
})[0]
print 'GP loss = ', loss
else: # other epoch train with gp evolution
gpIn1 = np.squeeze(image_array)
gpIn2 = np.zeros((28, 28))
image_gp = GP(gpIn1, gpIn2, seed=random.random())
image_gp2 = image_gp[np.newaxis, ...]
image_gp2 = image_gp2[..., np.newaxis]
loss = sess.run([loss_tensor, optim],
feed_dict={
input_ph: image_array,
label_ph: label_array
})[0]
print 'GP without nosiy loss = ', loss
loss = sess.run([loss_tensor, optim],
feed_dict={
input_ph: image_gp2,
label_ph: label_array
})[0]
print 'GP loss = ', loss
if loss < previous_loss:
for i in range(5):
loss = sess.run([loss_tensor, optim],
feed_dict={
input_ph: image_gp2,
label_ph: label_array
})[0]
gpIn1 = image_gp2
image_gp2[0, :, :,
0] = GP(gpIn1[0, :, :, 0],
gpIn2,
seed=random.random())
print 'GP EV loss = ', loss
previous_loss = loss
saver.save(sess,
os.path.join(save_dir, 'gp_model'),
global_step=e)
# close session
sess.close()
#
# encode audio feature
logit = get_logit(x, voca_size=voca_size)
# CTC loss
loss = logit.sg_ctc(target=y, seq_len=seq_len)
#
# run network
#
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
# init variables
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
# logging
tf.sg_info('Testing started on %s set at global step[%08d].' %
(tf.sg_arg().set.upper(), sess.run(tf.sg_global_step())))
with tf.sg_queue_context():
# create progress bar
iterator = tqdm(range(0, int(data.num_batch * tf.sg_arg().frac)), total=int(data.num_batch * tf.sg_arg().frac),
initial=0, desc='test', ncols=70, unit='b', leave=False)
batch_size = 1 # batch size
voca_size = data.voca_size
x = tf.placeholder(dtype=tf.sg_floatx, shape=(batch_size, None, 20))
# sequence length except zero-padding
seq_len = tf.not_equal(x.sg_sum(axis=2), 0.).sg_int().sg_sum(axis=1)
# encode audio feature
logit = get_logit(x, voca_size)
# ctc decoding
decoded, _ = tf.nn.ctc_beam_search_decoder(logit.sg_transpose(perm=[1, 0, 2]), seq_len, merge_repeated=False)
# to dense tensor
y = tf.add(tf.sparse_to_dense(decoded[0].indices, decoded[0].dense_shape, decoded[0].values), 1, name="output")
with tf.Session() as sess:
tf.sg_init(sess)
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
graph = tf.get_default_graph()
input_graph_def = graph.as_graph_def()
with tf.Session() as sess:
tf.sg_init(sess)
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
# Output model's graph details for reference.
tf.train.write_graph(sess.graph_def, '/root/speech-to-text-wavenet/asset/train', 'graph.txt', as_text=True)
# Freeze the output graph.
output_graph_def = graph_util.convert_variables_to_constants(sess,input_graph_def,"output".split(","))
# Write it into .pb file.
def test2():
'''
Predicts sequentially.
'''
X, Y = preprocess()
g = Graph(is_train=False)
with tf.Session() as sess:
tf.sg_init(sess)
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
total_blanks, total_hits = 0, 0
for x_3d, y_2d in zip(X, Y): # problem-wise x: (9, 9, 1), y: (9, 9)
x_2d = np.squeeze(x_3d, -1) #(9, 9)
x_4d = np.expand_dims(x_3d, 0) # (1, 9, 9, 1)
_x_2d = np.copy(x_2d) # (9, 9)
while 1:
logits = sess.run(g.logits,
{g.X: x_4d}) # (1, 9, 9, 10) float32
def softmax(x):
"""Compute softmax values for each sets of scores in x."""
e_x = np.exp(x - np.max(x, -1, keepdims=True))
return e_x / e_x.sum(axis=-1, keepdims=True)
activated = softmax(logits) # (1, 9, 9, 10) float32
preds = np.squeeze(np.argmax(activated, axis=-1),
0) # (9, 9) # most probable numbers
preds_prob = np.squeeze(
np.max(activated, axis=-1),
0) # (9, 9) # highest probabilities for blanks
preds_prob = np.where(x_2d == 0, preds_prob, 0) # (9, 9)
top1 = np.argmax(
preds_prob
) # the index of the most confident number amongst all predictions
ind = np.unravel_index(top1, (9, 9))
got = preds[ind] # the most confident number
x_2d[ind] = got # result
x_4d = np.expand_dims(np.expand_dims(x_2d, 0), -1)
if len(x_2d[x_2d == 0]) == 0:
expected = y_2d[_x_2d == 0]
got = x_2d[_x_2d == 0]
hits = np.equal(expected, got).sum()
result = np.where(_x_2d == 0, x_2d, y_2d).astype(int)
print result
print "Acc.=%d/%d=%.2f" % (hits, len(expected),
float(hits) / len(expected))
total_blanks += len(expected)
total_hits += hits
break
print "Total Accuracy = %d/%d=%.2f" % (
total_hits, total_blanks, float(total_hits) / total_blanks)
def train(self):
''' Network '''
batch_pred_feats, batch_pred_coords, batch_pred_confs, self.final_state = self.LSTM(
'lstm', self.x)
iou_predict_truth, intersection = self.iou(batch_pred_coords,
self.y[:, 0:4])
should_exist = I = tf.cast(
tf.reduce_sum(self.y[:, 0:4], axis=1) > 0., tf.float32)
no_I = tf.ones_like(I, dtype=tf.float32) - I
object_loss = tf.nn.l2_loss(
I * (batch_pred_confs - iou_predict_truth)) * self.object_scale
noobject_loss = tf.nn.l2_loss(
no_I *
(batch_pred_confs - iou_predict_truth)) * self.noobject_scale
p_sqrt_w = tf.sqrt(
tf.minimum(1.0, tf.maximum(0.0, batch_pred_coords[:, 2])))
p_sqrt_h = tf.sqrt(
tf.minimum(1.0, tf.maximum(0.0, batch_pred_coords[:, 3])))
sqrt_w = tf.sqrt(tf.abs(self.y[:, 2]))
sqrt_h = tf.sqrt(tf.abs(self.y[:, 3]))
loss = (tf.nn.l2_loss(I * (batch_pred_coords[:, 0] - self.y[:, 0])) +
tf.nn.l2_loss(I * (batch_pred_coords[:, 1] - self.y[:, 1])) +
tf.nn.l2_loss(I * (p_sqrt_w - sqrt_w)) +
tf.nn.l2_loss(I * (p_sqrt_h - sqrt_h))) * self.coord_scale
#max_iou = tf.nn.l2_loss(I*(tf.ones_like(iou_predict_truth, dtype=tf.float32) - iou_predict_truth))
total_loss = loss + object_loss + noobject_loss #+ max_iou
''' Optimizer '''
optimizer = tf.train.AdamOptimizer(
learning_rate=self.learning_rate).minimize(
total_loss) # Adam Optimizer
''' Summary for tensorboard analysis '''
dataset_loss = -1
dataset_loss_best = 100
test_writer = tf.summary.FileWriter('summary/test')
tf.summary.scalar('dataset_loss', dataset_loss)
summary_op = tf.summary.merge_all()
''' Initializing the variables '''
self.saver = tf.train.Saver()
batch_states = np.zeros((self.batchsize, 2 * self.len_vec))
# TODO: make this a command line argument, etc.
# training set loader
batch_loader = BatchLoader(
"./DATA/TRAINING/",
seq_len=self.nsteps,
batch_size=self.batchsize,
step_size=1,
folders_to_use=[
"GOPR0005", "GOPR0006", "GOPR0008", "GOPR0008_2", "GOPR0009",
"GOPR0009_2", "GOPR0010", "GOPR0011", "GOPR0012", "GOPR0013",
"GOPR0014", "GOPR0015", "GOPR0016", "MVI_8607", "MVI_8609",
"MVI_8610", "MVI_8612", "MVI_8614", "MVI_8615", "MVI_8616"
])
validation_set_loader = BatchLoader(
"./DATA/VALID/",
seq_len=self.nsteps,
batch_size=self.batchsize,
step_size=1,
folders_to_use=[
"bbd_2017__2017-01-09-21-40-02_cam_flimage_raw",
"bbd_2017__2017-01-09-21-44-31_cam_flimage_raw",
"bbd_2017__2017-01-09-21-48-46_cam_flimage_raw",
"bbd_2017__2017-01-10-16-07-49_cam_flimage_raw",
"bbd_2017__2017-01-10-16-21-01_cam_flimage_raw",
"bbd_2017__2017-01-10-16-31-57_cam_flimage_raw",
"bbd_2017__2017-01-10-21-43-03_cam_flimage_raw",
"bbd_2017__2017-01-11-20-21-32_cam_flimage_raw",
"bbd_2017__2017-01-11-21-02-37_cam_flimage_raw"
])
print("%d available training batches" % len(batch_loader.batches))
print("%d available validation batches" %
len(validation_set_loader.batches))
''' Launch the graph '''
with tf.Session() as sess:
if self.restore_weights == True and os.path.isfile(
self.rolo_current_save + ".index"):
# sess.run(init)
tf.sg_init(sess)
self.saver.restore(sess, self.rolo_current_save)
print("Weight loaded, finetuning")
else:
# sess.run(init)
tf.sg_init(sess)
print("Training from scratch")
epoch_loss = []
for self.iter_id in range(self.n_iters):
''' Load training data & ground truth '''
batch_id = self.iter_id - self.batch_offset
batch_xs, batch_ys, _ = batch_loader.load_batch(batch_id)
''' Update weights by back-propagation '''
sess.run(optimizer,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
if self.iter_id % self.display_step == 0:
''' Calculate batch loss '''
batch_loss = sess.run(total_loss,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
epoch_loss.append(batch_loss)
print("Total Batch loss for iteration %d: %.9f" %
(self.iter_id, batch_loss))
if self.iter_id % self.display_step == 0:
''' Calculate batch loss '''
batch_loss = sess.run(loss,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print(
"Bounding box coord error loss for iteration %d: %.9f"
% (self.iter_id, batch_loss))
if self.display_object_loss and self.iter_id % self.display_step == 0:
''' Calculate batch object loss '''
batch_o_loss = sess.run(object_loss,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print("Object loss for iteration %d: %.9f" %
(self.iter_id, batch_o_loss))
if self.display_object_loss and self.iter_id % self.display_step == 0:
''' Calculate batch object loss '''
batch_noo_loss = sess.run(noobject_loss,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print("No Object loss for iteration %d: %.9f" %
(self.iter_id, batch_noo_loss))
if self.iou_with_ground_truth and self.iter_id % self.display_step == 0:
''' Calculate batch object loss '''
batch_o_loss = sess.run(tf.reduce_mean(iou_predict_truth),
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print("Average IOU with ground for iteration %d: %.9f" %
(self.iter_id, batch_o_loss))
if self.display_coords is True and self.iter_id % self.display_step == 0:
''' Caculate predicted coordinates '''
coords_predict = sess.run(batch_pred_coords,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
print("predicted coords:" + str(coords_predict[0]))
print("ground truth coords:" + str(batch_ys[0]))
''' Save model '''
if self.iter_id % self.save_step == 1:
self.saver.save(sess, self.rolo_current_save)
print("\n Model saved in file: %s" %
self.rolo_current_save)
''' Validation '''
if self.validate == True and self.iter_id % self.validate_step == 0 and self.iter_id > 0:
# Run validation set
dataset_loss = self.test(sess, total_loss,
validation_set_loader,
batch_pred_feats,
batch_pred_coords,
batch_pred_confs,
self.final_state)
''' Early-stop regularization '''
if dataset_loss <= dataset_loss_best:
dataset_loss_best = dataset_loss
self.saver.save(sess, self.rolo_weights_file)
print("\n Better Model saved in file: %s" %
self.rolo_weights_file)
''' Write summary for tensorboard '''
summary = sess.run(summary_op,
feed_dict={
self.x: batch_xs,
self.y: batch_ys
})
test_writer.add_summary(summary, self.iter_id)
print("Average total loss %f" % np.mean(epoch_loss))
return
def main():
g = ModelGraph(mode="test")
with tf.Session() as sess:
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('asset/train'))
print("Restored!")
mname = open('asset/train/checkpoint',
'r').read().split('"')[1] # model name
X, Y = load_test_data()
char2idx, idx2char = load_char_vocab()
word2idx, idx2word = load_word_vocab()
results = []
rk = 0
num_para = 1
num_char = 1
for x, y in zip(X, Y):
stop_counting = False
x = np.concatenate(
(np.zeros((Hyperparams.seqlen - 1, )),
x[-np.count_nonzero(x):])) # lstrip and zero-pad
para = "".join([idx2char[idx] for idx in x])
chars, targets = [], [] # targets: the word that the char composes
for word in "".join(para).split():
chars.append(" ")
targets.append(word)
for char in word:
chars.append(char)
targets.append(word)
prefix = ""
preds = set()
for i, char_target in enumerate(zip(chars, targets)):
char, target = char_target
oov = ""
if target not in word2idx:
oov = "oov"
if i > Hyperparams.seqlen:
ctx = np.array(x[i - Hyperparams.seqlen:i], np.int32) #
if char == " ":
stop_counting = False
preds = set()
if not stop_counting:
logits = sess.run(
g.logits,
{g.x: np.expand_dims(ctx, 0)}) #(1, 20970)
while 1:
pred = np.argmax(logits, -1)[0] # (1,)
if pred in preds:
logits[:, pred] = -100000000
else:
break
rk += 1
predword = idx2word.get(pred)
if predword == target: # S
stop_counting = True
preds.add(pred)
results.append(u"{},{},{},{},{},{},{}".format(
num_char, num_para, char, target, oov, predword, rk))
num_char += 1
num_para += 1
with codecs.open('data/output_{}_rk_{}.csv'.format(mname, rk), 'w',
'utf-8') as fout:
fout.write("\n".join(results))
def main(argv):
# set log level to debug
tf.sg_verbosity(10)
#
# hyper parameters
#
size = 160, 147
batch_size = 1 # batch size
#
# inputs
#
pngName = argv
png = tf.read_file(pngName)
#png.thumbnail(size, Image.ANTIALIAS)
#png = tf.resize(png1, (14,14))
myPNG = tf.image.decode_png(png)
y = convert_image(pngName)
x = tf.reshape(y, [1, 28, 28, 1])
print(x)
# corrupted image
x_small = tf.image.resize_bicubic(x, (14, 14))
x_bicubic = tf.image.resize_bicubic(x_small, (28, 28)).sg_squeeze()
x_nearest = tf.image.resize_images(
x_small, (28, 28),
tf.image.ResizeMethod.NEAREST_NEIGHBOR).sg_squeeze()
#
# create generator
#
# I've used ESPCN scheme
# http://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Shi_Real-Time_Single_Image_CVPR_2016_paper.pdf
#
# generator network
with tf.sg_context(name='generator', act='relu', bn=True):
gen = (x.sg_conv(dim=32).sg_conv().sg_conv(
dim=4, act='sigmoid',
bn=False).sg_periodic_shuffle(factor=2).sg_squeeze())
#
# run generator
#
fileName = "inPython.png"
fig_name = "genImages/" + fileName
#fig_name2 = 'asset/train/sample2.png'
print("start")
with tf.Session() as sess:
with tf.sg_queue_context(sess):
tf.sg_init(sess)
# restore parameters
saver = tf.train.Saver()
#saver.restore(sess, tf.train.latest_checkpoint('asset/train/ckpt'))
saver.restore(
sess, tf.train.latest_checkpoint('python/asset/train/ckpt'))
# run generator
gt, low, bicubic, sr = sess.run(
[x.sg_squeeze(), x_nearest, x_bicubic, gen])
# plot result
#sr[0].thumbnail(size, Image.ANTIALIAS)
plt.figure(figsize=(1, 1))
#plt.set_axis_off()
hr = plt.imshow(sr[0], 'gray')
plt.axis('tight')
plt.axis('off')
#ax.set_axis_off()
#ax.thumbnail(size, Image.ANTIALIAS)
#plt.savefig(fig_name,bbox_inches='tight',pad_inches=0,dpi=600)
plt.savefig(fig_name, dpi=600)
#tf.sg_info('Sample image saved to "%s"' % fig_name)
plt.close()
##print (type (sr[0]))
##sourceImage = Image.fromarray(np.uint8(sr[0])
print("done")
