batch = 128
for count in range(5):
batch_num = int(mnist_data.train.num_examples / batch)
for i in range(batch_num):
feed, target = mnist_data.train.next_batch(batch)
feed = feed.reshape(batch, 28, 28, 1).astype(np.float64)
target = target.astype(np.float64)
pred, cost = net.train([fc2], {
image: feed,
label: target,
is_training: True
})
predict = pred.argmax(1)
truth = target.argmax(1)
accuracy = np.equal(predict, truth).mean()
print('Step {} Loss {} Accuracy {}'.format(i + 1 + count * batch_num,
cost, accuracy))
predict = net.forward([fc2], {
image: mnist_data.test.images.reshape((-1, 28, 28, 1)),
is_training: False
})[0]
true_labels = mnist_data.test.labels.argmax(1)
pred_labels = predict.argmax(1)
accuracy = np.equal(true_labels, pred_labels).mean()
print('Accuracy on test set:', accuracy)
image = net.portal((224, 224, 3))
question = net.portal((40, 300))
real_len = net.portal((1, ))
vgg16_feat = build_vgg16(net, image)
lstm3_feat = build_lstm3(net, question, real_len)
infer_feat = net.concat([lstm3_feat, vgg16_feat])
answer = build_infer(net, infer_feat)
image_feed = read_image('test.jpg')
queries = [u"What is the animal in the picture?"]
import time
query_feed = list()
start_time = time.time()
for query in queries:
query_feed.append(glove_embed(query))
image_feed = [image_feed] * len(queries)
image_feed = np.array(image_feed)
query_feed = np.array(query_feed)
predicts, = net.forward([answer], {
image: image_feed,
question: query_feed,
real_len: [30] * len(queries)
})
for i, predict in enumerate(predicts):
print('Q: {:<40}. A: {}'.format(queries[i], to_word(predict)))
print('Answered in {}s'.format(time.time() - start_time))
vanilla_loss = net.softmax_crossent(predict, y)
regularized_loss = net.weighted_loss((vanilla_loss, 1.0), (regularizer1, .1),
(regularizer2, .1))
net.optimize(regularized_loss, 'rmsprop', 1e-3)
# Helper function
def real_len(x_batch):
return [np.argmin(s + [0]) for s in x_batch]
# Training
batch = int(64)
epoch = int(15)
step = int(0)
for sentences, label in dat.yield_batch(batch, epoch):
pred, loss = net.train([predict], {
x: sentences,
y: label,
keep: .8,
lens: real_len(sentences),
center: 0.
})
acc = accuracy(pred, label)
print('Step {}, Loss {}, Accuracy {}%'.format(step + 1, loss, acc * 100))
step += 1
x_test, y_test = dat.yield_test()
pred = net.forward([predict], {x: x_test, keep: 1., lens: real_len(x_test)})[0]
acc = accuracy(pred, y_test)
print('Accuracy on test set: {}'.format(acc))
fc1 = net.matmul(image, w1)
bias = net.plus_b(fc1, b1)
relu = net.relu(bias)
dropped = net.dropout(relu, keep_prob)
fc2 = net.matmul(dropped, w2)
bias = net.plus_b(fc2, b2)
loss = net.softmax_crossent(bias, target)
net.optimize(loss, 'sgd', lr)
mnist_data = read_mnist()
s = time.time()
for count in range(30):
batch_num = int(mnist_data.train.num_examples / batch)
for i in range(batch_num):
img, lab = mnist_data.train.next_batch(batch)
loss = net.train([], {
image: img,
target: lab,
keep_prob: .5,
})[0]
print('Epoch {} loss {}'.format(count, loss))
print('Total time elapsed: {}'.format(time.time() - s))
bias_out = net.forward([bias], {image: mnist_data.test.images})[0]
true_labels = mnist_data.test.labels.argmax(1)
pred_labels = bias_out.argmax(1)
accuracy = np.equal(true_labels, pred_labels).mean()
print('Accuracy on test set', accuracy)
dropped = net.dropout(relu, keep_prob)
fc2 = net.matmul(dropped, w2)
bias = net.plus_b(fc2, b2)
loss = net.softmax_crossent(bias, target)
net.optimize(loss, 'sgd', lr)
mnist_data = read_mnist()
s = time.time()
for count in range(30):
batch_num = int(mnist_data.train.num_examples / batch)
for i in range(batch_num):
img, lab = mnist_data.train.next_batch(batch)
loss = net.train([], {
image: img,
target: lab,
keep_prob: .5,
})[0]
print('Epoch {} loss {}'.format(count, loss))
print('Total time elapsed: {}'.format(time.time() - s))
bias_out = net.forward([bias], {
image: mnist_data.test.images,
keep_prob: 1.0
})[0]
true_labels = mnist_data.test.labels.argmax(1)
pred_labels = bias_out.argmax(1)
accuracy = np.equal(true_labels, pred_labels).mean()
print('Accuracy on test set', accuracy)