def test_sqrt():
x = tf.placeholder(tf.float32)
y = tf.sqrt(x)
ax = np.ones((2, 3)) * 2
sess = tf.Session()
ay = sess.run(y, feed_dict={x: ax})
np.testing.assert_almost_equal(ay, np.sqrt(ax))
def test_softmax():
x = tf.placeholder(tf.float32)
y = tf.nn.softmax(x)
ax = np.ones((2, 4))
sess = tf.Session()
ay = sess.run(y, feed_dict={x: ax})
np.testing.assert_almost_equal(ay, ax / np.sum(ax, axis=1, keepdims=True))
def check_ewise_rscalar(ufunc):
x = 10
y = tf.placeholder(tf.float32)
z = ufunc(x, y)
ay = np.ones((2, 3))
sess = tf.Session()
az = sess.run(z, feed_dict={y: ay})
np.testing.assert_almost_equal(az, ufunc(x, ay))
def check_ewise_scalar(ufunc):
x = tf.placeholder(tf.float32)
y = 10
z = ufunc(x, y)
ax = np.ones((2, 3))
sess = tf.Session()
az = sess.run(z, feed_dict={x: ax})
np.testing.assert_almost_equal(az, ufunc(ax, y))
def test_mean_grad():
x = tf.placeholder(tf.float32)
ax = np.ones((2, 3))
z = -tf.reduce_mean(x) * 14
gx = tf.gradients(z, [x])[0]
sess = tf.Session()
agx = sess.run(gx, feed_dict={x: ax})
np.testing.assert_almost_equal(agx, -np.ones((2, 3)) * 14 / 6.0)
def test_argmax():
x = tf.placeholder(tf.float32)
y = tf.argmax(x, 1)
ax = np.random.uniform(size=(700, 10))
sess = tf.Session()
ay = sess.run(y, feed_dict={x:ax})
npy = np.argmax(ax, 1)
assert(np.mean(np.abs(ay - npy))) < 1e-6
def test_mean():
axis = [1, 3]
x = tf.placeholder(tf.float32)
y = tf.reduce_mean(x, reduction_indices=axis)
ax = np.random.uniform(size=(2, 4, 8, 7))
sess = tf.Session()
ay = sess.run(y, feed_dict={x:ax})
npy = ax.mean(axis=tuple(axis))
assert(np.mean(np.abs(ay - npy))) < 1e-6
def test_init():
x1 = tf.Variable(tf.ones(shape=[2, 3]))
x2 = tf.Variable(tf.zeros(shape=[2, 3]))
sess = tf.Session()
sess.run(tf.initialize_all_variables())
ax1 = sess.run(x1)
ax2 = sess.run(x2)
np.testing.assert_almost_equal(ax1, np.ones((2, 3)))
np.testing.assert_almost_equal(ax2, np.zeros((2, 3)))
def check_ewise(ufunc):
x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)
z = ufunc(x, y)
ax = np.ones((2, 3))
ay = np.ones((2, 3)) * 4
sess = tf.Session()
az = sess.run(z, feed_dict={x:ax, y:ay})
np.testing.assert_almost_equal(az, ufunc(ax, ay))
def test_matmul():
x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)
ax = np.ones((2, 3))
ay = np.ones((3, 4)) * 4
z = tf.matmul(x, y) * 4
sess = tf.Session()
az = sess.run(z, feed_dict={x: ax, y: ay})
np.testing.assert_almost_equal(az, np.dot(ax, ay) * 4)
def test_bias_add():
x = tf.placeholder(tf.float32)
b = tf.placeholder(tf.float32)
y = tf.nn.bias_add(x, b)
ax = np.random.uniform(size=(2, 3))
ab = np.random.uniform(size=(3, ))
sess = tf.Session()
ay = sess.run(y, feed_dict={x: ax, b: ab})
np.testing.assert_almost_equal(ay, ax + ab)
def test_add_grad():
x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)
ax = np.ones((2, 3))
ay = np.ones((2, 3)) * 4
z = x + y
gx, gy = tf.gradients(z, [x, y])
sess = tf.Session()
agx = sess.run(gx, feed_dict={x: ax, y: ay})
np.testing.assert_almost_equal(agx, np.ones((2, 3)))
def test_group():
x1 = tf.Variable(tf.zeros(shape=[2, 3]))
x2 = tf.Variable(tf.zeros(shape=[2, 3]))
a1 = tf.assign(x1, tf.zeros(shape=[2, 3]))
a2 = tf.assign(x2, tf.ones(shape=[2, 3]))
sess = tf.Session()
sess.run(tf.group(a1, a2))
ax1 = sess.run(x1)
ax2 = sess.run(x2)
np.testing.assert_almost_equal(ax1, np.zeros((2, 3)))
np.testing.assert_almost_equal(ax2, np.ones((2, 3)))
def test_matmul_grad():
x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)
ax = np.ones((2, 3))
ay = np.ones((3, 4)) * 4
z = tf.matmul(x, y) * 4
gx, gy = tf.gradients(z, [x, y])
sess = tf.Session()
agx = sess.run(gx, feed_dict={x: ax, y: ay})
agy = sess.run(gy, feed_dict={x: ax, y: ay})
np.testing.assert_almost_equal(agx, np.dot(np.ones((2, 4)), ay.T) * 4)
np.testing.assert_almost_equal(agy, np.dot(ax.T, np.ones((2, 4))) * 4)
def test_pad():
out_filter = 10
in_filter = 4
pad_width = (out_filter-in_filter)//2
x = tf.placeholder(tf.float32)
y = tf.pad(x, dim=1, pad=-pad_width)
z = tf.pad(y, dim=1, pad=pad_width)
nx = np.random.randn(100, 4, 28, 28)
npy = np.pad(nx, ((0, 0), (pad_width, pad_width), (0, 0), (0, 0)),
mode='constant', constant_values=0)
sess = tf.Session()
sess.run(tf.initialize_all_variables())
ay = sess.run(z, feed_dict={x : nx})
assert(np.mean(np.abs(ay - npy))) < 1e-6
ksize=[1, 5, 5, 1],
name="conv2",
no_bias=False)
tanh2 = tf.tanh(conv2)
pool2 = tf.nn.max_pool(tanh2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1])
flatten = tf.nn.flatten_layer(pool2)
fc1 = tf.nn.linear(flatten, num_hidden=500, name="fc1")
tanh3 = tf.tanh(fc1)
fc2 = tf.nn.linear(tanh3, num_hidden=10, name="fc2")
# define loss
label = tf.placeholder(tf.float32)
cross_entropy = tf.nn.mean_sparse_softmax_cross_entropy_with_logits(fc2, label)
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
sess = tf.Session(device='gpu')
# Auromatic variable shape inference API, infers the shape and initialize the weights.
known_shape = {x: [100, 1, 28, 28], label: [100]}
stdev = 0.01
init_step = []
for v, name, shape in tf.infer_variable_shapes(cross_entropy,
feed_dict=known_shape):
init_step.append(tf.assign(v, tf.normal(shape, stdev)))
print("shape[%s]=%s" % (name, str(shape)))
sess.run(init_step)
# get the mnist dataset
mnist = get_mnist(flatten=False, onehot=False)
print_period = 1000
def test_assign():
x = tf.Variable(tf.zeros(shape=[2, 3]))
sess = tf.Session()
sess.run(tf.assign(x, tf.zeros(shape=[2, 3])))
ax = sess.run(x)
np.testing.assert_almost_equal(ax, np.zeros((2, 3)))
import tinyflow as tf
from tinyflow.datasets import get_mnist
# Create the model
x = tf.placeholder(tf.float32, [None, 784])
W = tf.Variable(tf.zeros([784, 10]))
y = tf.nn.softmax(tf.matmul(x, W))
# Define loss and optimizer
y_ = tf.placeholder(tf.float32, [None, 10])
cross_entropy = tf.reduce_mean(
-tf.reduce_sum(y_ * tf.log(y), reduction_indices=[1]))
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
sess = tf.Session()
sess.run(tf.initialize_all_variables())
# get the mnist dataset
mnist = get_mnist(flatten=True, onehot=True)
print("minist download is completed!")
for i in range(1000):
batch_xs, batch_ys = mnist.train.next_batch(100)
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(correct_prediction)
print(
tanh1 = tf.tanh(conv1)
pool1 = tf.nn.max_pool(tanh1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1])
conv2 = tf.nn.conv2d(pool1, num_filter=50, ksize=[1, 5, 5, 1], name="conv2", no_bias=False)
tanh2 = tf.tanh(conv2)
pool2 = tf.nn.max_pool(tanh2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1])
flatten = tf.nn.flatten_layer(pool2)
fc1 = tf.nn.linear(flatten, num_hidden=500, name="fc1")
tanh3 = tf.tanh(fc1)
fc2 = tf.nn.linear(tanh3, num_hidden=10, name="fc2")
# define loss
label = tf.placeholder(tf.float32)
cross_entropy = tf.nn.mean_sparse_softmax_cross_entropy_with_logits(fc2, label)
train_step = tf.train.AdamOptimizer(0.005).minimize(cross_entropy)
sess = tf.Session(config='gpu')
# Auromatic variable shape inference API, infers the shape and initialize the weights.
known_shape = {x: [100, 1, 28, 28], label: [100]}
stdev = 0.01
init_step = []
for v, name, shape in tf.infer_variable_shapes(
cross_entropy, feed_dict=known_shape):
init_step.append(tf.assign(v, tf.normal(shape, stdev)))
print("shape[%s]=%s" % (name, str(shape)))
sess.run(init_step)
sess.run(tf.initialize_all_variables())
# get the mnist dataset
mnist = get_mnist(flatten=False, onehot=False)
