In [1]:
%matplotlib inline
import importlib
import utils2; importlib.reload(utils2)
from utils2 import *
Using TensorFlow backend. /home/jhoward/anaconda3/lib/python3.5/site-packages/sklearn/cross_validation.py:44: DeprecationWarning: This module was deprecated in version 0.18 in favor of the model_selection module into which all the refactored classes and functions are moved. Also note that the interface of the new CV iterators are different from that of this module. This module will be removed in 0.20. "This module will be removed in 0.20.", DeprecationWarning)
In [2]:
cfg = tf.ConfigProto()
cfg.gpu_options.allow_growth = True
sess = tf.Session(config=cfg)
In [4]:
a = tf.placeholder("float")
b = tf.placeholder("float")
y = tf.multiply(a, b)
print (sess.run(y, feed_dict={a: 3, b: 3}))
9.0
Simple regression¶
In [5]:
num_points = 1000
x_data= np.random.normal(0.0, 0.55, (num_points,))
y_data= x_data * 0.1 + 0.3 + np.random.normal(0.0, 0.03, (num_points,))
In [6]:
plt.scatter(x_data, y_data);
In [7]:
loss = tf.reduce_mean(tf.square(y - y_data))
optimizer = tf.train.GradientDescentOptimizer(0.5)
train = optimizer.minimize(loss)
In [8]:
init = tf.global_variables_initializer()
In [9]:
sess = tf.InteractiveSession(config=cfg)
In [10]:
init.run()
for step in range(8):
train.run()
print (step, W.eval(), b.eval())
0 [ 0.48098135] [ 0.30073947] 1 [ 0.36553362] [ 0.30032972] 2 [ 0.28506982] [ 0.30004594] 3 [ 0.22898886] [ 0.29984814] 4 [ 0.18990201] [ 0.2997103] 5 [ 0.16265959] [ 0.29961422] 6 [ 0.14367238] [ 0.29954726] 7 [ 0.13043883] [ 0.29950058]
In [11]:
plt.scatter(x_data, y_data)
plt.plot(x_data, W.eval() * x_data + b.eval());
Variable scope¶
In [18]:
tf.reset_default_graph()
In [19]:
with tf.variable_scope("foo"):
v = tf.get_variable("v", [2,3])
v.name
Out[19]:
'foo/v:0'
In [20]:
v.get_shape()
Out[20]:
TensorShape([Dimension(2), Dimension(3)])
In [21]:
with tf.variable_scope("foo", reuse=True):
v1 = tf.get_variable("v", [2,3])
In [22]:
v1==v
Out[22]:
True
In [23]:
v2 = tf.get_variable("v2", [1])
In [24]:
v2.name
Out[24]:
'v2:0'
K-means clustering¶
In [25]:
def plot_data(data, centroids):
colour = plt.cm.rainbow(np.linspace(0,1,len(centroids)))
for i, centroid in enumerate(centroids):
samples = data[i*n_samples:(i+1)*n_samples]
plt.scatter(samples[:,0], samples[:,1], c=colour[i])
plt.plot(centroid[0], centroid[1], markersize=15, marker="x", color='k', mew=10)
plt.plot(centroid[0], centroid[1], markersize=10, marker="x", color='m', mew=5)
plt.show()
In [26]:
n_clusters=10
n_samples =250
centroids = np.random.uniform(-35, 35, (n_clusters, 2))
slices = [np.random.multivariate_normal(centroids[i], np.diag([5., 5.]), n_samples)
for i in range(n_clusters)]
data = np.concatenate(slices).astype(np.float32)
In [27]:
plot_data(data, centroids)
In [28]:
# Numpy Version
def find_initial_centroids_numpy(data, k):
r_index = np.random.randint(data.shape[0])
r = data[r_index, :][np.newaxis]
initial_centroids = []
for i in range(k):
diff = data - np.expand_dims(r,1)
dist = np.linalg.norm(diff, axis=2) # 100x2 5x2 --> 100x5x2 --> 100x5
farthest_index = np.argmax(np.min(dist, axis=0))
farthest_point = data[farthest_index]
initial_centroids.append(farthest_point)
r = np.array(initial_centroids)
return r
In [29]:
def find_initial_centroids(data, k):
r_index = tf.random_uniform([1], 0, tf.shape(data)[0], dtype=tf.int32)
r = tf.expand_dims(data[tf.squeeze(r_index)], dim=1)
initial_centroids = []
for i in range(k):
diff = tf.squared_difference(tf.expand_dims(data, 0), tf.expand_dims(r,1))
dist = tf.reduce_sum(diff, axis=2)
farthest_index = tf.argmax(tf.reduce_min(dist, axis=0), 0)
farthest_point = data[tf.to_int32(farthest_index)]
initial_centroids.append(farthest_point)
r = tf.pack(initial_centroids)
return r
In [30]:
samples = tf.placeholder(tf.float32, (None, None))
In [31]:
initial_centroids = find_initial_centroids(samples, n_clusters).eval({samples: data})
--------------------------------------------------------------------------- ValueError Traceback (most recent call last) <ipython-input-31-f15b046ca073> in <module>() ----> 1 initial_centroids = find_initial_centroids(samples, n_clusters).eval({samples: data}) /home/jhoward/anaconda3/lib/python3.5/site-packages/tensorflow/python/framework/ops.py in eval(self, feed_dict, session) 573 574 """ --> 575 return _eval_using_default_session(self, feed_dict, self.graph, session) 576 577 /home/jhoward/anaconda3/lib/python3.5/site-packages/tensorflow/python/framework/ops.py in _eval_using_default_session(tensors, feed_dict, graph, session) 3622 "`eval(session=sess)`") 3623 if session.graph is not graph: -> 3624 raise ValueError("Cannot use the default session to evaluate tensor: " 3625 "the tensor's graph is different from the session's " 3626 "graph. Pass an explicit session to " ValueError: Cannot use the default session to evaluate tensor: the tensor's graph is different from the session's graph. Pass an explicit session to `eval(session=sess)`.
In [32]:
def choose_random_centroids(samples, n_clusters):
n_samples = tf.shape(samples)[0]
random_indices = tf.random_shuffle(tf.range(0, n_samples))
centroid_indices = random_indices[:n_clusters]
return tf.gather(samples, centroid_indices)
In [33]:
initial_centroids = find_initial_centroids_numpy(data, n_clusters)
In [756]:
plot_data(data, initial_centroids)
In [763]:
def assign_to_nearest(samples, centroids):
dim_dists = tf.squared_difference(tf.expand_dims(samples, 0), tf.expand_dims(centroids, 1))
return tf.argmin(tf.reduce_sum(dim_dists , 2), 0)
In [764]:
def update_centroids(samples, nearest_indices, n_clusters):
partitions = tf.dynamic_partition(samples, tf.to_int32(nearest_indices), n_clusters)
return tf.concat(0, [tf.expand_dims(tf.reduce_mean(partition, 0), 0)
for partition in partitions])
In [772]:
initial_centroids = choose_random_centroids(samples, n_clusters).eval({samples: data})
In [773]:
curr_centroids = tf.Variable(initial_centroids)
In [774]:
nearest_indices = assign_to_nearest(samples, curr_centroids)
updated_centroids = update_centroids(samples, nearest_indices, n_clusters)
In [775]:
tf.global_variables_initializer().run()
In [776]:
c = initial_centroids
for i in range(10):
#TODO animate
c2 = curr_centroids.assign(updated_centroids).eval({samples: data})
if np.allclose(c,c2):
break
c=c2
In [777]:
plot_data(data, curr_centroids.eval())
Tf->Keras LR (not working)¶
In [38]:
class LinRegr(Layer):
def __init__(self, **kwargs):
super(LinRegr, self).__init__(**kwargs)
def build(self, input_shape):
input_dims = input_shape[1:]
self.W = tf.Variable(tf.random_uniform([1], -1.0, 1.0))
self.b = tf.Variable(tf.zeros([1]))
self.built = True
def call(self, x, mask=None):
return self.W * x + self.b
def get_output_shape_for(self, input_shape):
return input_shape
In [39]:
inp = Input((1,))
res = LinRegr()(inp)
In [33]:
model = Model(inp, res)
In [34]:
model.compile('adam', 'mse')
In [35]:
model.fit(x_data, y_data, verbose=2)
Epoch 1/10 0s - loss: 0.3151 Epoch 2/10 0s - loss: 0.3151 Epoch 3/10 0s - loss: 0.3151 Epoch 4/10 0s - loss: 0.3151 Epoch 5/10 0s - loss: 0.3151 Epoch 6/10 0s - loss: 0.3151 Epoch 7/10 0s - loss: 0.3151 Epoch 8/10 0s - loss: 0.3151 Epoch 9/10 0s - loss: 0.3151 Epoch 10/10 0s - loss: 0.3151
Out[35]:
<keras.callbacks.History at 0x7f32f8260748>