markdown
stringlengths 0
37k
| code
stringlengths 1
33.3k
| path
stringlengths 8
215
| repo_name
stringlengths 6
77
| license
stringclasses 15
values |
|---|---|---|---|---|
In such a case, the condition evaluates to False and the print call included in the indented statement is simply skipped. However, showing (or rather creating) no output at all is not always desirable, and in the majority of use cases, there will definitely be a pool of two or more possibilities that need to be taken into consideration.
In order to overcome such limitations of simple if statements, we could insert an else clause followed by a code body that should be executed once the initial condition evaluates to False ('else code' in the above figure) in order for our code to behave more informative. | if n > 0:
print("Larger than zero.")
else:
print("Smaller than or equal to zero.") | docs/mpg-if_error_continue/examples/e-02-2_conditionals.ipynb | marburg-open-courseware/gmoc | mit |
At this point, be aware that the lines including if and else are not indented, whereas the related code bodies are. Due to the black-and-white nature of such an if-else statement, exactly one out of two possible blocks is executed. Starting from the top,
the if statement is evaluated and returns False (because n is not larger than zero),
hence the indented 'if code' underneath is skipped.
Our code now jumps directly to the next statement that is not indented
and evaluates the else statement included therein.
Since else means: "Perform the following operation if all previous conditional statements (plural!) failed", which evaluates to True in our case,
the subsequent print operation is executed.
Now, what if we wanted to know if a value is larger than, smaller than, or equal to zero, i.e. add another layer of information to our initial condition "Is a value larger than zero or not?". In order to solve this, elif (short for 'else if' in other languages) is the right way to go as it lets you insert an arbitrary number of additional conditions between if and else that go beyond the rather basic capabilities of else. | if n > 0:
print("Larger than zero.")
elif n < 0:
print("Smaller than zero.")
else:
print("Exactly zero.") | docs/mpg-if_error_continue/examples/e-02-2_conditionals.ipynb | marburg-open-courseware/gmoc | mit |
And similarly, | p = 0
if p > 0:
print("Larger than zero.")
elif p < 0:
print("Smaller than zero.")
else:
print("Exactly zero.") | docs/mpg-if_error_continue/examples/e-02-2_conditionals.ipynb | marburg-open-courseware/gmoc | mit |
Of course, indented blocks can have more than one statement, i.e. consist of multiple indented lines of code. In addition, they can embrace, or be embraced by, for or while loops. For example, if we wanted to count all the non-negative entries in a list, the following code snippet would be a proper solution that relies on both of the aforementioned features. | x = [0, 3, -6, -2, 7, 1, -4]
## set a counter
n = 0
for i in range(len(x)):
# if a non-negative integer is found, increment the counter by 1
if x[i] >= 0:
print("The value at position", i, "is larger than or equal to zero.")
n += 1
# else do not increment the counter
else:
print("The value at position", i, "is smaller than zero.")
if i == (len(x)-1):
print("\n")
print(n, "out of", len(x), "elements are larger than or equal to zero.") | docs/mpg-if_error_continue/examples/e-02-2_conditionals.ipynb | marburg-open-courseware/gmoc | mit |
<hr>
Brief digression: continue and break
There are (at least) two key words that allow for an even finer control of what happens inside a for loop, viz.
continue and
break.
As the name implies, continue moves directly on to the next iteration of a loop without executing the remaining code body. | for i in range(5):
if i in [1, 3]:
continue
print(i) | docs/mpg-if_error_continue/examples/e-02-2_conditionals.ipynb | marburg-open-courseware/gmoc | mit |
break, on the other hand, breaks out of the innermost loop. Here, (i) the remaining code body following the break statement in the current iteration, but also (ii) any outstanding iterations are not executed anymore. | for i in range(5):
if i == 2:
break
print(i) | docs/mpg-if_error_continue/examples/e-02-2_conditionals.ipynb | marburg-open-courseware/gmoc | mit |
Breadth-First Search
Breadt-first search (BFS) is algorithm that can find the closest members in a graph that match a certain search criterion.
BFS requires that we model our problem as a graph (nodes connected through edges). BFS can be applied to directed and undirected graph, where it can be applied to answer to types of question:
Is there are connection between a particular pair of nodes?
Which is the closest node to a given node that satisfies a certain criterion?
To answer these questions, BFS starts by checking all direct neighbors of a given node -- neighbors are nodes that have a direct connection to a particular node. Then, if none of those neighbors satisfies the criterion that we are looking for, the search is expanded to the neighbors of the nodes we just checked, and so on, until a match is found or all nodes in the graph were checked.
To keep track of the nodes that we have already checked and that we are going to check, we need two additional data structures:
1) A hash table to keep track of nodes we have already checked. If we don't check for previously checked nodes, we may end up in cycles depending on the structure of the graph.
2) A queue that stores the items to be checked.
Representing the graph
To represent the graph, its nodes and edges, we can simply use a hash table such as Python's built-in dictionaries. Imagine we have an undirected, social network graph that lists our direct friends (Elijah, Marissa, Nikolai) and friends of friends:
<img src="images/breadth-first-search/friend-graph-1.jpg" alt="" style="width: 400px;"/>
Say we are going to move to a new apartment next weekend, and we want to ask our friends if they have a pick-up truck that can be helpful in this endeavor. First, we would reach out to our directed friends (or 1st degree connections). If none of these have a pick-up truck, we ask them to ask their 1st degree connections (which are our 2nd degree connections), and so forth:
<img src="images/breadth-first-search/friend-graph-2.jpg" alt="" style="width: 600px;"/>
We can represent such a graph using a simple hash table (here: Python dictionary) as follows: | graph = {}
graph['You'] = ['Elijah', 'Marissa', 'Nikolai', 'Cassidy']
graph['Elijah'] = ['You']
graph['Marissa'] = ['You']
graph['Nikolai'] = ['John', 'Thomas', 'You']
graph['Cassidy'] = ['John', 'You']
graph['John'] = ['Cassidy', 'Nikolai']
graph['Thomas'] = ['Nikolai', 'Mario']
graph['Mario'] = ['Thomas'] | ipython_nbs/search/breadth-first-search.ipynb | rasbt/algorithms_in_ipython_notebooks | gpl-3.0 |
The Queue data structure
Next, let's setup a simple queue data structure. Of course, we can also use a regular Python list like a queue (using .insert(0, x) and .pop(), but this way, our breadth-first search implementation is maybe more illustrative. For more information about queues, please see the Queues and Deques notebook. | class QueueItem():
def __init__(self, value, pointer=None):
self.value = value
self.pointer = pointer
class Queue():
def __init__(self):
self.last = None
self.first = None
self.length = 0
def enqueue(self, value):
item = QueueItem(value, None)
if self.last:
self.last.pointer = item
if not self.first:
self.first = item
self.last = item
self.length += 1
def dequeue(self):
if self.first is not None:
value = self.first.value
self.first = self.first.pointer
self.length -= 1
else:
value = None
return value
qe = Queue()
qe.enqueue('a')
print('First element:', qe.first.value)
print('Last element:', qe.last.value)
print('Queue length:', qe.length)
qe.enqueue('b')
print('First element:', qe.first.value)
print('Last element:', qe.last.value)
print('Queue length:', qe.length)
qe.enqueue('c')
print('First element:', qe.first.value)
print('Last element:', qe.last.value)
print('Queue length:', qe.length)
val = qe.dequeue()
print('Dequeued value:', val)
print('Queue length:', qe.length)
val = qe.dequeue()
print('Dequeued value:', val)
print('Queue length:', qe.length)
val = qe.dequeue()
print('Dequeued value:', val)
print('Queue length:', qe.length)
val = qe.dequeue()
print('Dequeued value:', val)
print('Queue length:', qe.length)
qe.enqueue('c')
print('First element:', qe.first.value)
print('Last element:', qe.last.value)
print('Queue length:', qe.length) | ipython_nbs/search/breadth-first-search.ipynb | rasbt/algorithms_in_ipython_notebooks | gpl-3.0 |
Implementing freadth-first search to find the shortest path
Now, back to the graph, where we want to identify the closest connection that owns a truck, which can be helpful for moving (if we are allowed to borrow it, that is):
<img src="images/breadth-first-search/friend-graph-2.jpg" alt="" style="width: 600px;"/> | graph = {}
graph['You'] = ['Elijah', 'Marissa', 'Nikolai', 'Cassidy']
graph['Elijah'] = ['You']
graph['Marissa'] = ['You']
graph['Nikolai'] = ['John', 'Thomas', 'You']
graph['Cassidy'] = ['John', 'You']
graph['John'] = ['Cassidy', 'Nikolai']
graph['Thomas'] = ['Nikolai', 'Mario']
graph['Mario'] = ['Thomas'] | ipython_nbs/search/breadth-first-search.ipynb | rasbt/algorithms_in_ipython_notebooks | gpl-3.0 |
For simplicity, let's assume we have function that checks if a person ows a pick-up truck. (Say, Mario owns a pick-up truck, the check function knows it but we don't know it.) | def has_truck(person):
if person == 'Mario':
return True
else:
return False | ipython_nbs/search/breadth-first-search.ipynb | rasbt/algorithms_in_ipython_notebooks | gpl-3.0 |
Now, the breadth_first_search implementation below will check our closest neighbors first, then, it will check the neighnors of our neighbors and so forth. We will make use both of the graph we constructed and the Queue data structure that we implemented. Also, note that we are keeping track of people we already checked to prevent cycles in our search: | def breadth_first_search(graph):
# initialize queue
queue = Queue()
for person in graph['You']:
queue.enqueue(person)
people_checked = set()
degree = 0
while queue.length:
person = queue.dequeue()
if has_truck(person):
return person
else:
degree += 1
people_checked.add(person)
for next_person in graph[person]:
# check to prevent endless cycles
if next_person not in people_checked:
queue.enqueue(next_person)
breadth_first_search(graph) | ipython_nbs/search/breadth-first-search.ipynb | rasbt/algorithms_in_ipython_notebooks | gpl-3.0 |
Length range
Print out the gene names for all genes between 90 and 110 bases long. | import csv
with open('data.csv') as csvfile:
raw_data = csv.reader(csvfile)
for row in raw_data:
if len(row[1]) >= 90 or len(row[1]) <= 110:
print(row[2]) | Week_06/Week06 - 01 - Homework Solutions.ipynb | biof-309-python/BIOF309-2016-Fall | mit |
AT content
Print out the gene names for all genes whose AT content is less than 0.5 and whose expression level is greater than 200. | def is_at_rich(dna):
length = len(dna)
a_count = dna.upper().count('A')
t_count = dna.upper().count('T')
at_content = (a_count + t_count) / length
return at_content < 0.5
import csv
with open('data.csv') as csvfile:
raw_data = csv.reader(csvfile)
for row in raw_data:
if is_at_rich(row[1]) and int(row[3]) > 200:
print(row[2]) | Week_06/Week06 - 01 - Homework Solutions.ipynb | biof-309-python/BIOF309-2016-Fall | mit |
Complex condition
Print out the gene names for all genes whose name begins with “k” or “h” except those belonging to Drosophila melanogaster. | import csv
with open('data.csv') as csvfile:
raw_data = csv.reader(csvfile)
for row in raw_data:
if (row[2].startswith('k') or row[2].startswith('h')) and row[0] != 'Drosophila melanogaster':
print(row[2]) | Week_06/Week06 - 01 - Homework Solutions.ipynb | biof-309-python/BIOF309-2016-Fall | mit |
High low medium
For each gene, print out a message giving the gene name and saying whether its AT content is high (greater than 0.65), low (less than 0.45) or medium (between 0.45 and 0.65). | def at_percentage(dna):
length = len(dna)
a_count = dna.upper().count('A')
t_count = dna.upper().count('T')
at_content = (a_count + t_count) / length
return at_content
import csv
with open('data.csv') as csvfile:
raw_data = csv.reader(csvfile)
for row in raw_data:
at_percent = at_percentage(row[1])
if at_percent > 0.65:
print('AT content is high')
elif at_percent < 0.45:
print('AT content is high')
else:
print('AT content is medium') | Week_06/Week06 - 01 - Homework Solutions.ipynb | biof-309-python/BIOF309-2016-Fall | mit |
定义要训练的模型 | import collections
import time
import tensorflow as tf
import tensorflow_federated as tff
source, _ = tff.simulation.datasets.emnist.load_data()
def map_fn(example):
return collections.OrderedDict(
x=tf.reshape(example['pixels'], [-1, 784]), y=example['label'])
def client_data(n):
ds = source.create_tf_dataset_for_client(source.client_ids[n])
return ds.repeat(10).batch(20).map(map_fn)
train_data = [client_data(n) for n in range(10)]
input_spec = train_data[0].element_spec
def model_fn():
model = tf.keras.models.Sequential([
tf.keras.layers.InputLayer(input_shape=(784,)),
tf.keras.layers.Dense(units=10, kernel_initializer='zeros'),
tf.keras.layers.Softmax(),
])
return tff.learning.from_keras_model(
model,
input_spec=input_spec,
loss=tf.keras.losses.SparseCategoricalCrossentropy(),
metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])
trainer = tff.learning.build_federated_averaging_process(
model_fn, client_optimizer_fn=lambda: tf.keras.optimizers.SGD(0.02))
def evaluate(num_rounds=10):
state = trainer.initialize()
for round in range(num_rounds):
t1 = time.time()
state, metrics = trainer.next(state, train_data)
t2 = time.time()
print('Round {}: loss {}, round time {}'.format(round, metrics.loss, t2 - t1)) | site/zh-cn/federated/tutorials/high_performance_simulation_with_kubernetes.ipynb | tensorflow/docs-l10n | apache-2.0 |
设置远程执行器
默认情况下，TFF 在本地执行所有计算。在此步骤中，我们指示 TFF 连接到我们在上面设置的 Kubernetes 服务。确保在此处复制服务的 IP 地址。 | import grpc
ip_address = '0.0.0.0' #@param {type:"string"}
port = 80 #@param {type:"integer"}
channels = [grpc.insecure_channel(f'{ip_address}:{port}') for _ in range(10)]
tff.backends.native.set_remote_execution_context(channels) | site/zh-cn/federated/tutorials/high_performance_simulation_with_kubernetes.ipynb | tensorflow/docs-l10n | apache-2.0 |
运行训练 | evaluate() | site/zh-cn/federated/tutorials/high_performance_simulation_with_kubernetes.ipynb | tensorflow/docs-l10n | apache-2.0 |
The toy data created above consists of 4 gaussian blobs, having 200 points each, centered around the vertices of a rectancle. Let's plot it for convenience. | import matplotlib.pyplot as plt
%matplotlib inline
figure,axis = plt.subplots(1,1)
axis.plot(rectangle[0], rectangle[1], 'o', color='r', markersize=5)
axis.set_xlim(-5,15)
axis.set_ylim(-50,150)
axis.set_title('Toy data : Rectangle')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
With data at our disposal, it is time to apply KMeans to it using the KMeans class in Shogun. First we construct Shogun features from our data: | train_features = sg.create_features(rectangle) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Next we specify the number of clusters we want and create a distance object specifying the distance metric to be used over our data for our KMeans training: | # number of clusters
k = 2
# distance metric over feature matrix - Euclidean distance
distance = sg.create_distance('EuclideanDistance')
distance.init(train_features, train_features) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Next, we create a KMeans object with our desired inputs/parameters and train: | # KMeans object created
kmeans = sg.create_machine("KMeans", k=k, distance=distance)
# KMeans training
kmeans.train() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Now that training has been done, let's get the cluster centers and label for each data point | # cluster centers
centers = kmeans.get("cluster_centers")
# Labels for data points
result = kmeans.apply() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Finally let us plot the centers and the data points (in different colours for different clusters): | def plotResult(title = 'KMeans Plot'):
figure,axis = plt.subplots(1,1)
for i in range(totalPoints):
if result.get("labels")[i]==0.0:
axis.plot(rectangle[0,i], rectangle[1,i], 'go', markersize=3)
else:
axis.plot(rectangle[0,i], rectangle[1,i], 'yo', markersize=3)
axis.plot(centers[0,0], centers[1,0], 'go', markersize=10)
axis.plot(centers[0,1], centers[1,1], 'yo', markersize=10)
axis.set_xlim(-5,15)
axis.set_ylim(-50,150)
axis.set_title(title)
plt.show()
plotResult('KMeans Results') | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
<b>Note:</b> You might not get the perfect result always. That is an inherent flaw of KMeans algorithm. In subsequent sections, we will discuss techniques which allow us to counter this.<br>
Now that we have already worked out a simple KMeans implementation, it's time to understand certain specifics of KMeans implementaion and the options provided by Shogun to its users.
Initialization of cluster centers
The KMeans algorithm requires that the cluster centers are initialized with some values. Shogun offers 3 ways to initialize the clusters. <ul><li>Random initialization (default)</li><li>Initialization by hand</li><li>Initialization using <a href="http://en.wikipedia.org/wiki/K-means%2B%2B">KMeans++ algorithm</a></li></ul>Unless the user supplies initial centers or tells Shogun to use KMeans++, Random initialization is the default method used for cluster center initialization. This was precisely the case in the example discussed above.
Initialization by hand
There are 2 ways to initialize centers by hand. One way is to pass on the centers during KMeans object creation, as follows: | initial_centers = np.array([[0.,10.],[50.,50.]])
# initial centers passed
kmeans = sg.create_machine("KMeans", k=k, distance=distance, initial_centers=initial_centers) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Now, let's first get results by repeating the rest of the steps: | # KMeans training
kmeans.train(train_features)
# cluster centers
centers = kmeans.get("cluster_centers")
# Labels for data points
result = kmeans.apply()
# plot the results
plotResult('Hand initialized KMeans Results 1') | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
The other way to initialize centers by hand is as follows: | new_initial_centers = np.array([[5.,5.],[0.,100.]])
# set new initial centers
kmeans.put("initial_centers", new_initial_centers) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Let's complete the rest of the code to get results. | # KMeans training
kmeans.train(train_features)
# cluster centers
centers = kmeans.get("cluster_centers")
# Labels for data points
result = kmeans.apply()
# plot the results
plotResult('Hand initialized KMeans Results 2') | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Note the difference that inititial cluster centers can have on final result.
Initializing using KMeans++ algorithm
In Shogun, a user can also use <a href="http://en.wikipedia.org/wiki/K-means%2B%2B">KMeans++ algorithm</a> for center initialization. Using KMeans++ for center initialization is beneficial because it reduces total iterations used by KMeans and also the final centers mostly correspond to the global minima, which is often not the case with KMeans with random initialization. One of the ways to use KMeans++ is to set flag as <i>true</i> during KMeans object creation, as follows: | # set flag for using KMeans++
kmeans = sg.create_machine("KMeans", k=k, distance=distance, kmeanspp=True) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Completing rest of the steps to get result: | # KMeans training
kmeans.train(train_features)
# cluster centers
centers = kmeans.get("cluster_centers")
# Labels for data points
result = kmeans.apply()
# plot the results
plotResult('KMeans with KMeans++ Results') | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Training Methods
Shogun offers 2 training methods for KMeans clustering:<ul><li><a href='http://en.wikipedia.org/wiki/K-means_clustering#Standard_algorithm'>Classical Lloyd's training</a> (default)</li><li><a href='http://www.eecs.tufts.edu/~dsculley/papers/fastkmeans.pdf'>mini-batch KMeans training</a></li></ul>Lloyd's training method is used by Shogun by default unless user switches to mini-batch training method.
Mini-Batch KMeans
Mini-batch KMeans is very useful in case of extremely large datasets and/or very high dimensional data which is often the case in text mining. One can switch to Mini-batch KMeans training while creating KMeans object as follows: | # set training method to mini-batch
kmeans = sg.create_machine("KMeansMiniBatch", k=k, distance=distance) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Completing the code to get results: | # KMeans training
kmeans.train(train_features)
# cluster centers
centers = kmeans.get("cluster_centers")
# Labels for data points
result = kmeans.apply()
# plot the results
plotResult('Mini-batch KMeans Results') | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Applying KMeans on Real Data
In this section we see how useful KMeans can be in classifying the different varieties of Iris plant. For this purpose, we make use of Fisher's Iris dataset borrowed from the <a href='http://archive.ics.uci.edu/ml/datasets/Iris'>UCI Machine Learning Repository</a>. There are 3 varieties of Iris plants
<ul><li>Iris Sensosa</li><li>Iris Versicolour</li><li>Iris Virginica</li></ul>
The Iris dataset enlists 4 features that can be used to segregate these varieties, namely
<ul><li>sepal length</li><li>sepal width</li><li>petal length</li><li>petal width</li></ul>
It is additionally acknowledged that petal length and petal width are the 2 most important features (ie. features with very high class correlations)[refer to <a href='http://archive.ics.uci.edu/ml/machine-learning-databases/iris/iris.names'>summary statistics</a>]. Since the entire feature vector is impossible to plot, we only plot these two most important features in order to understand the dataset (at least partially). Note that we could have extracted the 2 most important features by applying PCA (or any one of the many dimensionality reduction methods available in Shogun) as well. | with open(os.path.join(SHOGUN_DATA_DIR, 'uci/iris/iris.data')) as f:
feats = []
# read data from file
for line in f:
words = line.rstrip().split(',')
feats.append([float(i) for i in words[0:4]])
# create observation matrix
obsmatrix = np.array(feats).T
# plot the data
figure,axis = plt.subplots(1,1)
# First 50 data belong to Iris Sentosa, plotted in green
axis.plot(obsmatrix[2,0:50], obsmatrix[3,0:50], 'o', color='green', markersize=5)
# Next 50 data belong to Iris Versicolour, plotted in red
axis.plot(obsmatrix[2,50:100], obsmatrix[3,50:100], 'o', color='red', markersize=5)
# Last 50 data belong to Iris Virginica, plotted in blue
axis.plot(obsmatrix[2,100:150], obsmatrix[3,100:150], 'o', color='blue', markersize=5)
axis.set_xlim(-1,8)
axis.set_ylim(-1,3)
axis.set_title('3 varieties of Iris plants')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
In the above plot we see that the data points labelled Iris Sentosa form a nice separate cluster of their own. But in case of other 2 varieties, while the data points of same label do form clusters of their own, there is some mixing between the clusters at the boundary. Now let us apply KMeans algorithm and see how well we can extract these clusters. | def apply_kmeans_iris(data):
# wrap to Shogun features
train_features = sg.create_features(data)
# number of cluster centers = 3
k = 3
# distance function features - euclidean
distance = sg.create_distance('EuclideanDistance')
distance.init(train_features, train_features)
# initialize KMeans object, use kmeans++ to initialize centers [play around: change it to False and compare results]
kmeans = sg.create_machine("KMeans", k=k, distance=distance, kmeanspp=True)
# training method is Lloyd by default [play around: change it to mini-batch by uncommenting the following lines]
#kmeans = sg.create_machine("KMeansMiniBatch", k=k, distance=distance)
# training kmeans
kmeans.train(train_features)
# labels for data points
result = kmeans.apply()
return result
result = apply_kmeans_iris(obsmatrix) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Now let us create a 2-D plot of the clusters formed making use of the two most important features (petal length and petal width) and compare it with the earlier plot depicting the actual labels of data points. | # plot the clusters over the original points in 2 dimensions
figure,axis = plt.subplots(1,1)
for i in range(150):
if result.get("labels")[i]==0.0:
axis.plot(obsmatrix[2,i],obsmatrix[3,i],'ro', markersize=5)
elif result.get("labels")[i]==1.0:
axis.plot(obsmatrix[2,i],obsmatrix[3,i],'go', markersize=5)
else:
axis.plot(obsmatrix[2,i],obsmatrix[3,i],'bo', markersize=5)
axis.set_xlim(-1,8)
axis.set_ylim(-1,3)
axis.set_title('Iris plants clustered based on attributes')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
From the above plot, it can be inferred that the accuracy of KMeans algorithm is very high for Iris dataset. Don't believe me? Alright, then let us make use of one of Shogun's clustering evaluation techniques to formally validate the claim. But before that, we have to label each sample in the dataset with a label corresponding to the class to which it belongs. | # first 50 are iris sensosa labelled 0, next 50 are iris versicolour labelled 1 and so on
labels = np.concatenate((np.zeros(50),np.ones(50),2.*np.ones(50)),0)
# bind labels assigned to Shogun multiclass labels
ground_truth = sg.create_labels(np.array(labels,dtype='float64')) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Now we can compute clustering accuracy making use of the ClusteringAccuracy class in Shogun | def analyzeResult(result):
# shogun object for clustering accuracy
AccuracyEval = sg.create_evaluation("ClusteringAccuracy")
# evaluates clustering accuracy
accuracy = AccuracyEval.evaluate(result, ground_truth)
# find out which sample points differ from actual labels (or ground truth)
compare = result.get("labels")-labels
diff = np.nonzero(compare)
return (diff,accuracy)
(diff,accuracy_4d) = analyzeResult(result)
print('Accuracy : ' + str(accuracy_4d))
# plot the difference between ground truth and predicted clusters
figure,axis = plt.subplots(1,1)
axis.plot(obsmatrix[2,:],obsmatrix[3,:],'x',color='black', markersize=5)
axis.plot(obsmatrix[2,diff],obsmatrix[3,diff],'x',color='r', markersize=7)
axis.set_xlim(-1,8)
axis.set_ylim(-1,3)
axis.set_title('Difference')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
In the above plot, wrongly clustered data points are marked in red. We see that the Iris Sentosa plants are perfectly clustered without error. The Iris Versicolour plants and Iris Virginica plants are also clustered with high accuracy, but there are some plant samples of either class that have been clustered with the wrong class. This happens near the boundary of the 2 classes in the plot and was well expected. Having mastered KMeans, it's time to move on to next interesting topic.
PCA as a preprocessor to KMeans
KMeans is highly affected by the <i>curse of dimensionality</i>. So, dimension reduction becomes an important preprocessing step. Shogun offers a variety of dimension reduction techniques to choose from. Since our data is not very high dimensional, PCA is a good choice for dimension reduction. We have already seen the accuracy of KMeans when all four dimensions are used. In the following exercise we shall see how the accuracy varies as one chooses lower dimensions to represent data.
1-Dimensional representation
Let us first apply PCA to reduce training features to 1 dimension | def apply_pca_to_data(target_dims):
train_features = sg.create_features(obsmatrix)
submean = sg.create_transformer("PruneVarSubMean", divide_by_std=False)
submean.fit(train_features)
submean.transform(train_features)
preprocessor = sg.create_transformer("PCA", target_dim=target_dims)
preprocessor.fit(train_features)
pca_transform = preprocessor.get("transformation_matrix")
new_features = np.dot(pca_transform.T, train_features.get("feature_matrix"))
return new_features
oneD_matrix = apply_pca_to_data(1) | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Next, let us get an idea of the data in 1-D by plotting it. | figure,axis = plt.subplots(1,1)
# First 50 data belong to Iris Sentosa, plotted in green
axis.plot(oneD_matrix[0,0:50], np.zeros(50), 'go', markersize=5)
# Next 50 data belong to Iris Versicolour, plotted in red
axis.plot(oneD_matrix[0,50:100], np.zeros(50), 'ro', markersize=5)
# Last 50 data belong to Iris Virginica, plotted in blue
axis.plot(oneD_matrix[0,100:150], np.zeros(50), 'bo', markersize=5)
axis.set_xlim(-5,5)
axis.set_ylim(-1,1)
axis.set_title('3 varieties of Iris plants')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Now that we have the results, the inevitable step is to check how good these results are. | (diff,accuracy_1d) = analyzeResult(result)
print('Accuracy : ' + str(accuracy_1d))
# plot the difference between ground truth and predicted clusters
figure,axis = plt.subplots(1,1)
axis.plot(oneD_matrix[0,:],np.zeros(150),'x',color='black', markersize=5)
axis.plot(oneD_matrix[0,diff],np.zeros(len(diff)),'x',color='r', markersize=7)
axis.set_xlim(-5,5)
axis.set_ylim(-1,1)
axis.set_title('Difference')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
2-Dimensional Representation
We follow the same steps as above and get the clustering accuracy.
STEP 1 : Apply PCA and plot the data (plotting is optional) | twoD_matrix = apply_pca_to_data(2)
figure,axis = plt.subplots(1,1)
# First 50 data belong to Iris Sentosa, plotted in green
axis.plot(twoD_matrix[0,0:50], twoD_matrix[1,0:50], 'go', markersize=5)
# Next 50 data belong to Iris Versicolour, plotted in red
axis.plot(twoD_matrix[0,50:100], twoD_matrix[1,50:100], 'ro', markersize=5)
# Last 50 data belong to Iris Virginica, plotted in blue
axis.plot(twoD_matrix[0,100:150], twoD_matrix[1,100:150], 'bo', markersize=5)
axis.set_title('3 varieties of Iris plants')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
STEP 3: Get the accuracy of the results | (diff,accuracy_2d) = analyzeResult(result)
print('Accuracy : ' + str(accuracy_2d))
# plot the difference between ground truth and predicted clusters
figure,axis = plt.subplots(1,1)
axis.plot(twoD_matrix[0,:],twoD_matrix[1,:],'x',color='black', markersize=5)
axis.plot(twoD_matrix[0,diff],twoD_matrix[1,diff],'x',color='r', markersize=7)
axis.set_title('Difference')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
STEP 3: Get accuracy of results. In this step, the 'difference' plot positions data points based petal length
and petal width in the original data. This will enable us to visually compare these results with that of KMeans applied
to 4-Dimensional data (ie. our first result on Iris dataset) | (diff,accuracy_3d) = analyzeResult(result)
print('Accuracy : ' + str(accuracy_3d))
# plot the difference between ground truth and predicted clusters
figure,axis = plt.subplots(1,1)
axis.plot(obsmatrix[2,:],obsmatrix[3,:],'x',color='black', markersize=5)
axis.plot(obsmatrix[2,diff],obsmatrix[3,diff],'x',color='r', markersize=7)
axis.set_title('Difference')
axis.set_xlim(-1,8)
axis.set_ylim(-1,3)
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
Finally, let us plot clustering accuracy vs. number of dimensions to consolidate our results. | from scipy.interpolate import interp1d
x = np.array([1, 2, 3, 4])
y = np.array([accuracy_1d, accuracy_2d, accuracy_3d, accuracy_4d])
f = interp1d(x, y)
xnew = np.linspace(1,4,10)
plt.plot(x,y,'o',xnew,f(xnew),'-')
plt.xlim([0,5])
plt.xlabel('no. of dims')
plt.ylabel('Clustering Accuracy')
plt.title('PCA Results')
plt.show() | doc/ipython-notebooks/clustering/KMeans.ipynb | geektoni/shogun | bsd-3-clause |
iii. Explain what private and public do
The private and public, also protected restrict the access to the class members.
A private member variable or function cannot be accessed, or even viewed from outside the class. Only the class and friend functions can access private members.
A public member is accessible from anywhere outside the class but within a program. You can set and get the value of public variables without any member function.
A protected member variable or function is very similar to a private member but it provided one additional benefit that they can be accessed in child classes which are called derived classes.
//From: https://www.tutorialspoint.com/cplusplus/cpp_class_access_modifiers.htm | //*Quote from comments:*
// This structure type is private to the class, and used as a form of
// linked list in order to contain the actual (static) data stored by the Stack class | Stack.ipynb | chapman-cs510-2016f/cw-12-redyellow | mit |
iv. Explain what size_t is used for
It is a type that can represent the size of any object in bytes: size_t is the type returned by the sizeof operator and is widely used in the standard library to represent sizes and counts.
//From: http://www.cplusplus.com/reference/cstring/size_t/ | //*Quote from comments:*
// Size method
// Specifying const tells the compiler that the method will not change the
// internal state of the instance of the class | Stack.ipynb | chapman-cs510-2016f/cw-12-redyellow | mit |
v. Explain why this code avoids the use of C pointers
First, raw pointers must under no circumstances own memory. That means you must delete after use it.
Second, most uses of pointers in C++ are unnecessary. C++ has very strong support for value semantics, you can use smart pointer, container classes, design patterns like RAII, ect, instead of pointer.
In computer science, a smart pointer is an abstract data type that simulates a pointer while providing additional features, such as automatic garbage collection or bounds checking. These additional features are intended to reduce bugs caused by the misuse of pointers while retaining efficiency. Smart pointers typically keep track of the objects they point to for the purpose of memory management.
The misuse of pointers is a major source of bugs: the constant allocation, deallocation and referencing that must be performed by a program written using pointers introduces the risk that memory leaks will occur. Smart pointers try to prevent memory leaks by making the resource deallocation automatic: when the pointer (or the last in a series of pointers) to an object is destroyed, for example because it goes out of scope, the pointed object is destroyed too.
//From: http://softwareengineering.stackexchange.com/questions/56935/why-are-pointers-not-recommended-when-coding-with-c
vi. Explain what new and delete do in C++, and how they relate to what you have done in C
"New" creates a pointer to an allocated memory block. "Delete" deallocates the memory that is allocated by "new".
It works differently from the way in C:
Allocate memory: <br>
C++: Node *n = new Node(); <br>
C : Node *n = (Node *)calloc(1, sizeof(Node));
Deallocate memory: <br>
C++: delete n; <br>
C : free(n);
vii. Explain what a memory leak is, and what you should do to avoid it
Memory leak means running out of system memory. When a program needs to store some temporary information during execution, it can dynamically request a chunk of memory from the system. However, the system has a fixed amount of total memory available. If one application uses up all of the system’s free memory, then other applications will not be able to obtain the memory that they require.
//From: https://msdn.microsoft.com/en-us/library/ms859408.aspx
I got three ways to avoid memory leak: <br>
1.free(C) or delete(C++) the memory you allocated after finishing use it; <br>
2.use smart pointer(C++) or other "garbage collector" to deallocate memory automatically after finishing use; <br>
3.use fewer pointers if it is possible.
viii. Explain what a unique_ptr is and how it relates to both new and C pointers
std::unique_ptr is a smart pointer that owns and manages another object through a pointer and disposes of that object when the unique_ptr goes out of scope.
The object is disposed of using the associated deleter when either of the following happens: <br>
the managing unique_ptr object is destroyed <br>
the managing unique_ptr object is assigned another pointer via operator= or reset().
It uses "new Node()" to allocate a new pointer, new_node_ptr, whose type is std::unique_ptr. It is a pointer but would deallocate the memory automatically when it is useless.
//From: http://en.cppreference.com/w/cpp/memory/unique_ptr | //*Quote from comments:*
// However, by using the "unique_ptr" type above, we carefully avoid any
// explicit memory allocation by using the allocation pre-defined inside the
// unique_ptr itself. By using memory-safe structures in this way, we are using
// the "Rule of Zero" and simplifying our life by defining ZERO of them:
// https://rmf.io/cxx11/rule-of-zero/
// http://www.cplusplus.com/reference/memory/unique_ptr/ | Stack.ipynb | chapman-cs510-2016f/cw-12-redyellow | mit |
ix. Explain what a list initializer does
Constructor is a special non-static member function of a class that is used to initialize objects of its class type.
In the definition of a constructor of a class, member initializer list specifies the initializers for direct and virtual base subobjects and non-static data members.
The order of member initializers in the list is irrelevant: the actual order of initialization is as follows:
1) If the constructor is for the most-derived class, virtual base classes are initialized in the order in which they appear in depth-first left-to-right traversal of the base class declarations (left-to-right refers to the appearance in base-specifier lists). <br>
2) Then, direct base classes are initialized in left-to-right order as they appear in this class's base-specifier list. <br>
3) Then, non-static data members are initialized in order of declaration in the class definition. <br>
4) Finally, the body of the constructor is executed.
//From:http://en.cppreference.com/w/cpp/language/initializer_list | //*Quote from comments*
// Implementation of default constructor
Stack::Stack()
: depth(0) // internal depth is 0
, head(nullptr) // internal linked list is null to start
{};
// The construction ": var1(val1), var2(val2) {}" is called a
// "list initializer" for a constructor, and is the preferred
// way of setting default field values for a class instance
// Here 0 is the default value for Stack::depth
// and nullptr is the default value for Stack::head | Stack.ipynb | chapman-cs510-2016f/cw-12-redyellow | mit |
x. Explain what the "Rule of Zero" is, and how it relates to the "Rule of Three"
Rule of Zero: Classes that have custom destructors, copy/move constructors or copy/move assignment operators should deal exclusively with ownership (which follows from the Single Responsibility Principle). Other classes should not have custom destructors, copy/move constructors or copy/move assignment operators.
Rule of Three: a class requires a user-defined destructor, a user-defined copy constructor, or a user-defined copy assignment operator. It almost certainly requires all three.
Rule of Zero does not need those three functions, but Rule of Three requires them.
//From: http://en.cppreference.com/w/cpp/language/rule_of_three | //*Quote from comments:*
// Normally we would have to implement the following things in C++ here:
// 1) Class Destructor : to deallocate memory when a Stack is deleted
// ~Stack();
//
// 2) Copy Constructor : to define what Stack b(a) does when a is a Stack
// This should create a copy b of the Stack a, but
// should be defined appropriately to do that
// Stack(const Stack&);
//
// 3) Copy Assignment : to define what b = a does when a is a Stack
// This should create a shallow copy of the outer
// structure of a, but leave the inner structure as
// pointers to the memory contained in a, and should
// be defined appropriately to do that
// Stack& operator=(const Stack&);
//
// The need for defining ALL THREE of these things when managing memory for a
// class explicitly is known as the "Rule of Three", and is standard
// http://stackoverflow.com/questions/4172722/what-is-the-rule-of-three
//
// However, by using the "unique_ptr" type above, we carefully avoid any
// explicit memory allocation by using the allocation pre-defined inside the
// unique_ptr itself. By using memory-safe structures in this way, we are using
// the "Rule of Zero" and simplifying our life by defining ZERO of them:
// https://rmf.io/cxx11/rule-of-zero/
// http://www.cplusplus.com/reference/memory/unique_ptr/ | Stack.ipynb | chapman-cs510-2016f/cw-12-redyellow | mit |
I accomplished the above by running this command at the command prompt:
THEANO_FLAGS='mode=FAST_RUN,device=gpu,floatX=float32' jupyter notebook | #import theano
from theano import function, config, sandbox, shared
import theano.tensor as T
import numpy as np
import scipy
import time | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
More theano setup in jupyter notebook boilerplate | print( theano.config.device )
print( theano.config.lib.cnmem) # cf. http://deeplearning.net/software/theano/library/config.html
print( theano.config.print_active_device)# Print active device at when the GPU device is initialized.
import os, sys
os.getcwd()
os.listdir( os.getcwd() )
%run gpu_test.py THEANO_FLAGS='mode=FAST_RUN,device=gpu,floatX=float32,lib.cnmem=0.85' # note lib.cnmem option for CnMem | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
sample data boilerplate | # Load the diabetes dataset
diabetes = sklearn.datasets.load_diabetes()
diabetes_X = diabetes.data
diabetes_Y = diabetes.target
#diabetes_X1 = diabetes_X[:,np.newaxis,2]
diabetes_X1 = diabetes_X[:,np.newaxis, 2].astype(theano.config.floatX)
#diabetes_Y = diabetes_Y.reshape( diabetes_Y.shape[0], 1)
diabetes_Y = diabetes_Y.astype(theano.config.floatX) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Linear regression
cf. Linear Regression In Theano
1_linear_regression.py from github Newmu/Theano-Tutorials
Train on $m$ number of input data points | m_lin = diabetes_X1.shape[0] | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
input, output variables $x$, $y$ for Theano | #x1 = T.vector('x1') # X1, input data, with only 1 feature, i.e. X \in \mathbb{R}^N, d=1
#ylin = T.vector('ylin') # target variable for linear regression, so that Y \in \mathbb{R}
x1 = T.scalar('x1') # X1, input data, with only 1 feature, i.e. X \in \mathbb{R}^N, d=1
ylin = T.scalar('ylin') # target variable for linear regression, so that Y \in \mathbb{R} | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Parameters (for a linear slope)
$$
(\theta^0, \theta^1) \in \mathbb{R}^2
$$ | thet0_init_val = np.random.randn()
thet1_init_val = np.random.randn()
thet0 = theano.shared( value=thet0_init_val, name='thet0', borrow=True) # \theta^0
thet1 = theano.shared( thet1_init_val, name='thet1', borrow=True) # \theta^1
| theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
hypothesis function $h_{\theta}$
$$
h_{\theta}(x) = \theta_1 x + \theta_0
$$ | #h_thet = T.dot( thet1, x1) + thet0
# whereas, Newmu uses
h_thet = thet1 * x1 + thet0 | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Cost function $J(\theta)$ | # roshansanthosh uses
#Jthet = T.sum( T.pow(h_thet-ylin,2))/(2*m_lin)
# whereas, Newmu uses
# Jthet = T.mean( T.sqr( thet_1*x1 + thet_0 - ylin ))
Jthet = T.mean( T.pow( h_thet-ylin,2))/2
#Jthet = sandbox.cuda.basic_ops.gpu_from_host( T.mean(
# sandbox.cuda.basic_ops.gpu_from_host( T.pow( h_thet-ylin,2))))/2 | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
$$
\text{grad}{\theta}J(\theta) = ( \text{grad}{\theta^0} J , \text{grad}_{\theta^1} J )
$$ | grad_thet0 = T.grad(Jthet, thet0)
grad_thet1 = T.grad(Jthet, thet1)
# so-called "learning rate"
gamma = 0.01 | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Note that "updates (iterable over pairs (shared_variable, new_expression) List, tuple or dict.) – expressions for new SharedVariable values" cf. Theano doc | train_lin = theano.function(inputs = [x1,ylin], outputs=Jthet,
updates=[[thet1,thet1-gamma*grad_thet1],[thet0,thet0-gamma*grad_thet0]])
test_lin = theano.function([x1],h_thet)
#X1_lin_in = shared( diabetes_X1 ,'float32')
#Y_lin_out = shared( diabetes_Y, 'float32')
training_steps = 1000 # 10000
sh_diabetes_X1 = shared( diabetes_X1 , borrow=True)
sh_diabetes_Y = shared( diabetes_Y, borrow=True)
"""
for i in range(training_steps):
for x,y in zip( diabetes_X1, diabetes_Y):
Jthet_val = train_lin( x, y )
"""
for i in range(training_steps):
# for x,y in zip( sh_diabetes_X1, sh_diabetes_Y) :
# Jthet_val = train_lin( x,y)
Jthet_val = train_lin( sh_diabetes_X1, sh_diabetes_Y)
print(Jthet_val)
print( thet0.get_value() ); print( thet1.get_value() )
test_lin_out = np.array( [ test_lin( x ) for x in diabetes_X1 ] )
plt.plot(diabetes_X1,diabetes_Y,'ro')
plt.plot(diabetes_X1,test_lin_out)
if any([x.op.__class__.__name__ in ['GpuGemm','GpuGemv'] for x in train_lin.maker.fgraph.toposort()]):
print("Used the gpu")
else:
print(train_lin.maker.fgraph.toposort())
if np.any([isinstance(x.op,T.Elemwise) for x in train_lin.maker.fgraph.toposort()]):
print("Used the cpu") | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Linear Algebra and theano
cf. Week 1, Linear Algebra Review, Coursera, Machine Learning with Ng
I'll take this opportunity to provide a dictionary between the syntax of linear algebra math and numpy.
Essentially, what I did was take Coursera's Week 1, Linear Algebra Review and then translated the math into theano, and in particular, running theano on the GPU.
Other reference that I used was
https://simplyml.com/linear-algebra-shootout-numpy-vs-theano-vs-tensorflow-2/
Linear Algebra Shootout: NumPy vs. Theano vs. TensorFlow by Charanpal Dhanjal - 14/07/16
Matrix addition
cf. Coursera, Intro. to Machine Learning, Linear Algebra Review, Addition and Scalar Multiplication | A = T.matrix('A')
B = T.matrix('B')
#matadd = function([A,B], A+B)
#matadd = function([A,B],sandbox.cuda.basic_ops.gpu_from_host(A+B) )
# Note: we are just defining the expressions, nothing is evaluated here!
C = sandbox.cuda.basic_ops.gpu_from_host(A+B)
matadd = function([A,B], C)
#A = T.dmatrix('A')
#B = T.dmatrix('B')
A = T.matrix('A')
B = T.matrix('B')
C_out = A + B
matadd_CPU = function([A,B], C_out)
A_eg = shared( np.array([[8,6,9],[10,1,10]]), 'float32')
B_eg = shared( np.array([[3,10,2],[6,1,-1]]), 'float32')
A_eg_CPU = np.array([[8,6,9],[10,1,10]])
B_eg_CPU = np.array([[3,10,2],[6,1,-1]])
print(A_eg_CPU)
print( type( A_eg_CPU ))
print( A_eg_CPU.shape)
print( B_eg_CPU.shape)
print( matadd.maker.fgraph.toposort() )
print( matadd_CPU.maker.fgraph.toposort() )
matadd( A_eg, B_eg) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
The way to do it, to "force" on the GPU, is like this (cf. Speeding up your Neural Network with Theano and the GPU - Wild ML): | np.random.randn( *A_eg_CPU.shape )
C_out = theano.shared( np.random.randn( *A_eg_CPU.shape).astype('float32') )
C_out.type()
#A_in = shared( A_eg_CPU, "float32")
#A_in = shared( A_eg_CPU, "float32")
A_in = shared( A_eg_CPU.astype("float32"), "float32")
B_in = shared( B_eg_CPU.astype("float32"), "float32")
#C_out_GPU = A_in + B_in
C_out_GPU = sandbox.cuda.basic_ops.gpu_from_host(A_in+B_in)
matadd_GPU = theano.function( [], C_out_GPU)
C_out_GPU_result = matadd_GPU()
C_out_GPU_result | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Notice how DIFFERENT this setup or syntax is: we have to set up tensor or matrix shared variables A_n, B_in, which are then used to define the theano function, theano.function. "By using shared variables we ensure that they are present in the GPU memory". cf. Linear Algebra Shootout: NumPy vs. Theano vs. TensorFlow | print( matadd_GPU.maker.fgraph.toposort() )
#if np.any([isinstance(C_out_GPU.op, tensor.Elemwise ) and
if np.any([isinstance( C_out_GPU.op, T.Elemwise ) and
('Gpu' not in type( C_out_GPU.op).__name__) for x in matadd_GPU.maker.fgraph.toposort()]) :
print('Used the cpu')
else:
print('Used the gpu')
matadd_CPU( A_eg_CPU.astype("float32"), B_eg_CPU.astype("float32") )
type(A_eg)
print( type( numpy.asarray(rng.rand(2000)) ) )
numpy.asarray(rng.rand(2000)).shape | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Bottom Line: there are 2 ways of doing linear algebra on the GPU
symbolic computation with the usual arguments
$$
A + B = C \in \text{Mat}_{\mathbb{R}}(M,N)
$$
$ \forall \, A, B \in \text{Mat}_{\mathbb{R}}(M,N)$ | A = T.matrix('A')
B = T.matrix('B')
C = sandbox.cuda.basic_ops.gpu_from_host( A + B ) # vs.
# C = A + B # this will result in an output array on the host, as opposed to CudaNdarray on device
matadd = function([A,B], C)
print( matadd.maker.fgraph.toposort() )
matadd( A_eg_CPU.astype("float32"), B_eg_CPU.astype("float32") ) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
with shared variables | A_in = shared( A_eg_CPU.astype("float32"), "float32") # initialize with the input values, A_eg_CPU, anyway
B_in = shared( B_eg_CPU.astype("float32"), "float32") # initialize with the input values B_eg_CPU, anyway
# C_out = A_in + B_in # this version will output to the host as a numpy.ndarray
# indeed, reading the graph,
"""
[GpuElemwise{add,no_inplace}(float32, float32), HostFromGpu(GpuElemwise{add,no_inplace}.0)]
"""
# this version immediately below, in 1 line, will result in a CudaNdarray on device
C_out = sandbox.cuda.basic_ops.gpu_from_host(A_in+B_in)
matadd_GPU = theano.function( [], C_out)
print( matadd_GPU.maker.fgraph.toposort() )
C_out_result = matadd_GPU()
C_out_result | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Scalar Multiplication (on the GPU)
cf. Scalar Multiplication of Linear Algebra Review, coursera, Machine Learning Intro by Ng | A_2 = np.array( [[4,5],[1,7] ])
a = T.scalar('a')
F = sandbox.cuda.basic_ops.gpu_from_host( a*A )
scalarmul = theano.function([a,A],F)
print( scalarmul.maker.fgraph.toposort() )
scalarmul( np.float32( 2.), A_2.astype("float32")) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Composition; Confirming that you can do composition of scalar multiplication on a matrix (or ring) addition
Being able to do composition is very important in math | scalarmul( np.float32(2.), matadd( A_eg_CPU.astype("float32"), B_eg_CPU.astype("float32") ) )
u = T.vector('u')
v = T.vector('v')
w = sandbox.cuda.basic_ops.gpu_from_host( u + v)
vecadd = theano.function( [u,v],w)
t = sandbox.cuda.basic_ops.gpu_from_host( a * u)
scalarmul_vec = theano.function([a,u], t)
print(vecadd.maker.fgraph.toposort())
print(scalarmul_vec.maker.fgraph.toposort())
u_eg = np.array( [4,6,7], dtype="float32")
v_eg = np.array( [2,1,0], dtype="float32")
print( u_eg.shape)
scalarmul_vec( np.float32(0.5), u_eg )
vecadd( scalarmul_vec( np.float32(0.5), u_eg ) , scalarmul_vec( np.float32(-3.), v_eg ) ) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
This was the computer equivalent to mathematical expression:
$$
\left[ \begin{matrix} 4 \ 6 \ 7 \end{matrix} \right] /2 - 3 * \left[ \begin{matrix} 2 \ 1 \ 0 \end{matrix} \right]
$$
sAxy or A-V multiplication or so-called "Gemv", or Matrix Multiplication on a vector, or linear transformation on a R-module, or vector space
i.e.
$$
Av = B
$$ | B_out = sandbox.cuda.basic_ops.gpu_from_host( T.dot(A,v))
AVmul = theano.function([A,v], B_out)
print(AVmul.maker.fgraph.toposort())
AVmul( np.array([[1,0,3],[2,1,5],[3,1,2]]).astype("float32"), np.array([1,6,2]).astype("float32"))
AVmul( np.array([[1,0,0],[0,1,0],[0,0,1]]).astype("float32"), np.array([1,6,2]).astype("float32")) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
AB or Gemm or Matrix Multiplication, i.e. Ring multiplication
i.e.
$$
A*B = C
$$ | C_f = sandbox.cuda.basic_ops.gpu_from_host( T.dot(A,B))
matmul = theano.function([A,B], C_f)
print( matmul.maker.fgraph.toposort())
matmul( np.array( [[1,3],[2,4],[0,5]] ).astype("float32"), np.array([[1,0],[2,3]]).astype("float32") ) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Inverse and Transpose
cf. Inverse and Transpose | Ainverse = sandbox.cuda.basic_ops.gpu_from_host( T.inv(A))
Ainv = theano.function([A], Ainverse)
print(Ainv.maker.fgraph.toposort())
Atranspose = sandbox.cuda.basic_ops.gpu_from_host( A.T)
AT = theano.function([A],Atranspose)
print(AT.maker.fgraph.toposort()) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Summation, sum, mean, scan
Linear Regression (again), via Coursera's Machine Learning Intro by Ng, Programming Exercise 1 for Week 2
Boilerplate, load sample data | linregdata = pd.read_csv('./coursera_Ng/machine-learning-ex1/ex1/ex1data1.txt', header=None)
X_linreg_training = linregdata.as_matrix([0]) # pandas.DataFrame.as_matrix convert frame to its numpy-array representation
y_linreg_training = linregdata.as_matrix([1])
m_linreg_training = len(y_linreg_training) # number of training examples
print( X_linreg_training.shape, type(X_linreg_training))
print( y_linreg_training.shape, type(y_linreg_training))
print m_linreg_training | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Try representing $\theta$, parameters or "weights", of size $|\theta|$ which should be equal to the number of features $n$ (or $d$). | # theta_linreg = T.vector('theta_linreg')
d = X_linreg_training.shape[1] # d = features
# Declare Theano symbolic variables
X = T.matrix('x')
y = T.vector('y') | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Preprocess training data (due to numpy's treatment of arrays) (note, this is not needed, if you use pandas to choose which column(s) you want to make into a numpy array) | #X_linreg_training = X_linreg_training.reshape( m_linreg_training,1)
#y_linreg_training = y_linreg_training.reshape( m_linreg_training,1)
# Instead, the training data X and test data values y are going to be represented by Theano symbolic variable above
#X_linreg = theano.shared(X_linreg_training.astype("float32"),"float32")
#y_linreg = theano.shared(y_linreg_training.astype("float32"),"float32")
#theta_0 = np.zeros( ( d+1,1)); print(theta_0)
theta_0 = np.zeros( d+1); print(theta_0)
theta = theano.shared( theta_0.astype("float32"), "theta")
alpha = np.float32(0.01) # learning rate gamma or alpha
# Construct Theano "expression graph"
predicted_vals = sandbox.cuda.basic_ops.gpu_from_host( T.dot(X,theta) ) # h_{\theta}
m = np.float32( y_linreg_training.shape[0] )
J_theta = sandbox.cuda.basic_ops.gpu_from_host(
T.dot( (T.dot(X,theta) - y).T, T.dot(X,theta) - y) * np.float32( 0.5 ) * np.float32( 1./ m )
) # cost function
update_theta = sandbox.cuda.basic_ops.gpu_from_host(
theta - alpha * T.grad( J_theta, theta) )
gradientDescent = theano.function(
inputs=[X,y],
outputs=[predicted_vals,J_theta],
updates=[(theta, update_theta)],
name = "gradientDescent")
print( gradientDescent.maker.fgraph.toposort() )
num_iters = 1500
J_History = [] | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Preprocess X to include intercepts | input_X_linreg = np.hstack( ( np.ones((m_linreg_training,1)), X_linreg_training ) ).astype("float32")
y_linreg_training_processed = y_linreg_training.reshape( m_linreg_training,).astype("float32")
J_History = [0 for iter in range(num_iters)]
for iter in range(num_iters):
predicted_vals_out, J_out = \
gradientDescent(input_X_linreg.astype("float32"), y_linreg_training_processed.astype("float32") )
J_History[iter] = J_out
Deg = (np.random.randn(40,10).astype("float32"), np.random.randint(size=40,low=0,high=2).astype("float32") )
Deg[0].shape
Deg[1].shape
theta.get_value()
dir( J_History[0] )
J_History[-5].gpudata
plt.plot( [ele.gpudata for ele in J_History]) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Denny Britz's way:
http://www.wildml.com/2015/09/speeding-up-your-neural-network-with-theano-and-the-gpu/
Speeding up your Neural Network with Theano and the GPU
and his jupyter notebook
https://github.com/dennybritz/nn-theano/blob/master/nn-theano-gpu.ipynb
nn-theano/nn-theano-gpu.ipynb | input_X_linreg.shape
# GPU NOTE: Conversion to float32 to store them on the GPU!
X = theano.shared( input_X_linreg.astype('float32'), name='X' )
y = theano.shared( y_linreg_training.astype('float32'), name='y')
# GPU NOTE: Conversion to float32 to store them on the GPU!
theta = theano.shared( np.vstack(theta_0).astype("float32"), name='theta')
# Construct Theano "expression graph"
predicted_vals = sandbox.cuda.basic_ops.gpu_from_host(
T.dot(X,theta) ) # h_{\theta}
m = np.float32( y_linreg_training.shape[0] )
# cost function J_theta, J_{\theta}
J_theta = sandbox.cuda.basic_ops.gpu_from_host(
(
T.dot( (T.dot(X,theta) - y).T, T.dot(X,theta) - y) * np.float32(0.5) * np.float32( 1./m)
).reshape([]) ) # cost function # reshape is to force "broadcast" into 0-dim. scalar for cost function
update_theta = sandbox.cuda.basic_ops.gpu_from_host(
theta - alpha * T.grad( J_theta, theta) )
# Note that we removed the input values because we will always use the same shared variable
# GPU Note: Removed the input values to avoid copying data to the GPU.
gradientDescent = theano.function(
inputs=[],
# outputs=[predicted_vals,J_theta],
updates=[(theta, update_theta)],
name = "gradientDescent")
print( gradientDescent.maker.fgraph.toposort() )
#J_History = [0 for iter in range(num_iters)]
for iter in range(num_iters):
gradientDescent( )
print( np.vstack( theta_0).shape )
print( y_linreg_training.shape )
theta.get_value()
# Profiling
print( theano.config.profile ) # Do the vm/cvm linkers profile the execution time of Theano functions?
print( theano.config.profile_memory ) # Do the vm/cvm linkers profile the memory usage of Theano functions? It only works when profile=True.
theano.printing.debugprint(gradientDescent)
#print( gradientDescent.profile.print_summary() )
dir( gradientDescent.profile) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Testing the Linear Regression with (Batch) Gradient Descent classes in ./ML/ | import sys
import os
#sys.path.append( os.getcwd() + '/ML')
sys.path.append( os.getcwd() + '/ML' )
from linreg_gradDes import LinearReg, LinearReg_loaded
#from ML import LinearReg, LinearReg_loaded | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Boilerplate for sample input data | linregdata1 = pd.read_csv('./coursera_Ng/machine-learning-ex1/ex1/ex1data1.txt', header=None)
linregdata1.as_matrix([0]).shape
linregdata1.as_matrix([1]).shape
features = linregdata1.as_matrix([0]).shape[1]
numberoftraining = linregdata1.as_matrix([0]).shape[0]
LinReg_housing = LinearReg( features, numberoftraining , 0.01)
Xin = LinReg_housing.preprocess_X( linregdata1.as_matrix([0]))
ytest = linregdata1.as_matrix([1]).flatten()
%time LinReg_housing.build_model( Xin, ytest )
LinRegloaded_housing = LinearReg_loaded( linregdata1.as_matrix([0]), linregdata1.as_matrix([1]),
features, numberoftraining )
%time LinRegloaded_housing.build_model()
print( LinReg_housing.gradientDescent.maker.fgraph.toposort() )
print( LinRegloaded_housing.gradientDescent.maker.fgraph.toposort() )
| theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Other (sample) datasets
Consider feature normalization | def featureNormalize(X):
"""
FEATURENORMALIZE Normalizes the features in X
FEATURENORMALIZE(X) returns a normalized version of X where
the mean value of each feature is 0 and the standard deviation
is 1. This is often a good preprocessing step to do when
working with learning algorithms.
"""
# You need to set these values correctly
X_norm = (X-X.mean(axis=0))/X.std(axis=0)
mu = X.mean(axis=0)
sigma = X.std(axis=0)
return [X_norm, mu, sigma]
linregdata2 = pd.read_csv('./coursera_Ng/machine-learning-ex1/ex1/ex1data2.txt', header=None)
features = linregdata2.as_matrix().shape[1] - 1
numberoftraining = linregdata2.as_matrix().shape[0]
Xdat = linregdata2.as_matrix( range(features) )
ytest = linregdata2.as_matrix( [features])
[Xnorm, mus,sigmas] = featureNormalize(Xdat)
LinReg_housing2 = LinearReg( features, numberoftraining, 0.01)
processed_X = LinReg_housing2.preprocess_X( Xnorm )
%time LinReg_housing2.build_model( processed_X, ytest.flatten(), 400)
LinRegloaded_housing2 = LinearReg_loaded( Xnorm, ytest,
features, numberoftraining )
%time LinRegloaded_housing2.build_model( 400) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Diabetes data from sklearn, sci-kit learn | # Load the diabetes dataset
diabetes = sklearn.datasets.load_diabetes()
diabetes_X = diabetes.data
diabetes_Y = diabetes.target
#diabetes_X1 = diabetes_X[:,np.newaxis,2]
diabetes_X1 = diabetes_X[:,np.newaxis, 2].astype(theano.config.floatX)
#diabetes_Y = diabetes_Y.reshape( diabetes_Y.shape[0], 1)
diabetes_Y = np.vstack( diabetes_Y.astype(theano.config.floatX) )
features1 = 1
numberoftraining = diabetes_Y.shape[0]
LinReg_diabetes = LinearReg( features1, numberoftraining, 0.01)
processed_X = LinReg_diabetes.preprocess_X( diabetes_X1 )
%time LinReg_diabetes.build_model( processed_X, diabetes_Y.flatten(), 10000)
LinRegloaded_diabetes = LinearReg_loaded( diabetes_X1, diabetes_Y,
features1, numberoftraining )
%time LinRegloaded_diabetes.build_model( 10000) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Multiple number of features case: | features = diabetes_X.shape[1]
LinReg_diabetes = LinearReg( features, numberoftraining, 0.01)
processed_X = LinReg_diabetes.preprocess_X( diabetes_X )
%time LinReg_diabetes.build_model( processed_X, diabetes_Y.flatten(), 10000)
LinRegloaded_diabetes = LinearReg_loaded( diabetes_X, diabetes_Y,
features, numberoftraining )
%time LinRegloaded_diabetes.build_model( 10000) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
ex2 Linear Regression, on d=2 features | data_ex1data2 = pd.read_csv('./coursera_Ng/machine-learning-ex1/ex1/ex1data2.txt', header=None)
X_ex1data2 = data_ex1data2.iloc[:,0:2]
y_ex1data2 = data_ex1data2.iloc[:,2]
m_ex1data2 = y_ex1data2.shape[0]
X_ex1data2=X_ex1data2.values.astype(np.float32)
y_ex1data2=y_ex1data2.values.reshape((m_ex1data2,1)).astype(np.float32)
print(type(X_ex1data2))
print(type(y_ex1data2))
print(X_ex1data2.shape)
print(y_ex1data2.shape)
print(m_ex1data2)
print(X_ex1data2[:5])
print(y_ex1data2[:5])
((X_ex1data2[:,1] - X_ex1data2[:,1].mean())/( X_ex1data2[:,1].std()) ).std()
# feature Normalize
#X_ex1data2_norm = sklearn.preprocessing.Normalizer.transform(X_ex1data2 )
X_ex1data2_norm = (X_ex1data2 - np.mean(X_ex1data2, axis=0)) / np.std(X_ex1data2, axis=0)
print(X_ex1data2_norm[:,0].mean())
print(X_ex1data2_norm[:,0].std())
print(X_ex1data2_norm[:,1].mean())
print(X_ex1data2_norm[:,1].std())
# X_ex1data2_norm[:5];
X=T.matrix(dtype=theano.config.floatX)
y=T.matrix(dtype=theano.config.floatX)
Theta=theano.shared(np.zeros((2,1)).astype(theano.config.floatX))
b = theano.shared(np.zeros(1).astype(theano.config.floatX))
print(b.get_value().shape)
yhat = T.dot( X, Theta) + b
# L2 norm
J = np.cast[theano.config.floatX](0.5)*T.mean( T.sqr( yhat-y))
alpha=0.01 # learning rate
# sandbox.cuda.basic_ops.gpu_from_host
updateThetab = [ Theta-np.float32(alpha)*T.grad(J,Theta), b-np.float32(alpha)*T.grad(J,b)]
gradientDescent_step = theano.function(inputs=[X,y],
outputs=J,
updates = zip([Theta,b],updateThetab) )
num_iters =400
JList=[]
for iter in range(num_iters):
err = gradientDescent_step(X_ex1data2_norm,y_ex1data2)
JList.append(err)
# Final mode:
print(Theta.get_value())
print(b.get_value())
# JList[-10:]
plt.plot(JList)
plt.show() | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Multi-class Classification
cf. ex3, Programming Exercise 3: Multi-class Classification and Neural Networks, Machine Learning
1 Multi-class Classification | os.getcwd()
os.listdir( './coursera_Ng/machine-learning-ex3/' )
os.listdir( './coursera_Ng/machine-learning-ex3/ex3' )
# Load saved matrices from file
multiclscls_data = scipy.io.loadmat('./coursera_Ng/machine-learning-ex3/ex3/ex3data1.mat') | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
import the classes from ML | import sys
import os
os.getcwd()
#sys.path.append( os.getcwd() + '/ML')
sys.path.append( os.getcwd() + '/ML' )
from gradDes import LogReg
# Test case for Cost function J_{\theta} with regularization
theta_t = np.vstack( np.array( [-2, -1, 1, 2]) )
X_t = np.array( [i/10. for i in range(1,16)]).reshape((3,5)).T
#X_t = np.hstack( ( np.ones((5,1)), X_t) ) # no need to preprocess the input data X with column of 1's
y_t = np.vstack( np.array( [1,0,1,0,1]))
MulClsCls_digits = LogReg( X_t, y_t, 3,5,0.01, 3. )
MulClsCls_digits.calculate_cost()
MulClsCls_digits.z.get_value()
print( MulClsCls_digits.X.get_value() )
MulClsCls_digits.y.get_value()
calc_z_test = theano.function([], MulClsCls_digits.z)
calc_z_test()
MulClsCls_digits.theta.set_value( theta_t.astype('float32') )
calc_z_test()
MulClsCls_digits.calculate_cost()
print( 1/(1+np.exp( np.dot( -np.hstack( ( np.ones((5,1)), X_t) ), theta_t) ) ) )
h_test = 1/(1+np.exp( np.dot( -np.hstack( ( np.ones((5,1)), X_t) ), theta_t) ) )
print( np.dot( (h_test - y_t).T, h_test- y_t) * 0.5/5 ) # non-regularized J_theta cost term
np.dot( theta_t[1:].T, theta_t[1:]) * 3 / (2.* 5)
MulClsCls_digits.predict()
MulClsCls_digit
theano.config.floatX | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Neural Networks
Model representation
cf. 2 Neural Networks, 2.1 Model representation, ex3.pdf | os.getcwd()
os.listdir( './coursera_Ng/machine-learning-ex3/' )
os.listdir( './coursera_Ng/machine-learning-ex3/ex3/' ) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
$ \Theta_1, \Theta_2 $ | # Load saved matrices from file
nn3_data = scipy.io.loadmat('./coursera_Ng/machine-learning-ex3/ex3/ex3weights.mat')
print( nn3_data.keys() )
print( type( nn3_data['Theta1']) )
print( type( nn3_data['Theta2']) )
print( nn3_data['Theta1'].shape )
print( nn3_data['Theta2'].shape )
Theta1[0] | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Feedforward | %load_ext tikzmagic | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
$$
\begin{tikzpicture}
\matrix (m) [matrix of math nodes, row sep=3em, column sep=4em, minimum width=2em]
{
\mathbb{R}^{s_l} & \mathbb{R}^{ s_l +1 } & \mathbb{R}^{s_{l+1} } & \mathbb{R}^{s_{l+1} } \
a^{(l)} & (a_0^{(l)} = 1, a^{(l)} ) & z^{(l+1)} & g(z^{(l+1)}) = a^{(l+1)} \
};
\path[->]
(m-1-1) edge node [above] {$a_0^{(l)}=1$} (m-1-2)
(m-1-2) edge node [above] {$\Theta^{(l)}$} (m-1-3)
(m-1-3) edge node [above] {$g$} (m-1-4)
;
\path[|->]
(m-2-1) edge node [above] {$a_0^{(l)}=1$} (m-2-2)
(m-2-2) edge node [above] {$\Theta^{(l)}$} (m-2-3)
(m-2-3) edge node [above] {$g$} (m-2-4)
;
\end{tikzpicture}
$$ | np.random.seed(0)
s_l = 400 # (layer) size of layer l, i.e. number of nodes, units in layer l
s_lp1 = 25
al = theano.shared( np.random.randn(s_l+1,1).astype('float32'), name="al")
#alp1 = theano.shared( np.random.randn(s_lp1,1).astype('float32'), name="al")
#Thetal = theano.shared( np.random.randn( s_lp1,s_l+1).astype('float32') , name="Thetal")
# Feedforward, forward propagation
#z = T.dot( Thetal, al)
#g = T.nnet.sigmoid( z)
s_l = 25
s_lp1 = 10
rng = np.random.RandomState(99)
Theta_values = np.asarray( rng.uniform(
low=-np.sqrt( 6. / (s_l+ s_lp1)),
high=np.sqrt( 6./(s_l + s_lp1)), size=(s_lp1,s_l+1)), dtype=theano.config.floatX )
print( Theta_values.shape )
print( Theta_values.dtype )
#Theta_values *= np.float32(4)
Theta_values *= 4.
print( Theta_values.dtype)
Theta_values.shape
np.float32( 4) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
From Deep Learning Tutorials of LISA lab of University of Montreal; logistic_sgd.py, mlp.py | %env
os.getcwd()
print( sys.path )
#sys.path.append( os.getcwd() + '/ML')
sys.path.append( '../DeepLearningTutorials/code/' )
#from logistic_sgd import LogisticRegression, load_data, sgd_optimization_mnist, predict
import logistic_sgd
MNIST_MTLdat = logistic_sgd.load_data("../DeepLearningTutorials/data/mnist.pkl.gz") # list of training data
print(len(MNIST_MTLdat))
print(type(MNIST_MTLdat))
for ele in MNIST_MTLdat: print type(ele), len(ele) # test_set_x, test_set_y, valid_set_x, valid_set_y, train_set_x,
print( MNIST_MTLdat[0][0].get_value().shape)
print( type(MNIST_MTLdat[0][1]))
print( MNIST_MTLdat[0][1].get_scalar_constant_value )
print( type( MNIST_MTLdat[1][1] ) )
MNIST_MTLdat[1][1].shape
dir(MNIST_MTLdat[0][1]) ;
import gzip
import six.moves.cPickle as pickle
with gzip.open("../DeepLearningTutorials/data/mnist.pkl.gz", 'rb') as f:
try:
train_set, valid_set, test_set = pickle.load(f, encoding='latin1')
except:
train_set, valid_set, test_set = pickle.load(f)
print( type( train_set[0] ))
print( train_set[0].shape )
print( type( train_set[1]))
print( train_set[1].shape )
print( type( valid_set[0] ))
print( valid_set[0].shape )
print( type( valid_set[1]))
print( valid_set[1].shape )
print( type( test_set[0] ))
print( test_set[0].shape )
print( type( test_set[1]))
print( test_set[1].shape )
X = train_set[0].T
pd.DataFrame(X.T).describe()
28*28
X_i = theano.shared( X.astype("float32"))
m = X_i.get_value().shape[1]
a1 = T.stack( [ theano.shared( np.ones((1,m)).astype("float32") ) , X_i ] , axis=1 )
print( type(a1) )
#print( a1.get_scalar_constant_value() )
dir(a1)
a1.get_parents()
a1.ndim
a1_0 = theano.shared( np.ones((1,m)).astype("float32"),name='a1_0')
a1 = T.stack( [a1_0,X_i], axis=0)
d = X_i.get_value().shape[0]
s_2 = d/2
rng1 = np.random.RandomState(1234)
Theta1_values = np.asarray( rng1.uniform( low=-np.sqrt(6./(d+s_2)),high=np.sqrt(6./(d+s_2)),size=(s_2,d+1)),
dtype=theano.config.floatX)
Theta1 = theano.shared(value=Theta1_values, name="Theta",borrow=True)
#rng1.uniform( low=-np.sqrt(6./(d+s_2)),high=np.sqrt(6./(d+s_2)),size=(s_2,d+1))
z1 = T.dot( Theta1, a1)
a2 = T.tanh(z1)
passthru1 = theano.function( [], a2)
print(d)
passthru1()
print(X.shape)
X_i = theano.shared( X.astype("float32"))
#m = X_i.get_value().shape[1]
m = X.shape[1]
print(m)
a1_0 = theano.shared( np.ones((1,m)).astype("float32"),name='a1_0')
print(a1_0.get_value().shape)
a1 = T.stack( [a1_0,X_i], axis=0)
addintercept = theano.function([],a1)
addintercept()
d = X_i.get_value().shape[0]
print(d)
s_2 = d/2
print(s_2)
rng1 = np.random.RandomState(1234)
Theta1_values = np.asarray( rng1.uniform( low=-np.sqrt(6./(d+s_2)),high=np.sqrt(6./(d+s_2)),size=(s_2,d)),
dtype=theano.config.floatX)
Theta1 = theano.shared(value=Theta1_values, name="Theta1",borrow=True)
b_values = np.vstack( np.zeros(s_2) ).astype(theano.config.floatX)
b1 = theano.shared(value=b_values, name='b1',borrow=True)
a1_values=np.array( np.zeros( (d,m)), dtype=theano.config.floatX)
a1 = theano.shared(value=a1_values, name='a1', borrow=True)
lin_z2 = T.dot( Theta1, a1) + T.tile(b1,(1,m))
#lin_z2 = T.dot( Theta1, a1)
test_mult = theano.function([],lin_z2)
print( type(b_values))
b_values.dtype
test_mult()
print( b1.get_value().shape )
T.tile( b1, (0,m)) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
NN.py, load NN.py for Layer class for Neural Net for Multiple Layers | import sys
import os
#sys.path.append( os.getcwd() + '/ML')
sys.path.append( os.getcwd() + '/ML' )
from NN import Layer, cost_functional, cost_functional_noreg, gradientDescent_step
| theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Boilerplate sample data, from Coursera's Machine Learning Introduction | # Load Training Data
print("Loading and Visualizing Data ... \n")
ex4data1 = scipy.io.loadmat('./coursera_Ng/machine-learning-ex4/ex4/ex4data1.mat')
ex4data1.keys()
print( ex4data1['X'].shape )
print( ex4data1['y'].shape )
test_rng = np.random.RandomState(1234)
#Theta1 = Layer( test_rng, 1, 400,25, 5000)
#help(Theta1.al.set_value); # Beginning with Theano 0.3.1, set_value will work in-place on the GPU, if ... source on CPU
Theta1.al.set_value( ex4data1['X'].T.astype(theano.config.floatX))
Theta1.alp1
print( type( Theta1.alp1 ) )
Theta2 = Layer( test_rng, 2, 25,10,5000, al=Theta1.alp1 )
Theta2.alp1
predicted = theano.function([],sandbox.cuda.basic_ops.gpu_from_host( Theta2.alp1 ) )
predicted().shape
print( ex4data1['y'].shape )
pd.DataFrame( ex4data1['y']).describe()
# recall that whereas the original labels (in the variable y) were 1, 2, ..., 10, for the purpose of training a
# neural network, we need to recode the labels as vectors containing only values 0 or 1
K=10
m = ex4data1['y'].shape[0]
y_prob = [np.zeros(K) for row in ex4data1['y']] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_prob[i][ ex4data1['y'][i]-1] = 1
y_prob = np.array(y_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print(y_prob.shape)
print( type(y_prob) )
type( np.asarray( y_prob, dtype=theano.config.floatX) )
help( T.nlinalg.trace )
y_sh_var = theano.shared( np.asarray( y_prob,dtype=theano.config.floatX),name='y')
h_test = Theta2.alp1
J = sandbox.cuda.basic_ops.gpu_from_host(
(-T.nlinalg.trace( T.dot( T.log( h_test ), y_sh_var.T)) - T.nlinalg.trace(
T.dot( T.log( np.float32(1.)-h_test),(np.float32(1.)- y_sh_var.T ) )))/np.float32(m)
)
print(type(J))
test_cost_func = theano.function([],J)
test_cost_func()
J_test_build = sandbox.cuda.basic_ops.gpu_from_host( -T.nlinalg.trace( T.dot( T.log(h_test),y_sh_var.T) ) )
test_cost_build_func = theano.function([], J_test_build)
test_cost_build_func() | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Sanity check using ex4.m, Exercise 4 or Programming Exercise 4 from Coursera's Machine Learning Introduction by Ng | Theta_testvals = scipy.io.loadmat('./coursera_Ng/machine-learning-ex4/ex4/ex4weights.mat')
print( Theta_testvals.keys() )
print( Theta_testvals['Theta1'].shape )
print( Theta_testvals['Theta2'].shape )
Theta1_testval = Theta_testvals['Theta1'][:,1:]
b1_testval = Theta_testvals['Theta1'][:,0:1]
print( Theta1_testval.shape )
print( b1_testval.shape )
Theta2_testval = Theta_testvals['Theta2'][:,1:]
b2_testval = Theta_testvals['Theta2'][:,0:1]
print( Theta2_testval.shape )
print( b2_testval.shape )
Theta1 = Layer( test_rng, 1, 400,25, 5000, activation=T.nnet.sigmoid)
Theta1.Theta.set_value( Theta1_testval.astype("float32"))
Theta1.b.set_value( b1_testval.astype('float32') )
Theta1.al.set_value( ex4data1['X'].T.astype('float32')) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
For $\Theta^{(2)}$, the key to connecting $\Theta^{(2)}$ with $\Theta^{(1)}$ is to set the argument in class Layer with al=Theta1.alp1, | Theta2 = Layer( test_rng, 2, 25,10,5000, al=Theta1.alp1 , activation=T.nnet.sigmoid)
Theta2.Theta.set_value( Theta2_testval.astype('float32'))
Theta2.b.set_value( b2_testval.astype('float32'))
h_test = Theta2.alp1
J = sandbox.cuda.basic_ops.gpu_from_host(
T.mean( T.sum(
- y_sh_var * T.log( h_test ) - ( np.float32( 1) - y_sh_var) * T.log( np.float32(1) - h_test), axis =0), axis=0)
)
#J = sandbox.cuda.basic_ops.gpu_from_host(
# T.log(h_test) * y_sh_var
# )
test_cost_func = theano.function([],J)
test_cost_func()
print(type( y_sh_var) )
print( y_sh_var.get_value().shape )
print( type( h_test ))
checklayer2 = theano.function([], sandbox.cuda.basic_ops.gpu_from_host(Theta1.alp1))
checklayer2()
testreg = theano.function([], T.sum( Theta1.Theta * Theta1.Theta ) )
testreg()
range(1,3)
Thetas_lst = [ Theta1.Theta, Theta2.Theta ]
T.sum( [ T.sum( theta*theta) for theta in Thetas_lst] )
cost_func_test = cost_functional(3, 1, y_prob, Theta2.alp1, [Theta1.Theta, Theta2.Theta])
cost_test = theano.function([], cost_func_test)
cost_test() # (this value should be about 0.383770)
grad_test = T.grad( cost_func_test,[Theta1.Theta, Theta2.Theta])
grad_test_test = theano.function([], grad_test)
print( type(grad_test_test() ) )
print( len( grad_test_test() ))
print( type(grad_test_test()[0] ))
print( grad_test_test()[0].shape )
print( grad_test_test()[1].shape )
print( range(6))
print( list( "Ernest") )
zip( range(6), list("Ernest"))
print( type(grad_test))
print( grad_test_test.maker.fgraph.toposort() )
0.01 * grad_test
test_update = [(Theta,sandbox.cuda.basic_ops.gpu_from_host( Theta - np.float32(0.01)*T.grad(cost_func_test, Theta)+0.0001*Theta ) ) for Theta in [Theta1.Theta, Theta2.Theta] ]
test_gradDes_step = theano.function( inputs=[], updates= test_update )
test_gradDes_step()
print( Theta1.Theta.get_value() )
print( Theta2.Theta.get_value() )
test_gradDes_step()
print( Theta1.Theta.get_value() )
print( Theta2.Theta.get_value() )
gradDes_test_res = gradientDescent_step(cost_func_test, [Theta1.Theta, Theta2.Theta], 0.01, 0.00001 )
print( type(gradDes_test_res) )
gradDes_step_test = gradDes_test_res[1]
gradDes_step_test()
print( Theta1.Theta.get_value() )
print( Theta2.Theta.get_value() )
gradDes_step_test()
print( Theta1.Theta.get_value() )
print( Theta2.Theta.get_value() )
y_prob.shape
ex4data1['y'].shape
pd.DataFrame( ex4data1['y']).describe()
print( Theta2.alp1.shape )
print( Theta2.alp1.shape.ndim )
# Theta2.alp1.shape.get_scalar_constant_value()
predicted_logreg = theano.function([],Theta2.alp1)
pd.DataFrame( predicted_logreg().T ).describe()
pd.DataFrame(predicted_logreg().T).describe().iloc[1:-1,:].plot()
print( np.argmax( predicted_logreg(), axis=0).shape )
np.vstack( np.argmax( predicted_logreg(),axis=0) ).shape
pd.DataFrame( np.vstack( np.argmax(predicted_logreg(),axis=0)) + 1).describe()
res = np.float32( ( np.vstack( np.argmax( predicted_logreg(),axis=0)) + 1 ) == ex4data1['y'] )
pd.DataFrame(res).describe()
range(1,3)
predicted_logreg().shape
print(y_prob.shape); print( np.argmax( y_prob,axis=0 ).shape) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Summary for Neural Net with Multiple Layers for logistic regression (but can be extended to linear regression)
Load boilerplate training data: | sys.path.append( os.getcwd() + '/ML' )
from NN import Layer, cost_functional, cost_functional_noreg, gradientDescent_step, MLP
# Load Training Data
print("Loading and Visualizing Data ... \n")
ex4data1 = scipy.io.loadmat('./coursera_Ng/machine-learning-ex4/ex4/ex4data1.mat')
# recall that whereas the original labels (in the variable y) were 1, 2, ..., 10, for the purpose of training a
# neural network, we need to recode the labels as vectors containing only values 0 or 1
K=10
m = ex4data1['y'].shape[0]
y_prob = [np.zeros(K) for row in ex4data1['y']] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_prob[i][ ex4data1['y'][i]-1] = 1
y_prob = np.array(y_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print(ex4data1['X'].T.shape)
print(y_prob.shape)
digitsMLP = MLP(3,[400,25,10], 5000, ex4data1['X'].T, y_prob, T.nnet.sigmoid, 1., 0.1, 0.0000)
digitsMLP.train_model(100000)
digitsMLP.accuracy_log_reg()
print( digitsMLP.Thetas[0].Theta.get_value() )
digitsMLP.Thetas[1].Theta.get_value()
digitsMLP.predicted_vals_logreg()
testL1a2 = theano.function([], digitsMLP.Thetas[0].alp1 )
print( testL1a2() )
testL2a2 = theano.function([], digitsMLP.Thetas[1].al )
print( testL2a2() )
[1,2,3,4,5] + [8,1,5]
print( digitsMLP.y.shape )
y_cls_test = np.vstack( np.argmax( digitsMLP.y, axis=0) )
print( y_cls_test.shape )
pd.DataFrame( y_cls_test ).describe()
pred_y_cls_test = np.vstack( np.argmax( digitsMLP.predicted_vals_logreg() , axis=0))
print( pred_y_cls_test.shape )
pd.DataFrame( pred_y_cls_test ).describe()
np.mean( pred_y_cls_test == y_cls_test ) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Testing on MNIST, from University of Montreal, Deep Learning Tutorial, data | K=10
m = len(train_set[1])
y_train_prob = [np.zeros(K) for row in train_set[1]] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_train_prob[i][ train_set[1][i]] = 1
y_train_prob = np.array(y_train_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print( y_train_prob.shape )
print( pd.DataFrame( y_train_prob).describe() )
m,d= train_set[0].shape
MNIST_MTL = MLP(3,[d,25,10], m, train_set[0].T, y_train_prob, T.nnet.sigmoid, 1., 0.1, 0.00001)
MNIST_MTL.accuracy_log_reg()
print( MNIST_MTL.Thetas[0].Theta.get_value() )
MNIST_MTL.Thetas[1].Theta.get_value()
MNIST_MTL.predicted_vals_logreg()
MNIST_MTL.train_model(100000)
MNIST_MTL.accuracy_log_reg()
print( MNIST_MTL.Thetas[0].Theta.get_value() )
MNIST_MTL.Thetas[1].Theta.get_value()
MNIST_MTL.predicted_vals_logreg() | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Save the mode; cf. Getting Started, DeepLearning 0.1 documentation, Loading and Saving Models | import cPickle
save_file = open('./saved_models/MNIST_MTL_log_reg','wb')
for Thet in MNIST_MTL.Thetas:
cPickle.dump( Thet.Theta.get_value(borrow=True), save_file,-1) # the -1 is for HIGHEST priority
cPickle.dump( Thet.b.get_value(borrow=True), save_file,-1)
save_file.close()
MNIST_MTL.Thetas[0].al.set_value( valid_set[0].T.astype(theano.config.floatX) )
K=10
m = len(valid_set[1])
y_valid_prob = [np.zeros(K) for row in valid_set[1]] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_valid_prob[i][ valid_set[1][i]] = 1
y_valid_prob = np.array(y_valid_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print( y_valid_prob.shape )
MNIST_MTL.y = y_valid_prob
MNIST_MTL.predicted_vals_logreg()
theano.function([], MNIST_MTL.Thetas[0].alp1)()
Layer1 = MNIST_MTL.Thetas[0]
Layer2 = MNIST_MTL.Thetas[1]
m = valid_set[0].shape[0]
print(m)
a2 = T.nnet.sigmoid( T.dot( Layer1.Theta, Layer1.al) + T.tile( Layer1.b, (1,m)) )
a3 = T.nnet.sigmoid( T.dot( Layer2.Theta, a2) + T.tile( Layer2.b, (1,m)) )
valid_pred = theano.function([], a3)()
print( valid_pred.shape)
pd.DataFrame( valid_pred.T).describe()
np.mean( np.vstack( np.argmax( valid_pred,axis=0)) == np.vstack( valid_set[1] ) )
X_in = T.matrix()
X_in.set_value( valid_set[0].T.astype(theano.config.floatX))
a2_giv = T.nnet.sigmoid( T.dot( Layer1.Theta, X_in) + T.tile(Layer1.b, (1,m)))
a3_giv = T.nnet.sigmoid( T.dot( Layer2.Theta, a2_giv) + T.tile( Layer2.b, (1,m)) )
valid_pred_givens = theano.function([], outputs=a3_giv, givens={ X_in: valid_set[0].T.astype(theano.config.floatX)} )
print( valid_pred_givens().shape )
pd.DataFrame( valid_pred_givens().T).describe()
np.mean( np.vstack( np.argmax( valid_pred_givens(),axis=0)) == np.vstack( valid_set[1] ) )
test_pred_givens = theano.function([], outputs=a3_giv, givens={ X_in: test_set[0].T.astype(theano.config.floatX)} )
np.mean( np.vstack( np.argmax( test_pred_givens(),axis=0)) == np.vstack( test_set[1] ) )
range(1,3)
range(3)
range(1,3-1) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
cf. Glass Classification | gls_data = pd.read_csv( "./kaggle/glass.csv")
gls_data.describe()
gls_data.get_values().shape
X_gls = gls_data.get_values()[:,:-1]
print(X_gls.shape)
y_gls = gls_data.get_values()[:,-1]
print(y_gls.shape)
print( y_gls[:10])
X_gls_train = gls_data.get_values()[:-14,:-1]
print(X_gls_train.shape)
y_gls_train = gls_data.get_values()[:-14,-1]
print(y_gls_train.shape)
K=7
m = len(y_gls_train)
y_gls_train_prob = [np.zeros(K) for row in y_gls_train] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_gls_train_prob[i][ y_gls_train[i]-1] = 1
y_gls_train_prob = np.array(y_gls_train_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print( y_gls_train_prob.shape )
gls_MLP = MLP( 3, [9,8,7],200, X_gls_train.T, y_gls_train_prob, T.nnet.sigmoid, 0.01,0.05,0.0001 )
gls_MLP.accuracy_log_reg()
gls_MLP.train_model(10000)
gls_MLP.accuracy_log_reg()
gls_MLP.predicted_vals_logreg()
gls_MLP.train_model(10000)
gls_MLP.accuracy_log_reg()
ga
X_gls_test = gls_data.get_values()[-14:,:-1]
print( X_gls_test.shape )
y_gls_test = gls_data.get_values()[-14:,-1]
print( y_gls_test.shape)
gls_predict_on_test = gls_MLP.predict_on( 14, X_gls_test.T )
np.mean( np.vstack( np.argmax( gls_predict_on_test(), axis=0) ) == (y_gls_test-1) )
y_gls_test
np.vstack( np.argmax( gls_predict_on_test(), axis=0))
X_sym = T.matrix()
rng = np.random.RandomState(1234)
Thetab1 = Layer( rng, 1, 4,3,2, al = X_sym, activation=T.nnet.sigmoid)
Thetab1.alp1
Thetab1.Theta.get_value().shape
Thetab2 = Layer( rng, 2, 3,2,2, al=Thetab1.alp1, activation=T.nnet.sigmoid)
Thetab2.al = Thetab1.alp1
X_sym.shape[0]
T.tile( Thetab1.b, (1, X_sym.shape[0]))
test12comp = theano.function( [], outputs=Thetab2.alp1, givens={ X_sym : X42test} )
X42test = np.array([1,2,3,4,5,6,7,8]).reshape((4,2)).astype(theano.config.floatX)
test12comp()
X43test = np.array(range(1,13)).reshape((4,3)).astype(theano.config.floatX)
X43test
test43comp = theano.function( [], outputs=Thetab2.alp1, givens={ X_sym : X43test} )
test43comp()
print( type(Thetab1.al ))
lin_zlp1 = T.dot(Thetab1.Theta, Thetab1.al)+T.tile( Thetab1.b, (1,Thetab1.al.shape[1]))
a1p1 = Thetab1.g( lin_zlp1 )
Thetab1.al = X_sym
Thetab2.al = a1p1
lin_z2p1 = T.dot(Thetab2.Theta, Thetab2.al)+T.tile( Thetab2.b, (1, Thetab2.al.shape[1]))
a2p1 = Thetab2.g( lin_z2p1 )
test_gen_conn = theano.function([], outputs=a2p1, givens={ Thetab1.al : X42test })
test_gen_conn()
test_gen_conn = theano.function([], outputs=a2p1, givens={ Thetab1.al : X43test })
test_gen_conn() | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
GPU test | test_gen_conn = theano.function([], outputs=sandbox.cuda.basic_ops.gpu_from_host(a2p1), givens={ Thetab1.al : X42test })
test_gen_conn()
test_gen_conn = theano.function([], outputs=sandbox.cuda.basic_ops.gpu_from_host(a2p1), givens={ Thetab1.al : X43test })
test_gen_conn() | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Summary for Neural Net with Multiple Layers for logistic regression (but can be extended to linear regression) | sys.path.append( os.getcwd() + '/ML' )
from NN import MLP
# Load Training Data
print("Loading and Visualizing Data ... \n")
ex4data1 = scipy.io.loadmat('./coursera_Ng/machine-learning-ex4/ex4/ex4data1.mat')
# recall that whereas the original labels (in the variable y) were 1, 2, ..., 10, for the purpose of training a
# neural network, we need to recode the labels as vectors containing only values 0 or 1
K=10
m = ex4data1['y'].shape[0]
y_prob = [np.zeros(K) for row in ex4data1['y']] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_prob[i][ ex4data1['y'][i]-1] = 1
y_prob = np.array(y_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print(ex4data1['X'].T.shape)
print(y_prob.shape)
digitsMLP = MLP( 3, [400,25,10], ex4data1['X'].T, y_prob, T.nnet.sigmoid, 1.)
digitsMLP.build_update(ex4data1['X'].T, y_prob, 0.01, 0.00001)
digitsMLP.predicted_vals_logreg()
digitsMLP.accuracy_logreg( ex4data1['X'].T, y_prob)
digitsMLP.train_model(10000)
digitsMLP.accuracy_logreg( ex4data1['X'].T, y_prob)
digitsMLP.train_model(50000)
digitsMLP.accuracy_logreg( ex4data1['X'].T, y_prob) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Testing on University of Montreal LISA lab MNIST data | import gzip
import six.moves.cPickle as pickle
with gzip.open("../DeepLearningTutorials/data/mnist.pkl.gz", 'rb') as f:
try:
train_set, valid_set, test_set = pickle.load(f, encoding='latin1')
except:
train_set, valid_set, test_set = pickle.load(f)
K=10
m = len(train_set[1])
y_train_prob = [np.zeros(K) for row in train_set[1]] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_train_prob[i][ train_set[1][i]] = 1
y_train_prob = np.array(y_train_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print( y_train_prob.shape )
MNIST_MLP = MLP( 3,[784,49,10], train_set[0].T, y_train_prob, T.nnet.sigmoid, 1.)
MNIST_MLP.build_update( train_set[0].T, y_train_prob, 0.01, 0.0001)
MNIST_MLP.accuracy_logreg( train_set[0].T, y_train_prob)
MNIST_MLP.train_model(50000)
MNIST_MLP.accuracy_logreg( train_set[0].T, y_train_prob)
%time MNIST_MLP.train_model(100000)
MNIST_MLP.accuracy_logreg( train_set[0].T,y_train_prob)
m = len(valid_set[1])
y_valid_prob = [np.zeros(K) for row in valid_set[1]] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_valid_prob[i][ valid_set[1][i]] = 1
y_valid_prob = np.array(y_valid_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print( y_valid_prob.shape )
m = len(test_set[1])
y_test_prob = [np.zeros(K) for row in test_set[1]] # list of 5000 numpy arrays of size dims. (10,)
for i in range( m):
y_test_prob[i][ test_set[1][i]] = 1
y_test_prob = np.array(y_test_prob).T.astype(theano.config.floatX) # size dims. (K,m)
print( y_test_prob.shape )
MNIST_MLP.accuracy_logreg( valid_set[0].T,y_valid_prob)
MNIST_MLP.accuracy_logreg( test_set[0].T,y_test_prob)
MNIST_d = train_set[0].T.shape[0]
print(MNIST_d)
MNIST_MLP = MLP( 3,[MNIST_d,25,10], train_set[0].T, y_train_prob, T.nnet.sigmoid, 1.)
MNIST_MLP.build_update( train_set[0].T, y_train_prob, 0.1, 0.00001)
MNIST_MLP.accuracy_logreg( train_set[0].T, y_train_prob)
MNIST_MLP.train_model(150000)
MNIST_MLP.accuracy_logreg( train_set[0].T, y_train_prob)
MNIST_MLP.accuracy_logreg( valid_set[0].T, y_valid_prob)
MNIST_MLP.accuracy_logreg( test_set[0].T, y_test_prob) | theano_ML.ipynb | ernestyalumni/MLgrabbag | mit |
Reading in markers, Calculating decompressed length: We use the (very awesome) itertools module to do the iterating and filtering for us.
We use an iterator to go over the input values, so that we can use the itertools functions such as takewhile that selects characters as long as a condition is fulfilled.
def takewhile(condition, data):
filtered_data = []
for item in data:
if condition(data):
filtered_data.append(item)
else:
break
return filtered_data
We use takewhile to swallow characters until we reach the markers, and then to get the marker itself. Using regular expressions, we extract the two values from the marker. Since we are using iterators, we need to skip the next A characters, which we do using a for loop.
At the end, the answer is in the count variable. | from itertools import islice, takewhile
import re
numbers = re.compile(r'(\d+)')
def decompress(data_iterator):
'''parses markers and returns index of last character and length of decompressed data'''
count = 0
index = 0
while True:
# handle single tokens that decompress to length 1 until start of marker
count += len(list(takewhile(lambda character: character != '(', data_iterator)))
# extract marker
marker = ''.join(takewhile(lambda character: character != ')', data_iterator))
# extract A and B
try:
a, b = map(int, numbers.findall(marker))
except ValueError:
# EOF or no other markers present
break
# skip the next a characters
for i in range(a):
next(data_iterator)
# increment count
count += a * b
return count
print(decompress(iter(data))) | 2016/python3/Day09.ipynb | coolharsh55/advent-of-code | mit |
Part Two
Apparently, the file actually uses version two of the format.
In version two, the only difference is that markers within decompressed data are decompressed. This, the documentation explains, provides much more substantial compression capabilities, allowing many-gigabyte files to be stored in only a few kilobytes.
For example:
(3x3)XYZ still becomes XYZXYZXYZ, as the decompressed section contains no markers.
X(8x2)(3x3)ABCY becomes XABCABCABCABCABCABCY, because the decompressed data from the (8x2) marker is then further decompressed, thus triggering the (3x3) marker twice for a total of six ABC sequences.
(27x12)(20x12)(13x14)(7x10)(1x12)A decompresses into a string of A repeated 241920 times.
(25x3)(3x3)ABC(2x3)XY(5x2)PQRSTX(18x9)(3x2)TWO(5x7)SEVEN becomes 445 characters long.
Unfortunately, the computer you brought probably doesn't have enough memory to actually decompress the file; you'll have to come up with another way to get its decompressed length.
What is the decompressed length of the file using this improved format?
Solution logic
In this part, we need to keep track of the markers within the skipped marked from part one. As an assumption, we take the approach that no internal marker will extend the limits of the external marker. If it does, we will need to take a different approach to scan the string over and over again. Instead, we use a recursive approach to parse the string by marker and return the correct length.
X(8x2)(3x3)ABCY
(8x2) --> 8 characters: (3x3)ABC multiplied by 2
--> 2 x decompressed (3x3)ABC
--> 2 x 3 x ABC
For this, we extend the decompress function so that it will return the length of the string plus recursively scan any marker within it and return the final index of the last character scanned. This is the same function as in part one, except that it recusively checks for markers.
The recursive part of this approach is to further decompress the string (or characters) that were skipped in the first part. For this, we use islice to extract part of the string specified by the markers and recursively call the function on it to get its decompressed length. | def decompress(data_iterator):
count = 0
'''parses markers and returns index of last character and length of decompressed data'''
while(True):
# handle all single characters
count += len(list(takewhile(lambda character: character != '(', data_iterator)))
# marker occurs here, extract marker
marker = ''.join(takewhile(lambda character: character != ')', data_iterator))
# extract A and B
try:
a, b = map(int, numbers.findall(marker))
except ValueError:
break
count += b * decompress(islice(data_iterator, a))
return count
print(decompress(iter(data))) | 2016/python3/Day09.ipynb | coolharsh55/advent-of-code | mit |