codeparrot/github-jupyter-text-code-pairs · Datasets at Hugging Face

markdown stringlengths 0 37k	code stringlengths 1 33.3k	path stringlengths 8 215	repo_name stringlengths 6 77	license stringclasses 15 values
그리고 각 노드에 대해 생성된 dict들을 저장하자.	for user in users: user["shortest_paths"] = shortest_paths_from(user)	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
그러면 이제 매개 중심성을 구할 준비가 다 되었다. 이제 각각의 최단 경로에 포함되는 각 노드의 매개 중심성에 $1/n$을 더해 주자.	for user in users: user["betweenness_centrality"] = 0.0 for source in users: source_id = source["id"] for target_id, paths in source["shortest_paths"].items(): # python2에서는 items 대신 iteritems 사용 if source_id < target_id: # 잘못해서 두 번 세지 않도록 주의하자 num_paths = len(paths) # 최단 경로가 몇 개 존재하는가? contrib = 1 / num_paths # 중심성에 기여하는 값 for path in paths: for id in path: if id not in [source_id, target_id]: users[id]["betweenness_centrality"] += contrib for user in users: print(user["id"], user["betweenness_centrality"])	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
사용자 0과 9의 최단 경로 사이에는 다른 사용자가 없으므로 매개 중심성이 0이다. 반면 사용자 3, 4, 5는 최단 경로상에 무척 빈번하게 위치하기 때문에 높은 매개 중심성을 가진다. 대게 중심성의 절댓값 자체는 큰 의미를 가지지 않고, 상대값만이 의미를 가진다. 그 외에 살펴볼 수 있는 중심성 지표 중 하나는 근접 중심성(closeness centrality)이다. 먼저 각 사용자의 원접성(farness)을 계산한다. 원접성이란 from_user와 다른 모든 사용자의 최단 경로를 합한 값이다.	# # closeness centrality # def farness(user): """모든 사용자와의 최단 거리 합""" return sum(len(paths[0]) for paths in user["shortest_paths"].values())	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
이제 근접 중심성은 간단히 계산할 수 있다.	for user in users: user["closeness_centrality"] = 1 / farness(user) for user in users: print(user["id"], user["closeness_centrality"])	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
계산된 근접 중심성의 편차는 더욱 작다. 네트워크 중심에 있는 노드조차 외곽에 위치한 노드들로부터 멀리 떨어져 있기 때문이다. 여기서 봤듯이 최단 경로를 계산하는 것은 꽤나 복잡하다. 그렇기 때문에 큰 네트워크에서는 근접 중심성을 자주 사용하지 않는다. 덜 직관적이지만 보통 더 쉽게 계산할 수 있는 고유벡터 중심성(eigenvector centrality)을 더 자주 사용한다. 21.2 고유벡터 중심성 고유벡터 중심성에 대해 알아보기 전에 먼저 고유벡터가 무엇인지 살펴봐야 하고, 고유벡터가 무엇인지 알기 위해서는 먼저 행렬 연산에 대해 알아봐야 한다. 21.2.1 행렬 연산	def matrix_product_entry(A, B, i, j): return dot(get_row(A, i), get_column(B, j)) def matrix_multiply(A, B): n1, k1 = shape(A) n2, k2 = shape(B) if k1 != n2: raise ArithmeticError("incompatible shapes!") return make_matrix(n1, k2, partial(matrix_product_entry, A, B)) def vector_as_matrix(v): """(list 형태의) 벡터 v를 n x 1 행렬로 변환""" return [[v_i] for v_i in v] def vector_from_matrix(v_as_matrix): """n x 1 행렬을 리스트로 변환""" return [row[0] for row in v_as_matrix] def matrix_operate(A, v): v_as_matrix = vector_as_matrix(v) product = matrix_multiply(A, v_as_matrix) return vector_from_matrix(product)	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
행렬 A의 고유 벡터를 찾기 위해, 임의의 벡터 $v$를 골라 matrix_operate를 수행하고, 결과값의 크기가 1이 되게 재조정하는 과정을 반복 수행한다.	def find_eigenvector(A, tolerance=0.00001): guess = [1 for __ in A] while True: result = matrix_operate(A, guess) length = magnitude(result) next_guess = scalar_multiply(1/length, result) if distance(guess, next_guess) < tolerance: return next_guess, length # eigenvector, eigenvalue guess = next_guess	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
결과값으로 반환되는 guess를 matrix_operate를 통해 결과값의 크기가 1인 벡터로 재조정하면, 자기 자신이 반환된다. 즉, 여기서 guess는 고유벡터라는 것을 의미한다. 모든 실수 행렬에 고유벡터와 고유값이 있는 것은 아니다. 예를 들어 시계 방향으로 90도 회전하는 연산을 하는 다음 행렬에는 곱했을 때 가지 자신이 되는 벡터는 영벡터밖에 없다.	rotate = [[0, 1], [-1, 0]]	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
이 행렬로 앞서 구현한 find_eignevector(rotate)를 수행하면, 영원히 끝나지 않을 것이다. 한편, 고유벡터가 있는 행렬도 때로는 무한루프에 빠질 수 있다.	flip = [[0, 1], [1, 0]]	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
이 행렬은 모든 벡터 [x, y]를 [y, x]로 변환한다. 따라서 [1, 1]은 고유값이 1인 고유벡터가 된다. 하지만 x, y값이 다른 임의의 벡터에서 출발해서 find_eigenvector를 수행하면 x, y값을 바꾸는 연산만 무한히 수행할 것이다. (NumPy같은 라이브러리에는 이런 케이스까지 다룰 수 있는 다양한 방법들이 구현되어 있다.) 이런 사소한 문제에도 불구하고, 어쨌든 find_eigenvector가 결과값을 반환한다면, 그 결과값은 곧 고유벡터이다. 21.2.2 중심성 고유벡터가 데이터 네트워크를 이해하는데 어떻게 도움을 줄까? 얘기를 하기 전에 먼저 네트워크를 인접행렬(adjacency matrix)의 형태로 나타내 보자. 이 행렬은 사용자 i와 사용자 j가 친구인 경우 (i, j)번째 항목에 1이 있고, 친구가 아닌 경우 0이 있는 행렬이다.	# # eigenvector centrality # def entry_fn(i, j): return 1 if (i, j) in friendships or (j, i) in friendships else 0 n = len(users) adjacency_matrix = make_matrix(n, n, entry_fn) adjacency_matrix	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
각 사용자의 고유벡터 중심성이란 find_eigenvector로 찾은 사용자의 고유벡터가 된다.	eigenvector_centralities, _ = find_eigenvector(adjacency_matrix) for user_id, centrality in enumerate(eigenvector_centralities): print(user_id, centrality)	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
연결의 수가 많고, 중심성이 높은 사용자들한테 연결된 사용자들은 고유벡터 중심성이 높다. 앞의 결과에 따르면 사용자 1, 사용자 2의 중심성이 가장 높은데, 이는 중심성이 높은 사람들과 세번이나 연결되었기 때문이다. 이들로부터 멀어질수록 사용자들의 중심성은 점차 줄어든다. 21.3 방향성 그래프(Directed graphs)와 페이지랭크 데이텀이 인기를 별로 끌지 못하자, 순이익 팀의 부사장은 친구 모델에서 보증(endorsement)모델로 전향하는 것을 고려 중이다. 알고 보니 사람들은 어떤 데이터 과학자들끼리 친구인지에 대해서는 별로 관심이 없었지만, 헤드헌터들은 다른 데이터 과학자로부터 존경 받는 데이터 과학자가 누구인지에 대해 관심이 많다. 이 새로운 모델에서 관계는 상호적인 것이 아니라, 한 사람(source)이 다른 멋진 한 사람(target)의 실력에 보증을 서주는 (source, target) 쌍으로 비대칭적인 관계를 표현하게 된다.	# # directed graphs # endorsements = [(0, 1), (1, 0), (0, 2), (2, 0), (1, 2), (2, 1), (1, 3), (2, 3), (3, 4), (5, 4), (5, 6), (7, 5), (6, 8), (8, 7), (8, 9)] for user in users: user["endorses"] = [] # add one list to track outgoing endorsements user["endorsed_by"] = [] # and another to track endorsements for source_id, target_id in endorsements: users[source_id]["endorses"].append(users[target_id]) users[target_id]["endorsed_by"].append(users[source_id])	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
그리고 가장 보증을 많이 받은 데이터 과학자들의 데이터를 수집해서, 그것을 헤드헌터들한테 팔면 된다.	endorsements_by_id = [(user["id"], len(user["endorsed_by"])) for user in users] sorted(endorsements_by_id, key=lambda x: x[1], # (user_id, num_endorsements) reverse=True)	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
사실 '보증의 수'와 같은 숫자는 조작하기가 매우 쉽다. 가장 간단한 방법 중 하나는, 가짜 계정을 여러 개 만들어서 그것들로 내 계정에 대한 보증을 서는 것이다. 또 다른 방법은, 친구들끼리 짜고 서로가 서로를 보증해 주는 것이다. (아마 사용자 0, 1, 2가 이런 관계일 가능성이 크다.) 좀 더 나은 지수는, '누가' 보증을 서는지를 고려하는 것이다. 보증을 많이 받은 사용자가 보증을 설 때는, 보증을 적게 받은 사용자가 보증을 설 때보다 더 중요한 것으로 받아들여지는 것이 타당하다. 그리고 사실 이것은 유명한 페이지랭크(PageRank) 알고리즘의 기본 철학이기도 하다. 1. 네트워크 전체에는 1.0(또는 100%)의 페이지랭크가 있다. 2. 초기에 이 페이지랭크를 모든 노드에 고르게 배당한다. 3. 각 스텝을 거칠 때마다 각 노드에 배당된 페이지랭크의 대부분은 외부로 향하는 링크에 균등하게 배당한다. 4. 각 스텝을 거칠 때마다 각 노드에 남아 있는 페이지랭크를 모든 노드에 고르게 배당한다.	def page_rank(users, damping = 0.85, num_iters = 100): # 먼저 페이지랭크를 모든 노드에 고르게 배당 num_users = len(users) pr = { user["id"] : 1 / num_users for user in users } # 매 스텝마다 각 노드가 받는 # 적은 양의 페이지랭크 base_pr = (1 - damping) / num_users for __ in range(num_iters): next_pr = { user["id"] : base_pr for user in users } for user in users: # 페이지랭크를 외부로 향하는 링크에 배당한다. links_pr = pr[user["id"]] * damping for endorsee in user["endorses"]: next_pr[endorsee["id"]] += links_pr / len(user["endorses"]) pr = next_pr return pr for user_id, pr in page_rank(users).items(): print(user_id, pr)	notebook/ch21_network_analysis.ipynb	rnder/data-science-from-scratch	unlicense
Generating some data Lets generate a classification dataset that is not easily linearly separable. Our favorite example is the spiral dataset, which can be generated as follows:	N = 100 #The number of "training examples" per class D = 2 # The dimension i.e i think the feature set K = 3 # The number of classes to classify into X = np.zeros((NK, D)) # The data matrix , each row is a single example # In our case 1003 amount of examples and each example has two dimensions # so the Matrix is 300 * 2 Y = np.zeros(NK , dtype = 'uint8') #Class Labels for j in range(K): ix = range (Nj,N(j+1)) #filling the matrices made above with datapoints r = np.linspace(0.0,1,N) # radius t = np.linspace(j4,(j+1)4,N) + np.random.randn(N) 0.2 #theta #numpy.linspace(start, stop, num = 50, endpoint = True, retstep = False, dtype = None) #: Returns number spaces evenly w.r.t interval. X[ix] = np.c_[rnp.sin(t),rnp.cos(t)] # the above line is written so as to give the dataset a spiral structure # which will be clear in the dataset plot below # i think this can be played with different kinds of datasets and check how our neural network reacts # np.c_ slices objects to concatenation along the second axis. Y[ix] = j plt.scatter(X[:, 0], X[:, 1], c=Y, s=40, cmap = plt.cm.Spectral) plt.xlim([-1,1]) plt.ylim([-1,1]) plt.show()	Deep Learning.ai/Numpy - Understanding Backprop.ipynb	Shashi456/Artificial-Intelligence	gpl-3.0
Training a Simple Softmax Linear Classifer Initialize the Parameters	# initialize parameters randomly W = 0.01 * np.random.randn(D,K) #randn(D,K) D * K Matrix # in our case 2 * 3 Matrix b = np.zeros((1,K)) # in our case 1 * 3 # some hyperparameters step_size = 1e-0 reg = 1e-3 # regularization strength	Deep Learning.ai/Numpy - Understanding Backprop.ipynb	Shashi456/Artificial-Intelligence	gpl-3.0
Compute the Class Scores	# compute class scores for a linear classif scores = np.dot(X,W) + b #scores will be 300 *3 and remember numpy broadcasting for b	Deep Learning.ai/Numpy - Understanding Backprop.ipynb	Shashi456/Artificial-Intelligence	gpl-3.0
Compute the loss	# Using cross - entropy loss # Softmax loss = data loss + regularization loss # Study about the Softmax loss and how it works before seeing how it works #in this case it could be intuitive num_examples = X.shape[0] # get unnormalized probabilities exp_scores = np.exp(scores) # normalize them for each example '''We now have an array probs of size [300 x 3], where each row now contains the class probabilities. In particular, since we’ve normalized them every row now sums to one. We can now query for the log probabilities assigned to the correct classes in each example:''' probs = exp_scores / np.sum(exp_scores, axis=1 , keepdims=True) # Calculating the correct_log probability '''The array correct_logprobs is a 1D array of just the probabilities assigned to the correct classes for each example. The full loss is then the average of these log probabilities and the regularization loss:''' correct_logprobs = -np.log(probs[range(num_examples),y]) # compute ths loss : average cross-entropy loss and regularization data_loss = np.sum(correct_logprobs)/ num_examples reg_loss = 0.5regnp.sum(WW) '''∂Li∂fk=pk−1(yi=k) Notice how elegant and simple this expression is. Suppose the probabilities we computed were p = [0.2, 0.3, 0.5], and that the correct class was the middle one (with probability 0.3). According to this derivation the gradient on the scores would be df = [0.2, -0.7, 0.5]. Recalling what the interpretation of the gradient, we see that this result is highly intuitive: increasing the first or last element of the score vector f (the scores of the incorrect classes) leads to an increased loss (due to the positive signs +0.2 and +0.5) - and increasing the loss is bad, as expected. However, increasing the score of the correct class has negative influence on the loss. The gradient of -0.7 is telling us that increasing the correct class score would lead to a decrease of the loss Li, which makes sense.''' dscores = probs dscores[range(num_examples),y]-=1 dscores/=num_examples dW = np.dot(X.T,dscores) db = np.sum(dscores, axis=0, keepdims=True) dW += regW # Not forgetting the regularization gradient #Updating Parameters W += -step_size dW b += -step_size db	Deep Learning.ai/Numpy - Understanding Backprop.ipynb	Shashi456/Artificial-Intelligence	gpl-3.0
Putting it all Together Till now we just looked at individual elements	#Train a Linear Classifier # initialize parameters randomly W = 0.01 * np.random.randn(D,K) b = np.zeros((1,K)) # some hyperparameters step_size = 1e-0 reg = 1e-3 # regularization strength # gradient descent loop num_examples = X.shape[0] for i in range(200): # evaluate class scores, [N x K] scores = np.dot(X, W) + b # compute the class probabilities exp_scores = np.exp(scores) probs = exp_scores / np.sum(exp_scores, axis=1, keepdims=True) # [N x K] # compute the loss: average cross-entropy loss and regularization corect_logprobs = -np.log(probs[range(num_examples),Y]) data_loss = np.sum(corect_logprobs)/num_examples reg_loss = 0.5regnp.sum(WW) loss = data_loss + reg_loss if i % 10 == 0: print ("iteration %d: loss %f" % (i, loss)) # compute the gradient on scores dscores = probs dscores[range(num_examples),Y] -= 1 dscores /= num_examples # backpropate the gradient to the parameters (W,b) dW = np.dot(X.T, dscores) db = np.sum(dscores, axis=0, keepdims=True) dW += regW # regularization gradient # perform a parameter update W += -step_size * dW b += -step_size * db # evaluate training set Accuracy scores = np.dot(X,W) + b predicted_class = np.argmax(scores, axis=1) print("training_accuracy: %.2f" % (np.mean(predicted_class==Y))) # plot the resulting classifier h = 0.02 x_min, x_max = X[:, 0].min() - 1, X[:, 0].max() + 1 y_min, y_max = X[:, 1].min() - 1, X[:, 1].max() + 1 xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h)) Z = np.dot(np.c_[xx.ravel(), yy.ravel()], W) + b Z = np.argmax(Z, axis=1) Z = Z.reshape(xx.shape) plt.contourf(xx, yy, Z, cmap=plt.cm.Spectral, alpha=0.8) plt.scatter(X[:, 0], X[:, 1], c=Y, s=40, cmap=plt.cm.Spectral) plt.xlim(xx.min(), xx.max()) plt.ylim(yy.min(), yy.max()) plt.show()	Deep Learning.ai/Numpy - Understanding Backprop.ipynb	Shashi456/Artificial-Intelligence	gpl-3.0
Using SQL for Queries Note that SQL is case-insensitive, but it is traditional to use ALL CAPS for SQL keywords. It is also standard to end SQL statements with a semi-colon. Simple Queries	pdsql('SELECT * FROM tips LIMIT 5;') q = """ SELECT sex as gender, smoker, day, size FROM tips WHERE size <= 3 AND smoker = 'Yes' LIMIT 5""" pdsql(q)	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Filtering on strings	q = """ SELECT sex as gender, smoker, day, size , time FROM tips WHERE time LIKE '%u_c%' LIMIT 5""" pdsql(q)	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Ordering	q = """ SELECT sex as gender, smoker, day, size , time FROM tips WHERE time LIKE '%u_c%' ORDER BY size DESC LIMIT 5""" pdsql(q)	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Aggregate queries	pdsql('select * from tips limit 5;') q = """ SELECT sex, smoker, sum(total_bill) as total, max(tip) as max_tip FROM tips WHERE time = 'Lunch' GROUP BY sex, smoker ORDER BY max_tip DESC """ pdsql(q)	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Matching students and majors Inner join	pdsql(""" SELECT * from student s INNER JOIN major m ON s.major_id = m.major_id; """)	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Left outer join SQL also has RIGHT OUTER JOIN and FULL OUTER JOIN but these are not currently supported by SQLite3 (the database engine used by pdsql).	pdsql(""" SELECT s.*, m.name from student s LEFT JOIN major m ON s.major_id = m.major_id; """)	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Emulating a full outer join with UNION ALL Only necessary if the database does not proivde FULL OUTER JOIN Using linker tables to match students to classes (a MANY TO MANY join) Using SQLite3 SQLite3 is part of the standard library. However, the mechanics of using essentially any database in Python is similar, because of the Python DB-API.	import sqlite3 c = sqlite3.connect('data/Chinook_Sqlite.sqlite')	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Standard SQL statements with parameter substitution Note: Using Python string substitution for Python defined parameters is dangerous because of the risk of SQL injection attacks. Use parameter substitution with ? instead. Do this	t = ['%rock%', 10] list(c.execute("SELECT * FROM Album WHERE Title like ? AND ArtistID > ? LIMIT 5;", t))	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Not this	t = ("'%rock%'", 10) list(c.execute("SELECT * FROM Album WHERE Title like %s AND ArtistID > %d LIMIT 5;" % t))	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
User defined functions Sometimes it is useful to have custom functions that run on the database server rather than on the client. These are called User Defined Functions or UDF. How do to do this varies with the database used, but it is fairly simple with Python and SQLite. A standard UDF	def encode(text, offset): """Caesar cipher of text with given offset.""" from string import ascii_lowercase, ascii_uppercase tbl = dict(zip(map(ord, ascii_lowercase + ascii_uppercase), ascii_lowercase[offset:] + ascii_lowercase[:offset] + ascii_uppercase[offset:] + ascii_uppercase[:offset])) return text.translate(tbl) c.create_function("encode", 2, encode) list(c.execute("SELECT Title, encode(Title, 2) FROM Album limit 5;"))	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
Using SQL magic functions We will use the ipython-sql notebook extension for convenience. This will only work in notebooks and IPython scripts with the .ipy extension.	%load_ext sql	scratch/Lecture08.ipynb	cliburn/sta-663-2017	mit
3. Open the file and use .read() to save the contents of the file to a string called fields. Make sure the file is closed at the end.	# Write your code here: # Run fields to see the contents of contacts.txt: fields	nlp/UPDATED_NLP_COURSE/00-Python-Text-Basics/03-Python-Text-Basics-Assessment.ipynb	rishuatgithub/MLPy	apache-2.0
Working with PDF Files 4. Use PyPDF2 to open the file Business_Proposal.pdf. Extract the text of page 2.	# Perform import # Open the file as a binary object # Use PyPDF2 to read the text of the file # Get the text from page 2 (CHALLENGE: Do this in one step!) page_two_text = # Close the file # Print the contents of page_two_text print(page_two_text)	nlp/UPDATED_NLP_COURSE/00-Python-Text-Basics/03-Python-Text-Basics-Assessment.ipynb	rishuatgithub/MLPy	apache-2.0
5. Open the file contacts.txt in append mode. Add the text of page 2 from above to contacts.txt. CHALLENGE: See if you can remove the word "AUTHORS:"	# Simple Solution: # CHALLENGE Solution (re-run the %%writefile cell above to obtain an unmodified contacts.txt file):	nlp/UPDATED_NLP_COURSE/00-Python-Text-Basics/03-Python-Text-Basics-Assessment.ipynb	rishuatgithub/MLPy	apache-2.0
Regular Expressions 6. Using the page_two_text variable created above, extract any email addresses that were contained in the file Business_Proposal.pdf.	import re # Enter your regex pattern here. This may take several tries! pattern = re.findall(pattern, page_two_text)	nlp/UPDATED_NLP_COURSE/00-Python-Text-Basics/03-Python-Text-Basics-Assessment.ipynb	rishuatgithub/MLPy	apache-2.0
Exercise: Compute the velocity of the penny after t seconds. Check that the units of the result are correct.	# Solution goes here	notebooks/chap01.ipynb	AllenDowney/ModSimPy	mit
Exercise: Why would it be nonsensical to add a and t? What happens if you try?	# Solution goes here	notebooks/chap01.ipynb	AllenDowney/ModSimPy	mit
Exercise: Suppose you bring a 10 foot pole to the top of the Empire State Building and use it to drop the penny from h plus 10 feet. Define a variable named foot that contains the unit foot provided by UNITS. Define a variable named pole_height and give it the value 10 feet. What happens if you add h, which is in units of meters, to pole_height, which is in units of feet? What happens if you write the addition the other way around?	# Solution goes here # Solution goes here	notebooks/chap01.ipynb	AllenDowney/ModSimPy	mit
Exercise: In reality, air resistance limits the velocity of the penny. At about 18 m/s, the force of air resistance equals the force of gravity and the penny stops accelerating. As a simplification, let's assume that the acceleration of the penny is a until the penny reaches 18 m/s, and then 0 afterwards. What is the total time for the penny to fall 381 m? You can break this question into three parts: How long until the penny reaches 18 m/s with constant acceleration a. How far would the penny fall during that time? How long to fall the remaining distance with constant velocity 18 m/s? Suggestion: Assign each intermediate result to a variable with a meaningful name. And assign units to all quantities!	# Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here	notebooks/chap01.ipynb	AllenDowney/ModSimPy	mit
Filtering and resampling data Some artifacts are restricted to certain frequencies and can therefore be fixed by filtering. An artifact that typically affects only some frequencies is due to the power line. Power-line noise is a noise created by the electrical network. It is composed of sharp peaks at 50Hz (or 60Hz depending on your geographical location). Some peaks may also be present at the harmonic frequencies, i.e. the integer multiples of the power-line frequency, e.g. 100Hz, 150Hz, ... (or 120Hz, 180Hz, ...). This tutorial covers some basics of how to filter data in MNE-Python. For more in-depth information about filter design in general and in MNE-Python in particular, check out sphx_glr_auto_tutorials_plot_background_filtering.py.	import numpy as np import mne from mne.datasets import sample data_path = sample.data_path() raw_fname = data_path + '/MEG/sample/sample_audvis_raw.fif' proj_fname = data_path + '/MEG/sample/sample_audvis_eog_proj.fif' tmin, tmax = 0, 20 # use the first 20s of data # Setup for reading the raw data (save memory by cropping the raw data # before loading it) raw = mne.io.read_raw_fif(raw_fname) raw.crop(tmin, tmax).load_data() raw.info['bads'] = ['MEG 2443', 'EEG 053'] # bads + 2 more fmin, fmax = 2, 300 # look at frequencies between 2 and 300Hz n_fft = 2048 # the FFT size (n_fft). Ideally a power of 2 # Pick a subset of channels (here for speed reason) selection = mne.read_selection('Left-temporal') picks = mne.pick_types(raw.info, meg='mag', eeg=False, eog=False, stim=False, exclude='bads', selection=selection) # Let's first check out all channel types raw.plot_psd(area_mode='range', tmax=10.0, picks=picks, average=False)	0.17/_downloads/1a57f50035589ab531155bdb55bcbcb5/plot_artifacts_correction_filtering.ipynb	mne-tools/mne-tools.github.io	bsd-3-clause
The same care must be taken with these results as with partial derivatives. The formula for $Y$ is ostensibly $$X_1 + X_2 = X_1 + X^2 + X_1 = 2 X_1 + X^2$$ Or $2X_1$ plus a parabola. However, the coefficient of $X_1$ is 1. That is because $Y$ changes by 1 if we change $X_1$ by 1 <i>while holding $X_2$ constant</i>. Multiple linear regression separates out contributions from different variables. Similarly, running a linear regression on two securities might give a high $\beta$. However, if we bring in a third security (like SPY, which tracks the S&P 500) as an independent variable, we may find that the correlation between the first two securities is almost entirely due to them both being correlated with the S&P 500. This is useful because the S&P 500 may then be a more reliable predictor of both securities than they were of each other. This method allows us to better gauge the significance between the two securities and prevent confounding the two variables.	# Load pricing data for two arbitrarily-chosen assets and SPY start = '2014-01-01' end = '2015-01-01' asset1 = get_pricing('DTV', fields='price', start_date=start, end_date=end) asset2 = get_pricing('FISV', fields='price', start_date=start, end_date=end) benchmark = get_pricing('SPY', fields='price', start_date=start, end_date=end) # First, run a linear regression on the two assets slr = regression.linear_model.OLS(asset1, sm.add_constant(asset2)).fit() print 'SLR beta of asset2:', slr.params[1] # Run multiple linear regression using asset2 and SPY as independent variables mlr = regression.linear_model.OLS(asset1, sm.add_constant(np.column_stack((asset2, benchmark)))).fit() prediction = mlr.params[0] + mlr.params[1]asset2 + mlr.params[2]benchmark prediction.name = 'Prediction' print 'MLR beta of asset2:', mlr.params[1], '\nMLR beta of S&P 500:', mlr.params[2]	notebooks/lectures/Multiple_Linear_Regression/notebook.ipynb	quantopian/research_public	apache-2.0
The Laplacian Matrix is a matrix that is extremely important in graph theory and numerical analysis. It is defined as $L=D-A$. Where $D$ is the degree matrix and $A$ is the adjecency matrix. For the purpose of this problem you don't need to understand the details of these matrices, although their definitions are relatively simple. The degree matrix for $K_n$ is an $n \times n$ diagonal matrix with the value $n-1$ along the diagonal and zeros everywhere else. Write a function to compute the degree matrix for $K_n$ using NumPy.	def complete_deg(n): a = np.identity((n), dtype = np.int) for element in np.nditer(a, op_flags=['readwrite']): if element > 0: element[...] = element + n - 2 return a complete_deg(3) D = complete_deg(5) assert D.shape==(5,5) assert D.dtype==np.dtype(int) assert np.all(D.diagonal()==4*np.ones(5)) assert np.all(D-np.diag(D.diagonal())==np.zeros((5,5),dtype=int))	assignments/assignment03/NumpyEx04.ipynb	geoneill12/phys202-2015-work	mit
The adjacency matrix for $K_n$ is an $n \times n$ matrix with zeros along the diagonal and ones everywhere else. Write a function to compute the adjacency matrix for $K_n$ using NumPy.	def complete_adj(n): b = np.identity((n), dtype = np.int) for element in np.nditer(b, op_flags=['readwrite']): if element == 0: element[...] = 1 else: element[...] = 0 return b complete_adj(3) A = complete_adj(5) assert A.shape==(5,5) assert A.dtype==np.dtype(int) assert np.all(A+np.eye(5,dtype=int)==np.ones((5,5),dtype=int))	assignments/assignment03/NumpyEx04.ipynb	geoneill12/phys202-2015-work	mit
Use NumPy to explore the eigenvalues or spectrum of the Laplacian L of $K_n$. What patterns do you notice as $n$ changes? Create a conjecture about the general Laplace spectrum of $K_n$.	def L(n): L = complete_deg(n) - complete_adj(n) return L print L(1) print L(2) print L(3) print L(4) print L(5) print L(6)	assignments/assignment03/NumpyEx04.ipynb	geoneill12/phys202-2015-work	mit
在新三板挂牌的公司还包括了原两网的公司（NET和STAQ），两网的公司是在90年代有中国人民银行批准设立的，总用9家公司（粤传媒(400003)上市后剩余8家公司），由于政策方面的约束，此类公司即便在三板市场上依然不被允许进行新的融资而只能股份转让。这些公司包含了： * 大自然（400001），该股票已经在在新三板退市，并以834019为代码成为全国股转公司的挂牌企业。可能出现的问题是，数据库把400001公布的财务信息也作为了新挂牌公司的财务信息，因此，处理时直接以新挂牌的公司转让公告书中披露的财务信息作为检验标准（To-do: 作为对比直接踢除这家公司的数据。） * 长白5 (400002) * 海国实5(400005) * 京中兴5(400006) * 华凯1 (400007) * 广建1 (400009) * 鹫峰5 (400010) * 港岳1 (400013)	stkcs_net_staq = ["400001.OC", "400002.OC", "400005.OC", "400006.OC", "400007.OC", "400009.OC", "400010.OC", "400013.OC"] stocks_neeq = stocks_neeq[~stocks_neeq["证券代码"].isin(stkcs_net_staq)] len(stocks_neeq) stocks_neeq[stocks_neeq["证券代码"] == "834019.OC"] stocks_neeq = stocks_neeq.drop(6644) stocks_neeq[stocks_neeq["证券代码"] == "834019.OC"]	src/Stocks.ipynb	januslian/neeqem	gpl-3.0
退市股票处理：11个转板上市公司，1个连续四年亏损，16个吸收合并的公司，3家公司将公司变为有限责任公司，6家公司未说明退市原因。在数据库退市股票列表里有两家公司400001和400003属于两网公司转板，不作为研究样本处理。	delisting = pandas.read_excel("../data/raw/退市股票一览.xlsx") #2016年4月15日之前退市的公司 delisting = delisting[:-2] delisting.head() delisting = delisting[~delisting["代码"].isin(["400001.OC", "400003.OC"])] stkcd = stocks_neeq["证券代码"].append(delisting["代码"]) stkcd[stkcd.duplicated()] delisting = pandas.read_excel("../data/delist.xls") delisting.head() stocks_neeq_all = pandas.concat([stocks_neeq, delisting], ignore_index=True) len(stocks_neeq_all) len(stocks_neeq), len(delisting) stocks_neeq_all.iloc[0] stocks_neeq_all[stocks_neeq_all["所属证监会行业"].isnull()] stocks_neeq_all.loc[stocks_neeq_all["证券代码"] == "835458.OC", "所属证监会行业"] = "货币金融服务" stocks_neeq_all.loc[stocks_neeq_all["证券代码"] == "830806.OC", "所属证监会行业"] = "软件和信息技术服务业" stocks_neeq_all.loc[stocks_neeq_all["证券代码"] == "430217.OC", "所属证监会行业"] = "互联网和相关服务业" stocks_neeq_all[stocks_neeq_all["所属证监会行业"].isnull()] finance_sector = ["货币金融服务", "资本市场服务", "保险业", "其他金融业"] print(len(stocks_neeq_all[stocks_neeq_all["所属证监会行业"].isin(finance_sector)])) stocks_neeq_all = stocks_neeq_all[~stocks_neeq_all["所属证监会行业"].isin(finance_sector)] stocks_neeq_all.to_csv("../data/NEEQ_sample.csv", index=False) a = len(stocks_neeq_all[stocks_neeq_all["挂牌日期"] <= pandas.Timestamp('2009-10-30 00:00:00')])	src/Stocks.ipynb	januslian/neeqem	gpl-3.0
由此得到的“新三板”公司的有效样本数量为{{len(stocks_neeq_all)}}。 “新三板”样本公司最早的挂牌时间为{{stocks_neeq_all["挂牌日期"].min()}}，所以在选择与之对比的A股主板、中小板和创业板公司时也选择从2006年以后，其中创业板开始于2009年10月，在2006到2009年10月之间挂牌的新三板公司有{{a}}家公司。由于所有适合本研究的“新三板”公司均为2006年之后挂牌交易的，所以对应的主板、中小板和创业板公司也将会选择2005年之后IPOs的公司作为对比的对象。后续的检验也会将主中创三个板块的定向增发作为对比对象，以保证在发行过程上有些许的可比性。 To-do: 需要确定估计样本的来源，备选的方案包括了：（1）可以使用当年所有已经上市的公司作为估计样本；（2）可以使用当年所有上市已经超过一年的公司作为样本；（3）使用规模分层的估计方法；	stocks_neeq_all.iloc[0] year = lambda x: x.year stocks_neeq_all["挂牌年度"] = stocks_neeq_all["挂牌日期"].apply(year) stocks_neeq_all["挂牌年度"].value_counts().sort_index() sns.barplot(stocks_neeq_all["挂牌年度"].value_counts().sort_index().index, stocks_neeq_all["挂牌年度"].value_counts().sort_index(), palette="Blues_d"); sns.countplot(x="挂牌年度", data=stocks_neeq_all, palette="Blues_d"); stocks_neeq_all["所属证监会行业"].value_counts()	src/Stocks.ipynb	januslian/neeqem	gpl-3.0
2. Hypothesis tests	#California burritos vs. Carnitas burritos TODO # Don Carlos 1 vs. Don Carlos 2 TODO # Bonferroni correction TODO	burrito/.ipynb_checkpoints/Burrito_bootcamp-checkpoint.ipynb	srcole/qwm	mit
3. Burrito dimension distributions Distribution of each burrito quality	import math def metrichist(metricname): if metricname == 'Volume': bins = np.arange(.375,1.225,.05) xticks = np.arange(.4,1.2,.1) xlim = (.4,1.2) else: bins = np.arange(-.25,5.5,.5) xticks = np.arange(0,5.5,.5) xlim = (-.25,5.25) plt.figure(figsize=(5,5)) n, _, _ = plt.hist(df[metricname].dropna(),bins,color='k') plt.xlabel(metricname + ' rating',size=20) plt.xticks(xticks,size=15) plt.xlim(xlim) plt.ylabel('Count',size=20) plt.yticks((0,int(math.ceil(np.max(n) / 5.)) * 5),size=15) plt.tight_layout() m_Hist = ['Hunger','Volume','Tortilla','Temp','Meat','Fillings', 'Meat:filling','Uniformity','Salsa','Synergy','Wrap','overall'] for m in m_Hist: metrichist(m)	burrito/.ipynb_checkpoints/Burrito_bootcamp-checkpoint.ipynb	srcole/qwm	mit
Test for normal distribution	TODO	burrito/.ipynb_checkpoints/Burrito_bootcamp-checkpoint.ipynb	srcole/qwm	mit
Changes implemented From most important to least important: * pre-compute PSF parameters across multiple spectra and wavelengths to enable vectorized calls to legval instead of many scalar calls [~25% faster] * hoist wavelength -> [-1,1] for legendre calculations out of loop [~8% faster] * replace np.outer with numba optimized version [~5% faster (?)] * cast astropy.io.fits.Header to dict() before many many key/value lookups [~4% faster] * Avoid intermediate memory allocations by using out=... [<~1% faster] Note: amount of improvement was different for Haswell vs. KNL vs. Core i7 Laptop and varies with extraction problem size; above numbers are approximate; see below for before/after Haswell and KNL results. Before / after comparison	h1 = Table.read('data/portland/hsw-before.txt', format='ascii.basic') h2 = Table.read('data/portland/hsw-after.txt', format='ascii.basic') k1 = Table.read('data/portland/knl-before.txt', format='ascii.basic') k2 = Table.read('data/portland/knl-after.txt', format='ascii.basic') plot(h1['nproc'], h1['rate'], 's:', color='C0', label='Haswell original') plot(h2['nproc'], h2['rate'], 's-', color='C0', label='Haswell') plot(h2['nproc'], h2['nproc']h2['rate'][0], color='0.5', alpha=0.5, label='_none_') plot(k1['nproc'], k1['rate'], 'd:', color='C1', label='KNL original') plot(k2['nproc'], k2['rate'], 'd-', color='C1', label='KNL') plot(k2['nproc'], k2['nproc']k2['rate'][0], color='0.5', alpha=0.5, label='perfect scaling') semilogx() semilogy() xticks([1,2,4,8,16,32,64,128], [1,2,4,8,16,32,64,128]) legend(loc='upper left') ylim(20, 20000) xlabel('Number of Processes') ylabel('Extraction Rate') title('25 spectra x 300 wavelengths') savefig('desi-extract-improvements.pdf')	doc/portland-extract.ipynb	sbailey/knltest	bsd-3-clause
Build in plots for storytelling	def labelit(title_=''): semilogx() semilogy() xticks([1,2,4,8,16,32,64,128], [1,2,4,8,16,32,64,128]) legend(loc='upper left') title(title_) ylim(20, 20000) xlabel('Number of Processes') ylabel('Extraction Rate') figure() plot(h1['nproc'], h1['rate'], 's-', color='C0', label='Haswell') plot(h1['nproc'], h1['nproc']h1['rate'][0], color='0.5', alpha=0.5, label='_none_') plot(k1['nproc'], k1['rate'][0]np.ones(len(k1['nproc'])), 'd-', color='C1', label='KNL') plot(k1['nproc'], k1['nproc']k1['rate'][0], color='0.5', alpha=0.5, label='_none_') labelit('Starting point: no KNL scaling') savefig('portland1.pdf') figure() plot(h1['nproc'], h1['rate'], 's-', color='C0', label='Haswell') plot(h1['nproc'], h1['nproc']h1['rate'][0], color='0.5', alpha=0.5, label='_none_') plot(k1['nproc'], k1['rate'], 'd-', color='C1', label='KNL') plot(k1['nproc'], k1['rate'][0]np.ones(len(k1['nproc'])), ':', color='C1', label='_none_', alpha=0.8) plot(k1['nproc'], k1['nproc']k1['rate'][0], color='0.5', alpha=0.5, label='_none_') labelit('After OpenMP bug workaround') savefig('portland2.pdf') figure() plot(h1['nproc'], h1['rate'], ':', color='C0', label='_none_', alpha=0.8) plot(h2['nproc'], h2['rate'], 's-', color='C0', label='Haswell') plot(h2['nproc'], h2['nproc']h2['rate'][0], color='0.5', alpha=0.5, label='_none_') plot(k1['nproc'], k1['rate'], ':', color='C1', label='_none_', alpha=0.8) plot(k2['nproc'], k2['rate'], 'd-', color='C1', label='KNL') plot(k2['nproc'], k2['nproc']k2['rate'][0], color='0.5', alpha=0.5, label='_none_') labelit('After Portland Hackfest') savefig('portland3.pdf') print('For 25 spectra x 300 wavelengths...') print('Starting point ratios of HSW/KNL') print(' per processes {:.2f}'.format(h1['rate'][0]/k1['rate'][0])) print(' per node {:.2f}'.format(np.max(h1['rate'])/np.max(k1['rate']))) print('Ending ratios of HSW/KNL') print(' per processes {:.2f}'.format(h2['rate'][0]/k2['rate'][0])) print(' per node {:.2f}'.format(np.max(h2['rate'])/np.max(k2['rate']))) print('Haswell per-node improvement {:.2f}'.format(np.max(h2['rate'])/np.max(h1['rate']))) print('KNL per-node improvement {:.2f}'.format(np.max(k2['rate'])/np.max(k1['rate']))) print('Original HSW / Current KNL {:.2f}'.format(np.max(h1['rate'])/np.max(k2['rate']))) print('Original HSW rate {:.1f}'.format(np.max(h1['rate'])))	doc/portland-extract.ipynb	sbailey/knltest	bsd-3-clause
Configuration parameters	import GAN.utils as utils # reload(utils) class Parameters(utils.Parameters): # dataset to be loaded load_datasets=utils.param(["moriond_v9","abs(ScEta) < 1.5"]) c_names = utils.param(['Phi','ScEta','Pt','rho','run']) x_names = utils.param(['EtaWidth','R9','SigmaIeIe','S4','PhiWidth','CovarianceIetaIphi', 'CovarianceIphiIphi']) # generate variables in MC to be distributed like in data sample_from_data = ['run'] # MC reweighting reweight = ['rewei_zee_5d_barrel_m_cleaning_xgb'] mcweight = utils.param('') # if empty ignore original weight use_abs_weight = utils.param(True) # negative weights are notoriously nasty # cleaning cuts cleaning = utils.param('cleaning_zee_m_barrel.hd5') # input features transformation feat_transform = utils.param('minmax') global_match = utils.param(False) # generator and discriminator (aka as 'critic' in wGAN) g_opts=utils.param(dict(name="G_512to64x8", kernel_sizes=[64,64,128,128,256,256,512,512], do_weight_reg=1e-5,do_last_l1reg=1e-5 )) pretrain_g = utils.param(False) d_opts=utils.param(dict(name="D_1024to32x6",kernel_sizes=[1024,512,128,64,32],do_bn=False, do_dropout=False,hidden_activation="relu", clip_weights=2.e-2,activation=None)) # weights clipping and no actication # optimizers dm_opts=utils.param(dict(optimizer="RMSprop",opt_kwargs=dict(lr=1e-5,decay=1e-6))) am_opts=utils.param(dict(optimizer="RMSprop",opt_kwargs=dict(lr=1e-5,decay=1e-6))) loss = utils.param("wgan_loss") # use WGAN loss gan_targets = utils.param('gan_targets_hinge') # hinge targets are 1, -1 instead of 0, 1 schedule = utils.param([0]5+[1]1) # number of critic iterations per generator iterations # schedule training epochs=utils.param(200) batch_size=utils.param(4096) plot_every=utils.param(10) frac_data=utils.param(10) # fraction of data to use monitor_dir = utils.param('log') # folder for logging batch = utils.param(False) # are we running in batch? class MyApp(utils.MyApp): classes = utils.List([Parameters]) # Read all parameters above from command line. # Note: names are all converted to be all capital notebook_parameters = Parameters(MyApp()).get_params() # copy parameters to global scope globals().update(notebook_parameters) DM_OPTS.update( {"loss":LOSS} ) AM_OPTS.update( {"loss":LOSS} ) plotting.batch = BATCH notebook_parameters	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Load datasets Apply cleaning and reweight MC to match data	data,mc = cms.load_zee(*LOAD_DATASETS) data.columns mc.columns	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Cleaning	if not CLEANING is None: print('cleaning data and mc') thr_up = pd.read_hdf(CLEANING,'thr_up') thr_down = pd.read_hdf(CLEANING,'thr_down') nevts_data = data.shape[0] nevts_mc = mc.shape[0] data = data[ ((data[thr_down.index] >= thr_down) & (data[thr_up.index] <= thr_up)).all(axis=1) ] mc = mc[ ((mc[thr_down.index] >= thr_down) & (mc[thr_up.index] <= thr_up)).all(axis=1) ] print('cleaning eff (data,mc): %1.2f % 1.2f' % ( data.shape[0] / nevts_data, mc.shape[0] / nevts_mc ))	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Generate extra variables for MC (eg run number) distributed like data	for col in SAMPLE_FROM_DATA: print('sampling', col) mc[col] = data[col].sample(mc.shape[0]).values	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Select target and conditional features	c_names = C_NAMES x_names = X_NAMES data_c = data[c_names] data_x = data[x_names] mc_c = mc[c_names] mc_x = mc[x_names] data_x.columns, data_x.shape, data_c.columns, data_c.shape data_x.columns, data_c.columns mc_x.columns, mc_c.columns	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Reweight MC	# reload(preprocessing) # reload(utils) # iniliatlize weights to default if MCWEIGHT == '': mc_w = np.ones(mc_x.shape[0]) else: mc_w = mc[MCWEIGHT].values print(mc_w[:10]) # take care of negative weights if USE_ABS_WEIGHT: mc_w = np.abs(mc_w) # reweighting if not REWEIGHT is None: for fil in REWEIGHT: # apply weights from n-dimnesional histograms if 'npy' in fil: info = np.load(fil) inputs = info[0] weights = info[1] bins = info[2:] # print(bins[1]) print('weighting',inputs) mc_w = preprocessing.reweight(mc,inputs,bins,weights,base=None) else: # or use a classifier clf = utils.read_pickle(fil,encoding='latin1') if hasattr(clf,'nthread'): clf.nthread = min(4,clf.nthread) clf.n_jobs = clf.nthread mc_w = preprocessing.reweight_multidim(mc[clf.features_],clf) data_w = np.ones(data_x.shape[0])	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Features transformation	# reload(preprocessing) if GLOBAL_MATCH: _,data_c,mc_x,mc_c,scaler_x,scaler_c = preprocessing.transform(None,data_c,mc_x,mc_c,FEAT_TRANSFORM) _,_,data_x,_,scaler_x_data,_ = preprocessing.transform(None,None,data_x,None,FEAT_TRANSFORM) else: data_x,data_c,mc_x,mc_c,scaler_x,scaler_c = preprocessing.transform(data_x,data_c,mc_x,mc_c,FEAT_TRANSFORM) data_x.shape,mc_x.shape data_x[2].max(),mc_x[2].max() print(mc_c.max(axis=0),mc_c.min(axis=0)) print(data_c.max(axis=0),data_c.min(axis=0)) print(mc_w.shape,mc_c.shape) for ic in range(len(c_names)): plotting.plot_hists(data_c[:,0,ic],mc_c[:,0,ic],generated_w=mc_w,bins=100,range=[-2.0,2.0]) plt.xlabel(c_names[ic]) plt.show() for ix in range(len(x_names)): plotting.plot_hists(data_x[:,0,ix],mc_x[:,0,ix],generated_w=mc_w,bins=100,range=[-3,3]) plt.xlabel(x_names[ix]) plt.show()	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Prepare train and test sample	nmax = min(data_x.shape[0]//FRAC_DATA,mc_x.shape[0]) data_x_train,data_x_test,data_c_train,data_c_test,data_w_train,data_w_test = cms.train_test_split(data_x[:nmax],data_c[:nmax],data_w[:nmax],test_size=0.1) mc_x_train,mc_x_test,mc_c_train,mc_c_test,mc_w_train,mc_w_test = cms.train_test_split(mc_x[:nmax],mc_c[:nmax],mc_w[:nmax],test_size=0.1) wscl = data_w[:nmax].sum() / mc_w[:nmax].sum() mc_w_train = wscl mc_w_test = wscl print(nmax,mc_w_train.sum()/data_w_train.sum())	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Instantiate the model Create the model and compile it.	# reload(models) xz_shape = (1,len(x_names)) c_shape = (1,len(c_names)) gan = models.MyFFGAN( xz_shape, xz_shape, c_shape=c_shape, g_opts=G_OPTS, d_opts=D_OPTS, dm_opts=DM_OPTS, am_opts=AM_OPTS, gan_targets=GAN_TARGETS )	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
This will actually trigger the instantiation of the generator (if not done here, it will happen befor compilation).	gan.get_generator()	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Same as above for the discriminator. Two discriminator models are created: one for the discriminator training phase and another for the generator trainin phase. The difference between the two is that the second does not contain dropout layers.	gan.get_discriminator()	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
model compilation	gan.adversarial_compile(loss=LOSS,schedule=SCHEDULE) gan.get_generator().summary() gan.get_discriminator()[0].summary() gan.get_discriminator()[1].summary() # sanity check set(gan.am[0].trainable_weights)-set(gan.am[1].trainable_weights)	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
Training Everything is ready. We start training.	from keras.optimizers import RMSprop if PRETRAIN_G: generator = gan.get_generator() generator.compile(loss="mse",optimizer=RMSprop(lr=1.e-3)) generator.fit( [mc_c_train,mc_x_train], [mc_c_train,mc_x_train], sample_weight=[mc_w_train,mc_w_train], epochs=1, batch_size=BATCH_SIZE ) # reload(base) initial_epoch = 0 # if hasattr(gan.model,"history"): # initial_epoch = gan.model.history.epoch[-1] + 1 do = dict( x_train=data_x_train, z_train=mc_x_train, c_x_train=data_c_train, c_z_train=mc_c_train, w_x_train = data_w_train, w_z_train = mc_w_train, x_test=data_x_test, z_test=mc_x_test, c_x_test=data_c_test, c_z_test=mc_c_test, w_x_test = data_w_test, w_z_test = mc_w_test, n_epochs=EPOCHS + initial_epoch - 1, initial_epoch=initial_epoch, batch_size=BATCH_SIZE, plot_every=PLOT_EVERY, monitor_dir=MONITOR_DIR ) exc = None try: gan.adversarial_fit(**do) except Exception as e: exc = e if not exc is None: print(exc)	cms_zee_conditional_wgan.ipynb	musella/GAN	gpl-3.0
### Clustering pearsons_r with HDBSCAN	# Clustering the pearsons_R with N/A vlaues removed hdb_t1 = time.time() hdb_pearson_r = hdbscan.HDBSCAN(metric = "precomputed", min_cluster_size=10).fit(df3_pearson_r) hdb_pearson_r_labels = hdb_pearson_r.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_pearson_r_labels)) # unique bins, zero is noise print(np.bincount(hdb_pearson_r_labels[hdb_pearson_r_labels!=-1])) pearson_clusters = {i: np.where(hdb_pearson_r_labels == i)[0] for i in range(2)} pearson_clusters #pd.set_option('display.height', 500) #These two commands allow for the display of max of 500 rows - exploring genes #pd.set_option('display.max_rows', 500) df2_TPM.iloc[pearson_clusters[1],:] #the genes that were clustered together [0,1]	data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb	gilmana/Cu_transition_time_course-	mit
Looks like there are two clusters, some expression and zero expression across samples. ### Clustering mean centered euclidean distance with with HDBSCAN	df3_euclidean_mean.hist() # Clustering the mean centered euclidean distance of TPM counts hdb_t1 = time.time() hdb_euclidean_mean = hdbscan.HDBSCAN(metric = "precomputed", min_cluster_size=10).fit(df3_euclidean_mean) hdb_euclidean_mean_labels = hdb_euclidean_mean.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_euclidean_mean_labels)) print(np.bincount(hdb_euclidean_mean_labels[hdb_euclidean_mean_labels!=-1])) euclidean_mean_clusters = {i: np.where(hdb_euclidean_mean_labels == i)[0] for i in range(2)} df2_TPM.iloc[euclidean_mean_clusters[1],:]	data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb	gilmana/Cu_transition_time_course-	mit
Looks like 2 clusters - both with zero expression. looks like wether it is a numpy array or pandas dataframe, the result is the same. lets now try to get index of the clustered points. ### Clustering log transformed euclidean distance with with HDBSCAN	df3_euclidean_log2 # Clustering the log2 transformed euclidean distance of TPM counts hdb_t1 = time.time() hdb_euclidean_log2 = hdbscan.HDBSCAN(metric = "precomputed", min_cluster_size=10).fit(df3_euclidean_log2) hdb_euclidean_log2_labels = hdb_euclidean_log2.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_euclidean_log2_labels)) print(np.bincount(hdb_euclidean_log2_labels[hdb_euclidean_log2_labels!=-1])) euclidean_log2_clusters = {i: np.where(hdb_euclidean_log2_labels == i)[0] for i in range(2)} df2_TPM.iloc[euclidean_log2_clusters[1],:]	data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb	gilmana/Cu_transition_time_course-	mit
Clustering using built-in HDBSCAN euclidean distance metric (mean centered and scaled to unit variance)	df2_TPM_values = df2_TPM.loc[:,"5GB1_FM40_T0m_TR2":"5GB1_FM40_T180m_TR1"] #isolating the data values df2_TPM_values_T = df2_TPM_values.T #transposing the data standard_scaler = StandardScaler() TPM_counts_mean_centered = standard_scaler.fit_transform(df2_TPM_values_T) #mean centering the data TPM_counts_mean_centered = pd.DataFrame(TPM_counts_mean_centered) #back to Dataframe #transposing back to original form and reincerting indeces and columns my_index = df2_TPM_values.index my_columns = df2_TPM_values.columns TPM_counts_mean_centered = TPM_counts_mean_centered.T TPM_counts_mean_centered.set_index(my_index, inplace=True) TPM_counts_mean_centered.columns = my_columns # Clustering the pearsons_R with N/A vlaues removed hdb_t1 = time.time() hdb_euclidean = hdbscan.HDBSCAN(metric = "euclidean", min_cluster_size=5).fit(TPM_counts_mean_centered) hdb_euclidean_labels = hdb_euclidean.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_euclidean_labels)) print(np.bincount(hdb_euclidean_labels[hdb_euclidean_labels!=-1]))	data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb	gilmana/Cu_transition_time_course-	mit
lets look at some clusters Euclidean_standard_scaled_clusters = {i: np.where(hdb_euclidean_labels == i)[0] for i in range(7)} df2_TPM.iloc[Euclidean_standard_scaled_clusters[0],:]	Euclidean_standard_scaled_clusters = {i: np.where(hdb_euclidean_labels == i)[0] for i in range(7)} df2_TPM.iloc[Euclidean_standard_scaled_clusters[1],:]	data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb	gilmana/Cu_transition_time_course-	mit
Euclidean_standard_scaled_clusters Clustering log2 transformed data using built-in HDBSCAN euclidean distance metric (mean centered and scaled to unit variance)	df2_TPM_log2_scale= df2_TPM_log2.T #transposing the data standard_scaler = StandardScaler() TPM_log2_mean_scaled = standard_scaler.fit_transform(df2_TPM_log2_scale) #mean centering the data TPM_log2_mean_scaled = pd.DataFrame(TPM_log2_mean_scaled) #back to Dataframe #transposing back to original form and reincerting indeces and columns my_index = df2_TPM_values.index my_columns = df2_TPM_values.columns TPM_log2_mean_scaled = TPM_log2_mean_scaled.T TPM_log2_mean_scaled.set_index(my_index, inplace=True) TPM_log2_mean_scaled.columns = my_columns # Clustering the pearsons_R with N/A vlaues removed hdb_t1 = time.time() hdb_log2_euclidean = hdbscan.HDBSCAN(metric = "euclidean", min_cluster_size=5).fit(TPM_log2_mean_scaled) hdb_log2_euclidean = hdb_log2_euclidean.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_log2_euclidean)) print(np.bincount(hdb_log2_euclidean[hdb_log2_euclidean!=-1]))	data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb	gilmana/Cu_transition_time_course-	mit
DM -- Piece by piece (as coded) $\rho_b = \Omega_b \rho_c (1+z)^3$ $\rho_{\rm diffuse} = \rho_b - (\rho_* + \rho_{\rm ISM})$ $\rho_*$ is the mass density in stars $\rho_{\rm ISM}$ is the mass density in the neutral ISM Number densities $n_{\rm H} = \rho_{\rm diffuse}/(m_p \, \mu)$ $\mu \approx 1.3$ accounts for Helium $n_{\rm He} = n_{\rm H}/12$ $n_e = n_{\rm H} [1-f_{\rm HI}] + n_{\rm He} Y$ $f_{\rm HI}$ is the fraction atomic Hydrogen [value betwee 0-1] $Y$ gives the number of free electrons per He nucleus [value between 0-2] Integrating $DM = \int \frac{n_e \, dr}{1+z} = \frac{c}{H_0} \int \frac{n_e \, dz}{(1+z)^2 \sqrt{(1+z)^3 \Omega_m + \Omega_\Lambda}}$ DM -- Altogether (using the code)	reload(igm) DM = igm.average_DM(1.) DM	docs/nb/DM_cosmic.ipynb	FRBs/FRB	bsd-3-clause
Cumulative plot	DM_cumul, zeval = igm.average_DM(1., cumul=True) # Inoue approximation DM_approx = 1000. * zeval * units.pc / units.cm**3 plt.clf() ax = plt.gca() ax.plot(zeval, DM_cumul, label='JXP') ax.plot(zeval, DM_approx, label='Approx') # Label ax.set_xlabel('z') ax.set_ylabel(r'${\rm DM}_{\rm IGM} [\rm pc / cm^3]$ ') # Legend legend = plt.legend(loc='lower right', scatterpoints=1, borderpad=0.2, handletextpad=0.1, fontsize='large') plt.show() plt.clf() ax = plt.gca() ax.plot(zeval, DM_approx/DM_cumul, label='Approx/JXP') #ax.plot(zeval, DM_approx, label='Approx') # Label ax.set_xlabel('z') ax.set_ylabel(r'Ratio of ${\rm DM}_{\rm IGM} [\rm pc / cm^3]$ ') # Legend legend = plt.legend(loc='upper right', scatterpoints=1, borderpad=0.2, handletextpad=0.1, fontsize='large') plt.show() DM_cumul[0:10] DM_approx[0:10] zeval[0]	docs/nb/DM_cosmic.ipynb	FRBs/FRB	bsd-3-clause
The Longley dataset is a time series dataset:	data = sm.datasets.longley.load() data.exog = sm.add_constant(data.exog) print(data.exog[:5])	examples/notebooks/gls.ipynb	ChadFulton/statsmodels	bsd-3-clause
While we don't have strong evidence that the errors follow an AR(1) process we continue	rho = resid_fit.params[1]	examples/notebooks/gls.ipynb	ChadFulton/statsmodels	bsd-3-clause
2x2 Grid You can easily create a layout with 4 widgets arranged on 2x2 matrix using the TwoByTwoLayout widget:	from ipywidgets import TwoByTwoLayout TwoByTwoLayout(top_left=top_left_button, top_right=top_right_button, bottom_left=bottom_left_button, bottom_right=bottom_right_button)	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
If you don't define a widget for some of the slots, the layout will automatically re-configure itself by merging neighbouring cells	TwoByTwoLayout(top_left=top_left_button, bottom_left=bottom_left_button, bottom_right=bottom_right_button)	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
You can pass merge=False in the argument of the TwoByTwoLayout constructor if you don't want this behavior	TwoByTwoLayout(top_left=top_left_button, bottom_left=bottom_left_button, bottom_right=bottom_right_button, merge=False)	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
You can add a missing widget even after the layout initialization:	layout_2x2 = TwoByTwoLayout(top_left=top_left_button, bottom_left=bottom_left_button, bottom_right=bottom_right_button) layout_2x2 layout_2x2.top_right = top_right_button	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
You can easily create more complex layouts with custom widgets. For example, you can use bqplot Figure widget to add plots:	import bqplot as bq import numpy as np size = 100 np.random.seed(0) x_data = range(size) y_data = np.random.randn(size) y_data_2 = np.random.randn(size) y_data_3 = np.cumsum(np.random.randn(size) * 100.) x_ord = bq.OrdinalScale() y_sc = bq.LinearScale() bar = bq.Bars(x=np.arange(10), y=np.random.rand(10), scales={'x': x_ord, 'y': y_sc}) ax_x = bq.Axis(scale=x_ord) ax_y = bq.Axis(scale=y_sc, tick_format='0.2f', orientation='vertical') fig = bq.Figure(marks=[bar], axes=[ax_x, ax_y], padding_x=0.025, padding_y=0.025, layout=Layout(width='auto', height='90%')) from ipywidgets import FloatSlider max_slider = FloatSlider(min=0, max=10, default_value=2, description="Max: ", layout=Layout(width='auto', height='auto')) min_slider = FloatSlider(min=-1, max=10, description="Min: ", layout=Layout(width='auto', height='auto')) app = TwoByTwoLayout(top_left=min_slider, bottom_left=max_slider, bottom_right=fig, align_items="center", height='700px') jslink((y_sc, 'max'), (max_slider, 'value')) jslink((y_sc, 'min'), (min_slider, 'value')) jslink((min_slider, 'max'), (max_slider, 'value')) jslink((max_slider, 'min'), (min_slider, 'value')) max_slider.value = 1.5 app	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
AppLayout AppLayout is a widget layout template that allows you to create an application-like widget arrangements. It consist of a header, a footer, two sidebars and a central pane:	from ipywidgets import AppLayout, Button, Layout header_button = create_expanded_button('Header', 'success') left_button = create_expanded_button('Left', 'info') center_button = create_expanded_button('Center', 'warning') right_button = create_expanded_button('Right', 'info') footer_button = create_expanded_button('Footer', 'success') AppLayout(header=header_button, left_sidebar=left_button, center=center_button, right_sidebar=right_button, footer=footer_button)	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
However with the automatic merging feature, it's possible to achieve many other layouts:	AppLayout(header=None, left_sidebar=None, center=center_button, right_sidebar=None, footer=None) AppLayout(header=header_button, left_sidebar=left_button, center=center_button, right_sidebar=right_button, footer=None) AppLayout(header=None, left_sidebar=left_button, center=center_button, right_sidebar=right_button, footer=None) AppLayout(header=header_button, left_sidebar=left_button, center=center_button, right_sidebar=None, footer=footer_button) AppLayout(header=header_button, left_sidebar=None, center=center_button, right_sidebar=right_button, footer=footer_button) AppLayout(header=header_button, left_sidebar=None, center=center_button, right_sidebar=None, footer=footer_button) AppLayout(header=header_button, left_sidebar=left_button, center=None, right_sidebar=right_button, footer=footer_button)	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
You can also modify the relative and absolute widths and heights of the panes using pane_widths and pane_heights arguments. Both accept a sequence of three elements, each of which is either an integer (equivalent to the weight given to the row/column) or a string in the format '1fr' (same as integer) or '100px' (absolute size).	AppLayout(header=header_button, left_sidebar=left_button, center=center_button, right_sidebar=right_button, footer=footer_button, pane_widths=[3, 3, 1], pane_heights=[1, 5, '60px'])	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
Grid layout GridspecLayout is a N-by-M grid layout allowing for flexible layout definitions using an API similar to matplotlib's GridSpec. You can use GridspecLayout to define a simple regularly-spaced grid. For example, to create a 4x3 layout:	from ipywidgets import GridspecLayout grid = GridspecLayout(4, 3) for i in range(4): for j in range(3): grid[i, j] = create_expanded_button('Button {} - {}'.format(i, j), 'warning') grid	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
To make a widget span several columns and/or rows, you can use slice notation:	grid = GridspecLayout(4, 3, height='300px') grid[:3, 1:] = create_expanded_button('One', 'success') grid[:, 0] = create_expanded_button('Two', 'info') grid[3, 1] = create_expanded_button('Three', 'warning') grid[3, 2] = create_expanded_button('Four', 'danger') grid	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
You can still change properties of the widgets stored in the grid, using the same indexing notation.	grid = GridspecLayout(4, 3, height='300px') grid[:3, 1:] = create_expanded_button('One', 'success') grid[:, 0] = create_expanded_button('Two', 'info') grid[3, 1] = create_expanded_button('Three', 'warning') grid[3, 2] = create_expanded_button('Four', 'danger') grid grid[0, 0].description = "I am the blue one"	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
Note: It's enough to pass an index of one of the grid cells occupied by the widget of interest. Slices are not supported in this context. If there is already a widget that conflicts with the position of the widget being added, it will be removed from the grid:	grid = GridspecLayout(4, 3, height='300px') grid[:3, 1:] = create_expanded_button('One', 'info') grid[:, 0] = create_expanded_button('Two', 'info') grid[3, 1] = create_expanded_button('Three', 'info') grid[3, 2] = create_expanded_button('Four', 'info') grid grid[3, 1] = create_expanded_button('New button!!', 'danger')	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
Creating scatter plots using GridspecLayout In these examples, we will demonstrate how to use GridspecLayout and bqplot widget to create a multipanel scatter plot. To run this example you will need to install the bqplot package. For example, you can use the following snippet to obtain a scatter plot across multiple dimensions:	import bqplot as bq import numpy as np from ipywidgets import GridspecLayout, Button, Layout n_features = 5 data = np.random.randn(100, n_features) data[:50, 2] += 4 * data[:50, 0] **2 data[50:, :] += 4 A = np.random.randn(n_features, n_features)/5 data = np.dot(data,A) scales_x = [bq.LinearScale() for i in range(n_features)] scales_y = [bq.LinearScale() for i in range(n_features)] gs = GridspecLayout(n_features, n_features) for i in range(n_features): for j in range(n_features): if i != j: sc_x = scales_x[j] sc_y = scales_y[i] scatt = bq.Scatter(x=data[:, j], y=data[:, i], scales={'x': sc_x, 'y': sc_y}, default_size=1) gs[i, j] = bq.Figure(marks=[scatt], layout=Layout(width='auto', height='auto'), fig_margin=dict(top=0, bottom=0, left=0, right=0)) else: sc_x = scales_x[j] sc_y = bq.LinearScale() hist = bq.Hist(sample=data[:,i], scales={'sample': sc_x, 'count': sc_y}) gs[i, j] = bq.Figure(marks=[hist], layout=Layout(width='auto', height='auto'), fig_margin=dict(top=0, bottom=0, left=0, right=0)) gs	docs/source/examples/Layout Templates.ipynb	SylvainCorlay/ipywidgets	bsd-3-clause
<a id='index'></a> Index Back to top 1 Introduction 2 ScatterPlot components 2.1 The scatter plot marker 2.2 Internal structure 2.3 Data structures 2.3.1 Original data 2.3.2 Tooltip data 2.3.3 Mapper data 2.3.2 Output data 3 ScatterPlot interface 3.1 Data mapper 3.2 Tooltip selector 3.3 Colors in Hex format 4 Taking a snapshot of the current plot 5 Plotting pandas Panel and Panel4D <a id='introduction'></a> 1 Introduction Now that we are familiar with the framework's basics, we can start showing the full capabilities of Shaolin. In order to do that I will rewiev one of the most "simple" and widely used plots in data science: The scatter plot. We provide in the dashboards section of the Shaolin framework several Dashboards suited for complex data processing, and the Bokeh Scatterplot is the one on which we are going to center this tutorial. All the individual components of this Dashboard will be explained deeply in further tutorials. <a id='components'></a> 2 ScatterPlot components <a id='scatter_marker'></a> 2.1 The scatter plot marker Back to top A scatter plot, as we all know is a kind of plot in which we represent two datapoint vectors (x and y) against each other and assign a marker(by now just a circle) to each pair of data points. Althoug the x and y coordinates of the marker are the only two compulsory parameters, it is also possible to customize the following parameters for a circle marker: x: x coordinate of the marker. y: y coordinate of the marker. size: Marker size. fill_alpha: Transparency value of the interior of the marker. fill_color: Color of the interior of the marker. line_color: Color of the marker's border. line_alpha: Transparency of the marker's border. line_width: Width of the marker's boder. It is possible to fully customize what data from the data structure we want to plot willn be mapped to a marker parameter and how that mapping will be performed. In order to assign values to a marker parameter we have to follow this process: Select a chunk of data from your pandas data structure that has the correct shape. (In this case each parameter must be a datapoint vector) Select how the data will be scaled to fit in the range of values that the marker parameter can have. (For example, for the line_width param all the values should be between 1 and 4.) Select if the values of the parameter will be a mapping of the data or a default value that will be the same for all the data points. This means that we could theoretically plot 8 dimensional data by mapping each parameter to a coordinate of a data point, but in practise it is sometime more usefull to use more than one parameter to map the same vector of data points in order to emphatise some feature of the data we are plotting. For example, we could map the fill_color parameter and the fill_alpha parameter to the same feature so it would be easy to emphatise the higher values of the plotted vector. <a id='internals'></a> 2.2 Internal structure Back to top The scatter plot is a Dashboard with the following attributes: data: The pandas data structure that we will use for the plot. widget: GUI for selecting the plot parameters. plot: Bokeh plot where the data is displayed. mapper: Dashboard in charge of mapping data to every marker parameter. tooltip: Dashboard in charge of managing the information displayed on the tooltip. output: DataFrame with all the information available to the plot. bokeh_source: Bokeh DataSource that mimics the infromation contained in the source df. In the following diagram you can se the process of how data is mapped into visual information. <img src="scatter_data/structure.svg"></img> For this example we will use the classic Iris dataset imported from the bokeh data samples.	from IPython.display import Image #this is for displaying the widgets in the web version of the notebook from shaolin.dashboards.bokeh import ScatterPlot from bokeh.sampledata.iris import flowers scplot = ScatterPlot(flowers)	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
<a id='data_structures'></a> 2.3 Data structures Back to top The data contained in the blocks described in the above diagram gcan be accessed the following way: <a id='original_data'></a> 2.3.1 Original data	scplot.data.head()	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
<a id='tooltip_data'></a> 2.3.2 Tooltip data Back to top	scplot.tooltip.output.head()	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
<a id='mapper_data'></a> 2.3.3 Mapper data	scplot.mapper.output.head()	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
<a id='output_data'></a> 2.3.4 output data	scplot.output.head()	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
<a id='plot_interface'></a> 3 Scatter plot Interface Back to top The scatter plot Dashboard contains the bokeh scatter plot and a widget. That widget is a toggle menu that can display two Dashboards: - Mapper: This dashboard is in charge of managing how the data is displayed. - Tooltip: The BokehTooltip Dashboard allows to select what information will be displayed on the plot tooltips. The complete plot interface can be displayed calling the function show. As you will see, the interface layout has not been yet customized, so any suggestion regarding interface desing will be appreciated. <a id='data_mapper'></a> 3.1 Data Mapper Back to top This is the Dashboard that allows to customize how the data will be plotted. We will color each of its components so its easier to locate them. This is a good example of a complex Dashboard comprised of multiple Dashboards.	mapper = scplot.mapper mapper.buttons.widget.layout.border = "blue solid" mapper.buttons.value = 'line_width' mapper.line_width.data_scaler.widget.layout.border = 'yellow solid' mapper.line_width.data_slicer.widget.layout.border = 'red solid 0.4em' mapper.line_width.data_slicer.columns_slicer.widget.layout.border = 'green solid 0.4em' mapper.line_width.data_slicer.index_slicer.widget.layout.border = 'green solid 0.4em' mapper.line_width.default_value.widget.layout.border = 'purple solid 0.4em' mapper.line_width.apply_row.widget.layout.border = "pink solid 0.4em" scplot.widget Image(filename='scatter_data/img_1.png')	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
A plot mapper has the following components: - Marker parameter selector(Blue): A dropdown that allows to select which marker parameter that is going to be changed. - Data slicer(Red): A dashoard in charge of selecting a datapoint vector from a pandas data structure. We can slice each of the dimensions of the data structure thanks to an AxisSlicer(Green) Dashboard. - Data scaler(Yellow): Dashboard in charge of scaling the data. Similar to the data scaler from the tutorials. - Activate mapping(pink): If the value of the checkbox is True the value of the marker parameter will be the output of the scaler, otherwise it will be the default value(Purple) for every data point. <a id='tooltip_selector'></a> 3.1 Tooltip selector Back to top It is possible to choose what information from the data attribute of the ScatterPlot will be shown when hovering above a marker. In the above cell we click in the "tooltip" button of the toggleButtons in order to make the widget visible. As we can see there is a SelectMultiple widget for every column of the original DataFrame.	scplot.widget Image(filename='scatter_data/img_2.png')	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
<a id='snapshot'></a> 4 Taking a snapshot of the current plot Back to top Although it is possible to save the bokeh plot with any of the standard methods that the bokeh library offers by accessing the plot attribute of the ScatterPlot, shaolin offers the possibility of saving an snapshot of the plot as a shaolin widget compatible with the framework, this way it can be included in a Dashboard for displaying purposes. This process is done by accessing the snapshot attribute of the scatterPlot. This way the current plot is exported and we can keep working with the ScatterPlot Dashboard in case we need to make more plots. An snapshot is an HTML widget which value is an exported notebook_div of the plot.	widget_plot = scplot.snapshot widget_plot.widget Image(filename='scatter_data/img_3.png')	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
<a id='plot_pandas'></a> 5 Plotting pandas Panel and Panel4D Back to top It is also possible to plot a pandas Panel or a Panel4d the same way as a DataFrame. The only resctriction for now is that the axis that will be used as index must be the major_axis in case of a Panel and the items axis in case of a Panel4D. The tooltips are disabled, custom tooltips will be available in the next release. It would be nice to have feedback on how would you like to display and select the tooltips.	from pandas.io.data import DataReader# I know its deprecated but i can't make the pandas_datareader work :P import datetime symbols_list = ['ORCL', 'TSLA', 'IBM','YELP', 'MSFT'] start = datetime.datetime(2010, 1, 1) end = datetime.datetime(2013, 1, 27) panel = DataReader( symbols_list, start=start, end=end,data_source='yahoo') panel sc_panel = ScatterPlot(panel) #sc_panel.show() Image(filename='scatter_data/img_4.png')	examples/Scatter Plot introduction.ipynb	HCsoft-RD/shaolin	agpl-3.0
The following blocks will be used to define the functions that computes the accumulative sum by the chosen three approaches. Native implementation using Python for command	def native_acc(x): """ Calculate accumulative sum using Python native loop approach :param x: Numpy array :return: Accumulative sum (list) """ native_acc_sum = [] sum_aux = 0 for val in x: native_acc_sum.append(val + sum_aux) sum_aux += val return native_acc_sum	accumulative_sum_benchmark.ipynb	ESSS/notebooks	mit
Just a binding for numpy.cumsum	def np_acc(x): """ Calculate accumulative sum using numpy.cumsum :param x: Numpy array :return: Accumulative sum (Numpy array) """ return np.cumsum(x)	accumulative_sum_benchmark.ipynb	ESSS/notebooks	mit
Now using sci20 implementation	def sci20_acc(x): """ Calculate accumulative sum using sci20 Array :param x: Numpy array :return: Accumulative sum (sci20 array) """ x_array = Array(FromNumPy(x)) return AccumulateArrayReturning(x_array)	accumulative_sum_benchmark.ipynb	ESSS/notebooks	mit
Here we use the timeit standard lib to obtain the elapsed time of each method	def do_benchmark(n=1000, k=10): """ Compute elapsed time for each accumulative sum implementation (Native loop, Numpy and sci20). :param n: Number of array elements. Default is 1000. :param k: Number of averages used for timing. Default is 10. :param dtype: Array data type. Default is np.double. :return: A tuple (dt_native, dt_sci20, dt_np) containing the elapsed time for each method. """ x = np.linspace(1, 100, n) dt_native = timeit.Timer(functools.partial(native_acc, x,)).timeit(k) dt_sci20 = timeit.Timer(functools.partial(sci20_acc, x,)).timeit(k) dt_np = timeit.Timer(functools.partial(np_acc, x,)).timeit(k) return dt_native, dt_sci20, dt_np	accumulative_sum_benchmark.ipynb	ESSS/notebooks	mit
And finally we have the results. sci20 and Numpy implementations are almost equivalent, while native loop has demonstrated its inefficiency	""" Computes and prints the elapsed time for each accumulative sum implementation (Native loop, Numpy and sci20). """ n_values = [10**x for x in range(1, 8)] # [10, 100, ..., 10^7] for n in n_values: print("Computing benchmark for n={}...".format(n)) dt_native, dt_sci20, dt_np = do_benchmark(n) print("Native: {:.8f}s / sci20: {:.8f}s / Numpy: {:.8f}s.".format(dt_native, dt_sci20, dt_np))	accumulative_sum_benchmark.ipynb	ESSS/notebooks	mit

그리고 각 노드에 대해 생성된 dict들을 저장하자.

for user in users: user["shortest_paths"] = shortest_paths_from(user)

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

그러면 이제 매개 중심성을 구할 준비가 다 되었다. 이제 각각의 최단 경로에 포함되는 각 노드의 매개 중심성에 $1/n$을 더해 주자.

for user in users: user["betweenness_centrality"] = 0.0 for source in users: source_id = source["id"] for target_id, paths in source["shortest_paths"].items(): # python2에서는 items 대신 iteritems 사용 if source_id < target_id: # 잘못해서 두 번 세지 않도록 주의하자 num_paths = len(paths) # 최단 경로가 몇 개 존재하는가? contrib = 1 / num_paths # 중심성에 기여하는 값 for path in paths: for id in path: if id not in [source_id, target_id]: users[id]["betweenness_centrality"] += contrib for user in users: print(user["id"], user["betweenness_centrality"])

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

사용자 0과 9의 최단 경로 사이에는 다른 사용자가 없으므로 매개 중심성이 0이다. 반면 사용자 3, 4, 5는 최단 경로상에 무척 빈번하게 위치하기 때문에 높은 매개 중심성을 가진다. 대게 중심성의 절댓값 자체는 큰 의미를 가지지 않고, 상대값만이 의미를 가진다. 그 외에 살펴볼 수 있는 중심성 지표 중 하나는 근접 중심성(closeness centrality)이다. 먼저 각 사용자의 원접성(farness)을 계산한다. 원접성이란 from_user와 다른 모든 사용자의 최단 경로를 합한 값이다.

# # closeness centrality # def farness(user): """모든 사용자와의 최단 거리 합""" return sum(len(paths[0]) for paths in user["shortest_paths"].values())

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

이제 근접 중심성은 간단히 계산할 수 있다.

for user in users: user["closeness_centrality"] = 1 / farness(user) for user in users: print(user["id"], user["closeness_centrality"])

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

계산된 근접 중심성의 편차는 더욱 작다. 네트워크 중심에 있는 노드조차 외곽에 위치한 노드들로부터 멀리 떨어져 있기 때문이다. 여기서 봤듯이 최단 경로를 계산하는 것은 꽤나 복잡하다. 그렇기 때문에 큰 네트워크에서는 근접 중심성을 자주 사용하지 않는다. 덜 직관적이지만 보통 더 쉽게 계산할 수 있는 고유벡터 중심성(eigenvector centrality)을 더 자주 사용한다. 21.2 고유벡터 중심성 고유벡터 중심성에 대해 알아보기 전에 먼저 고유벡터가 무엇인지 살펴봐야 하고, 고유벡터가 무엇인지 알기 위해서는 먼저 행렬 연산에 대해 알아봐야 한다. 21.2.1 행렬 연산

def matrix_product_entry(A, B, i, j): return dot(get_row(A, i), get_column(B, j)) def matrix_multiply(A, B): n1, k1 = shape(A) n2, k2 = shape(B) if k1 != n2: raise ArithmeticError("incompatible shapes!") return make_matrix(n1, k2, partial(matrix_product_entry, A, B)) def vector_as_matrix(v): """(list 형태의) 벡터 v를 n x 1 행렬로 변환""" return [[v_i] for v_i in v] def vector_from_matrix(v_as_matrix): """n x 1 행렬을 리스트로 변환""" return [row[0] for row in v_as_matrix] def matrix_operate(A, v): v_as_matrix = vector_as_matrix(v) product = matrix_multiply(A, v_as_matrix) return vector_from_matrix(product)

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

행렬 A의 고유 벡터를 찾기 위해, 임의의 벡터 $v$를 골라 matrix_operate를 수행하고, 결과값의 크기가 1이 되게 재조정하는 과정을 반복 수행한다.

def find_eigenvector(A, tolerance=0.00001): guess = [1 for __ in A] while True: result = matrix_operate(A, guess) length = magnitude(result) next_guess = scalar_multiply(1/length, result) if distance(guess, next_guess) < tolerance: return next_guess, length # eigenvector, eigenvalue guess = next_guess

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

결과값으로 반환되는 guess를 matrix_operate를 통해 결과값의 크기가 1인 벡터로 재조정하면, 자기 자신이 반환된다. 즉, 여기서 guess는 고유벡터라는 것을 의미한다. 모든 실수 행렬에 고유벡터와 고유값이 있는 것은 아니다. 예를 들어 시계 방향으로 90도 회전하는 연산을 하는 다음 행렬에는 곱했을 때 가지 자신이 되는 벡터는 영벡터밖에 없다.

rotate = [[0, 1], [-1, 0]]

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

이 행렬로 앞서 구현한 find_eignevector(rotate)를 수행하면, 영원히 끝나지 않을 것이다. 한편, 고유벡터가 있는 행렬도 때로는 무한루프에 빠질 수 있다.

flip = [[0, 1], [1, 0]]

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

이 행렬은 모든 벡터 [x, y]를 [y, x]로 변환한다. 따라서 [1, 1]은 고유값이 1인 고유벡터가 된다. 하지만 x, y값이 다른 임의의 벡터에서 출발해서 find_eigenvector를 수행하면 x, y값을 바꾸는 연산만 무한히 수행할 것이다. (NumPy같은 라이브러리에는 이런 케이스까지 다룰 수 있는 다양한 방법들이 구현되어 있다.) 이런 사소한 문제에도 불구하고, 어쨌든 find_eigenvector가 결과값을 반환한다면, 그 결과값은 곧 고유벡터이다. 21.2.2 중심성 고유벡터가 데이터 네트워크를 이해하는데 어떻게 도움을 줄까? 얘기를 하기 전에 먼저 네트워크를 인접행렬(adjacency matrix)의 형태로 나타내 보자. 이 행렬은 사용자 i와 사용자 j가 친구인 경우 (i, j)번째 항목에 1이 있고, 친구가 아닌 경우 0이 있는 행렬이다.

# # eigenvector centrality # def entry_fn(i, j): return 1 if (i, j) in friendships or (j, i) in friendships else 0 n = len(users) adjacency_matrix = make_matrix(n, n, entry_fn) adjacency_matrix

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

각 사용자의 고유벡터 중심성이란 find_eigenvector로 찾은 사용자의 고유벡터가 된다.

eigenvector_centralities, _ = find_eigenvector(adjacency_matrix) for user_id, centrality in enumerate(eigenvector_centralities): print(user_id, centrality)

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

연결의 수가 많고, 중심성이 높은 사용자들한테 연결된 사용자들은 고유벡터 중심성이 높다. 앞의 결과에 따르면 사용자 1, 사용자 2의 중심성이 가장 높은데, 이는 중심성이 높은 사람들과 세번이나 연결되었기 때문이다. 이들로부터 멀어질수록 사용자들의 중심성은 점차 줄어든다. 21.3 방향성 그래프(Directed graphs)와 페이지랭크 데이텀이 인기를 별로 끌지 못하자, 순이익 팀의 부사장은 친구 모델에서 보증(endorsement)모델로 전향하는 것을 고려 중이다. 알고 보니 사람들은 어떤 데이터 과학자들끼리 친구인지에 대해서는 별로 관심이 없었지만, 헤드헌터들은 다른 데이터 과학자로부터 존경 받는 데이터 과학자가 누구인지에 대해 관심이 많다. 이 새로운 모델에서 관계는 상호적인 것이 아니라, 한 사람(source)이 다른 멋진 한 사람(target)의 실력에 보증을 서주는 (source, target) 쌍으로 비대칭적인 관계를 표현하게 된다.

# # directed graphs # endorsements = [(0, 1), (1, 0), (0, 2), (2, 0), (1, 2), (2, 1), (1, 3), (2, 3), (3, 4), (5, 4), (5, 6), (7, 5), (6, 8), (8, 7), (8, 9)] for user in users: user["endorses"] = [] # add one list to track outgoing endorsements user["endorsed_by"] = [] # and another to track endorsements for source_id, target_id in endorsements: users[source_id]["endorses"].append(users[target_id]) users[target_id]["endorsed_by"].append(users[source_id])

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

그리고 가장 보증을 많이 받은 데이터 과학자들의 데이터를 수집해서, 그것을 헤드헌터들한테 팔면 된다.

endorsements_by_id = [(user["id"], len(user["endorsed_by"])) for user in users] sorted(endorsements_by_id, key=lambda x: x[1], # (user_id, num_endorsements) reverse=True)

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

사실 '보증의 수'와 같은 숫자는 조작하기가 매우 쉽다. 가장 간단한 방법 중 하나는, 가짜 계정을 여러 개 만들어서 그것들로 내 계정에 대한 보증을 서는 것이다. 또 다른 방법은, 친구들끼리 짜고 서로가 서로를 보증해 주는 것이다. (아마 사용자 0, 1, 2가 이런 관계일 가능성이 크다.) 좀 더 나은 지수는, '누가' 보증을 서는지를 고려하는 것이다. 보증을 많이 받은 사용자가 보증을 설 때는, 보증을 적게 받은 사용자가 보증을 설 때보다 더 중요한 것으로 받아들여지는 것이 타당하다. 그리고 사실 이것은 유명한 페이지랭크(PageRank) 알고리즘의 기본 철학이기도 하다. 1. 네트워크 전체에는 1.0(또는 100%)의 페이지랭크가 있다. 2. 초기에 이 페이지랭크를 모든 노드에 고르게 배당한다. 3. 각 스텝을 거칠 때마다 각 노드에 배당된 페이지랭크의 대부분은 외부로 향하는 링크에 균등하게 배당한다. 4. 각 스텝을 거칠 때마다 각 노드에 남아 있는 페이지랭크를 모든 노드에 고르게 배당한다.

def page_rank(users, damping = 0.85, num_iters = 100): # 먼저 페이지랭크를 모든 노드에 고르게 배당 num_users = len(users) pr = { user["id"] : 1 / num_users for user in users } # 매 스텝마다 각 노드가 받는 # 적은 양의 페이지랭크 base_pr = (1 - damping) / num_users for __ in range(num_iters): next_pr = { user["id"] : base_pr for user in users } for user in users: # 페이지랭크를 외부로 향하는 링크에 배당한다. links_pr = pr[user["id"]] * damping for endorsee in user["endorses"]: next_pr[endorsee["id"]] += links_pr / len(user["endorses"]) pr = next_pr return pr for user_id, pr in page_rank(users).items(): print(user_id, pr)

notebook/ch21_network_analysis.ipynb

rnder/data-science-from-scratch

unlicense

Generating some data Lets generate a classification dataset that is not easily linearly separable. Our favorite example is the spiral dataset, which can be generated as follows:

N = 100 #The number of "training examples" per class D = 2 # The dimension i.e i think the feature set K = 3 # The number of classes to classify into X = np.zeros((N*K, D)) # The data matrix , each row is a single example # In our case 100*3 amount of examples and each example has two dimensions # so the Matrix is 300 * 2 Y = np.zeros(N*K , dtype = 'uint8') #Class Labels for j in range(K): ix = range (N*j,N*(j+1)) #filling the matrices made above with datapoints r = np.linspace(0.0,1,N) # radius t = np.linspace(j*4,(j+1)*4,N) + np.random.randn(N) * 0.2 #theta #numpy.linspace(start, stop, num = 50, endpoint = True, retstep = False, dtype = None) #: Returns number spaces evenly w.r.t interval. X[ix] = np.c_[r*np.sin(t),r*np.cos(t)] # the above line is written so as to give the dataset a spiral structure # which will be clear in the dataset plot below # i think this can be played with different kinds of datasets and check how our neural network reacts # np.c_ slices objects to concatenation along the second axis. Y[ix] = j plt.scatter(X[:, 0], X[:, 1], c=Y, s=40, cmap = plt.cm.Spectral) plt.xlim([-1,1]) plt.ylim([-1,1]) plt.show()

Deep Learning.ai/Numpy - Understanding Backprop.ipynb

Shashi456/Artificial-Intelligence

gpl-3.0

Training a Simple Softmax Linear Classifer Initialize the Parameters

# initialize parameters randomly W = 0.01 * np.random.randn(D,K) #randn(D,K) D * K Matrix # in our case 2 * 3 Matrix b = np.zeros((1,K)) # in our case 1 * 3 # some hyperparameters step_size = 1e-0 reg = 1e-3 # regularization strength

Deep Learning.ai/Numpy - Understanding Backprop.ipynb

Shashi456/Artificial-Intelligence

gpl-3.0

Compute the Class Scores

# compute class scores for a linear classif scores = np.dot(X,W) + b #scores will be 300 *3 and remember numpy broadcasting for b

Deep Learning.ai/Numpy - Understanding Backprop.ipynb

Shashi456/Artificial-Intelligence

gpl-3.0

Compute the loss

# Using cross - entropy loss # Softmax loss = data loss + regularization loss # Study about the Softmax loss and how it works before seeing how it works #in this case it could be intuitive num_examples = X.shape[0] # get unnormalized probabilities exp_scores = np.exp(scores) # normalize them for each example '''We now have an array probs of size [300 x 3], where each row now contains the class probabilities. In particular, since we’ve normalized them every row now sums to one. We can now query for the log probabilities assigned to the correct classes in each example:''' probs = exp_scores / np.sum(exp_scores, axis=1 , keepdims=True) # Calculating the correct_log probability '''The array correct_logprobs is a 1D array of just the probabilities assigned to the correct classes for each example. The full loss is then the average of these log probabilities and the regularization loss:''' correct_logprobs = -np.log(probs[range(num_examples),y]) # compute ths loss : average cross-entropy loss and regularization data_loss = np.sum(correct_logprobs)/ num_examples reg_loss = 0.5*reg*np.sum(W*W) '''∂Li∂fk=pk−1(yi=k) Notice how elegant and simple this expression is. Suppose the probabilities we computed were p = [0.2, 0.3, 0.5], and that the correct class was the middle one (with probability 0.3). According to this derivation the gradient on the scores would be df = [0.2, -0.7, 0.5]. Recalling what the interpretation of the gradient, we see that this result is highly intuitive: increasing the first or last element of the score vector f (the scores of the incorrect classes) leads to an increased loss (due to the positive signs +0.2 and +0.5) - and increasing the loss is bad, as expected. However, increasing the score of the correct class has negative influence on the loss. The gradient of -0.7 is telling us that increasing the correct class score would lead to a decrease of the loss Li, which makes sense.''' dscores = probs dscores[range(num_examples),y]-=1 dscores/=num_examples dW = np.dot(X.T,dscores) db = np.sum(dscores, axis=0, keepdims=True) dW += reg*W # Not forgetting the regularization gradient #Updating Parameters W += -step_size *dW b += -step_size *db

Deep Learning.ai/Numpy - Understanding Backprop.ipynb

Shashi456/Artificial-Intelligence

gpl-3.0

Putting it all Together Till now we just looked at individual elements

#Train a Linear Classifier # initialize parameters randomly W = 0.01 * np.random.randn(D,K) b = np.zeros((1,K)) # some hyperparameters step_size = 1e-0 reg = 1e-3 # regularization strength # gradient descent loop num_examples = X.shape[0] for i in range(200): # evaluate class scores, [N x K] scores = np.dot(X, W) + b # compute the class probabilities exp_scores = np.exp(scores) probs = exp_scores / np.sum(exp_scores, axis=1, keepdims=True) # [N x K] # compute the loss: average cross-entropy loss and regularization corect_logprobs = -np.log(probs[range(num_examples),Y]) data_loss = np.sum(corect_logprobs)/num_examples reg_loss = 0.5*reg*np.sum(W*W) loss = data_loss + reg_loss if i % 10 == 0: print ("iteration %d: loss %f" % (i, loss)) # compute the gradient on scores dscores = probs dscores[range(num_examples),Y] -= 1 dscores /= num_examples # backpropate the gradient to the parameters (W,b) dW = np.dot(X.T, dscores) db = np.sum(dscores, axis=0, keepdims=True) dW += reg*W # regularization gradient # perform a parameter update W += -step_size * dW b += -step_size * db # evaluate training set Accuracy scores = np.dot(X,W) + b predicted_class = np.argmax(scores, axis=1) print("training_accuracy: %.2f" % (np.mean(predicted_class==Y))) # plot the resulting classifier h = 0.02 x_min, x_max = X[:, 0].min() - 1, X[:, 0].max() + 1 y_min, y_max = X[:, 1].min() - 1, X[:, 1].max() + 1 xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h)) Z = np.dot(np.c_[xx.ravel(), yy.ravel()], W) + b Z = np.argmax(Z, axis=1) Z = Z.reshape(xx.shape) plt.contourf(xx, yy, Z, cmap=plt.cm.Spectral, alpha=0.8) plt.scatter(X[:, 0], X[:, 1], c=Y, s=40, cmap=plt.cm.Spectral) plt.xlim(xx.min(), xx.max()) plt.ylim(yy.min(), yy.max()) plt.show()

Deep Learning.ai/Numpy - Understanding Backprop.ipynb

Shashi456/Artificial-Intelligence

gpl-3.0

Using SQL for Queries Note that SQL is case-insensitive, but it is traditional to use ALL CAPS for SQL keywords. It is also standard to end SQL statements with a semi-colon. Simple Queries

pdsql('SELECT * FROM tips LIMIT 5;') q = """ SELECT sex as gender, smoker, day, size FROM tips WHERE size <= 3 AND smoker = 'Yes' LIMIT 5""" pdsql(q)

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Filtering on strings

q = """ SELECT sex as gender, smoker, day, size , time FROM tips WHERE time LIKE '%u_c%' LIMIT 5""" pdsql(q)

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Ordering

q = """ SELECT sex as gender, smoker, day, size , time FROM tips WHERE time LIKE '%u_c%' ORDER BY size DESC LIMIT 5""" pdsql(q)

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Aggregate queries

pdsql('select * from tips limit 5;') q = """ SELECT sex, smoker, sum(total_bill) as total, max(tip) as max_tip FROM tips WHERE time = 'Lunch' GROUP BY sex, smoker ORDER BY max_tip DESC """ pdsql(q)

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Matching students and majors Inner join

pdsql(""" SELECT * from student s INNER JOIN major m ON s.major_id = m.major_id; """)

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Left outer join SQL also has RIGHT OUTER JOIN and FULL OUTER JOIN but these are not currently supported by SQLite3 (the database engine used by pdsql).

pdsql(""" SELECT s.*, m.name from student s LEFT JOIN major m ON s.major_id = m.major_id; """)

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Emulating a full outer join with UNION ALL Only necessary if the database does not proivde FULL OUTER JOIN Using linker tables to match students to classes (a MANY TO MANY join) Using SQLite3 SQLite3 is part of the standard library. However, the mechanics of using essentially any database in Python is similar, because of the Python DB-API.

import sqlite3 c = sqlite3.connect('data/Chinook_Sqlite.sqlite')

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Standard SQL statements with parameter substitution Note: Using Python string substitution for Python defined parameters is dangerous because of the risk of SQL injection attacks. Use parameter substitution with ? instead. Do this

t = ['%rock%', 10] list(c.execute("SELECT * FROM Album WHERE Title like ? AND ArtistID > ? LIMIT 5;", t))

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Not this

t = ("'%rock%'", 10) list(c.execute("SELECT * FROM Album WHERE Title like %s AND ArtistID > %d LIMIT 5;" % t))

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

User defined functions Sometimes it is useful to have custom functions that run on the database server rather than on the client. These are called User Defined Functions or UDF. How do to do this varies with the database used, but it is fairly simple with Python and SQLite. A standard UDF

def encode(text, offset): """Caesar cipher of text with given offset.""" from string import ascii_lowercase, ascii_uppercase tbl = dict(zip(map(ord, ascii_lowercase + ascii_uppercase), ascii_lowercase[offset:] + ascii_lowercase[:offset] + ascii_uppercase[offset:] + ascii_uppercase[:offset])) return text.translate(tbl) c.create_function("encode", 2, encode) list(c.execute("SELECT Title, encode(Title, 2) FROM Album limit 5;"))

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

Using SQL magic functions We will use the ipython-sql notebook extension for convenience. This will only work in notebooks and IPython scripts with the .ipy extension.

%load_ext sql

scratch/Lecture08.ipynb

cliburn/sta-663-2017

mit

3. Open the file and use .read() to save the contents of the file to a string called fields. Make sure the file is closed at the end.

# Write your code here: # Run fields to see the contents of contacts.txt: fields

nlp/UPDATED_NLP_COURSE/00-Python-Text-Basics/03-Python-Text-Basics-Assessment.ipynb

rishuatgithub/MLPy

apache-2.0

Working with PDF Files 4. Use PyPDF2 to open the file Business_Proposal.pdf. Extract the text of page 2.

# Perform import # Open the file as a binary object # Use PyPDF2 to read the text of the file # Get the text from page 2 (CHALLENGE: Do this in one step!) page_two_text = # Close the file # Print the contents of page_two_text print(page_two_text)

nlp/UPDATED_NLP_COURSE/00-Python-Text-Basics/03-Python-Text-Basics-Assessment.ipynb

rishuatgithub/MLPy

apache-2.0

5. Open the file contacts.txt in append mode. Add the text of page 2 from above to contacts.txt. CHALLENGE: See if you can remove the word "AUTHORS:"

# Simple Solution: # CHALLENGE Solution (re-run the %%writefile cell above to obtain an unmodified contacts.txt file):

nlp/UPDATED_NLP_COURSE/00-Python-Text-Basics/03-Python-Text-Basics-Assessment.ipynb

rishuatgithub/MLPy

apache-2.0

Regular Expressions 6. Using the page_two_text variable created above, extract any email addresses that were contained in the file Business_Proposal.pdf.

import re # Enter your regex pattern here. This may take several tries! pattern = re.findall(pattern, page_two_text)

nlp/UPDATED_NLP_COURSE/00-Python-Text-Basics/03-Python-Text-Basics-Assessment.ipynb

rishuatgithub/MLPy

apache-2.0

Exercise: Compute the velocity of the penny after t seconds. Check that the units of the result are correct.

# Solution goes here

notebooks/chap01.ipynb

AllenDowney/ModSimPy

mit

Exercise: Why would it be nonsensical to add a and t? What happens if you try?

# Solution goes here

notebooks/chap01.ipynb

AllenDowney/ModSimPy

mit

Exercise: Suppose you bring a 10 foot pole to the top of the Empire State Building and use it to drop the penny from h plus 10 feet. Define a variable named foot that contains the unit foot provided by UNITS. Define a variable named pole_height and give it the value 10 feet. What happens if you add h, which is in units of meters, to pole_height, which is in units of feet? What happens if you write the addition the other way around?

# Solution goes here # Solution goes here

notebooks/chap01.ipynb

AllenDowney/ModSimPy

mit

Exercise: In reality, air resistance limits the velocity of the penny. At about 18 m/s, the force of air resistance equals the force of gravity and the penny stops accelerating. As a simplification, let's assume that the acceleration of the penny is a until the penny reaches 18 m/s, and then 0 afterwards. What is the total time for the penny to fall 381 m? You can break this question into three parts: How long until the penny reaches 18 m/s with constant acceleration a. How far would the penny fall during that time? How long to fall the remaining distance with constant velocity 18 m/s? Suggestion: Assign each intermediate result to a variable with a meaningful name. And assign units to all quantities!

# Solution goes here # Solution goes here # Solution goes here # Solution goes here # Solution goes here

notebooks/chap01.ipynb

AllenDowney/ModSimPy

mit

Filtering and resampling data Some artifacts are restricted to certain frequencies and can therefore be fixed by filtering. An artifact that typically affects only some frequencies is due to the power line. Power-line noise is a noise created by the electrical network. It is composed of sharp peaks at 50Hz (or 60Hz depending on your geographical location). Some peaks may also be present at the harmonic frequencies, i.e. the integer multiples of the power-line frequency, e.g. 100Hz, 150Hz, ... (or 120Hz, 180Hz, ...). This tutorial covers some basics of how to filter data in MNE-Python. For more in-depth information about filter design in general and in MNE-Python in particular, check out sphx_glr_auto_tutorials_plot_background_filtering.py.

import numpy as np import mne from mne.datasets import sample data_path = sample.data_path() raw_fname = data_path + '/MEG/sample/sample_audvis_raw.fif' proj_fname = data_path + '/MEG/sample/sample_audvis_eog_proj.fif' tmin, tmax = 0, 20 # use the first 20s of data # Setup for reading the raw data (save memory by cropping the raw data # before loading it) raw = mne.io.read_raw_fif(raw_fname) raw.crop(tmin, tmax).load_data() raw.info['bads'] = ['MEG 2443', 'EEG 053'] # bads + 2 more fmin, fmax = 2, 300 # look at frequencies between 2 and 300Hz n_fft = 2048 # the FFT size (n_fft). Ideally a power of 2 # Pick a subset of channels (here for speed reason) selection = mne.read_selection('Left-temporal') picks = mne.pick_types(raw.info, meg='mag', eeg=False, eog=False, stim=False, exclude='bads', selection=selection) # Let's first check out all channel types raw.plot_psd(area_mode='range', tmax=10.0, picks=picks, average=False)

0.17/_downloads/1a57f50035589ab531155bdb55bcbcb5/plot_artifacts_correction_filtering.ipynb

mne-tools/mne-tools.github.io

bsd-3-clause

The same care must be taken with these results as with partial derivatives. The formula for $Y$ is ostensibly $$X_1 + X_2 = X_1 + X^2 + X_1 = 2 X_1 + X^2$$ Or $2X_1$ plus a parabola. However, the coefficient of $X_1$ is 1. That is because $Y$ changes by 1 if we change $X_1$ by 1 <i>while holding $X_2$ constant</i>. Multiple linear regression separates out contributions from different variables. Similarly, running a linear regression on two securities might give a high $\beta$. However, if we bring in a third security (like SPY, which tracks the S&P 500) as an independent variable, we may find that the correlation between the first two securities is almost entirely due to them both being correlated with the S&P 500. This is useful because the S&P 500 may then be a more reliable predictor of both securities than they were of each other. This method allows us to better gauge the significance between the two securities and prevent confounding the two variables.

# Load pricing data for two arbitrarily-chosen assets and SPY start = '2014-01-01' end = '2015-01-01' asset1 = get_pricing('DTV', fields='price', start_date=start, end_date=end) asset2 = get_pricing('FISV', fields='price', start_date=start, end_date=end) benchmark = get_pricing('SPY', fields='price', start_date=start, end_date=end) # First, run a linear regression on the two assets slr = regression.linear_model.OLS(asset1, sm.add_constant(asset2)).fit() print 'SLR beta of asset2:', slr.params[1] # Run multiple linear regression using asset2 and SPY as independent variables mlr = regression.linear_model.OLS(asset1, sm.add_constant(np.column_stack((asset2, benchmark)))).fit() prediction = mlr.params[0] + mlr.params[1]*asset2 + mlr.params[2]*benchmark prediction.name = 'Prediction' print 'MLR beta of asset2:', mlr.params[1], '\nMLR beta of S&P 500:', mlr.params[2]

notebooks/lectures/Multiple_Linear_Regression/notebook.ipynb

quantopian/research_public

apache-2.0

The Laplacian Matrix is a matrix that is extremely important in graph theory and numerical analysis. It is defined as $L=D-A$. Where $D$ is the degree matrix and $A$ is the adjecency matrix. For the purpose of this problem you don't need to understand the details of these matrices, although their definitions are relatively simple. The degree matrix for $K_n$ is an $n \times n$ diagonal matrix with the value $n-1$ along the diagonal and zeros everywhere else. Write a function to compute the degree matrix for $K_n$ using NumPy.

def complete_deg(n): a = np.identity((n), dtype = np.int) for element in np.nditer(a, op_flags=['readwrite']): if element > 0: element[...] = element + n - 2 return a complete_deg(3) D = complete_deg(5) assert D.shape==(5,5) assert D.dtype==np.dtype(int) assert np.all(D.diagonal()==4*np.ones(5)) assert np.all(D-np.diag(D.diagonal())==np.zeros((5,5),dtype=int))

assignments/assignment03/NumpyEx04.ipynb

geoneill12/phys202-2015-work

mit

The adjacency matrix for $K_n$ is an $n \times n$ matrix with zeros along the diagonal and ones everywhere else. Write a function to compute the adjacency matrix for $K_n$ using NumPy.

def complete_adj(n): b = np.identity((n), dtype = np.int) for element in np.nditer(b, op_flags=['readwrite']): if element == 0: element[...] = 1 else: element[...] = 0 return b complete_adj(3) A = complete_adj(5) assert A.shape==(5,5) assert A.dtype==np.dtype(int) assert np.all(A+np.eye(5,dtype=int)==np.ones((5,5),dtype=int))

assignments/assignment03/NumpyEx04.ipynb

geoneill12/phys202-2015-work

mit

Use NumPy to explore the eigenvalues or spectrum of the Laplacian L of $K_n$. What patterns do you notice as $n$ changes? Create a conjecture about the general Laplace spectrum of $K_n$.

def L(n): L = complete_deg(n) - complete_adj(n) return L print L(1) print L(2) print L(3) print L(4) print L(5) print L(6)

assignments/assignment03/NumpyEx04.ipynb

geoneill12/phys202-2015-work

mit

在新三板挂牌的公司还包括了原两网的公司（NET和STAQ），两网的公司是在90年代有中国人民银行批准设立的，总用9家公司（粤传媒(400003)上市后剩余8家公司），由于政策方面的约束，此类公司即便在三板市场上依然不被允许进行新的融资而只能股份转让。这些公司包含了： * 大自然（400001），该股票已经在在新三板退市，并以834019为代码成为全国股转公司的挂牌企业。可能出现的问题是，数据库把400001公布的财务信息也作为了新挂牌公司的财务信息，因此，处理时直接以新挂牌的公司转让公告书中披露的财务信息作为检验标准（To-do: 作为对比直接踢除这家公司的数据。） * 长白5 (400002) * 海国实5(400005) * 京中兴5(400006) * 华凯1 (400007) * 广建1 (400009) * 鹫峰5 (400010) * 港岳1 (400013)

stkcs_net_staq = ["400001.OC", "400002.OC", "400005.OC", "400006.OC", "400007.OC", "400009.OC", "400010.OC", "400013.OC"] stocks_neeq = stocks_neeq[~stocks_neeq["证券代码"].isin(stkcs_net_staq)] len(stocks_neeq) stocks_neeq[stocks_neeq["证券代码"] == "834019.OC"] stocks_neeq = stocks_neeq.drop(6644) stocks_neeq[stocks_neeq["证券代码"] == "834019.OC"]

src/Stocks.ipynb

januslian/neeqem

gpl-3.0

退市股票处理：11个转板上市公司，1个连续四年亏损，16个吸收合并的公司，3家公司将公司变为有限责任公司，6家公司未说明退市原因。在数据库退市股票列表里有两家公司400001和400003属于两网公司转板，不作为研究样本处理。

delisting = pandas.read_excel("../data/raw/退市股票一览.xlsx") #2016年4月15日之前退市的公司 delisting = delisting[:-2] delisting.head() delisting = delisting[~delisting["代码"].isin(["400001.OC", "400003.OC"])] stkcd = stocks_neeq["证券代码"].append(delisting["代码"]) stkcd[stkcd.duplicated()] delisting = pandas.read_excel("../data/delist.xls") delisting.head() stocks_neeq_all = pandas.concat([stocks_neeq, delisting], ignore_index=True) len(stocks_neeq_all) len(stocks_neeq), len(delisting) stocks_neeq_all.iloc[0] stocks_neeq_all[stocks_neeq_all["所属证监会行业"].isnull()] stocks_neeq_all.loc[stocks_neeq_all["证券代码"] == "835458.OC", "所属证监会行业"] = "货币金融服务" stocks_neeq_all.loc[stocks_neeq_all["证券代码"] == "830806.OC", "所属证监会行业"] = "软件和信息技术服务业" stocks_neeq_all.loc[stocks_neeq_all["证券代码"] == "430217.OC", "所属证监会行业"] = "互联网和相关服务业" stocks_neeq_all[stocks_neeq_all["所属证监会行业"].isnull()] finance_sector = ["货币金融服务", "资本市场服务", "保险业", "其他金融业"] print(len(stocks_neeq_all[stocks_neeq_all["所属证监会行业"].isin(finance_sector)])) stocks_neeq_all = stocks_neeq_all[~stocks_neeq_all["所属证监会行业"].isin(finance_sector)] stocks_neeq_all.to_csv("../data/NEEQ_sample.csv", index=False) a = len(stocks_neeq_all[stocks_neeq_all["挂牌日期"] <= pandas.Timestamp('2009-10-30 00:00:00')])

src/Stocks.ipynb

januslian/neeqem

gpl-3.0

由此得到的“新三板”公司的有效样本数量为{{len(stocks_neeq_all)}}。 “新三板”样本公司最早的挂牌时间为{{stocks_neeq_all["挂牌日期"].min()}}，所以在选择与之对比的A股主板、中小板和创业板公司时也选择从2006年以后，其中创业板开始于2009年10月，在2006到2009年10月之间挂牌的新三板公司有{{a}}家公司。由于所有适合本研究的“新三板”公司均为2006年之后挂牌交易的，所以对应的主板、中小板和创业板公司也将会选择2005年之后IPOs的公司作为对比的对象。后续的检验也会将主中创三个板块的定向增发作为对比对象，以保证在发行过程上有些许的可比性。 To-do: 需要确定估计样本的来源，备选的方案包括了：（1）可以使用当年所有已经上市的公司作为估计样本；（2）可以使用当年所有上市已经超过一年的公司作为样本；（3）使用规模分层的估计方法；

stocks_neeq_all.iloc[0] year = lambda x: x.year stocks_neeq_all["挂牌年度"] = stocks_neeq_all["挂牌日期"].apply(year) stocks_neeq_all["挂牌年度"].value_counts().sort_index() sns.barplot(stocks_neeq_all["挂牌年度"].value_counts().sort_index().index, stocks_neeq_all["挂牌年度"].value_counts().sort_index(), palette="Blues_d"); sns.countplot(x="挂牌年度", data=stocks_neeq_all, palette="Blues_d"); stocks_neeq_all["所属证监会行业"].value_counts()

src/Stocks.ipynb

januslian/neeqem

gpl-3.0

2. Hypothesis tests

#California burritos vs. Carnitas burritos TODO # Don Carlos 1 vs. Don Carlos 2 TODO # Bonferroni correction TODO

burrito/.ipynb_checkpoints/Burrito_bootcamp-checkpoint.ipynb

srcole/qwm

mit

3. Burrito dimension distributions Distribution of each burrito quality

import math def metrichist(metricname): if metricname == 'Volume': bins = np.arange(.375,1.225,.05) xticks = np.arange(.4,1.2,.1) xlim = (.4,1.2) else: bins = np.arange(-.25,5.5,.5) xticks = np.arange(0,5.5,.5) xlim = (-.25,5.25) plt.figure(figsize=(5,5)) n, _, _ = plt.hist(df[metricname].dropna(),bins,color='k') plt.xlabel(metricname + ' rating',size=20) plt.xticks(xticks,size=15) plt.xlim(xlim) plt.ylabel('Count',size=20) plt.yticks((0,int(math.ceil(np.max(n) / 5.)) * 5),size=15) plt.tight_layout() m_Hist = ['Hunger','Volume','Tortilla','Temp','Meat','Fillings', 'Meat:filling','Uniformity','Salsa','Synergy','Wrap','overall'] for m in m_Hist: metrichist(m)

burrito/.ipynb_checkpoints/Burrito_bootcamp-checkpoint.ipynb

srcole/qwm

mit

Test for normal distribution

TODO

burrito/.ipynb_checkpoints/Burrito_bootcamp-checkpoint.ipynb

srcole/qwm

mit

Changes implemented From most important to least important: * pre-compute PSF parameters across multiple spectra and wavelengths to enable vectorized calls to legval instead of many scalar calls [~25% faster] * hoist wavelength -> [-1,1] for legendre calculations out of loop [~8% faster] * replace np.outer with numba optimized version [~5% faster (?)] * cast astropy.io.fits.Header to dict() before many many key/value lookups [~4% faster] * Avoid intermediate memory allocations by using out=... [<~1% faster] Note: amount of improvement was different for Haswell vs. KNL vs. Core i7 Laptop and varies with extraction problem size; above numbers are approximate; see below for before/after Haswell and KNL results. Before / after comparison

h1 = Table.read('data/portland/hsw-before.txt', format='ascii.basic') h2 = Table.read('data/portland/hsw-after.txt', format='ascii.basic') k1 = Table.read('data/portland/knl-before.txt', format='ascii.basic') k2 = Table.read('data/portland/knl-after.txt', format='ascii.basic') plot(h1['nproc'], h1['rate'], 's:', color='C0', label='Haswell original') plot(h2['nproc'], h2['rate'], 's-', color='C0', label='Haswell') plot(h2['nproc'], h2['nproc']*h2['rate'][0], color='0.5', alpha=0.5, label='_none_') plot(k1['nproc'], k1['rate'], 'd:', color='C1', label='KNL original') plot(k2['nproc'], k2['rate'], 'd-', color='C1', label='KNL') plot(k2['nproc'], k2['nproc']*k2['rate'][0], color='0.5', alpha=0.5, label='perfect scaling') semilogx() semilogy() xticks([1,2,4,8,16,32,64,128], [1,2,4,8,16,32,64,128]) legend(loc='upper left') ylim(20, 20000) xlabel('Number of Processes') ylabel('Extraction Rate') title('25 spectra x 300 wavelengths') savefig('desi-extract-improvements.pdf')

doc/portland-extract.ipynb

sbailey/knltest

bsd-3-clause

Build in plots for storytelling

def labelit(title_=''): semilogx() semilogy() xticks([1,2,4,8,16,32,64,128], [1,2,4,8,16,32,64,128]) legend(loc='upper left') title(title_) ylim(20, 20000) xlabel('Number of Processes') ylabel('Extraction Rate') figure() plot(h1['nproc'], h1['rate'], 's-', color='C0', label='Haswell') plot(h1['nproc'], h1['nproc']*h1['rate'][0], color='0.5', alpha=0.5, label='_none_') plot(k1['nproc'], k1['rate'][0]*np.ones(len(k1['nproc'])), 'd-', color='C1', label='KNL') plot(k1['nproc'], k1['nproc']*k1['rate'][0], color='0.5', alpha=0.5, label='_none_') labelit('Starting point: no KNL scaling') savefig('portland1.pdf') figure() plot(h1['nproc'], h1['rate'], 's-', color='C0', label='Haswell') plot(h1['nproc'], h1['nproc']*h1['rate'][0], color='0.5', alpha=0.5, label='_none_') plot(k1['nproc'], k1['rate'], 'd-', color='C1', label='KNL') plot(k1['nproc'], k1['rate'][0]*np.ones(len(k1['nproc'])), ':', color='C1', label='_none_', alpha=0.8) plot(k1['nproc'], k1['nproc']*k1['rate'][0], color='0.5', alpha=0.5, label='_none_') labelit('After OpenMP bug workaround') savefig('portland2.pdf') figure() plot(h1['nproc'], h1['rate'], ':', color='C0', label='_none_', alpha=0.8) plot(h2['nproc'], h2['rate'], 's-', color='C0', label='Haswell') plot(h2['nproc'], h2['nproc']*h2['rate'][0], color='0.5', alpha=0.5, label='_none_') plot(k1['nproc'], k1['rate'], ':', color='C1', label='_none_', alpha=0.8) plot(k2['nproc'], k2['rate'], 'd-', color='C1', label='KNL') plot(k2['nproc'], k2['nproc']*k2['rate'][0], color='0.5', alpha=0.5, label='_none_') labelit('After Portland Hackfest') savefig('portland3.pdf') print('For 25 spectra x 300 wavelengths...') print('Starting point ratios of HSW/KNL') print(' per processes {:.2f}'.format(h1['rate'][0]/k1['rate'][0])) print(' per node {:.2f}'.format(np.max(h1['rate'])/np.max(k1['rate']))) print('Ending ratios of HSW/KNL') print(' per processes {:.2f}'.format(h2['rate'][0]/k2['rate'][0])) print(' per node {:.2f}'.format(np.max(h2['rate'])/np.max(k2['rate']))) print('Haswell per-node improvement {:.2f}'.format(np.max(h2['rate'])/np.max(h1['rate']))) print('KNL per-node improvement {:.2f}'.format(np.max(k2['rate'])/np.max(k1['rate']))) print('Original HSW / Current KNL {:.2f}'.format(np.max(h1['rate'])/np.max(k2['rate']))) print('Original HSW rate {:.1f}'.format(np.max(h1['rate'])))

doc/portland-extract.ipynb

sbailey/knltest

bsd-3-clause

Configuration parameters

import GAN.utils as utils # reload(utils) class Parameters(utils.Parameters): # dataset to be loaded load_datasets=utils.param(["moriond_v9","abs(ScEta) < 1.5"]) c_names = utils.param(['Phi','ScEta','Pt','rho','run']) x_names = utils.param(['EtaWidth','R9','SigmaIeIe','S4','PhiWidth','CovarianceIetaIphi', 'CovarianceIphiIphi']) # generate variables in MC to be distributed like in data sample_from_data = ['run'] # MC reweighting reweight = ['rewei_zee_5d_barrel_m_cleaning_xgb'] mcweight = utils.param('') # if empty ignore original weight use_abs_weight = utils.param(True) # negative weights are notoriously nasty # cleaning cuts cleaning = utils.param('cleaning_zee_m_barrel.hd5') # input features transformation feat_transform = utils.param('minmax') global_match = utils.param(False) # generator and discriminator (aka as 'critic' in wGAN) g_opts=utils.param(dict(name="G_512to64x8", kernel_sizes=[64,64,128,128,256,256,512,512], do_weight_reg=1e-5,do_last_l1reg=1e-5 )) pretrain_g = utils.param(False) d_opts=utils.param(dict(name="D_1024to32x6",kernel_sizes=[1024,512,128,64,32],do_bn=False, do_dropout=False,hidden_activation="relu", clip_weights=2.e-2,activation=None)) # weights clipping and no actication # optimizers dm_opts=utils.param(dict(optimizer="RMSprop",opt_kwargs=dict(lr=1e-5,decay=1e-6))) am_opts=utils.param(dict(optimizer="RMSprop",opt_kwargs=dict(lr=1e-5,decay=1e-6))) loss = utils.param("wgan_loss") # use WGAN loss gan_targets = utils.param('gan_targets_hinge') # hinge targets are 1, -1 instead of 0, 1 schedule = utils.param([0]*5+[1]*1) # number of critic iterations per generator iterations # schedule training epochs=utils.param(200) batch_size=utils.param(4096) plot_every=utils.param(10) frac_data=utils.param(10) # fraction of data to use monitor_dir = utils.param('log') # folder for logging batch = utils.param(False) # are we running in batch? class MyApp(utils.MyApp): classes = utils.List([Parameters]) # Read all parameters above from command line. # Note: names are all converted to be all capital notebook_parameters = Parameters(MyApp()).get_params() # copy parameters to global scope globals().update(notebook_parameters) DM_OPTS.update( {"loss":LOSS} ) AM_OPTS.update( {"loss":LOSS} ) plotting.batch = BATCH notebook_parameters

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Load datasets Apply cleaning and reweight MC to match data

data,mc = cms.load_zee(*LOAD_DATASETS) data.columns mc.columns

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Cleaning

if not CLEANING is None: print('cleaning data and mc') thr_up = pd.read_hdf(CLEANING,'thr_up') thr_down = pd.read_hdf(CLEANING,'thr_down') nevts_data = data.shape[0] nevts_mc = mc.shape[0] data = data[ ((data[thr_down.index] >= thr_down) & (data[thr_up.index] <= thr_up)).all(axis=1) ] mc = mc[ ((mc[thr_down.index] >= thr_down) & (mc[thr_up.index] <= thr_up)).all(axis=1) ] print('cleaning eff (data,mc): %1.2f % 1.2f' % ( data.shape[0] / nevts_data, mc.shape[0] / nevts_mc ))

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Generate extra variables for MC (eg run number) distributed like data

for col in SAMPLE_FROM_DATA: print('sampling', col) mc[col] = data[col].sample(mc.shape[0]).values

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Select target and conditional features

c_names = C_NAMES x_names = X_NAMES data_c = data[c_names] data_x = data[x_names] mc_c = mc[c_names] mc_x = mc[x_names] data_x.columns, data_x.shape, data_c.columns, data_c.shape data_x.columns, data_c.columns mc_x.columns, mc_c.columns

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Reweight MC

# reload(preprocessing) # reload(utils) # iniliatlize weights to default if MCWEIGHT == '': mc_w = np.ones(mc_x.shape[0]) else: mc_w = mc[MCWEIGHT].values print(mc_w[:10]) # take care of negative weights if USE_ABS_WEIGHT: mc_w = np.abs(mc_w) # reweighting if not REWEIGHT is None: for fil in REWEIGHT: # apply weights from n-dimnesional histograms if 'npy' in fil: info = np.load(fil) inputs = info[0] weights = info[1] bins = info[2:] # print(bins[1]) print('weighting',inputs) mc_w *= preprocessing.reweight(mc,inputs,bins,weights,base=None) else: # or use a classifier clf = utils.read_pickle(fil,encoding='latin1') if hasattr(clf,'nthread'): clf.nthread = min(4,clf.nthread) clf.n_jobs = clf.nthread mc_w *= preprocessing.reweight_multidim(mc[clf.features_],clf) data_w = np.ones(data_x.shape[0])

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Features transformation

# reload(preprocessing) if GLOBAL_MATCH: _,data_c,mc_x,mc_c,scaler_x,scaler_c = preprocessing.transform(None,data_c,mc_x,mc_c,FEAT_TRANSFORM) _,_,data_x,_,scaler_x_data,_ = preprocessing.transform(None,None,data_x,None,FEAT_TRANSFORM) else: data_x,data_c,mc_x,mc_c,scaler_x,scaler_c = preprocessing.transform(data_x,data_c,mc_x,mc_c,FEAT_TRANSFORM) data_x.shape,mc_x.shape data_x[2].max(),mc_x[2].max() print(mc_c.max(axis=0),mc_c.min(axis=0)) print(data_c.max(axis=0),data_c.min(axis=0)) print(mc_w.shape,mc_c.shape) for ic in range(len(c_names)): plotting.plot_hists(data_c[:,0,ic],mc_c[:,0,ic],generated_w=mc_w,bins=100,range=[-2.0,2.0]) plt.xlabel(c_names[ic]) plt.show() for ix in range(len(x_names)): plotting.plot_hists(data_x[:,0,ix],mc_x[:,0,ix],generated_w=mc_w,bins=100,range=[-3,3]) plt.xlabel(x_names[ix]) plt.show()

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Prepare train and test sample

nmax = min(data_x.shape[0]//FRAC_DATA,mc_x.shape[0]) data_x_train,data_x_test,data_c_train,data_c_test,data_w_train,data_w_test = cms.train_test_split(data_x[:nmax],data_c[:nmax],data_w[:nmax],test_size=0.1) mc_x_train,mc_x_test,mc_c_train,mc_c_test,mc_w_train,mc_w_test = cms.train_test_split(mc_x[:nmax],mc_c[:nmax],mc_w[:nmax],test_size=0.1) wscl = data_w[:nmax].sum() / mc_w[:nmax].sum() mc_w_train *= wscl mc_w_test *= wscl print(nmax,mc_w_train.sum()/data_w_train.sum())

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Instantiate the model Create the model and compile it.

# reload(models) xz_shape = (1,len(x_names)) c_shape = (1,len(c_names)) gan = models.MyFFGAN( xz_shape, xz_shape, c_shape=c_shape, g_opts=G_OPTS, d_opts=D_OPTS, dm_opts=DM_OPTS, am_opts=AM_OPTS, gan_targets=GAN_TARGETS )

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

This will actually trigger the instantiation of the generator (if not done here, it will happen befor compilation).

gan.get_generator()

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Same as above for the discriminator. Two discriminator models are created: one for the discriminator training phase and another for the generator trainin phase. The difference between the two is that the second does not contain dropout layers.

gan.get_discriminator()

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

model compilation

gan.adversarial_compile(loss=LOSS,schedule=SCHEDULE) gan.get_generator().summary() gan.get_discriminator()[0].summary() gan.get_discriminator()[1].summary() # sanity check set(gan.am[0].trainable_weights)-set(gan.am[1].trainable_weights)

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

Training Everything is ready. We start training.

from keras.optimizers import RMSprop if PRETRAIN_G: generator = gan.get_generator() generator.compile(loss="mse",optimizer=RMSprop(lr=1.e-3)) generator.fit( [mc_c_train,mc_x_train], [mc_c_train,mc_x_train], sample_weight=[mc_w_train,mc_w_train], epochs=1, batch_size=BATCH_SIZE ) # reload(base) initial_epoch = 0 # if hasattr(gan.model,"history"): # initial_epoch = gan.model.history.epoch[-1] + 1 do = dict( x_train=data_x_train, z_train=mc_x_train, c_x_train=data_c_train, c_z_train=mc_c_train, w_x_train = data_w_train, w_z_train = mc_w_train, x_test=data_x_test, z_test=mc_x_test, c_x_test=data_c_test, c_z_test=mc_c_test, w_x_test = data_w_test, w_z_test = mc_w_test, n_epochs=EPOCHS + initial_epoch - 1, initial_epoch=initial_epoch, batch_size=BATCH_SIZE, plot_every=PLOT_EVERY, monitor_dir=MONITOR_DIR ) exc = None try: gan.adversarial_fit(**do) except Exception as e: exc = e if not exc is None: print(exc)

cms_zee_conditional_wgan.ipynb

musella/GAN

gpl-3.0

### Clustering pearsons_r with HDBSCAN

# Clustering the pearsons_R with N/A vlaues removed hdb_t1 = time.time() hdb_pearson_r = hdbscan.HDBSCAN(metric = "precomputed", min_cluster_size=10).fit(df3_pearson_r) hdb_pearson_r_labels = hdb_pearson_r.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_pearson_r_labels)) # unique bins, zero is noise print(np.bincount(hdb_pearson_r_labels[hdb_pearson_r_labels!=-1])) pearson_clusters = {i: np.where(hdb_pearson_r_labels == i)[0] for i in range(2)} pearson_clusters #pd.set_option('display.height', 500) #These two commands allow for the display of max of 500 rows - exploring genes #pd.set_option('display.max_rows', 500) df2_TPM.iloc[pearson_clusters[1],:] #the genes that were clustered together [0,1]

data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb

gilmana/Cu_transition_time_course-

mit

Looks like there are two clusters, some expression and zero expression across samples. ### Clustering mean centered euclidean distance with with HDBSCAN

df3_euclidean_mean.hist() # Clustering the mean centered euclidean distance of TPM counts hdb_t1 = time.time() hdb_euclidean_mean = hdbscan.HDBSCAN(metric = "precomputed", min_cluster_size=10).fit(df3_euclidean_mean) hdb_euclidean_mean_labels = hdb_euclidean_mean.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_euclidean_mean_labels)) print(np.bincount(hdb_euclidean_mean_labels[hdb_euclidean_mean_labels!=-1])) euclidean_mean_clusters = {i: np.where(hdb_euclidean_mean_labels == i)[0] for i in range(2)} df2_TPM.iloc[euclidean_mean_clusters[1],:]

data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb

gilmana/Cu_transition_time_course-

mit

Looks like 2 clusters - both with zero expression. looks like wether it is a numpy array or pandas dataframe, the result is the same. lets now try to get index of the clustered points. ### Clustering log transformed euclidean distance with with HDBSCAN

df3_euclidean_log2 # Clustering the log2 transformed euclidean distance of TPM counts hdb_t1 = time.time() hdb_euclidean_log2 = hdbscan.HDBSCAN(metric = "precomputed", min_cluster_size=10).fit(df3_euclidean_log2) hdb_euclidean_log2_labels = hdb_euclidean_log2.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_euclidean_log2_labels)) print(np.bincount(hdb_euclidean_log2_labels[hdb_euclidean_log2_labels!=-1])) euclidean_log2_clusters = {i: np.where(hdb_euclidean_log2_labels == i)[0] for i in range(2)} df2_TPM.iloc[euclidean_log2_clusters[1],:]

data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb

gilmana/Cu_transition_time_course-

mit

Clustering using built-in HDBSCAN euclidean distance metric (mean centered and scaled to unit variance)

df2_TPM_values = df2_TPM.loc[:,"5GB1_FM40_T0m_TR2":"5GB1_FM40_T180m_TR1"] #isolating the data values df2_TPM_values_T = df2_TPM_values.T #transposing the data standard_scaler = StandardScaler() TPM_counts_mean_centered = standard_scaler.fit_transform(df2_TPM_values_T) #mean centering the data TPM_counts_mean_centered = pd.DataFrame(TPM_counts_mean_centered) #back to Dataframe #transposing back to original form and reincerting indeces and columns my_index = df2_TPM_values.index my_columns = df2_TPM_values.columns TPM_counts_mean_centered = TPM_counts_mean_centered.T TPM_counts_mean_centered.set_index(my_index, inplace=True) TPM_counts_mean_centered.columns = my_columns # Clustering the pearsons_R with N/A vlaues removed hdb_t1 = time.time() hdb_euclidean = hdbscan.HDBSCAN(metric = "euclidean", min_cluster_size=5).fit(TPM_counts_mean_centered) hdb_euclidean_labels = hdb_euclidean.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_euclidean_labels)) print(np.bincount(hdb_euclidean_labels[hdb_euclidean_labels!=-1]))

data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb

gilmana/Cu_transition_time_course-

mit

lets look at some clusters Euclidean_standard_scaled_clusters = {i: np.where(hdb_euclidean_labels == i)[0] for i in range(7)} df2_TPM.iloc[Euclidean_standard_scaled_clusters[0],:]

Euclidean_standard_scaled_clusters = {i: np.where(hdb_euclidean_labels == i)[0] for i in range(7)} df2_TPM.iloc[Euclidean_standard_scaled_clusters[1],:]

data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb

gilmana/Cu_transition_time_course-

mit

Euclidean_standard_scaled_clusters Clustering log2 transformed data using built-in HDBSCAN euclidean distance metric (mean centered and scaled to unit variance)

df2_TPM_log2_scale= df2_TPM_log2.T #transposing the data standard_scaler = StandardScaler() TPM_log2_mean_scaled = standard_scaler.fit_transform(df2_TPM_log2_scale) #mean centering the data TPM_log2_mean_scaled = pd.DataFrame(TPM_log2_mean_scaled) #back to Dataframe #transposing back to original form and reincerting indeces and columns my_index = df2_TPM_values.index my_columns = df2_TPM_values.columns TPM_log2_mean_scaled = TPM_log2_mean_scaled.T TPM_log2_mean_scaled.set_index(my_index, inplace=True) TPM_log2_mean_scaled.columns = my_columns # Clustering the pearsons_R with N/A vlaues removed hdb_t1 = time.time() hdb_log2_euclidean = hdbscan.HDBSCAN(metric = "euclidean", min_cluster_size=5).fit(TPM_log2_mean_scaled) hdb_log2_euclidean = hdb_log2_euclidean.labels_ hdb_elapsed_time = time.time() - hdb_t1 print("time to cluster", hdb_elapsed_time) print(np.unique(hdb_log2_euclidean)) print(np.bincount(hdb_log2_euclidean[hdb_log2_euclidean!=-1]))

data_explore_failed_clstr_mthds/HDBSCAN_clustering.ipynb

gilmana/Cu_transition_time_course-

mit

DM -- Piece by piece (as coded) $\rho_b = \Omega_b \rho_c (1+z)^3$ $\rho_{\rm diffuse} = \rho_b - (\rho_* + \rho_{\rm ISM})$ $\rho_*$ is the mass density in stars $\rho_{\rm ISM}$ is the mass density in the neutral ISM Number densities $n_{\rm H} = \rho_{\rm diffuse}/(m_p \, \mu)$ $\mu \approx 1.3$ accounts for Helium $n_{\rm He} = n_{\rm H}/12$ $n_e = n_{\rm H} [1-f_{\rm HI}] + n_{\rm He} Y$ $f_{\rm HI}$ is the fraction atomic Hydrogen [value betwee 0-1] $Y$ gives the number of free electrons per He nucleus [value between 0-2] Integrating $DM = \int \frac{n_e \, dr}{1+z} = \frac{c}{H_0} \int \frac{n_e \, dz}{(1+z)^2 \sqrt{(1+z)^3 \Omega_m + \Omega_\Lambda}}$ DM -- Altogether (using the code)

reload(igm) DM = igm.average_DM(1.) DM

docs/nb/DM_cosmic.ipynb

FRBs/FRB

bsd-3-clause

Cumulative plot

DM_cumul, zeval = igm.average_DM(1., cumul=True) # Inoue approximation DM_approx = 1000. * zeval * units.pc / units.cm**3 plt.clf() ax = plt.gca() ax.plot(zeval, DM_cumul, label='JXP') ax.plot(zeval, DM_approx, label='Approx') # Label ax.set_xlabel('z') ax.set_ylabel(r'${\rm DM}_{\rm IGM} [\rm pc / cm^3]$ ') # Legend legend = plt.legend(loc='lower right', scatterpoints=1, borderpad=0.2, handletextpad=0.1, fontsize='large') plt.show() plt.clf() ax = plt.gca() ax.plot(zeval, DM_approx/DM_cumul, label='Approx/JXP') #ax.plot(zeval, DM_approx, label='Approx') # Label ax.set_xlabel('z') ax.set_ylabel(r'Ratio of ${\rm DM}_{\rm IGM} [\rm pc / cm^3]$ ') # Legend legend = plt.legend(loc='upper right', scatterpoints=1, borderpad=0.2, handletextpad=0.1, fontsize='large') plt.show() DM_cumul[0:10] DM_approx[0:10] zeval[0]

docs/nb/DM_cosmic.ipynb

FRBs/FRB

bsd-3-clause

The Longley dataset is a time series dataset:

data = sm.datasets.longley.load() data.exog = sm.add_constant(data.exog) print(data.exog[:5])

examples/notebooks/gls.ipynb

ChadFulton/statsmodels

bsd-3-clause

While we don't have strong evidence that the errors follow an AR(1) process we continue

rho = resid_fit.params[1]

examples/notebooks/gls.ipynb

ChadFulton/statsmodels

bsd-3-clause

2x2 Grid You can easily create a layout with 4 widgets arranged on 2x2 matrix using the TwoByTwoLayout widget:

from ipywidgets import TwoByTwoLayout TwoByTwoLayout(top_left=top_left_button, top_right=top_right_button, bottom_left=bottom_left_button, bottom_right=bottom_right_button)