4 ms·

Recently was trying to convert Python to C#... there are few things that are as absolutely incomprehensible as a Python programmer being clever with numpy. I ga

Recently was trying to convert Python to C#... there are few things that are as absolutely incomprehensible as a Python programmer being clever with numpy. I gave up on the conversion. Perhaps this is a point for Python’s flexibility, but I’d say it’s just a waste. Here’s the code. Try to convert this to any C style language. mins = X.min(axis=0) maxs = X.max(axis=0) idxs = np.random.choice(range(self.dim), self.dim-self.exlevel-1, replace=False) # Pick the indices for which the normal vector elements should be set to zero acccording to the extension level. self.n = np.random.normal(0,1,self.dim) # A random normal vector picked form a uniform n-sphere. Note that in order to pick uniformly from n-sphere, we need to pick a random normal for each component of this vector. self.n[idxs] = 0 self.p = np.random.uniform(mins,maxs) # Picking a random intercept point for the hyperplane splitting data. w = (X-self.p).dot(self.n) < 0 # Criteria that determines if a data point should go to the left or right child node. return Node(X, self.n, self.p, e,\ left=self.make_tree(X[w],e+1,l),\ right=self.make_tree(X[~w],e+1,l),\ node_type = 'inNode' )

adrianN 7y ago
Try converting typical Matlab code to a C-style language and you'll run into similar problems. That's just what you get when your base language isn't all that fast but you have a fast numerics library that allows you to do clever things with arrays and matrices.
TheDong 7y ago
Your challenge isn't possible because the code snippet isn't complete. It doesn't include the definition of 'make_tree', 'Node', or the broader context. That being said, I have no doubt that C++ with Boost would look quite similar to that. The boost libraries have very powerful numeric constructs and can easily handle matrices, uniform random distributions, dot products, and trees.
- lwansbrough 7y ago
  Well, yeah. My point is the syntax itself is very (too) clever. This syntax is probably possible in C# but you’d be a monster to write it (lots of operator overloads.) The whole implementation is from an implementation of an extended isolation forest. It’s easy enough to find on GitHub if you’re interested.
pjmlp 7y ago
Here is the C# version of that snippet, assuming the existence of such libraries. var mins = X.Min(axis: 0); var maxs = X.Max(axis: 0); var idxs = np.Random.Choice(Enunmeration.Range(this.dim), this.dim-this.exlevel-1, replace: false); // Pick the indices for which the normal vector elements should be set to zero acccording to the extension level. this.n = np.Random.Normal(0,1, this.dim); // A random normal vector picked form a uniform n-sphere. Note that in order to pick uniformly from n-sphere, we need to pick a random normal for each component of this vector. this.n[idxs] = 0; this.p = np.Random.Uniform(mins,maxs); // Picking a random intercept point for the hyperplane splitting data. var w = (X-this.p).Dot(this.n) < 0; // Criteria that determines if a data point should go to the left or right child node. return new Node(X, this.n, this.p, e, left: this.MakeTree(X[w],e+1,l), right: this.MakeTree(X[~w],e+1,l), node_type: "inNode") There is nothing about Python flexibility there other than a possible missing library implementation.

adrianN 7y ago

Try converting typical Matlab code to a C-style language and you'll run into similar problems. That's just what you get when your base language isn't all that fast but you have a fast numerics library that allows you to do clever things with arrays and matrices.

TheDong 7y ago

Your challenge isn't possible because the code snippet isn't complete. It doesn't include the definition of 'make_tree', 'Node', or the broader context. That being said, I have no doubt that C++ with Boost would look quite similar to that. The boost libraries have very powerful numeric constructs and can easily handle matrices, uniform random distributions, dot products, and trees.

lwansbrough 7y ago
Well, yeah. My point is the syntax itself is very (too) clever. This syntax is probably possible in C# but you’d be a monster to write it (lots of operator overloads.) The whole implementation is from an implementation of an extended isolation forest. It’s easy enough to find on GitHub if you’re interested.

pjmlp 7y ago

Here is the C# version of that snippet, assuming the existence of such libraries. var mins = X.Min(axis: 0); var maxs = X.Max(axis: 0); var idxs = np.Random.Choice(Enunmeration.Range(this.dim), this.dim-this.exlevel-1, replace: false); // Pick the indices for which the normal vector elements should be set to zero acccording to the extension level. this.n = np.Random.Normal(0,1, this.dim); // A random normal vector picked form a uniform n-sphere. Note that in order to pick uniformly from n-sphere, we need to pick a random normal for each component of this vector. this.n[idxs] = 0; this.p = np.Random.Uniform(mins,maxs); // Picking a random intercept point for the hyperplane splitting data. var w = (X-this.p).Dot(this.n) < 0; // Criteria that determines if a data point should go to the left or right child node. return new Node(X, this.n, this.p, e, left: this.MakeTree(X[w],e+1,l), right: this.MakeTree(X[~w],e+1,l), node_type: "inNode") There is nothing about Python flexibility there other than a possible missing library implementation.