Owl API Cheatsheet¶

Category	Description [1]	Function
Element-wise	basics	`+ - * /`
	exponential	`owl.elewise.exp(x)`
	ln	`owl.elewise.ln(x)`
	sigmoid	`owl.elewise.sigm(x)`
	sigmoid back	`owl.elewise.sigm_back(y)`
	tanh	`owl.elewise.tanh(x)`
	tanh back	`owl.elewise.tanh_back(y)`
	relu	`owl.elewise.relu(x)`
	relu back	`owl.elewise.relu_back(y, x)`
Matrix	multiplication	`*`
Matrix	transpose	`.trans()`
Reduction [2]	sum	`.sum(axis)`
	max	`.max(axis)`
	argmax	`.argmax(axis)`
	count zero	`.count_zero()`
Broadcast [3]	broadcast	`+ - * /`
Generator	fill zeros	`owl.zeros(shape)`
	fill ones	`owl.ones(shape)`
	normal distribution	`owl.randn(shape, mu, var)`
	bernoulli distribution	`owl.randb(shape, prob)`
Shape	reshape	`.reshape(shape)`
	slice	`owl.slice(src, dim, off, count)`
	concat	`owl.concat(arrays, dim)`
I/O	convert to numpy	`.to_numpy()`
I/O	convert from numpy	`owl.from_numpy(nparr)`
Convnet	convolution (ff/bp/grad)	`owl.conv.Convolver`
	pooling (ff/bp)	`owl.conv.Pooler`
	lrn (ff/bp)	`owl.conv.Lrner`
	softmax	`owl.conv.softmax(x, op)`
Model & Trainer	neural network definition	`owl.net.Net`
	build network from Caffe’s config	`owl.net.CaffeNetBuilder`
	multi-gpu trainer	`owl.net.NetTrainer`
	feature extractor	`owl.net.FeatureExtractor`
	multi-view tester	`owl.net.MultiviewTester`
System	initialize system	`owl.initialize(argv)`
	wait for all laziness to finish	`owl.wait_for_all()`
	create cpu device	`owl.create_cpu_device()`
	create gpu device	`owl.create_gpu_device(id)`
	choose device to compute	`owl.set_device(dev)`

[1]	Descriptions of non-lazy operations are in bold text.

[2]	Currently, reduction (`sum`, `max`, `argmax`) is only allowed on 2-D matrix along row or column.

[3]	Definition of broadcast. Currently, broadcast is only allowed on 2-D matrix along row or column.