Merge ca5fbea into 3f1fe33

PlotMNIST.jl

@@ -0,0 +1,90 @@
+### A Pluto.jl notebook ###
+# v0.14.5
+
+using Markdown
+using InteractiveUtils
+
+# ╔═╡ e17ac940-bdbc-11eb-1f8c-41b46d5552c8
+begin
+	using UMAP: umap
+	using MLDatasets
+	using Plots
+	using PyCall
+end
+
+# ╔═╡ 8e5644c2-219f-41f5-bd57-69a32fe0ea35
+md"""
+# UMAP for MNIST and FashionMNIST
+"""
+
+# ╔═╡ 6686a26a-8f33-4300-8dd2-80890f8138fc
+@pyimport umap as py_umap;
+
+# ╔═╡ 19e30f46-bcd1-4403-a986-eb5862c8ffa4
+md"""
+### 1) MNIST
+"""
+
+# ╔═╡ 2873c13b-f583-478a-a183-70a507e06aa2
+mnist_x = reshape(MNIST.traintensor(Float64),28*28,:);
+
+# ╔═╡ 4c632fe2-9213-48e7-b6bc-5d2fd1d45ec6
+mnist_y = MNIST.trainlabels(1:size(mnist_x, 2));
+
+# ╔═╡ 2eb7cd88-7b04-4ec1-9dca-35ae604c0311
+@time res_jl = umap(mnist_x; n_neighbors=10, min_dist=0.001, n_epochs=200);
+
+# ╔═╡ cf976fc2-5563-4f0c-9a53-26a8ca7fad5c
+scatter(res_jl[1,:], res_jl[2,:], zcolor=mnist_y, 
+        title="MNIST: Julia UMAP", marker=(2, 2, :auto, stroke(0)))
+
+# ╔═╡ b1bfb601-162d-4994-8e18-510ab13bdd22
+@time res_py = py_umap.UMAP(n_neighbors=10, min_dist=0.001, n_epochs=200)[:fit_transform](permutedims(mnist_x));
+
+# ╔═╡ 52885e4d-ebb1-4be9-af9c-3de6f30e33d2
+scatter(res_py[:,1], res_py[:,2], zcolor=mnist_y, 
+        title="MNIST: Python UMAP", marker=(2, 2, :auto, stroke(0)))
+
+# ╔═╡ 8bdd4ed8-4841-4634-a050-913cf08de1aa
+md"""
+### 2) FashionMNIST
+"""
+
+# ╔═╡ 2d98a3cd-3451-47e5-9a9a-9a0dde61dc42
+fmnist_x = reshape(FashionMNIST.traintensor(Float64),28*28,:);
+
+# ╔═╡ 9490cc37-fcbc-4b0b-8ba9-7a48f8ac3c3d
+fmnist_y = FashionMNIST.trainlabels(1:size(fmnist_x, 2));
+
+# ╔═╡ c441fcd2-334e-4a1c-b476-09ade0274cf5
+@time res_jl2 = umap(fmnist_x; n_neighbors=5, min_dist=0.1, n_epochs=200);
+
+# ╔═╡ eecd01b4-3bf2-48a4-8482-f9fc5d08e051
+scatter(res_jl2[1,:], res_jl2[2,:], zcolor=fmnist_y, 
+        title="FMNIST: Julia UMAP", marker=(2, 2, :auto, stroke(0)))
+
+# ╔═╡ 23358904-df72-4e7e-8b13-c97966407512
+@time res_py2 = py_umap.UMAP(n_neighbors=5, min_dist=0.1, n_epochs=200)[:fit_transform](permutedims(fmnist_x));
+
+# ╔═╡ faa1e53d-5d98-46bb-b0af-cac647e96e48
+scatter(res_py2[:,1], res_py2[:,2], zcolor=fmnist_y, 
+        title="FMNIST: Python UMAP", marker=(2, 2, :auto, stroke(0)))
+
+# ╔═╡ Cell order:
+# ╟─8e5644c2-219f-41f5-bd57-69a32fe0ea35
+# ╠═e17ac940-bdbc-11eb-1f8c-41b46d5552c8
+# ╠═6686a26a-8f33-4300-8dd2-80890f8138fc
+# ╟─19e30f46-bcd1-4403-a986-eb5862c8ffa4
+# ╠═2873c13b-f583-478a-a183-70a507e06aa2
+# ╠═4c632fe2-9213-48e7-b6bc-5d2fd1d45ec6
+# ╠═2eb7cd88-7b04-4ec1-9dca-35ae604c0311
+# ╠═cf976fc2-5563-4f0c-9a53-26a8ca7fad5c
+# ╠═b1bfb601-162d-4994-8e18-510ab13bdd22
+# ╠═52885e4d-ebb1-4be9-af9c-3de6f30e33d2
+# ╟─8bdd4ed8-4841-4634-a050-913cf08de1aa
+# ╠═2d98a3cd-3451-47e5-9a9a-9a0dde61dc42
+# ╠═9490cc37-fcbc-4b0b-8ba9-7a48f8ac3c3d
+# ╠═c441fcd2-334e-4a1c-b476-09ade0274cf5
+# ╠═eecd01b4-3bf2-48a4-8482-f9fc5d08e051
+# ╠═23358904-df72-4e7e-8b13-c97966407512
+# ╠═faa1e53d-5d98-46bb-b0af-cac647e96e48

README.md

@@ -67,7 +67,7 @@ The low-dimensional embedding is initialized (by default) with the eigenvectors
 - If you're familiar with [t-SNE](https://lvdmaaten.github.io/tsne/), then [this page](https://jlmelville.github.io/uwot/umap-for-tsne.html) describes UMAP with similar vocabulary to that dimension reduction algorithm
 
 ## Examples
-The full MNIST and FMNIST datasets are plotted below using both this implementation and the [Python implementation](github.com/lmcinnes/umap) for comparison. These were generated by [this notebook](PlotMNIST.ipynb).
+The full MNIST and FMNIST datasets are plotted below using both this implementation and the [Python implementation](github.com/lmcinnes/umap) for comparison. These were generated by [this Pluto notebook](PlotMNIST.jl).
 
 Note that the memory allocation for the Python UMAP is unreliable, as Julia's benchmarking doesn't count memory allocated within Python itself.
 ### MNIST