gaphex’s gists

gaphex / euclidean_distance.py

Created June 13, 2019 15:13

euclidean distance computation

	squared_distance = tf.reduce_sum(tf.pow(Q - S, 2), reduction_indices=1)
	distance = tf.sqrt(squared_distance)

gaphex / get_top_k.py

Last active June 20, 2019 13:46

top_k distances and indices retrieval

	top_k = 3

	top_neg_dists, top_indices = tf.math.top_k(tf.negative(distance), k=top_k)
	top_dists = tf.negative(top_neg_dists)

gaphex / tf_retriever.py

Last active June 20, 2019 12:41

Nearest Neighbour retriever

	class TFRanker:
	def __init__(self, dim, metric, top_k=3):
	self.dim = dim
	self.top_k = top_k
	self.metric = metric
	self.graph = tf.Graph()
	self.session = tf.Session(graph=self.graph)
	self.build_graph()

gaphex / euclidean.md

Last active June 14, 2019 08:26

>$$e^x=\sum_{i=0}^\infty \frac{1}{i!}x^i$$

gaphex / dot_l2_distances.py

Last active June 28, 2019 19:20

	Q = tf.placeholder("float", [dim])
	S = tf.placeholder("float", [None, dim])

	Qr = tf.reshape(Q, (1, -1))

	PP = tf.keras.backend.batch_dot(S, S, axes=1)
	QQ = tf.matmul(Qr, tf.transpose(Qr))
	PQ = tf.matmul(S, tf.transpose(Qr))

	distance = PP - 2 * PQ + QQ

gaphex / reuters_bert_embeddings.tsv

Created June 22, 2019 18:09

We can't make this file beautiful and searchable because it's too large.

-0.02492	-0.01458	0.1818	0.03638	0.4434	-0.299	-0.3193	0.04535	0.03632	0.3071	-0.1837	-0.2976	-0.006374	0.712	-0.06793	0.6587	0.3665	-0.02434	0.07965	0.2109	-0.02144	-0.3918	0.4436	0.256	0.1763	0.01851	0.2637	0.02402	-0.604	-0.007587	0.4468	0.564	-0.2263	-0.2507	0.1254	-0.11273	0.005077	-0.2065	-0.5293	-0.1367	-0.6157	-0.2585	-0.2081	-0.412	0.1055	-0.1884	0.4023	0.0169	-0.1935	-0.02063	-0.4822	-0.05426	-0.6025	-0.613	0.2583	0.738	-0.2817	-0.2483	0.1346	-0.09863	0.09424	-0.1578	0.3335	-0.1814	0.312	0.01714	-0.1401	0.09564	-0.1921	0.1261	-0.11487	-0.1396	-0.1423	0.301	0.08057	-0.1716	-0.2308	0.0813	-0.1829	0.3032	-0.209	-0.2654	0.10913	-0.4744	0.1678	0.2065	-0.04	0.2131	-0.0995	0.7812	0.04114	0.4329	-0.2094	0.1263	0.01222	-0.522	-0.4272	-0.291	0.05374	0.2788	0.05502	0.03622	0.3337	0.635	-0.1764	-0.397	-0.2678	-0.07465	-0.00988	0.1323	-0.2157	-0.0542	0.206	-0.213	-0.4614	-0.1632	0.252	0.288	0.246	0.2472	0.103	0.01074	0.2651	0.2769	0.1927	-0.1417	0.0927	-0.1571	0.003601	0.638	0.0479	-0.3318	-0.1776	-0.011246	-0.2

gaphex / reuters_bert_metadata.tsv

Last active June 22, 2019 19:43

	Label	Sentence
	wheat	SRI LANKA GETS USDA APPROVAL FOR WHEAT PRICE Food Department officials said the U . S . Department of Agriculture approved the Continental Grain Co sale of 52 , 500 tonnes of soft wheat at 89 U . S . Dlrs a tonne C and F from Pacific Northwest to Colombo .
	wheat	They said the shipment was for April 8 to 20 delivery .
	wheat	FURTHER ARGENTINE COARSE GRAIN LOSSES FEARED Argentine grain producers adjusted their yield estimates for the 1986 / 87 coarse grain crop downward in the week to yesterday after the heavy rains at the end of March and beginning of April , trade sources said .
	wheat	They said sunflower , maize and sorghum production estimates had been reduced despite some later warm , dry weather , which has allowed a return to harvesting in some areas .
	wheat	However , as showers fell intermittently after last weekend , producers feared another spell of prolonged and intense rain could cause more damage to crops already badly hit this season .
	wheat	Rains in the middle of last week reac

gaphex / reuters_bert_projector.json

Last active June 22, 2019 18:24

	{
	"embeddings": [
	{
	"tensorName": "My tensor",
	"tensorShape": [
	1134,
	768
	],
	"tensorPath": "https://gist.githubusercontent.com/gaphex/7f3daf793e56d6563bbf1df4ffddb88f/raw/2dba89b44e80c31aa82ee800d07d0c5d972b3275/reuters_bert_embeddings.tsv",
	"metadataPath": "https://gist.githubusercontent.com/gaphex/b3d2838e56349d5b53a3011000e2f35e/raw/5d3ce116fb08fb6fdc7fc286053f8e01a44da0fe/reuters_bert_metadata.tsv",

gaphex / reuters_bookmark.txt

Last active June 22, 2019 18:54

[{"label":"State 0","isSelected":true,"tSNEIteration":1553,"tSNEPerplexity":9,"tSNELearningRate":1,"tSNEis3d":true,"umapIs3d":true,"umapNeighbors":15,"pcaComponentDimensions":[0,1,2],"projections":[{"pca-0":0.5753942131996155,"pca-1":-7.144152641296387,"pca-2":2.135350465774536,"pca-3":0.30271515250205994,"pca-4":-7.306628227233887,"pca-5":2.3445184230804443,"pca-6":5.1897382736206055,"pca-7":-1.710606575012207,"pca-8":0.14948318898677826,"pca-9":-1.2810907363891602,"tsne-0":-19.17779998293605,"tsne-1":-8.737170856322848,"tsne-2":3.4405034737394518},{"pca-0":-3.3178958892822266,"pca-1":-1.894553303718567,"pca-2":-3.520504951477051,"pca-3":-1.990181565284729,"pca-4":0.030808638781309128,"pca-5":-2.9290878772735596,"pca-6":1.1588380336761475,"pca-7":-1.9897701740264893,"pca-8":-1.3593165874481201,"pca-9":-2.974226236343384,"tsne-0":-2.792864945677704,"tsne-1":46.16809832776983,"tsne-2":-34.11351794923433},{"pca-0":-1.2383508682250977,"pca-1":3.416457176208496,"pca-2":-2.6669344902038574,"pca-3":0.48395350575447

gaphex / reuters_proj_config.json

Last active June 22, 2019 19:44

	{
	"embeddings": [
	{
	"tensorName": "Reuters",
	"tensorShape": [
	1134,
	768
	],
	"tensorPath": "https://gist.githubusercontent.com/gaphex/7f3daf793e56d6563bbf1df4ffddb88f/raw/2dba89b44e80c31aa82ee800d07d0c5d972b3275/reuters_bert_embeddings.tsv",
	"metadataPath": "https://gist.githubusercontent.com/gaphex/b3d2838e56349d5b53a3011000e2f35e/raw/c597282ef1d5f238eb22a0371d321c88a3c7e0ce/reuters_bert_metadata.tsv",

Denis gaphex