diff --git a/docs/functions/convertHTMLSpecialChars.html b/docs/functions/convertHTMLSpecialChars.html
index d97e349..1dda1dc 100644
--- a/docs/functions/convertHTMLSpecialChars.html
+++ b/docs/functions/convertHTMLSpecialChars.html
@@ -4,10 +4,10 @@
The processed string.
-
JS-PyTorch is a neural net matrix multiplication library -with GPU.js acceleration (translates matmul into WebGPU shader code) -and using PyTorch API syntax. -torch -Tensor Creation and Manipulation:
+- Preparing search index...
- The search index is not available
ai-research-agentClass torch
Torch is a neural net matrix multiplication library
++- Uses PyTorch API syntax for tensors and neural nets.
+- Uses GPU.js acceleration to translate matmul into WebGL shader code.
+- Neural Net API: MultiHeadSelfAttention, FullyConnected, Block,
+Embedding, PositionalEmbedding, ReLU, Softmax, Dropout, LayerNorm, CrossEntropyLoss
+
+torch
Function
tensor(data, requires_grad = false, device = 'cpu') Creates a new Tensor filled with the given data
Function
zeros(*shape, requires_grad = false, device = 'cpu') Creates a new Tensor filled with zeros
Function
ones(*shape, requires_grad = false, device = 'cpu') Creates a new Tensor filled with ones
@@ -31,61 +34,62 @@Function<
Function
sqrt(a) Returns element-wise square root of the tensor
Function
exp(a) Returns element-wise exponentiation of the tensor
Function
log(a) Returns element-wise natural log of the tensor
-Neural Network Layers: -nn.Linear(in_size, out_size, device, bias, xavier) Applies a linear transformation to the input tensor -nn.MultiHeadSelfAttention(in_size, out_size, n_heads, n_timesteps, dropout_prob, device) Applies a self-attention layer on the input tensor -nn.FullyConnected(in_size, out_size, dropout_prob, device, bias) Applies a fully-connected layer on the input tensor -nn.Block(in_size, out_size, n_heads, n_timesteps, dropout_prob, device) Applies a transformer Block layer on the input tensor -nn.Embedding(in_size, embed_size) Creates an embedding table for vocabulary -nn.PositionalEmbedding(input_size, embed_size) Creates a positional embedding table -nn.ReLU() Applies Rectified Linear Unit activation function -nn.Softmax() Applies Softmax activation function -nn.Dropout(drop_prob) Applies dropout to input tensor -nn.LayerNorm(n_embed) Applies Layer Normalization to input tensor -nn.CrossEntropyLoss() Computes Cross Entropy Loss between target and input tensor
+torch.nn +Neural Network Layers:
+Method
nn.Linear(in_size, out_size, device, bias, xavier) Applies a linear transformation to the input tensor
+Method
nn.MultiHeadSelfAttention(in_size, out_size, n_heads, n_timesteps, dropout_prob, device) Applies a self-attention layer on the input tensor
+Function
nn.FullyConnected(in_size, out_size, dropout_prob, device, bias) Applies a fully-connected layer on the input tensor
+Function
nn.Block(in_size, out_size, n_heads, n_timesteps, dropout_prob, device) Applies a transformer Block layer on the input tensor
+Function
nn.Embedding(in_size, embed_size) Creates an embedding table for vocabulary
+Function
nn.PositionalEmbedding(input_size, embed_size) Creates a positional embedding table
+Function
nn.ReLU() Applies Rectified Linear Unit activation function
+Function
nn.Softmax() Applies Softmax activation function
+Function
nn.Dropout(drop_prob) Applies dropout to input tensor
+Function
nn.LayerNorm(n_embed) Applies Layer Normalization to input tensor
+Function
nn.CrossEntropyLoss() Computes Cross Entropy Loss between target and input tensor
Optimization: optim.Adam(params, lr, reg, betas, eps) Adam optimizer for updating model parameters
Utility Functions:
-Function
save(model, file) Saves the model reruning data blob (for you to save)
-Function
load(model, loadedData) Loads the model from saved data
+Function
save(model, file) Saves the model reruning data blob (for you to save)
+Function
load(model, loadedData) Loads the model from saved data
Author
PyTorch Contributors, Leao, E. et al (2022), See also: Brain.js
-Index
Properties
Index
Properties
Properties
Static_reshapeStaticaddStaticatStaticbroadcastStaticdivStaticexpStaticgetType declaration
import { GPU } from "@eduardoleao052/gpu"
-Parameters
Returns any[]
StaticloadStaticlogStaticmasked_StaticmatmulStaticmeanStaticmulStaticnegStaticnnBlock: typeof Block;
CrossEntropyLoss: typeof CrossEntropyLoss;
Dropout: typeof Dropout;
Embedding: typeof Embedding;
FullyConnected: typeof FullyConnected;
LayerNorm: typeof LayerNorm;
Linear: typeof Linear;
Module: typeof Module;
MultiHeadSelfAttention: typeof MultiHeadSelfAttention;
PositionalEmbedding: typeof PositionalEmbedding;
ReLU: typeof ReLU;
Softmax: typeof Softmax;
}
Add submodules:
-StaticonesStaticoptimAdam: typeof Adam;
}
StaticParameterStaticpowStaticrandStaticrandintStaticrandnStaticreshapeStaticsaveStaticsqrtStatictensorStaticTensorAdd methods from tensor.js (these methods are accessed with "torch."):
-StatictransposeStatictrilStaticvarianceStaticzerosSettings
On This Page
Properties
- Preparing search index...
- The search index is not available
ai-research-agentFunction addEmbeddingVectorsToIndex
VSEARCH: Vector Similarity Embedding Approximation in RAM-Limited Cluster Heirarchy
+- Compile hnswlib-node or NGT algorithm C++ to WASM JS for efficient similarity search.
+- Vector index is split by K-means into regional clusters, each being a
+specific size to fit in RAM. This is better than popular vector engines that
+require costly 100gb-RAM servers because they load all the vectors at once.
+- Vectors for centroids of each cluster are stored in a list in SQL, each
+cluster's binary quantized data is exported as base64 string to SQL, S3, etc.
+- Search: Embed Query, Compare to each cluster centroid to pick top clusters,
+download base64 strings for those clusters, load each into WASM, find top neighbors
+per cluster, merge results sorted by distance.
+
+NGT Algorithm +NGT Cluster
+Vald Vector Engine Docs +ANN Benchmarks
+Parameters
An array of document texts to be vectorized.
+Optionaloptions: {numDimensions: number;
maxElements: number;
} = {}
Optional parameters for vector generation and indexing.
+num
The length of data point vector that will be indexed.
+max
The maximum number of data points.
+Returns Promise<HierarchicalNSW>
The created HNSW index.
+Author
Malkov et al. (2016), +*
+Settings
On This Page
- Preparing search index...
- The search index is not available
ai-research-agentFunction convertEmbeddingsToHNSW
Generates vectors for a set of documents and creates an HNSW index using -hnswlib in C++ compiled to WASM JS for efficient similarity search.
-ANN Benchmarks
-https://github.com/brtholomy/hnsw -Pinecone - HNSW
-Parameters
An array of document texts to be vectorized.
-Optionaloptions: {maxElements: number;
numDimensions: number;
} = {}
Optional parameters for vector generation and indexing.
-max
The maximum number of data points.
-num
The length of data point vector that will be indexed.
-Returns Promise<HierarchicalNSW>
The created HNSW index.
-Author
Malkov, Y. et al (2016), -Tatsuma, Y. et al (2022) -*
-Settings
Understanding UMAP
-
+UMAP Algorithm Overview
Parameters
The dictionary of embeddings.
-Optionaloptions: {numberDimensions: number;
numberDistance: number;
numberNeighbors: number;
} = {}
number
[default=2] - The number of dimensions for UMAP output.
-number
[default=0.1] - The minimum distance parameter for UMAP.
+Optionaloptions: {numberDimensions: number;
numberNeighbors: number;
numberDistance: number;
} = {}
number
[default=2] - The number of dimensions for UMAP output.
number
[default=15] - The number of nearest neighbors for UMAP.
+number
[default=0.1] - The minimum distance parameter for UMAP.
Returns Promise<PlotDataPoint[]>
An array of plot data points.
Author
McInnes et al. (2018)
-Coenen et al. (2019)
Settings
On This Page