@stellarapp/tfjs-stellar 1.0.4 → 1.0.5

package/dist/src/layers/multihead_attention.js

@@ -1,269 +0,0 @@
-import * as tf from '@tensorflow/tfjs';
-import { causal as generateCausalMask } from "@/masks";
-/**
- * This MultiHead Attention layer implements the algorithm as described in
- * the paper "Attention is all you Need" Vaswani et al., 2017.
- *
- * @param numHeads number of attention heads to use
- * @param embedDim the embedding size of the input (input embeddings, typically the last dimension)
- * @param causal use causal masking, default `false`
- * @param dropout use dropout during the attention calculations, default `0.0`
- * @param useBias use bias for the dense sublayers, default `true`
- *
- * The TensorFlow version uses tf.einsum, whose gradient op has not yet been
- * implemented (https://github.com/tensorflow/tfjs/pull/4955#discussion_r619219334),
- * therefore we follow the PyTorch implementation described in:
- * https://docs.pytorch.org/tutorials/intermediate/transformer_building_blocks.html#multiheadattention
- * https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
- *
- * This implementation is different from TensorFlow's whose attention weights
- * are shaped [embed dim, heads, embed dim] where as PyTorch and OpenAI's attention weights
- * are shaped [embed dim, embed dim]
- * https://github.com/pytorch/pytorch/blob/134179474539648ba7dee1317959529fbd0e7f89/torch/nn/modules/activation.py#L1080
- * https://github.com/openai/gpt-2/blob/9b63575ef42771a015060c964af2c3da4cf7c8ab/src/model.py#L53
- *
- * TODO: implement a fast track for self attention (query = key = value)
- * where a single dense layer combines and replaces the query, key and projection layers
- *
- * TODO: add kDim and vDim to accept key and values whose embedding dimensions differ from query's.
- */
-export class MultiHeadAttention extends tf.layers.Layer {
-    static className = "MultiHeadAttention";
-    numHeads;
-    embedDim; // size of embedding dim of inputs, also per attention head
-    useBias;
-    dropout;
-    causal; // use causal attention to mask future words
-    // projection simply means matrix multiplying query, key, and value
-    // with weights to create a representation of the inputs
-    queryProjection;
-    keyProjection;
-    valueProjection;
-    outputProjection;
-    constructor({ numHeads, embedDim, useBias = true, dropout = 0.0, causal = false, ...args }) {
-        super(args);
-        if (embedDim % numHeads != 0) {
-            throw Error(`${this.getClassName()}::constructor ${this.name} embedDim (${embedDim}) is not divisible by numHeads (${numHeads})`);
-        }
-        this.numHeads = numHeads;
-        this.embedDim = embedDim;
-        this.useBias = useBias;
-        this.dropout = dropout;
-        this.causal = causal;
-        if (this.dropout >= 1) {
-            throw Error(`${this.getClassName()}::constructor dropout must be within [0, 1)`);
-        }
-        // intialize the projection weights, this should be in the
-        // build() function but is done here to avoid linting complaints
-        this.queryProjection = tf.layers.dense({ useBias, units: embedDim });
-        this.keyProjection = tf.layers.dense({ useBias, units: embedDim });
-        this.valueProjection = tf.layers.dense({ useBias, units: embedDim });
-        this.outputProjection = tf.layers.dense({ useBias, units: embedDim });
-    }
-    /**
-     * Forward propagation. Provide one input tensor or three identical tensors to self-attention.
-     * @param inputs a single tensor for self-attention or an array of exactly three
-     *   tensors that are either identical (self-attention) or different (cross-attention)
-     * @param kwargs.packingMask a mask to prevent tokens from attending across document boundaries
-     */
-    call(inputs, kwargs) {
-        // validate the input tensors
-        if (!Array.isArray(inputs)) {
-            inputs = [inputs];
-        }
-        // accept only 1 input (self attention) or 3 inputs (self or cross attention)
-        if (inputs.length != 1 && inputs.length != 3) {
-            throw Error(`${this.getClassName()}::call ${this.name} expects exactly one or three input tensors, ${inputs.length} were provided`);
-        }
-        for (const input of inputs) {
-            if (input.shape.length != 3) {
-                throw Error(`${this.getClassName()}::call ${this.name} expected input shapes of [batch, seq, embed_dim], got ${JSON.stringify(input.shape)}`);
-            }
-        }
-        const [query, key, value] = inputs;
-        const packingMask = kwargs.packingMask ?? null;
-        const causalMask = kwargs.causalMask ?? null;
-        return inputs.length == 3
-            // cross-attention
-            ? this.forward(query, key, value, packingMask, causalMask, kwargs)
-            // self-attention
-            : this.forward(query, query, query, packingMask, causalMask, kwargs);
-    }
-    /**
-     * Forward propagation
-     */
-    forward(query_input, key_input, value_input, packing_mask, causal_mask, kwargs) {
-        // dimensions abbreviations
-        // batch = the number of sequences in the input
-        // seq = the length of each sequence in the input
-        // dims = the size of each token's embedding
-        return tf.tidy(() => {
-            const { query, key, value } = this.applyInputProjections(query_input, key_input, value_input);
-            // swap the seq and heads dimensions: [batch, seq, heads, head_dim] -> [batch, heads, seq, head_dim]
-            const move_head_dim_forward = [0, 2, 1, 3];
-            const { query_split, key_split, value_split } = this.splitHeads(query, key, value, move_head_dim_forward);
-            // apply scaled dot production attention to get [batch, seq, numHeads, embedDim]
-            const spda = MultiHeadAttention.scaledDotProductionAttention(query_split, key_split, value_split, kwargs.attentionMask ?? null, packing_mask, causal_mask, this.dropout, this.causal, kwargs);
-            // concat heads and apply the output projection
-            const output = this.outputProjection.apply(spda.transpose(move_head_dim_forward).reshape([query_input.shape[0], -1, this.embedDim]));
-            return output;
-        });
-    }
-    applyInputProjections(query_input, key_input, value_input) {
-        // apply input projections, this is a batched matrix multiplication operated on the last
-        // dimension of query_input and first dimension of the dense layer weights,
-        // [batch, seq, dims] x [dims, dims] = [batch x seq, dims] x [dims, dims] = [batch x seq, dims] = [batch, seq, dims]
-        return tf.tidy(() => {
-            return {
-                query: this.queryProjection.apply(query_input),
-                key: this.keyProjection.apply(key_input),
-                value: this.valueProjection.apply(value_input)
-            };
-        });
-    }
-    splitHeads(query, key, value, shuffle) {
-        // split heads and prepare for scaled dot product attention by splitting the
-        // last dimension to get the heads, bring the heads forward
-        // [batch, seq, dims] -> [batch, seq, heads, dims / heads] -> [batch, heads, seq, head_dim]
-        const batch_size = query.shape[0];
-        const split_heads = [batch_size, -1, this.numHeads, this.embedDim / this.numHeads];
-        return tf.tidy(() => {
-            return {
-                query_split: query.reshape(split_heads).transpose(shuffle),
-                key_split: key.reshape(split_heads).transpose(shuffle),
-                value_split: value.reshape(split_heads).transpose(shuffle)
-            };
-        });
-    }
-    /**
-     * Applies the scaled dot-product formula: softmax(QK_t / sqrt(d_k))V,
-     * formula (1) of the 2017 paper Attention Is All You Need
-     *
-     * @param attentionMask a mask to prevent tokens from being
-     *   attended to (usually for padding tokens). It should have the shape
-     *   [batch, head, query_sequence_len, key_sequence_len]. To use in
-     *   conjunction with causal masking, the tensor should be a boolean type
-     *   where false indicates a masked token.
-     * @param packingMask a mask to prevent tokens from attending across document boundaries
-     */
-    static scaledDotProductionAttention(query, key, value, attentionMask, packingMask, causalMask, dropout, causal, kwargs = {}) {
-        return tf.tidy(() => {
-            const { training = false, scaling_factor } = kwargs;
-            key.shape.forEach((val, index) => {
-                if (key.shape[index] != value.shape[index]) {
-                    throw Error(`scaledDotProductionAttention: expected key and value` +
-                        ` to have the same shape, got ${JSON.stringify(key.shape)} (key) and` +
-                        ` ${JSON.stringify(value.shape)} (value)`);
-                }
-            });
-            // mask's shape is [..., seq, seq] where seq is the number of words/tokens in the input,
-            // not adding the batch dimension yet to lessen the calculations
-            const causal_mask_shape = [
-                query.shape[query.shape.length - 2],
-                key.shape[key.shape.length - 2]
-            ];
-            let mask = tf.zeros(causal_mask_shape);
-            if (causal && causal_mask_shape[0] > 1) {
-                if (attentionMask && attentionMask.dtype != "bool") {
-                    throw Error(`scaledDotProductionAttention: the attention mask must be undefined or a boolean type if used with causal attention`);
-                }
-                // apply a causal attention mask so that tokens can only attend to preceding tokens,
-                // prevents looking at head
-                if (causalMask) {
-                    mask = causalMask;
-                }
-                else {
-                    mask = generateCausalMask(causal_mask_shape[0], causal_mask_shape[1]);
-                }
-            }
-            if (attentionMask) {
-                if (attentionMask.dtype == "bool") {
-                    // convert the boolean mask to float
-                    // warning: do not use 1e9, it will overflow, use something smaller like 1e7
-                    mask = mask.add(attentionMask.cast("float32").sub(1).mul(1e7));
-                }
-                else {
-                    // this will occur only when not using causal masking,
-                    // if the attention mask is not boolean, it's assumed the masking is already calculated,
-                    mask = attentionMask;
-                }
-            }
-            // 1. matrix multiply query and transposed key
-            // 2. divide by scaling factor
-            // 3. apply softmax to the result
-            // 4. apply attention and/or causal mask
-            // 5. apply dropout
-            // 6. matrix multiply softmax result with value
-            let pre_softmax = query
-                .matMul(key, false, true)
-                .div(Math.sqrt(scaling_factor ?? key.shape[key.shape.length - 1]))
-                .add(mask);
-            if (packingMask) {
-                // packing mask is added separately because each mask within a batch may be different,
-                // so it cannot be broadcasted
-                pre_softmax = pre_softmax.add(packingMask);
-            }
-            const spda = tf.softmax(pre_softmax);
-            const spda_dropout = tf.dropout(spda, training ? dropout : 0);
-            const attention = spda_dropout.matMul(value);
-            return attention;
-        });
-    }
-    build(inputShape) {
-        let input_shape = [];
-        if (Array.isArray(inputShape) && Array.isArray(inputShape[0])) {
-            input_shape = inputShape;
-        }
-        else {
-            input_shape = [inputShape, inputShape, inputShape];
-        }
-        if (input_shape.length != 1 && input_shape.length != 3) {
-            throw Error(`${this.getClassName()}::build ${this.name} accepts either exactly one or three inputs, received ${JSON.stringify(inputShape)}`);
-        }
-        // initialize the sublayer weights
-        this.queryProjection.build(input_shape[0]);
-        this.keyProjection.build(input_shape[1]);
-        this.valueProjection.build(input_shape[2]);
-        this.outputProjection.build(input_shape[0]);
-        // the sublayer weights need to be tracked by this layer otherwise
-        // backpropagation will complain about no trainable parameters found,
-        // this is an extra step that TF's Python version does not need
-        this.trainableWeights = [
-            ...this.queryProjection.trainableWeights,
-            ...this.keyProjection.trainableWeights,
-            ...this.valueProjection.trainableWeights,
-            ...this.outputProjection.trainableWeights
-        ];
-        // rename the weights otherwise they'll take on the default naming and overlap
-        // each other which breaks model loading due to duplicate weight names
-        let indexing = 0;
-        for (const weight of this.trainableWeights) {
-            const unique_name = `${this.getClassName()}_${indexing}`;
-            weight.name += unique_name;
-            weight.originalName += unique_name;
-            indexing++;
-        }
-        super.build(inputShape);
-    }
-    /**
-     * MultiHead attention's output is the same shape the query's.
-     */
-    computeOutputShape(inputShape) {
-        return Array.isArray(inputShape) && Array.isArray(inputShape[0]) ? inputShape[0] : inputShape;
-    }
-    getConfig() {
-        const base_config = super.getConfig();
-        const config = {
-            numHeads: this.numHeads,
-            embedDim: this.embedDim,
-            useBias: this.useBias,
-            causal: this.causal,
-            dropout: this.dropout,
-            name: this.name,
-        };
-        Object.assign(config, base_config);
-        return config;
-    }
-}
-tf.serialization.registerClass(MultiHeadAttention);
-//# sourceMappingURL=multihead_attention.js.map

package/dist/src/layers/multihead_attention.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"multihead_attention.js","sourceRoot":"","sources":["../../../src/layers/multihead_attention.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,kBAAkB,CAAC;AAGvC,OAAO,EAAE,MAAM,IAAI,kBAAkB,EAAE,MAAM,SAAS,CAAC;AAoBvD;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AACH,MAAM,OAAO,kBAAmB,SAAQ,EAAE,CAAC,MAAM,CAAC,KAAK;IACnD,MAAM,CAAC,SAAS,GAAG,oBAAoB,CAAC;IACrB,QAAQ,CAAS;IACjB,QAAQ,CAAS,CAAC,2DAA2D;IAC7E,OAAO,CAAU;IACjB,OAAO,CAAS;IAChB,MAAM,CAAU,CAAC,4CAA4C;IAEhF,mEAAmE;IACnE,wDAAwD;IACrC,eAAe,CAAkB;IACjC,aAAa,CAAkB;IAC/B,eAAe,CAAkB;IACjC,gBAAgB,CAAkB;IAGrD,YAAY,EAAE,QAAQ,EAAE,QAAQ,EAAE,OAAO,GAAG,IAAI,EAAE,OAAO,GAAG,GAAG,EAAE,MAAM,GAAG,KAAK,EAAE,GAAG,IAAI,EAA0B;QAC9G,KAAK,CAAC,IAAI,CAAC,CAAC;QAEZ,IAAI,QAAQ,GAAG,QAAQ,IAAI,CAAC,EAAE,CAAC;YAC3B,MAAM,KAAK,CAAC,GAAG,IAAI,CAAC,YAAY,EAAE,iBAAiB,IAAI,CAAC,IAAI,cAAc,QAAQ,mCAAmC,QAAQ,GAAG,CAAC,CAAC;QACtI,CAAC;QAED,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,QAAQ,GAAG,QAAQ,CAAC;QACzB,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QAErB,IAAI,IAAI,CAAC,OAAO,IAAI,CAAC,EAAE,CAAC;YACpB,MAAM,KAAK,CAAC,GAAG,IAAI,CAAC,YAAY,EAAE,6CAA6C,CAAC,CAAC;QACrF,CAAC;QAED,0DAA0D;QAC1D,gEAAgE;QAChE,IAAI,CAAC,eAAe,GAAG,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;QACrE,IAAI,CAAC,aAAa,GAAG,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;QACnE,IAAI,CAAC,eAAe,GAAG,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;QACrE,IAAI,CAAC,gBAAgB,GAAG,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;IAC1E,CAAC;IAGD;;;;;OAKG;IACM,IAAI,CACT,MAA+B,EAC/B,MAGC;QAED,6BAA6B;QAC7B,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;YACzB,MAAM,GAAG,CAAC,MAAM,CAAC,CAAC;QACtB,CAAC;QAED,6EAA6E;QAC7E,IAAI,MAAM,CAAC,MAAM,IAAI,CAAC,IAAI,MAAM,CAAC,MAAM,IAAI,CAAC,EAAE,CAAC;YAC3C,MAAM,KAAK,CAAC,GAAG,IAAI,CAAC,YAAY,EAAE,UAAU,IAAI,CAAC,IAAI,gDAAgD,MAAM,CAAC,MAAM,gBAAgB,CAAC,CAAC;QACxI,CAAC;QAED,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;YACzB,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,IAAI,CAAC,EAAE,CAAC;gBAC1B,MAAM,KAAK,CAAC,GAAG,IAAI,CAAC,YAAY,EAAE,UAAU,IAAI,CAAC,IAAI,0DAA0D,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;YAClJ,CAAC;QACL,CAAC;QAED,MAAM,CAAC,KAAK,EAAE,GAAG,EAAE,KAAK,CAAC,GAAG,MAAM,CAAC;QACnC,MAAM,WAAW,GAAG,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC;QAC/C,MAAM,UAAU,GAAG,MAAM,CAAC,UAAU,IAAI,IAAI,CAAC;QAE7C,OAAO,MAAM,CAAC,MAAM,IAAI,CAAC;YACrB,kBAAkB;YAClB,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,KAAM,EAAE,GAAI,EAAE,KAAM,EAAE,WAAW,EAAE,UAAU,EAAE,MAAM,CAAC;YACrE,iBAAiB;YACjB,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,KAAM,EAAE,KAAM,EAAE,KAAM,EAAE,WAAW,EAAE,UAAU,EAAE,MAAM,CAAC,CAAC;IAChF,CAAC;IAGD;;OAEG;IACO,OAAO,CACb,WAAsB,EACtB,SAAoB,EACpB,WAAsB,EACtB,YAA8B,EAC9B,WAA6B,EAC7B,MAAc;QAEd,2BAA2B;QAC3B,+CAA+C;QAC/C,iDAAiD;QACjD,4CAA4C;QAC5C,OAAO,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE;YAChB,MAAM,EAAE,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC,qBAAqB,CAAC,WAAW,EAAE,SAAS,EAAE,WAAW,CAAC,CAAC;YAE9F,oGAAoG;YACpG,MAAM,qBAAqB,GAAG,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC;YAE3C,MAAM,EACF,WAAW,EAAE,SAAS,EAAE,WAAW,EACtC,GAAG,IAAI,CAAC,UAAU,CAAC,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,qBAAqB,CAAC,CAAC;YAE9D,gFAAgF;YAChF,MAAM,IAAI,GAAG,kBAAkB,CAAC,4BAA4B,CACxD,WAAW,EAAE,SAAS,EAAE,WAAW,EACnC,MAAM,CAAC,aAAa,IAAI,IAAI,EAAE,YAAY,EAAE,WAAW,EACvD,IAAI,CAAC,OAAO,EAAE,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;YAEvC,+CAA+C;YAC/C,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,KAAK,CACtC,IAAI,CAAC,SAAS,CAAC,qBAAqB,CAAC,CAAC,OAAO,CAAC,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC;YAE9F,OAAO,MAAmB,CAAC;QAC/B,CAAC,CAAC,CAAA;IACN,CAAC;IAGS,qBAAqB,CAAC,WAAsB,EAAE,SAAoB,EAAE,WAAsB;QAChG,wFAAwF;QACxF,2EAA2E;QAC3E,oHAAoH;QACpH,OAAO,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE;YAChB,OAAO;gBACH,KAAK,EAAE,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,WAAW,CAAc;gBAC3D,GAAG,EAAE,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,SAAS,CAAc;gBACrD,KAAK,EAAE,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,WAAW,CAAc;aAC9D,CAAA;QACL,CAAC,CAAC,CAAA;IACN,CAAC;IAGS,UAAU,CAAC,KAAgB,EAAE,GAAc,EAAE,KAAgB,EAAE,OAAiB;QACtF,4EAA4E;QAC5E,2DAA2D;QAC3D,2FAA2F;QAC3F,MAAM,UAAU,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAClC,MAAM,WAAW,GAAG,CAAC,UAAU,EAAE,CAAC,CAAC,EAAE,IAAI,CAAC,QAAQ,EAAE,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAC,CAAC;QAEnF,OAAO,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE;YAChB,OAAO;gBACH,WAAW,EAAE,KAAK,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC,SAAS,CAAC,OAAO,CAAgB;gBACzE,SAAS,EAAE,GAAG,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC,SAAS,CAAC,OAAO,CAAgB;gBACrE,WAAW,EAAE,KAAK,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC,SAAS,CAAC,OAAO,CAAgB;aAC5E,CAAA;QACL,CAAC,CAAC,CAAA;IACN,CAAC;IAGD;;;;;;;;;;OAUG;IACH,MAAM,CAAC,4BAA4B,CAC/B,KAAgB,EAChB,GAAc,EACd,KAAgB,EAChB,aAA+B,EAC/B,WAA6B,EAC7B,UAA4B,EAC5B,OAAe,EACf,MAAe,EACf,SAA6C,EAAE;QAE/C,OAAO,EAAE,CAAC,IAAI,CAAC,GAAG,EAAE;YAChB,MAAM,EAAE,QAAQ,GAAG,KAAK,EAAE,cAAc,EAAE,GAAG,MAAM,CAAC;YAEpD,GAAG,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,KAAK,EAAE,EAAE;gBAC7B,IAAI,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC;oBACzC,MAAM,KAAK,CAAC,sDAAsD;wBAC9D,gCAAgC,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,KAAK,CAAC,YAAY;wBACrE,IAAI,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC;gBACnD,CAAC;YACL,CAAC,CAAC,CAAA;YAGF,wFAAwF;YACxF,gEAAgE;YAChE,MAAM,iBAAiB,GAAG;gBACtB,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC;gBACnC,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC;aAAC,CAAC;YAErC,IAAI,IAAI,GAAG,EAAE,CAAC,KAAK,CAAC,iBAAiB,CAAC,CAAC;YAEvC,IAAI,MAAM,IAAI,iBAAiB,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC;gBACrC,IAAI,aAAa,IAAI,aAAa,CAAC,KAAK,IAAI,MAAM,EAAE,CAAC;oBACjD,MAAM,KAAK,CAAC,oHAAoH,CAAC,CAAC;gBACtI,CAAC;gBAED,oFAAoF;gBACpF,2BAA2B;gBAC3B,IAAI,UAAU,EAAE,CAAC;oBACb,IAAI,GAAG,UAAU,CAAC;gBACtB,CAAC;qBAAM,CAAC;oBACJ,IAAI,GAAG,kBAAkB,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,iBAAiB,CAAC,CAAC,CAAC,CAAC,CAAC;gBAC1E,CAAC;YACL,CAAC;YAED,IAAI,aAAa,EAAE,CAAC;gBAChB,IAAI,aAAa,CAAC,KAAK,IAAI,MAAM,EAAE,CAAC;oBAChC,oCAAoC;oBACpC,4EAA4E;oBAC5E,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,aAAa,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;gBACnE,CAAC;qBAAM,CAAC;oBACJ,sDAAsD;oBACtD,wFAAwF;oBACxF,IAAI,GAAG,aAAa,CAAC;gBACzB,CAAC;YACL,CAAC;YAED,8CAA8C;YAC9C,8BAA8B;YAC9B,iCAAiC;YACjC,wCAAwC;YACxC,mBAAmB;YACnB,+CAA+C;YAC/C,IAAI,WAAW,GAAG,KAAK;iBAClB,MAAM,CAAC,GAAG,EAAE,KAAK,EAAE,IAAI,CAAC;iBACxB,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,cAAc,IAAI,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC;iBACjE,GAAG,CAAC,IAAI,CAAC,CAAC;YAEf,IAAI,WAAW,EAAE,CAAC;gBACd,sFAAsF;gBACtF,8BAA8B;gBAC9B,WAAW,GAAG,WAAW,CAAC,GAAG,CAAC,WAAW,CAAC,CAAC;YAC/C,CAAC;YAED,MAAM,IAAI,GAAG,EAAE,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC;YAErC,MAAM,YAAY,GAAG,EAAE,CAAC,OAAO,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAC9D,MAAM,SAAS,GAAG,YAAY,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;YAE7C,OAAO,SAAS,CAAC;QACrB,CAAC,CAAC,CAAC;IACP,CAAC;IAGQ,KAAK,CAAC,UAAiC;QAC5C,IAAI,WAAW,GAAe,EAAE,CAAC;QAEjC,IAAI,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,IAAI,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YAC5D,WAAW,GAAG,UAAwB,CAAC;QAC3C,CAAC;aAAM,CAAC;YACJ,WAAW,GAAG,CAAC,UAAsB,EAAE,UAAsB,EAAE,UAAsB,CAAC,CAAC;QAC3F,CAAC;QAED,IAAI,WAAW,CAAC,MAAM,IAAI,CAAC,IAAI,WAAW,CAAC,MAAM,IAAI,CAAC,EAAE,CAAC;YACrD,MAAM,KAAK,CAAC,GAAG,IAAI,CAAC,YAAY,EAAE,WAAW,IAAI,CAAC,IAAI,yDAAyD,IAAI,CAAC,SAAS,CAAC,UAAU,CAAC,EAAE,CAAC,CAAC;QACjJ,CAAC;QAED,kCAAkC;QAClC,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;QAC3C,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;QACzC,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;QAC3C,IAAI,CAAC,gBAAgB,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;QAE5C,kEAAkE;QAClE,qEAAqE;QACrE,+DAA+D;QAC/D,IAAI,CAAC,gBAAgB,GAAG;YACpB,GAAG,IAAI,CAAC,eAAe,CAAC,gBAAgB;YACxC,GAAG,IAAI,CAAC,aAAa,CAAC,gBAAgB;YACtC,GAAG,IAAI,CAAC,eAAe,CAAC,gBAAgB;YACxC,GAAG,IAAI,CAAC,gBAAgB,CAAC,gBAAgB;SAC5C,CAAC;QAEF,8EAA8E;QAC9E,sEAAsE;QACtE,IAAI,QAAQ,GAAG,CAAC,CAAC;QAEjB,KAAK,MAAM,MAAM,IAAI,IAAI,CAAC,gBAAgB,EAAE,CAAC;YACzC,MAAM,WAAW,GAAG,GAAG,IAAI,CAAC,YAAY,EAAE,IAAI,QAAQ,EAAE,CAAC;YACxD,MAAc,CAAC,IAAI,IAAI,WAAW,CAAC;YACnC,MAAc,CAAC,YAAY,IAAI,WAAW,CAAC;YAC5C,QAAQ,EAAE,CAAC;QACf,CAAC;QAED,KAAK,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC;IAC5B,CAAC;IAGD;;OAEG;IACM,kBAAkB,CAAC,UAAiC;QACzD,OAAO,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,IAAI,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC;IAClG,CAAC;IAGQ,SAAS;QACd,MAAM,WAAW,GAAG,KAAK,CAAC,SAAS,EAAE,CAAC;QAEtC,MAAM,MAAM,GAAG;YACX,QAAQ,EAAE,IAAI,CAAC,QAAQ;YACvB,QAAQ,EAAE,IAAI,CAAC,QAAQ;YACvB,OAAO,EAAE,IAAI,CAAC,OAAO;YACrB,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,OAAO,EAAE,IAAI,CAAC,OAAO;YACrB,IAAI,EAAE,IAAI,CAAC,IAAI;SAClB,CAAA;QAED,MAAM,CAAC,MAAM,CAAC,MAAM,EAAE,WAAW,CAAC,CAAC;QAEnC,OAAO,MAAM,CAAC;IAClB,CAAC;;AAIL,EAAE,CAAC,aAAa,CAAC,aAAa,CAAC,kBAAkB,CAAC,CAAC"}

package/dist/src/layers/multihead_attention.test.d.ts

@@ -1,2 +0,0 @@
-export {};
-//# sourceMappingURL=multihead_attention.test.d.ts.map

package/dist/src/layers/multihead_attention.test.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"multihead_attention.test.d.ts","sourceRoot":"","sources":["../../../src/layers/multihead_attention.test.ts"],"names":[],"mappings":""}

package/dist/src/layers/multihead_attention.test.js

@@ -1,160 +0,0 @@
-import * as tf from '@tensorflow/tfjs';
-import { CachedRoPEMultiHeadAttention } from '@/layers/cached_rope_multihead_attention';
-import { causal as generateCausalMask } from "@/masks";
-import { MultiHeadAttention } from '@/layers/multihead_attention';
-// disables warning for using the faster node backend,
-// https://github.com/tensorflow/tfjs/issues/5349#issuecomment-885170504
-tf.env().set('IS_NODE', false);
-describe("MultiHeadAttention tests", () => {
-    it("should fail to instantiate a layer if heads count is not divisible by the input's embedding dimension", () => {
-        expect(() => new CachedRoPEMultiHeadAttention({ numHeads: 3, embedDim: 10 })).toThrow();
-        expect(() => new CachedRoPEMultiHeadAttention({ numHeads: 15, embedDim: 60 })).not.toThrow();
-    });
-    test("successfull forward calls", () => {
-        const input = tf.randomUniform([2, 3, 12]);
-        const attention = new CachedRoPEMultiHeadAttention({ numHeads: 2, embedDim: input.shape.at(-1) });
-        expect(() => attention.apply(input)).not.toThrow();
-        expect(() => attention.apply([input])).not.toThrow();
-        const causal = new CachedRoPEMultiHeadAttention({ numHeads: 2, embedDim: input.shape.at(-1), causal: true });
-        expect(() => causal.apply(input)).not.toThrow();
-        expect(() => causal.apply([input])).not.toThrow();
-    });
-    test("query and value must have the same shape for scaled dot product attention to succeed", () => {
-        const query = tf.randomUniform([2, 3, 12]);
-        const key = tf.randomUniform([2, 3, 12]);
-        const value = tf.randomUniform([2, 3, 12]);
-        const value_thats_too_long = tf.randomUniform([2, 100, 12]);
-        const attention = new CachedRoPEMultiHeadAttention({ numHeads: 2, embedDim: query.shape.at(-1) });
-        expect(() => attention.apply([query, key, value])).not.toThrow();
-        expect(() => attention.apply([query, key, value_thats_too_long])).toThrow();
-    });
-    it("should only accept rank 3 tensors", () => {
-        const embed_dims = 12;
-        const BAD_RANK2 = tf.randomUniform([2, embed_dims]);
-        const GOOD = tf.randomUniform([2, 3, embed_dims]);
-        const BAD_RANK4 = tf.randomUniform([2, 3, 10, embed_dims]);
-        const attention = new CachedRoPEMultiHeadAttention({ numHeads: 2, embedDim: embed_dims });
-        // BAD
-        expect(() => attention.apply(BAD_RANK2)).toThrow();
-        expect(() => attention.apply([BAD_RANK2])).toThrow();
-        expect(() => attention.apply([BAD_RANK2, BAD_RANK2, BAD_RANK2])).toThrow();
-        // OK
-        expect(() => attention.apply(GOOD)).not.toThrow();
-        expect(() => attention.apply([GOOD])).not.toThrow();
-        expect(() => attention.apply([GOOD, GOOD, GOOD])).not.toThrow();
-        // BAD
-        expect(() => attention.apply(BAD_RANK4)).toThrow();
-        expect(() => attention.apply([BAD_RANK4])).toThrow();
-        expect(() => attention.apply([BAD_RANK4, BAD_RANK4, BAD_RANK4])).toThrow();
-        // BAD
-        expect(() => attention.apply([GOOD, BAD_RANK2, BAD_RANK4])).toThrow();
-        expect(() => attention.apply([BAD_RANK2, GOOD, BAD_RANK4])).toThrow();
-        expect(() => attention.apply([BAD_RANK2, BAD_RANK4, GOOD])).toThrow();
-        expect(() => attention.apply([BAD_RANK2, GOOD, GOOD])).toThrow();
-        expect(() => attention.apply([GOOD, GOOD, BAD_RANK4])).toThrow();
-    });
-    it("should only 1 or 3 inputs total", () => {
-        const input = tf.randomUniform([2, 3, 12]);
-        let attention = new CachedRoPEMultiHeadAttention({ numHeads: 2, embedDim: input.shape.at(-1) });
-        // OK
-        expect(() => attention.apply(input, { packingMask: undefined })).not.toThrow();
-        expect(() => attention.apply([input])).not.toThrow();
-        // reinitialize to rerun build()
-        attention = new CachedRoPEMultiHeadAttention({ numHeads: 2, embedDim: input.shape.at(-1) });
-        expect(() => attention.apply([input, input, input])).not.toThrow();
-        // BAD
-        expect(() => attention.apply([])).toThrow();
-        expect(() => attention.apply([input, input])).toThrow();
-        // reinitialize to rerun build()
-        attention = new CachedRoPEMultiHeadAttention({ numHeads: 2, embedDim: input.shape.at(-1) });
-        expect(() => attention.apply([input, input, input, input])).toThrow();
-    });
-    test("attention masking", () => {
-        const query = tf.randomUniform([2, 3, 12]);
-        const key = tf.randomUniform([2, 3, 12]);
-        const value = tf.randomUniform([2, 3, 12]);
-        const attention = new CachedRoPEMultiHeadAttention({ numHeads: 2, embedDim: query.shape.at(-1), causal: true });
-        expect(() => attention.call(query, {})).not.toThrow();
-        // cross attention
-        expect(() => attention.call([query, key, value], {})).not.toThrow();
-        const query5 = tf.randomUniform([2, 5, 10]);
-        const key4 = tf.randomUniform([2, 4, 10]);
-        const value5 = tf.randomUniform([2, 4, 10]);
-        const expected_mask = tf.tensor([[
-                // vertical represents query, false means that token cannot attend to the keys
-                // horizontal represents key, false means that token cannot attend to the queries
-                [false, false, false, false],
-                [true, true, true, false,],
-                [true, true, true, false,],
-                [false, false, false, false],
-                [true, true, true, false,],
-            ]]);
-        const packing_mask = tf.tensor([
-            [0, 0, 0, -1e7, -1e7],
-            [0, 0, 0, -1e7, -1e7],
-            [0, 0, 0, -1e7, -1e7],
-            [-1e7, -1e7, -1e7, 0, 0],
-            [-1e7, -1e7, -1e7, 0, 0]
-        ]);
-        // for causal attention, the attention mask must be boolean
-        expect(() => MultiHeadAttention.scaledDotProductionAttention(query5, key4, value5, expected_mask.asType("float32"), null, null, 0.1, true, { scaling_factor: 10 })).toThrow();
-        // for causal attention, using pre-calculated causal mask
-        expect(() => MultiHeadAttention.scaledDotProductionAttention(query5, key4, value5, expected_mask.asType("float32"), null, generateCausalMask(query5.shape[1], key4.shape[1]), 0.2, true, { scaling_factor: 10 })).toThrow();
-        // when not using causal attention, the attention mask can be a float32 tensor
-        expect(() => MultiHeadAttention.scaledDotProductionAttention(query5, key4, value5, expected_mask.asType("float32"), null, null, 0, false)).not.toThrow();
-        // packing mask for self attention
-        expect(() => MultiHeadAttention.scaledDotProductionAttention(query5, query5, query5, null, packing_mask, null, 0.9, true)).not.toThrow();
-    });
-    it("should return a non-empty config dict", () => {
-        const input = tf.randomUniform([2, 3, 10]);
-        const attention = new CachedRoPEMultiHeadAttention({ numHeads: 1, embedDim: input.shape.at(-1) });
-        expect(Object.keys(attention.getConfig())).not.toBe(0);
-    });
-    test("causal attention hard coded values", () => {
-        // input and output shapes: [2, 3, 10]
-        const input = tf.tensor([
-            [[0.2109915, 0.6158954, 0.6012088, 0.9867562, 0.8728716, 0.7496274, 0.8173883, 0.2958342, 0.9650571, 0.2075207],
-                [0.2946285, 0.9779906, 0.3203818, 0.4037617, 0.3762881, 0.9863171, 0.6655593, 0.7707329, 0.3216831, 0.7984023],
-                [0.9080769, 0.0026282, 0.379492, 0.0162054, 0.1939302, 0.2201049, 0.8190675, 0.0203963, 0.0114392, 0.5015539]],
-            [[0.6241482, 0.7631097, 0.6687831, 0.7259795, 0.0457698, 0.6889264, 0.0853676, 0.8697655, 0.3637198, 0.2105307],
-                [0.5221761, 0.4476321, 0.1244729, 0.8863543, 0.7319002, 0.2954829, 0.3200496, 0.0905503, 0.607977, 0.1309131],
-                [0.4693873, 0.4609751, 0.9170766, 0.7065565, 0.4795104, 0.3225758, 0.1353116, 0.7083887, 0.1928891, 0.967386]]
-        ]);
-        const expected = tf.tensor([
-            [[0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344],
-                [0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376],
-                [0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539]],
-            [[0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718],
-                [0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268],
-                [0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877]]
-        ]);
-        const attention = new CachedRoPEMultiHeadAttention({ numHeads: 1, embedDim: input.shape.at(-1), causal: true });
-        attention.build(input.shape);
-        attention.setWeights(attention.getWeights().map(weight => tf.onesLike(weight).mul(0.05)));
-        expect(expected.sub(attention.apply(input)).sum().dataSync()[0]).toBeLessThan(1e-6);
-    });
-    test("non-causal attention hard coded values", () => {
-        // input and output shapes: [2, 3, 10]
-        const input = tf.tensor([
-            [[0.2109915, 0.6158954, 0.6012088, 0.9867562, 0.8728716, 0.7496274, 0.8173883, 0.2958342, 0.9650571, 0.2075207],
-                [0.2946285, 0.9779906, 0.3203818, 0.4037617, 0.3762881, 0.9863171, 0.6655593, 0.7707329, 0.3216831, 0.7984023],
-                [0.9080769, 0.0026282, 0.379492, 0.0162054, 0.1939302, 0.2201049, 0.8190675, 0.0203963, 0.0114392, 0.5015539]],
-            [[0.6241482, 0.7631097, 0.6687831, 0.7259795, 0.0457698, 0.6889264, 0.0853676, 0.8697655, 0.3637198, 0.2105307],
-                [0.5221761, 0.4476321, 0.1244729, 0.8863543, 0.7319002, 0.2954829, 0.3200496, 0.0905503, 0.607977, 0.1309131],
-                [0.4693873, 0.4609751, 0.9170766, 0.7065565, 0.4795104, 0.3225758, 0.1353116, 0.7083887, 0.1928891, 0.967386]]
-        ]);
-        const expected = tf.tensor([
-            [[0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344, 0.2055344],
-                [0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376, 0.205376],
-                [0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539, 0.2042539]],
-            [[0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718, 0.1966718],
-                [0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268, 0.1966268],
-                [0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877, 0.1966877]]
-        ]);
-        const attention = new CachedRoPEMultiHeadAttention({ numHeads: 1, embedDim: input.shape.at(-1), causal: false });
-        attention.build(input.shape);
-        attention.setWeights(attention.getWeights().map(weight => tf.onesLike(weight).mul(0.05)));
-        expect(expected.sub(attention.apply(input)).sum().dataSync()[0]).toBeLessThan(1e-6);
-    });
-});
-//# sourceMappingURL=multihead_attention.test.js.map

package/dist/src/layers/multihead_attention.test.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"multihead_attention.test.js","sourceRoot":"","sources":["../../../src/layers/multihead_attention.test.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,kBAAkB,CAAC;AAEvC,OAAO,EAAE,4BAA4B,EAAE,MAAM,0CAA0C,CAAC;AACxF,OAAO,EAAE,MAAM,IAAI,kBAAkB,EAAE,MAAM,SAAS,CAAC;AACvD,OAAO,EAAE,kBAAkB,EAAE,MAAM,8BAA8B,CAAC;AAElE,sDAAsD;AACtD,wEAAwE;AACxE,EAAE,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC,SAAS,EAAE,KAAK,CAAC,CAAC;AAG/B,QAAQ,CAAC,0BAA0B,EAAE,GAAG,EAAE;IACtC,EAAE,CAAC,uGAAuG,EAAE,GAAG,EAAE;QAC7G,MAAM,CAAC,GAAG,EAAE,CAAC,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACxF,MAAM,CAAC,GAAG,EAAE,CAAC,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,EAAE,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;IACjG,CAAC,CAAC,CAAA;IAGF,IAAI,CAAC,2BAA2B,EAAE,GAAG,EAAE;QACnC,MAAM,KAAK,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAE3C,MAAM,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC;QACnG,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QACnD,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAErD,MAAM,MAAM,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC;QAC9G,MAAM,CAAC,GAAG,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAChD,MAAM,CAAC,GAAG,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;IACtD,CAAC,CAAC,CAAA;IAGF,IAAI,CAAC,sFAAsF,EAAE,GAAG,EAAE;QAC9F,MAAM,KAAK,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAC3C,MAAM,GAAG,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QACzC,MAAM,KAAK,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAC3C,MAAM,oBAAoB,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,EAAE,CAAC,CAAC,CAAC;QAE5D,MAAM,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC;QACnG,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,GAAG,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QACjE,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,GAAG,EAAE,oBAAoB,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;IAChF,CAAC,CAAC,CAAA;IAGF,EAAE,CAAC,mCAAmC,EAAE,GAAG,EAAE;QACzC,MAAM,UAAU,GAAG,EAAE,CAAC;QAEtB,MAAM,SAAS,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC;QACpD,MAAM,IAAI,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC;QAClD,MAAM,SAAS,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,EAAE,UAAU,CAAC,CAAC,CAAC;QAE3D,MAAM,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,UAAU,EAAE,CAAC,CAAC;QAE1F,MAAM;QACN,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACnD,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACrD,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QAE3E,KAAK;QACL,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAClD,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QACpD,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAEhE,MAAM;QACN,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACnD,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACrD,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QAE3E,MAAM;QACN,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACtE,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,EAAE,IAAI,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACtE,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACtE,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,SAAS,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACjE,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,IAAI,EAAE,IAAI,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;IACrE,CAAC,CAAC,CAAA;IAGF,EAAE,CAAC,iCAAiC,EAAE,GAAG,EAAE;QACvC,MAAM,KAAK,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAE3C,IAAI,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC;QAEjG,KAAK;QACL,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,KAAK,EAAE,EAAE,WAAW,EAAE,SAAS,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAC/E,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QACrD,gCAAgC;QAChC,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC;QAC7F,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAEnE,MAAM;QACN,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QAC5C,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QACxD,gCAAgC;QAChC,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC;QAC7F,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;IAC1E,CAAC,CAAC,CAAA;IAGF,IAAI,CAAC,mBAAmB,EAAE,GAAG,EAAE;QAC3B,MAAM,KAAK,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAC3C,MAAM,GAAG,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QACzC,MAAM,KAAK,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAE3C,MAAM,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC;QAEjH,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,IAAI,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAEtD,kBAAkB;QAClB,MAAM,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,GAAG,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QAGpE,MAAM,MAAM,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAC5C,MAAM,IAAI,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAC1C,MAAM,MAAM,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAE5C,MAAM,aAAa,GAAG,EAAE,CAAC,MAAM,CAAC,CAAC;gBAC7B,8EAA8E;gBAC9E,iFAAiF;gBACjF,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC;gBAC5B,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE;gBAC1B,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE;gBAC1B,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,CAAC;gBAC5B,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE;aAC7B,CAAC,CAAC,CAAC;QAEJ,MAAM,YAAY,GAAG,EAAE,CAAC,MAAM,CAAC;YAC3B,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC;YACrB,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC;YACrB,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC;YACrB,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,CAAC,EAAE,CAAC,CAAC;YACxB,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,CAAC,GAAG,EAAE,CAAC,EAAE,CAAC,CAAC;SAC3B,CAAC,CAAA;QAEF,2DAA2D;QAC3D,MAAM,CAAC,GAAG,EAAE,CAAC,kBAAkB,CAAC,4BAA4B,CAAC,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC,SAAS,CAAC,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,EAAE,IAAI,EAAE,EAAE,cAAc,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QAC9K,yDAAyD;QACzD,MAAM,CAAC,GAAG,EAAE,CAAC,kBAAkB,CAAC,4BAA4B,CAAC,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC,SAAS,CAAC,EAAE,IAAI,EAAE,kBAAkB,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAE,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC,EAAE,GAAG,EAAE,IAAI,EAAE,EAAE,cAAc,EAAE,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC;QAC9N,8EAA8E;QAC9E,MAAM,CAAC,GAAG,EAAE,CAAC,kBAAkB,CAAC,4BAA4B,CAAC,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,aAAa,CAAC,MAAM,CAAC,SAAS,CAAC,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,EAAE,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;QACzJ,kCAAkC;QAClC,MAAM,CAAC,GAAG,EAAE,CAAC,kBAAkB,CAAC,4BAA4B,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,YAAY,EAAE,IAAI,EAAE,GAAG,EAAE,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC;IAC7I,CAAC,CAAC,CAAA;IAGF,EAAE,CAAC,uCAAuC,EAAE,GAAG,EAAE;QAC7C,MAAM,KAAK,GAAG,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC;QAE3C,MAAM,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,CAAC,CAAC;QACnG,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,SAAS,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;IAC3D,CAAC,CAAC,CAAA;IAGF,IAAI,CAAC,oCAAoC,EAAE,GAAG,EAAE;QAC5C,sCAAsC;QACtC,MAAM,KAAK,GAAG,EAAE,CAAC,MAAM,CAAC;YACpB,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC/G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC9G,CAAC,SAAS,EAAE,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC;YAE9G,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC/G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,QAAQ,EAAE,SAAS,CAAC;gBAC7G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,QAAQ,CAAC,CAAC;SACjH,CAAC,CAAC;QAEH,MAAM,QAAQ,GAAG,EAAE,CAAC,MAAM,CAAC;YACvB,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC/G,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAC;gBACpG,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC;YAE/G,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC/G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC9G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC;SAClH,CAAC,CAAC;QAGH,MAAM,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC;QACjH,SAAS,CAAC,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;QAC7B,SAAS,CAAC,UAAU,CAAC,SAAS,CAAC,UAAU,EAAE,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAE1F,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAC,KAAK,CAAC,KAAK,CAAc,CAAC,CAAC,GAAG,EAAE,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,YAAY,CAAC,IAAI,CAAC,CAAC;IACrG,CAAC,CAAC,CAAA;IAGF,IAAI,CAAC,wCAAwC,EAAE,GAAG,EAAE;QAChD,sCAAsC;QACtC,MAAM,KAAK,GAAG,EAAE,CAAC,MAAM,CAAC;YACpB,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC/G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC9G,CAAC,SAAS,EAAE,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC;YAE9G,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC/G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,QAAQ,EAAE,SAAS,CAAC;gBAC7G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,QAAQ,CAAC,CAAC;SACjH,CAAC,CAAC;QAGH,MAAM,QAAQ,GAAG,EAAE,CAAC,MAAM,CAAC;YACvB,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC/G,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAC;gBACpG,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC;YAE/G,CAAC,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC/G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC;gBAC9G,CAAC,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,EAAE,SAAS,CAAC,CAAC;SAClH,CAAC,CAAC;QAEH,MAAM,SAAS,GAAG,IAAI,4BAA4B,CAAC,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAE,EAAE,MAAM,EAAE,KAAK,EAAE,CAAC,CAAC;QAClH,SAAS,CAAC,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;QAC7B,SAAS,CAAC,UAAU,CAAC,SAAS,CAAC,UAAU,EAAE,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAE1F,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAC,KAAK,CAAC,KAAK,CAAc,CAAC,CAAC,GAAG,EAAE,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,YAAY,CAAC,IAAI,CAAC,CAAC;IACrG,CAAC,CAAC,CAAC;AACP,CAAC,CAAC,CAAC"}

package/dist/src/layers/positional_encoding.d.ts

@@ -1,37 +0,0 @@
-import * as tf from "@tensorflow/tfjs";
-import { type LayerArgs } from '@tensorflow/tfjs-layers/dist/engine/topology';
-import { type Kwargs } from "@tensorflow/tfjs-layers/dist/types";
-export interface PositionalEncodingArgs extends LayerArgs {
-    embedDim: number;
-    maxSequenceLength?: number;
-}
-/**
- * This class implements the position encoding logic described in the
- * 2017 paper "Attention Is All You Need".
- *
- * This layer is untrainable and accepts inputs of shape `[ batch, sequences, embedding dims ]`
- * and adds positional encoding to return an output tensor of the same shape.
- *
- * @param embedDim the size of each token/word's embedding
- * @param maxSequenceLength the max number of tokens (words) per input (sentence), default `5120`
- */
-export declare class PositionalEncoding extends tf.layers.Layer {
-    static className: string;
-    private readonly maxSequenceLength;
-    private readonly embedDim;
-    private positionalEncodings;
-    constructor(args: PositionalEncodingArgs);
-    /**
-     * Forward propagation. Injects positional encoding to the input embeddings
-     */
-    call(inputs: tf.Tensor | tf.Tensor[], kwargs: Kwargs): tf.Tensor | tf.Tensor[];
-    /**
-     * Generate the positional encoding from the paper Attention Is All You Need.
-     * Note that because the inner term of the position formula is the same for both even
-     * and odd indices, we only create half of it and apply sine and cosine individually.
-     */
-    build(inputShape: tf.Shape | tf.Shape[]): void;
-    computeOutputShape(inputShape: tf.Shape | tf.Shape[]): tf.Shape | tf.Shape[];
-    getConfig(): tf.serialization.ConfigDict;
-}
-//# sourceMappingURL=positional_encoding.d.ts.map

package/dist/src/layers/positional_encoding.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"positional_encoding.d.ts","sourceRoot":"","sources":["../../../src/layers/positional_encoding.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,kBAAkB,CAAC;AACvC,OAAO,EAAE,KAAK,SAAS,EAAE,MAAM,8CAA8C,CAAC;AAC9E,OAAO,EAAE,KAAK,MAAM,EAAE,MAAM,oCAAoC,CAAC;AAGjE,MAAM,WAAW,sBAAuB,SAAQ,SAAS;IAErD,QAAQ,EAAE,MAAM,CAAC;IAEjB,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC9B;AAGD;;;;;;;;;GASG;AACH,qBAAa,kBAAmB,SAAQ,EAAE,CAAC,MAAM,CAAC,KAAK;IACnD,MAAM,CAAC,SAAS,SAAwB;IACxC,OAAO,CAAC,QAAQ,CAAC,iBAAiB,CAAS;IAC3C,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAS;IAClC,OAAO,CAAC,mBAAmB,CAAmB;gBAGlC,IAAI,EAAE,sBAAsB;IAuBxC;;OAEG;IACM,IAAI,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,GAAG,EAAE,CAAC,MAAM,EAAE,EAAE,MAAM,EAAE,MAAM,GAAG,EAAE,CAAC,MAAM,GAAG,EAAE,CAAC,MAAM,EAAE;IAyBvF;;;;OAIG;IACM,KAAK,CAAC,UAAU,EAAE,EAAE,CAAC,KAAK,GAAG,EAAE,CAAC,KAAK,EAAE,GAAG,IAAI;IAmD9C,kBAAkB,CAAC,UAAU,EAAE,EAAE,CAAC,KAAK,GAAG,EAAE,CAAC,KAAK,EAAE,GAAG,EAAE,CAAC,KAAK,GAAG,EAAE,CAAC,KAAK,EAAE;IAK5E,SAAS,IAAI,EAAE,CAAC,aAAa,CAAC,UAAU;CAYpD"}

package/dist/src/layers/positional_encoding.js

@@ -1,115 +0,0 @@
-import * as tf from "@tensorflow/tfjs";
-/**
- * This class implements the position encoding logic described in the
- * 2017 paper "Attention Is All You Need".
- *
- * This layer is untrainable and accepts inputs of shape `[ batch, sequences, embedding dims ]`
- * and adds positional encoding to return an output tensor of the same shape.
- *
- * @param embedDim the size of each token/word's embedding
- * @param maxSequenceLength the max number of tokens (words) per input (sentence), default `5120`
- */
-export class PositionalEncoding extends tf.layers.Layer {
-    static className = "PositionalEncoding";
-    maxSequenceLength;
-    embedDim;
-    positionalEncodings;
-    constructor(args) {
-        super(args);
-        this.maxSequenceLength = args.maxSequenceLength ?? 5120;
-        this.embedDim = args.embedDim;
-        if (this.maxSequenceLength < 1) {
-            throw Error(`${this.getClassName()}::constructor ${this.name} maxSequenceLength` +
-                ` (${args.maxSequenceLength}) must be greater than 0`);
-        }
-        if (this.embedDim < 1) {
-            throw Error(`${this.getClassName()}::constructor ${this.name} embedDim` +
-                ` (${args.embedDim}) must be greater than 0`);
-        }
-        // positional encodings are not trainable
-        this.positionalEncodings = this.addWeight('positional_encodings', [this.maxSequenceLength, this.embedDim], "float32", tf.initializers.zeros(), undefined, false);
-    }
-    /**
-     * Forward propagation. Injects positional encoding to the input embeddings
-     */
-    call(inputs, kwargs) {
-        // validate the input tensors
-        const input = Array.isArray(inputs) ? inputs[0] : inputs;
-        const sequences = input.shape[1];
-        if (input.shape.length != 3 || input.shape[2] != this.embedDim) {
-            throw Error(`${this.getClassName()}::call ${this.name} expected an input shape of` +
-                ` [batch, (up to ${this.maxSequenceLength}), ${this.embedDim}], instead got ${input.shape}`);
-        }
-        if (sequences > this.maxSequenceLength) {
-            // unexpected sequence length
-            throw Error(`${this.getClassName()}::call ${this.name} received an input with` +
-                ` sequence length (${sequences}) which is greater than the max sequence length` +
-                ` ${this.maxSequenceLength}`);
-        }
-        // perform forward propagation
-        return tf.tidy(() => {
-            return input.add(this.positionalEncodings.read()
-                .slice([0, 0], [sequences, this.embedDim]) // gets the first "sequences" rows
-                .expandDims(0)); // introduce the batch dimension and let add() broadcast it
-        });
-    }
-    /**
-     * Generate the positional encoding from the paper Attention Is All You Need.
-     * Note that because the inner term of the position formula is the same for both even
-     * and odd indices, we only create half of it and apply sine and cosine individually.
-     */
-    build(inputShape) {
-        tf.tidy(() => {
-            const embedDimHalved = Math.ceil(this.embedDim / 2);
-            // create the position matrix as [ 0, 1, 2, 3, etc ],
-            // and broadcast it horizontally to match the number of embeddings,
-            const numerator = tf.range(0, this.maxSequenceLength, 1)
-                .reshape([this.maxSequenceLength, 1])
-                // this creates an extra, unsued positional encoding column later on for odd embedding sizes
-                .broadcastTo([this.maxSequenceLength, embedDimHalved]);
-            // the inner term's denominator's exponent's numerator is created as
-            // [ 0, 0, 2, 2, 4, 4, etc ] ( technically [0, 2, 4] as explained above ) and not
-            // [ 0, 2, 4, 6, 8, 10, etc ] because the even and odd indices are counted as pairs
-            // when incrementing "i",
-            // the denominator formula is 10_000^(2i/d_model) where each "i" is a sine cosine pair
-            const denominator = tf.pow(10_000, tf.range(0, this.embedDim, 2).div(this.embedDim));
-            const inner_term = numerator.div(denominator);
-            const sine = tf.sin(inner_term);
-            const cosine = tf.cos(inner_term);
-            // horizontally interweave the sine and cosine columns together to form
-            // [sin, cos, sin, cos, etc]
-            // [sin, cos, sin, cos, etc]
-            // etc
-            const interweaved = [];
-            const ALL_ROWS = -1;
-            const ONE_COL = 1;
-            const FIRST_ROW = 0;
-            for (let targetCol = 0; targetCol < this.embedDim / 2; targetCol++) {
-                interweaved.push(sine.slice([FIRST_ROW, targetCol], [ALL_ROWS, ONE_COL]));
-                if (targetCol != Math.floor(this.embedDim / 2)) {
-                    // for odd numbered embedDim sizes skip the last cosine column
-                    // e.g. if embedDim = 5, create [ i=0 (sin), i=0 (cos), i=1 (sin), i=1 (cos), i=2 (sin) ]
-                    // and the final i=2 (cos) is ignored
-                    interweaved.push(cosine.slice([FIRST_ROW, targetCol], [ALL_ROWS, ONE_COL]));
-                }
-            }
-            // add the positional encoding
-            this.setWeights([tf.concat(interweaved, 1)]);
-        });
-        super.build(inputShape);
-    }
-    computeOutputShape(inputShape) {
-        return inputShape;
-    }
-    getConfig() {
-        const base_config = super.getConfig();
-        const config = {
-            maxSequenceLength: this.maxSequenceLength,
-            embedDim: this.embedDim,
-        };
-        Object.assign(config, base_config);
-        return config;
-    }
-}
-tf.serialization.registerClass(PositionalEncoding);
-//# sourceMappingURL=positional_encoding.js.map