netron/source/ncnn.js

import * as base from './base.js';

const ncnn = {};
const pnnx = {};

// https://github.com/Tencent/ncnn/wiki/param-and-model-file-structure
// https://github.com/Tencent/ncnn/wiki/operation-param-weight-table
// https://github.com/Tencent/ncnn/wiki/operators

ncnn.ModelFactory = class {

    async match(context) {
        const identifier = context.identifier.toLowerCase();
        if (identifier.endsWith('.param.bin') || identifier.endsWith('.ncnnmodel')) {
            const stream = context.stream;
            if (stream.length >= 4) {
                const buffer = stream.peek(4);
                const view = new DataView(buffer.buffer, buffer.byteOffset, buffer.byteLength);
                const signature = view.getUint32(0, true);
                if (signature === 0x007685dd) {
                    return context.set('ncnn.model.bin');
                }
            }
        } else if (identifier.endsWith('.param') || identifier.endsWith('.cfg.ncnn')) {
            const reader = await context.read('text', 0x10000);
            if (reader) {
                let type = '';
                try {
                    let match = false;
                    const signature = reader.read('\n');
                    if (signature !== undefined) {
                        if (signature.trim() === '7767517') {
                            match = true;
                        } else {
                            const header = signature.trim().split(' ');
                            if (header.length === 2 && header.every((value) => value >>> 0 === parseFloat(value))) {
                                match = true;
                            }
                        }
                    }
                    if (match) {
                        type = 'ncnn.model';
                        for (let i = 0; i < 32; i++) {
                            const line = reader.read('\n');
                            if (!line) {
                                break;
                            }
                            if (line.startsWith('pnnx.') || line.startsWith('nn.') || line.startsWith('F.') || line.startsWith('torch.') || line.startsWith('Tensor.')) {
                                type = 'pnnx.model';
                                break;
                            }
                        }
                    }
                } catch {
                    // continue regardless of error
                }
                if (type) {
                    return context.set(type);
                }
            }
        } else if (identifier.endsWith('.ncnn.bin')) {
            return context.set('ncnn.weights');
        } else if (identifier.endsWith('.pnnx.bin')) {
            const entries = await context.peek('zip');
            if (entries) { // can be empty
                return context.set('pnnx.weights', entries);
            }
        } else if (identifier.endsWith('.bin') || identifier.endsWith('.weights.ncnn')) {
            const stream = context.stream;
            const length = Math.min(0x10000, stream.length);
            if (length > 4) {
                let buffer = stream.peek(length);
                const view = new DataView(buffer.buffer, buffer.byteOffset, buffer.byteLength);
                const signature = view.getUint32(0, true);
                if (signature === 0x00000000) {
                    const size = Math.min(buffer.length & 0xfffffffc, 128);
                    buffer = buffer.subarray(0, size);
                    const f32 = new Array(buffer.length >> 2);
                    for (let i = 1; i < f32.length; i++) {
                        f32[i] = view.getFloat32(i << 2, true);
                    }
                    if (f32.every((value) => !Number.isNaN(value) && Number.isFinite(value) && value > -20.0 && value < 20.0)) {
                        return context.set('ncnn.weights');
                    }
                } else {
                    for (let i = 0; i < buffer.length - 4; i++) {
                        const signature = view.getUint32(i, true);
                        if (signature === 0xdeadbeef) { // Core ML
                            return null;
                        }
                        if (signature === 0x01306b47 || signature === 0x000d4b38 || signature === 0x0002c056) { // ncnn
                            return context.set('ncnn.weights');
                        }
                    }
                }
            }
        }
        return null;
    }

    filter(context, match) {
        return (context.type !== 'ncnn.model' && context.type !== 'ncnn.model.bin') || match.type !== 'ncnn.weights';
    }

    async open(context) {
        let metadata = null;
        if (context.type.startsWith('pnnx.')) {
            metadata = await pnnx.Metadata.open(context);
        } else {
            metadata = await context.metadata('ncnn-metadata.json');
        }
        const identifier = context.identifier.toLowerCase();
        const format = context.type.split('.').shift();
        switch (context.type) {
            case 'pnnx.model':
            case 'ncnn.model': {
                let file = null;
                if (identifier.endsWith('.param')) {
                    const extension = context.type === 'pnnx.model' && !identifier.endsWith('.pnnx.param') ? '.pnnx.bin' : '.bin';
                    file = context.identifier.replace(/\.param$/, extension);
                } else if (identifier.endsWith('.cfg.ncnn')) {
                    file = context.identifier.replace(/\.cfg\.ncnn$/, '.weights.ncnn');
                }
                let content = null;
                try {
                    content = await context.fetch(file);
                } catch {
                    // continue regardless of error
                }
                const param = await context.read('text');
                const reader = new ncnn.TextParamReader(param);
                const blobs = await ncnn.BlobReader.open(content);
                return new ncnn.Model(metadata, format, reader, blobs);
            }
            case 'ncnn.model.bin': {
                const bin = `${context.identifier.substring(0, context.identifier.length - 10)}.bin`;
                let content = null;
                try {
                    content = await context.fetch(bin);
                } catch {
                    // continue regardless of error
                }
                const param = context.stream.peek();
                const reader = new ncnn.BinaryParamReader(param);
                const blobs = await ncnn.BlobReader.open(content);
                return new ncnn.Model(metadata, format, reader, blobs);
            }
            case 'pnnx.weights':
            case 'ncnn.weights': {
                let file = null;
                if (identifier.endsWith('.bin')) {
                    file = context.identifier.replace(/\.bin$/, '.param');
                } else if (identifier.endsWith('.weights.ncnn')) {
                    file = context.identifier.replace(/\.weights\.ncnn$/, '.cfg.ncnn');
                }
                let content = null;
                let message = null;
                try {
                    content = await context.fetch(file);
                } catch (error) {
                    message = error && error.message ? error.message : error.toString();
                    try {
                        content = await context.fetch(`${file}.bin`);
                    } catch {
                        // continue regardless of error
                    }
                }
                if (!content) {
                    throw new ncnn.Error(`Required ncnn model definition not found (${message.replace(/\.$/, '')}).`);
                }
                let reader = null;
                if (content.identifier.endsWith('.bin')) {
                    const data = content.stream.peek();
                    reader = new ncnn.BinaryParamReader(data);
                } else {
                    const text = await content.read('text');
                    reader = new ncnn.TextParamReader(text);
                }
                const blobs = await ncnn.BlobReader.open(context);
                return new ncnn.Model(metadata, format, reader, blobs);
            }
            default: {
                throw new ncnn.Error(`Unsupported ncnn format '${context.type}'.`);
            }
        }
    }
};

ncnn.Model = class {

    constructor(metadata, format, param, blobs) {
        this.format = format === 'pnnx' ? 'PNNX' : 'ncnn';
        this.modules = [new ncnn.Graph(metadata, format, param, blobs)];
    }
};

ncnn.Graph = class {

    constructor(metadata, format, param, blobs) {
        this.inputs = [];
        this.outputs = [];
        this.nodes = [];
        const layers = param.layers;
        const values = new Map();
        values.map = (name, type, tensor) => {
            if (name.length === 0 && tensor) {
                return new ncnn.Value(name, type, tensor);
            }
            if (!values.has(name)) {
                values.set(name, new ncnn.Value(name, type || null, tensor || null));
            } else if (tensor || (type && !type.equals(values.get(name).type))) {
                throw new ncnn.Error(`Duplicate value '${name}'.`);
            }
            return values.get(name);
        };
        for (const layer of layers) {
            const params = layer.params;
            if (params && params.size > 0) {
                for (const [key, list] of params) {
                    if (key === '30' && Array.isArray(list)) {
                        const value = list.map((item) => parseInt(item, 10));
                        for (const output of layer.outputs || []) {
                            if (value.length > 0 && value[0] <= value.length - 1) {
                                const shape = new Array(value.shift());
                                for (let i = 0; i < shape.length; i++) {
                                    shape[i] = value.shift();
                                }
                                const type = new ncnn.TensorType('float32', new ncnn.TensorShape(shape));
                                values.map(output, type);
                            }
                            params.delete(key);
                        }
                    }
                }
            }
        }
        for (const layer of layers) {
            if (layer.type === 'Input' || layer.type === 16) {
                const dimensions = Array.from(layer.params.values()).map((value) => isNaN(parseInt(value, 10)) ? value : parseInt(value, 10));
                const shape = new ncnn.TensorShape(dimensions);
                const type = new ncnn.TensorType('float32', shape);
                const argument = new ncnn.Argument(layer.name, layer.outputs.map((output) => values.map(output, type)));
                this.inputs.push(argument);
            } else if (layer.type === 'pnnx.Input' && layer.params) {
                const argument = new ncnn.Argument(layer.name, layer.outputs.map((output) => values.map(output, ncnn.Utility.route(layer.params, output))));
                this.inputs.push(argument);
            } else if (layer.type === 'pnnx.Output' && layer.params) {
                const argument = new ncnn.Argument(layer.name, layer.inputs.map((input) => values.map(input, ncnn.Utility.route(layer.params, input))));
                this.outputs.push(argument);
            } else {
                const node = new ncnn.Node(metadata, format, blobs, layer, values);
                this.nodes.push(node);
            }
        }
        blobs.validate();
    }
};

ncnn.Argument = class {

    constructor(name, value, type, visible) {
        this.name = name;
        this.value = value;
        this.type = type;
        this.visible = visible !== false;
    }
};

ncnn.Value = class {

    constructor(name, type, initializer = null) {
        if (typeof name !== 'string') {
            throw new ncnn.Error(`Invalid value identifier '${JSON.stringify(name)}'.`);
        }
        this.name = name;
        this.type = initializer ? initializer.type : type;
        this.initializer = initializer;
        this.quantization = initializer ? initializer.quantization : null;
    }
};

ncnn.Node = class {

    constructor(metadata, format, blobs, layer, values) {
        this.inputs = [];
        this.outputs = [];
        this.attributes = [];
        this.chain = [];
        this.name = layer.name || '';
        this.type = { ...metadata.type(layer.type) };
        delete this.type.identifier;
        const params = layer.params;
        const inputs = layer.inputs || [];
        let inputIndex = 0;
        const names = new Map();
        if (params) {
            for (const [key, value] of params) {
                if (key.startsWith('$')) {
                    names.set(value, key.substring(1));
                    params.delete(key);
                }
            }
        }
        if (this.type && Array.isArray(this.type.inputs)) {
            for (const input of this.type.inputs) {
                if (inputIndex < inputs.length || input.optional === false) {
                    const count = (input.type === 'Tensor[]') ? (inputs.length - inputIndex) : 1;
                    const list = inputs.slice(inputIndex, inputIndex + count).filter((id) => id !== '' || input.option !== 'optional').map((id) => values.map(id, ncnn.Utility.route(params, id)));
                    const argument = new ncnn.Argument(input.name, list);
                    this.inputs.push(argument);
                    inputIndex += count;
                }
            }
        }
        this.inputs.push(...inputs.slice(inputIndex).map((input, index) => {
            index = inputIndex + index;
            let name = 'input';
            if (names.has(input)) {
                name = names.get(input);
            } else if (index !== 0) {
                name = index.toString();
            }
            return new ncnn.Argument(name, [values.map(input, ncnn.Utility.route(params, input))]);
        }));
        const outputs = layer.outputs || [];
        let outputIndex = 0;
        if (this.type && Array.isArray(this.type.outputs)) {
            for (const output of this.type.outputs) {
                if (outputIndex < outputs.length || output.option !== 'optional') {
                    const count = (output.type === 'Tensor[]') ? (outputs.length - outputIndex) : 1;
                    const list = outputs.slice(outputIndex, outputIndex + count).map((id) => values.map(id, ncnn.Utility.route(params, id)));
                    const argument = new ncnn.Argument(output.name, list);
                    this.outputs.push(argument);
                    outputIndex += count;
                }
            }
        }
        this.outputs.push(...outputs.slice(outputIndex).map((output, index) => {
            const name = (outputIndex + index) === 0 ? 'output' : (outputIndex + index).toString();
            return new ncnn.Argument(name, [values.map(output, ncnn.Utility.route(params, output))]);
        }));
        blobs.weight = (name, shape, code) => {
            const blob = blobs.load(shape, code || 0);
            const dataType = blob ? (blob.dataType || '?') : (code || 0).toString();
            const data = blob ? blob.data : null;
            const quantization = blob && blob.quantization ? blob.quantization : null;
            const type = new ncnn.TensorType(dataType, new ncnn.TensorShape(shape));
            const tensor = new ncnn.Tensor(type, data, quantization);
            const argument = new ncnn.Argument(name, [values.map('', null, tensor)]);
            this.inputs.push(argument);
        };
        switch (this.type.name) {
            case 'BatchNorm': {
                const channels = parseInt(params.get('0') || 0, 10);
                blobs.weight('slope', [channels], 1);
                blobs.weight('mean', [channels], 1);
                blobs.weight('variance', [channels], 1);
                blobs.weight('bias', [channels], 1);
                break;
            }
            case 'InnerProduct': {
                const num_output = parseInt(params.get('0') || 0, 10);
                const bias_term = parseInt(params.get('1') || 0, 10);
                const weight_data_size = parseInt(params.get('2') || 0, 10);
                const int8_scale_term = parseInt(params.get('8') || 0, 10);
                const activation_type = parseInt(params.get('9') || 0, 10);
                const input_size = num_output > 0 ? Math.floor(weight_data_size / num_output) : 0;
                blobs.weight('weight', [num_output, input_size]);
                if (bias_term) {
                    blobs.weight('bias', [num_output], 1);
                }
                if (int8_scale_term) {
                    blobs.weight('weight_scales', [num_output], 1);
                    blobs.weight('bottom_scales', [1], 1);
                }
                const activation_names = ['', 'ReLU', 'Leaky ReLU', 'Clip', 'Sigmoid', 'Mish', 'HardSwish'];
                if (activation_type > 0 && activation_type < activation_names.length) {
                    const layer = { type: activation_names[activation_type] };
                    this.chain.push(new ncnn.Node(metadata, format, blobs, layer, values));
                }
                params.delete('2');
                break;
            }
            case 'Bias': {
                const bias_data_size = parseInt(params.get('0') || 0, 10);
                blobs.weight('bias', [bias_data_size], 1);
                break;
            }
            case 'Embed': {
                const num_output = parseInt(params.get('0') || 0, 10);
                const weight_data_size = parseInt(params.get('3') || 0, 10);
                blobs.weight('weight', [weight_data_size / num_output, num_output]);
                if (parseInt(params.get('2') || 0, 10) === 1) {
                    blobs.weight('bias', [num_output], 1);
                }
                params.get('3');
                break;
            }
            case 'Convolution':
            case 'ConvolutionDepthWise':
            case 'Deconvolution':
            case 'DeconvolutionDepthWise': {
                const num_output = parseInt(params.get('0') || 0, 10);
                const kernel_w = parseInt(params.get('1') || 0, 10);
                const kernel_h = parseInt(params.get('11') || kernel_w, 10);
                const weight_data_size = parseInt(params.get('6') || 0, 10);
                blobs.weight('weight', [num_output, weight_data_size / (num_output * kernel_w * kernel_h), kernel_h, kernel_w]);
                if (parseInt(params.get('5') || 0, 10) === 1) {
                    blobs.weight('bias', [num_output], 1);
                }
                const int8_scale_term = parseInt(params.get('8') || 0, 10);
                if (this.type.name === 'Convolution') {
                    if (int8_scale_term) {
                        blobs.weight('weight_scales', [num_output], 1);
                        blobs.weight('bottom_scales', [1], 1);
                    }
                    if (int8_scale_term > 100) {
                        blobs.weight('top_scales', [1], 1);
                    }
                } else if (this.type.name === 'ConvolutionDepthWise') {
                    const group =  parseInt(params.get('7') || 1, 10);
                    if (int8_scale_term === 1 || int8_scale_term === 101) {
                        blobs.weight('weight_scales', [group], 1);
                        blobs.weight('bottom_scales', [1], 1);
                    } else if (int8_scale_term === 2 || int8_scale_term === 102) {
                        blobs.weight('weight_scales', [1], 1);
                        blobs.weight('bottom_scales', [1], 1);
                    }
                    if (int8_scale_term > 100) {
                        blobs.weight('top_scales', [1], 1);
                    }
                }
                params.delete('6');
                const activation_names = ['', 'ReLU', 'LeakyReLU', 'Clip', 'Sigmoid', 'Mish', 'HardSwish'];
                const activation_type = parseInt(params.get('9') || 0, 10);
                if (activation_type > 0 && activation_type < activation_names.length) {
                    const layer = { type: activation_names[activation_type] };
                    this.chain.push(new ncnn.Node(metadata, format, blobs, layer, values));
                }
                break;
            }
            case 'Convolution1D':
            case 'ConvolutionDepthWise1D': {
                const num_output = parseInt(params.get('0') || 0, 10);
                const kernel_w = parseInt(params.get('1') || 0, 10);
                const dynamic_weight = parseInt(params.get('19') || 0, 10);
                if (!dynamic_weight) {
                    const weight_data_size = parseInt(params.get('6') || 0, 10);
                    blobs.weight('weight', [num_output, weight_data_size / (num_output * kernel_w), kernel_w]);
                    if (parseInt(params.get('5') || 0, 10) === 1) {
                        blobs.weight('bias', [num_output], 1);
                    }
                    params.delete('6');
                }
                params.delete('19');
                const activation_names = ['', 'ReLU', 'LeakyReLU', 'Clip', 'Sigmoid', 'Mish', 'HardSwish'];
                const activation_type = parseInt(params.get('9') || 0, 10);
                if (activation_type > 0 && activation_type < activation_names.length) {
                    const layer = { type: activation_names[activation_type] };
                    const node = new ncnn.Node(metadata, format, blobs, layer, values);
                    this.chain.push(node);
                }
                break;
            }
            case 'Deconvolution1D': {
                const activation_names = ['', 'ReLU', 'LeakyReLU', 'Clip', 'Sigmoid', 'Mish', 'HardSwish'];
                const activation_type = parseInt(params.get('9') || 0, 10);
                if (activation_type > 0 && activation_type < activation_names.length) {
                    const layer = { type: activation_names[activation_type] };
                    const node = new ncnn.Node(metadata, format, blobs, layer, values);
                    this.chain.push(node);
                }
                const num_output = parseInt(params.get('0') || 0, 10);
                const kernel_w = parseInt(params.get('1') || 0, 10);
                const dynamic_weight = parseInt(params.get('28') || 0, 10);
                if (!dynamic_weight) {
                    const weight_data_size = parseInt(params.get('6') || 0, 10);
                    blobs.weight('weight', [num_output, weight_data_size / (num_output * kernel_w), kernel_w]);
                    if (parseInt(params.get('5') || 0, 10) === 1) {
                        blobs.weight('bias', [num_output], 1);
                    }
                    params.delete('6');
                }
                params.delete('28');
                break;
            }
            case 'Convolution3D':
            case 'ConvolutionDepthWise3D': {
                const num_output = parseInt(params.get('0') || 0, 10);
                const kernel_w = parseInt(params.get('1') || 0, 10);
                const kernel_h = parseInt(params.get('11') || kernel_w, 10);
                const kernel_d = parseInt(params.get('21') || kernel_w, 10);
                const weight_data_size = parseInt(params.get('6') || 0, 10);
                blobs.weight('weight', [num_output, weight_data_size / (num_output * kernel_w * kernel_h * kernel_d), kernel_d, kernel_h, kernel_w]);
                if (parseInt(params.get('5') || 0, 10) === 1) {
                    blobs.weight('bias', [num_output], 1);
                }
                params.delete('6');
                const activation_names = ['', 'ReLU', 'LeakyReLU', 'Clip', 'Sigmoid', 'Mish', 'HardSwish'];
                const activation_type = parseInt(params.get('9') || 0, 10);
                if (activation_type > 0 && activation_type < activation_names.length) {
                    const layer = { type: activation_names[activation_type] };
                    this.chain.push(new ncnn.Node(metadata, format, blobs, layer, values));
                }
                break;
            }
            case 'Quantize': {
                const scale_data_size = parseInt(params.get('0') || 1, 10);
                blobs.weight('scale', [scale_data_size], 1);
                break;
            }
            case 'Dequantize': {
                const scale_data_size = parseInt(params.get('0') || 1, 10);
                const bias_data_size = parseInt(params.get('1') || 0, 10);
                blobs.weight('scale', [scale_data_size], 1);
                blobs.weight('bias', [bias_data_size], 1);
                break;
            }
            case 'Requantize': {
                const scale_in_data_size = parseInt(params.get('0') || 1, 10);
                const scale_out_data_size = parseInt(params.get('1') || 1, 10);
                const bias_data_size = parseInt(params.get('2') || 0, 10);
                blobs.weight('scale_in', [scale_in_data_size], 1);
                blobs.weight('scale_out', [scale_out_data_size], 1);
                blobs.weight('bias', [bias_data_size], 1);
                break;
            }
            case 'InstanceNorm': {
                const affine = parseInt(params.get('2') || 1, 10);
                if (affine === 1) {
                    const channels = parseInt(params.get('0') || 0, 10);
                    blobs.weight('gamma', [channels], 1);
                    blobs.weight('beta', [channels], 1);
                }
                break;
            }
            case 'Scale': {
                const scale_data_size = parseInt(params.get('0') || 0, 10);
                if (scale_data_size !== -233) {
                    blobs.weight('scale', [scale_data_size], 1);
                    if (params.get('1') === '1') {
                        blobs.weight('bias', [scale_data_size], 1);
                    }
                }
                break;
            }
            case 'Normalize': {
                const scale_data_size = parseInt(params.get('3') || 0, 10);
                blobs.weight('scale', [scale_data_size], 1);
                break;
            }
            case 'PReLU': {
                const num_slope = parseInt(params.get('0') || 0, 10);
                blobs.weight('slope', [num_slope], 1);
                break;
            }
            case 'Padding': {
                const per_channel_pad_data_size = parseInt(params.get('6') || 0, 10);
                blobs.weight('per_channel_pad_data', [per_channel_pad_data_size], 1);
                break;
            }
            case 'MemoryData': {
                const w = parseInt(params.get('0') || 0, 10);
                const h = parseInt(params.get('1') || 0, 10);
                const d = parseInt(params.get('11') || 0, 10);
                const c = parseInt(params.get('2') || 0, 10);
                const load_type = parseInt(params.get('21') || 1, 10);
                /* eslint-disable no-negated-condition */
                if (d !== 0) {
                    blobs.weight('data', [w, h, d, c], load_type);
                } else if (c !== 0) {
                    blobs.weight('data', [w, h, c], load_type);
                } else if (h !== 0) {
                    blobs.weight('data', [w, h], load_type);
                } else if (w !== 0) {
                    blobs.weight('data', [w], load_type);
                } else {
                    blobs.weight('data', [1], load_type);
                }
                /* eslint-enable no-negated-condition */
                break;
            }
            case 'GroupNorm': {
                const affine = parseInt(params.get('3') || 1, 10);
                if (affine === 1) {
                    const channels = parseInt(params.get('1') || 0, 10);
                    blobs.weight('gamma', [channels], 1);
                    blobs.weight('beta', [channels], 1);
                }
                break;
            }
            case 'LayerNorm': {
                const channels = parseInt(params.get('0') || 0, 10);
                blobs.weight('gamma', [channels], 1);
                blobs.weight('beta', [channels], 1);
                break;
            }
            case 'RNN': {
                const num_output = parseInt(params.get('0') || 0, 10);
                const weight_data_size = parseInt(params.get('1') || 0, 10);
                const direction = parseInt(params.get('2') || 0, 10);
                const num_directions = direction === 2 ? 2 : 1;
                blobs.weight('weight_xc', [num_directions, num_output, weight_data_size / num_directions / num_output]);
                blobs.weight('bias_c', [num_directions, num_output]);
                blobs.weight('weight_hc', [num_directions, num_output, num_output]);
                params.delete('1');
                break;
            }
            case 'LSTM': {
                const num_output = parseInt(params.get('0') || 0, 10);
                const weight_data_size = parseInt(params.get('1') || 0, 10);
                const direction = parseInt(params.get('2') || 0, 10);
                const num_directions = direction === 2 ? 2 : 1;
                blobs.weight('weight_xc', [num_directions, 4, num_output, weight_data_size / num_directions / num_output / 4]);
                blobs.weight('bias_c', [num_directions, 4, num_output]);
                blobs.weight('weight_hc', [num_directions, 4, num_output, num_output]);
                params.delete('1');
                break;
            }
            case 'GRU': {
                const num_output = parseInt(params.get('0') || 0, 10);
                const weight_data_size = parseInt(params.get('1') || 0, 10);
                const direction = parseInt(params.get('2') || 0, 10);
                const num_directions = direction === 2 ? 2 : 1;
                blobs.weight('weight_xc', [num_directions, 3, num_output, weight_data_size / num_directions / num_output / 3]);
                blobs.weight('bias_c', [num_directions, 4, num_output]);
                blobs.weight('weight_hc', [num_directions, 3, num_output, num_output]);
                params.delete('1');
                break;
            }
            case 'MultiHeadAttention': {
                const embed_dim = parseInt(params.get('0') || 0, 10);
                // const num_head = parseInt(params.get('1') || 0, 10);
                // const weight_data_size = parseInt(params.get('2') || 0, 10);
                blobs.weight('weight_q', [embed_dim, embed_dim]);
                blobs.weight('bias_q', [embed_dim], 1);
                blobs.weight('weight_k', [embed_dim, embed_dim]);
                blobs.weight('bias_k', [embed_dim], 1);
                blobs.weight('weight_v', [embed_dim, embed_dim]);
                blobs.weight('bias_v', [embed_dim], 1);
                blobs.weight('weight_out', [embed_dim, embed_dim]);
                blobs.weight('bias_out', [embed_dim], 1);
                params.delete('2');
                break;
            }
            case 'Gemm': {
                const transA = parseInt(params.get('2') || 0, 10);
                const transB = parseInt(params.get('3') || 0, 10);
                const constantA = parseInt(params.get('4') || 0, 10);
                const constantB = parseInt(params.get('5') || 0, 10);
                const constantC = parseInt(params.get('6') || 0, 10);
                const M = parseInt(params.get('7') || 0, 10);
                const N = parseInt(params.get('8') || 0, 10);
                const K = parseInt(params.get('9') || 0, 10);
                const constant_broadcast_type_C = parseInt(params.get('10') || 0, 10);
                if (constantA === 1) {
                    blobs.weight('A', transA === 0 ? [K, M] : [M, K]);
                }
                if (constantB === 1) {
                    blobs.weight('B', transB === 1 ? [N, K] : [K, N]);
                }
                if (constantC === 1 && constant_broadcast_type_C !== -1) {
                    let shape = null;
                    switch (constant_broadcast_type_C) {
                        case 0: shape = [1]; break;
                        case 1: shape = [M]; break;
                        case 2: shape = [1, M]; break;
                        case 3: shape = [N, M]; break;
                        case 4: shape = [N, 1]; break;
                        default: break;
                    }
                    if (shape) {
                        blobs.weight('C', shape);
                    }
                }
                break;
            }
            default: {
                break;
            }
        }
        if (params && params.size > 0) {
            for (const [key, signature] of params) {
                if (key.startsWith('@')) {
                    const name = key.substring(1);
                    const identifier = `${this.name}.${name}`;
                    const data = blobs.entry(identifier);
                    const type = ncnn.Utility.type(signature);
                    const tensor = new ncnn.Tensor(type, data, null);
                    const value = new ncnn.Value(identifier, null, tensor);
                    const argument = new ncnn.Argument(name, [value]);
                    this.inputs.push(argument);
                    params.delete(key);
                }
            }
            const attributes = Array.isArray(this.type.attributes) ? this.type.attributes : [];
            for (const [index, obj] of params) {
                const metadata = attributes[index];
                let name = index;
                let value = obj;
                let type = '';
                let visible = true;
                if (metadata) {
                    name = metadata.name;
                    type = metadata.type ? metadata.type : type;
                    switch (type) {
                        case 'int32': {
                            value = parseInt(obj, 10);
                            break;
                        }
                        case 'float32': {
                            value = parseFloat(obj);
                            break;
                        }
                        case 'float32[]': {
                            value = obj.map((v) => parseFloat(v));
                            break;
                        }
                        default: {
                            value = type ? ncnn.Utility.value(obj, type) : obj;
                            break;
                        }
                    }
                    if (metadata && metadata.visible === false) {
                        visible = false;
                    } else if (metadata.default !== undefined) {
                        if (value === metadata.default || (value && value.toString() === metadata.default.toString())) {
                            visible = false;
                        }
                    }
                }
                if (!type && typeof value === 'string') {
                    if (value === 'True') {
                        value = true;
                    } else if (value === 'False') {
                        value = false;
                    } else if (Number.isInteger(Number(value))) {
                        value = Number(value);
                    } else if (value.length > 3 && value.startsWith('(') && value.endsWith(')')) {
                        const list = value.substring(1, value.length - 1).split(',').map((item) => Number(item.trim()));
                        if (list.every((item) => Number.isInteger(item))) {
                            value = list.map((item) => parseInt(item, 10));
                        }
                    }
                }
                const argument = new ncnn.Argument(name, value, type, visible);
                this.attributes.push(argument);
            }
        }
    }
};

ncnn.Tensor = class {

    constructor(type, values, quantization) {
        this.type = type;
        this.values = values;
        this.quantization = quantization;
    }
};

ncnn.TensorType = class {

    constructor(dataType, shape) {
        this.dataType = dataType || '?';
        this.shape = shape;
    }

    equals(obj) {
        return obj && this.dataType === obj.dataType && this.shape && this.shape.equals(obj.shape);
    }

    toString() {
        return this.dataType + this.shape.toString();
    }
};

ncnn.TensorShape = class {

    constructor(dimensions) {
        this.dimensions = dimensions;
    }

    equals(obj) {
        return obj && Array.isArray(obj.dimensions) && Array.isArray(this.dimensions) && this.dimensions.length === obj.dimensions.length && obj.dimensions.every((value, index) => Object.is(this.dimensions[index], value));
    }

    toString() {
        return this.dimensions ? (`[${this.dimensions.map((dimension) => dimension ? dimension.toString() : '?').join(',')}]`) : '';
    }
};

ncnn.Utility = class {

    static value(value, type) {
        ncnn.Utility._enum = ncnn.Utility._enum || new Map([
            ['BinaryOpType', ['Add', 'Sub', 'Mul', 'Div', 'Max', 'Min', 'Pow', 'RSub', 'RDiv']],
            ['CastOpType', ['Auto', 'Float32', 'Float16', 'Int8', 'BFloat16']],
            ['EltwiseType', ['Prod', 'Sum', 'Max']],
            ['PaddingType', ['Constant', 'Replicate', 'Reflect']],
            ['PoolingType', ['Max', 'Average']],
            ['InterpResizeType', ['', 'Nearest', 'Bilinear', 'Bicubic']],
            ['PermuteOrderType', ['WH WHC WHDC', 'HW HWC HWDC', 'WCH WDHC', 'CWH DWHC', 'HCW HDWC', 'CHW DHWC', 'WHCD', 'HWCD', 'WCHD', 'CWHD', 'HCWD', 'CHWD', 'WDCH', 'DWCH', 'WCDH', 'CWDH', 'DCWH', 'CDWH', 'HDCW', 'DHCW', 'HCDW', 'CHDW', 'DCHW', 'CDHW']],
            ['ReductionOpType', ['Sum', 'ASum', 'SumSq', 'Mean', 'Max', 'Min', 'Prod', 'L1', 'L2', 'LogSum', 'LogSumExp']],
            ['UnaryOpType', ['Abs', 'Neg', 'Floor', 'Ceil', 'Square', 'Sqrt', 'Rsq', 'Exp', 'Log', 'Sin', 'Cos', 'Tan', 'ASin', 'ACos', 'ATan', 'Reciprocal', 'Tanh']]
        ]);
        if (ncnn.Utility._enum.has(type) && typeof value === 'string') {
            const index = parseInt(value, 10);
            const list = ncnn.Utility._enum.get(type);
            if (Number.isInteger(index) && index < list.length) {
                return list[index];
            }
        }
        return value;
    }

    static type(signature) {
        const match = signature.match(/\(([^)]+)\)(\w+)/);
        const shape = new ncnn.TensorShape(match[1].split(',').map((v) => parseInt(v, 10)));
        const dataTypes = new Map([['f32', 'float32'], ['f16', 'float16']]);
        const dataType = dataTypes.get(match[2]) || match[2];
        return new ncnn.TensorType(dataType, shape);
    }

    static route(params, id) {
        const key = `#${id}`;
        if (params && params.has(key)) {
            const signature = params.get(key);
            params.delete(key);
            return ncnn.Utility.type(signature);
        }
        return null;
    }
};

ncnn.TextParamReader = class {

    constructor(reader) {
        const lines = [];
        for (let line = reader.read('\n'); line !== undefined; line = reader.read('\n')) {
            line = line.trim();
            lines.push(line);
        }
        const signature = lines.shift();
        const header = (signature === '7767517' ? lines.shift() : signature).split(' ');
        if (header.length !== 2 || !header.every((value) => value >>> 0 === parseFloat(value))) {
            throw new ncnn.Error('Invalid header.');
        }
        this.layers = [];
        while (lines.length > 0) {
            const line = lines.shift();
            if (line.length > 0) {
                const columns = line.split(' ').filter((s) => s.length !== 0);
                const type = columns.shift();
                const name = columns.shift();
                const inputCount = parseInt(columns.shift(), 10);
                const outputCount = parseInt(columns.shift(), 10);
                const inputs = columns.splice(0, inputCount);
                const outputs = columns.splice(0, outputCount);
                const params = new Map();
                let index = 0;
                for (const column of columns) {
                    const parts = column.split('=');
                    if (parts.length > 2) {
                        throw new ncnn.Error(`Invalid attribute '${column}'.`);
                    }
                    let key = (parts.length === 2) ? parts[0].trim() : index.toString();
                    let value = (parts.length === 2) ? parts[1].trim() : parts[0].trim();
                    const keyInt = parseInt(key, 10);
                    if (keyInt < 0) {
                        value = value.split(',').map((v) => v.trim());
                        value.shift();
                        key = (-(keyInt + 23300)).toString();
                    }
                    params.set(key, value);
                    index++;
                }
                this.layers.push({ type, name, inputs, outputs, params });
            }
        }
    }
};

ncnn.BinaryParamReader = class {

    constructor(buffer) {
        const reader = base.BinaryReader.open(buffer);
        if (reader.int32() !== 0x007685dd) {
            throw new ncnn.Error('Invalid signature.');
        }
        const layerCount = reader.int32();
        /* const blobCount = */ reader.int32();
        this.layers = [];
        for (let i = 0; i < layerCount; i++) {
            const type = reader.int32();
            const name = i.toString();
            const inputs = new Array(reader.int32());
            const outputs = new Array(reader.int32());
            for (let j = 0; j < inputs.length; j++) {
                inputs[j] = reader.int32().toString();
            }
            for (let j = 0; j < outputs.length; j++) {
                outputs[j] = reader.int32().toString();
            }
            const params = new Map();
            let id = reader.int32();
            while (id !== -233) {
                const isArray = id <= -23300;
                if (isArray) {
                    id = -id - 23300;
                }
                const key = id.toString();
                if (isArray) {
                    const length = reader.int32();
                    const values = [];
                    for (let i = 0; i < length; i++) {
                        values.push(reader.int32());
                    }
                    params.set(key, values);
                } else {
                    const value = reader.int32();
                    params.set(key, value);
                }
                id = reader.int32();
            }
            this.layers.push({ type, name, inputs, outputs, params });
        }
    }
};

ncnn.BlobReader = class {

    static async open(context) {
        let identifier = null;
        let entries = new Map();
        let buffer = null;
        let position = 0;
        if (context) {
            identifier = context.identifier;
            if (identifier.toLowerCase().endsWith('.pnnx.bin')) {
                entries = await context.peek('zip');
            } else {
                buffer = context.stream.peek();
                position = 0;
            }
        }
        return new ncnn.BlobReader(identifier, entries, buffer, position);
    }

    constructor(identifier, entries, buffer, position) {
        this._identifier = identifier;
        this._entries = entries;
        this._buffer = buffer;
        this._position = position;
    }

    skip(length) {
        this._position += length;
        if (this._position > this._buffer.length) {
            throw new ncnn.Error('Unexpected end of file.');
        }
    }

    read(length) {
        const position = this._position;
        this.skip(length);
        return this._buffer.subarray(position, this._position);
    }

    align(size) {
        const remainder = this._position % size;
        if (remainder !== 0) {
            this.skip(size - remainder);
        }
    }

    load(shape, type) {
        if (!this._buffer) {
            return null;
        }
        const size = shape.reduce((a, b) => a * b, 1);
        if (type === 0) {
            const buffer = this.read(4);
            const [f0, f1, f2, f3] = buffer;
            const flag = (f0 | f1 << 8 | f2 << 16 | f3 << 24) >>> 0;
            // https://github.com/Tencent/ncnn/blob/master/src/modelbin.cpp
            if (flag === 0x01306B47) { // float16
                const data = this.read(size * 2);
                this.align(4);
                return { dataType: 'float16', data };
            } else if (flag === 0x000D4B38) { // int8
                const data = this.read(size);
                this.align(4);
                return { dataType: 'int8', data };
            } else if (flag === 0x00000001) { // qint8
                // this.skip(size + 1024);
                // data = null;
                // return { dataType: 'qint8', data };
                throw new ncnn.Error("Unsupported weight type '0x00000001'.");
            } else if (flag === 0x0002C056) {
                // size * sizeof(float) - raw data with extra scaling
                throw new ncnn.Error("Unsupported weight type '0x0002C056'.");
            } else if (flag === 0x00000000) { // float32
                const data = this.read(size * 4);
                return { dataType: 'float32', data };
            } else {
                const size = shape.reduce((a, b) => a * b, 1);
                const buffer = this.read(1024);
                const quantization = {
                    type: 'lookup',
                    value: Array.from(new Float32Array(buffer.buffer, buffer.bufferOffset, buffer.length / 4))
                };
                const data = this.read(size);
                this.align(4);
                return { dataType: 'uint8', data, quantization };
            }
        } else if (type === 1) {
            const data = this.read(size * 4);
            return { dataType: 'float32', data };
        }
        throw new ncnn.Error(`Load type '${type}' not supported.`);
    }

    validate() {
        const files = [
            ['encoder_jit_trace-pnnx.ncnn.bin', 139191256]
        ];
        if (this._buffer && this._buffer.length !== this._position &&
            !this._identifier.toLowerCase().endsWith('.pnnx.bin') &&
            !files.find((file) => file[0] === this._identifier && file[1] === this._buffer.length)) {
            throw new ncnn.Error('Invalid weights data size.');
        }
    }

    entry(identifier) {
        if (this._entries && this._entries.has(identifier)) {
            const reader = this._entries.get(identifier);
            return reader.peek();
        }
        return null;
    }
};

pnnx.Metadata = class {

    static async open(context) {
        if (!pnnx.Metadata._metadata) {
            let data = null;
            try {
                data = await context.request('pytorch-metadata.json');
            } catch {
                // continue regardless of error
            }
            pnnx.Metadata._metadata = new pnnx.Metadata(data);
        }
        return pnnx.Metadata._metadata;
    }

    constructor(data) {
        this._types = new Map();
        this._attributes = new Map();
        this._index = new Map();
        if (data) {
            const items = JSON.parse(data);
            for (const item of items) {
                item.name = item.name.replace(/^torch\.nn\.modules\.(\w)+\./, 'nn.');
                const match = item.name.match(/^aten::([a-z_]+)/);
                if (match) {
                    const name = match[1];
                    if (item.category) {
                        if (!this._types.has(`torch.${name}`)) {
                            this._types.set(`torch.${name}`, { name: `torch.${name}`, category: item.category });
                        }
                        if (!this._types.has(`F.${name}`)) {
                            this._types.set(`F.${name}`, { name: `F.${name}`, category: item.category });
                        }
                    }
                }
                this._types.set(item.name, { name: item.name, category: item.category });
            }
        }
    }

    type(name) {
        if (!this._types.has(name)) {
            this._types.set(name, { name: name.toString() });
        }
        return this._types.get(name);
    }

    attribute(type, name) {
        const key = `${type}:${name}`;
        if (!this._attributes.has(key)) {
            this._attributes.set(key, null);
            const metadata = this.type(type);
            if (metadata) {
                if (metadata.inputs) {
                    for (const input of metadata.inputs) {
                        this._attributes.set(`${type}:${input.name}`, input);
                    }
                }
                if (metadata.attributes) {
                    for (const attribute of metadata.attributes) {
                        this._attributes.set(`${type}:${attribute.name}`, attribute);
                    }
                }
            }
        }
        return this._attributes.get(key);
    }
};

ncnn.Error = class extends Error {

    constructor(message) {
        super(message);
        this.name = 'Error loading ncnn model.';
    }
};

export const ModelFactory = ncnn.ModelFactory;