netron/source/darknet.js

const darknet = {};

darknet.ModelFactory = class {

    async match(context) {
        const identifier = context.identifier;
        const extension = identifier.lastIndexOf('.') > 0 ? identifier.split('.').pop().toLowerCase() : '';
        if (extension === 'weights' && !identifier.toLowerCase().endsWith('.espresso.weights')) {
            const weights = await darknet.Weights.open(context);
            if (weights) {
                return context.set('darknet.weights', weights);
            }
            return null;
        }
        const reader = await context.read('text', 65536);
        if (reader) {
            try {
                for (let line = reader.read('\n'); line !== undefined; line = reader.read('\n')) {
                    const content = line.trim();
                    if (content.length > 0 && !content.startsWith('#')) {
                        if (content.startsWith('[') && content.endsWith(']')) {
                            return context.set('darknet.model');
                        }
                        return null;
                    }
                }
            } catch {
                // continue regardless of error
            }
        }
        return null;
    }

    async open(context) {
        const metadata = await context.metadata('darknet-metadata.json');
        const identifier = context.identifier;
        const parts = identifier.split('.');
        parts.pop();
        const basename = parts.join('.');
        switch (context.type) {
            case 'darknet.weights': {
                const weights = context.value;
                const name = `${basename}.cfg`;
                const content = await context.fetch(name);
                const reader = await content.read('text');
                const configuration = new darknet.Configuration(reader, content.identifier);
                return new darknet.Model(metadata, configuration, weights);
            }
            case 'darknet.model': {
                try {
                    const name = `${basename}.weights`;
                    const content = await context.fetch(name);
                    const weights = await darknet.Weights.open(content);
                    const reader = await context.read('text');
                    const configuration = new darknet.Configuration(reader, context.identifier);
                    return new darknet.Model(metadata, configuration, weights);
                } catch {
                    const reader = await context.read('text');
                    const configuration = new darknet.Configuration(reader, context.identifier);
                    return new darknet.Model(metadata, configuration, null);
                }
            }
            default: {
                throw new darknet.Error(`Unsupported Darknet format '${context.type}'.`);
            }
        }
    }
};

darknet.Model = class {

    constructor(metadata, configuration, weights) {
        this.format = 'Darknet';
        this.modules = [new darknet.Graph(metadata, configuration, weights)];
    }
};

darknet.Graph = class {

    constructor(metadata, configuration, weights) {
        this.inputs = [];
        this.outputs = [];
        this.nodes = [];
        const params = {};
        const sections = configuration.read();
        const globals = new Map();
        const net = sections.shift();
        const option_find_int = (options, key, defaultValue) => {
            let value = options[key];
            if (typeof value === 'string' && value.startsWith('$')) {
                const key = value.substring(1);
                value = globals.has(key) ? globals.get(key) : value;
            }
            if (value !== undefined) {
                const number = parseInt(value, 10);
                if (!Number.isInteger(number)) {
                    throw new darknet.Error(`Invalid int option '${JSON.stringify(options[key])}'.`);
                }
                return number;
            }
            return defaultValue;
        };
        const option_find_str = (options, key, defaultValue) => {
            const value = options[key];
            return value === undefined ? defaultValue : value;
        };
        const make_shape = (dimensions, source) => {
            if (dimensions.some((dimension) => dimension === 0 || dimension === undefined || isNaN(dimension))) {
                throw new darknet.Error(`Invalid tensor shape '${JSON.stringify(dimensions)}' in '${source}'.`);
            }
            return new darknet.TensorShape(dimensions);
        };
        const load_weights = (name, shape, visible) => {
            const data = weights ? weights.read(4 * shape.reduce((a, b) => a * b, 1)) : null;
            const type = new darknet.TensorType('float32', make_shape(shape, 'load_weights'));
            const initializer = new darknet.Tensor(type, data);
            const value = new darknet.Value('', null, initializer);
            return new darknet.Argument(name, [value], null, visible !== false);
        };
        const load_batch_normalize_weights = (layer, prefix, size) => {
            layer.weights.push(load_weights(`${prefix}scale`, [size], prefix === ''));
            layer.weights.push(load_weights(`${prefix}mean`, [size], prefix === ''));
            layer.weights.push(load_weights(`${prefix}variance`, [size], prefix === ''));
        };
        const make_convolutional_layer = (layer, prefix, w, h, c, n, groups, size, stride_x, stride_y, padding, batch_normalize) => {
            layer.out_w = Math.floor((w + 2 * padding - size) / stride_x) + 1;
            layer.out_h = Math.floor((h + 2 * padding - size) / stride_y) + 1;
            layer.out_c = n;
            layer.out = layer.out_w * layer.out_h * layer.out_c;
            layer.weights.push(load_weights(`${prefix}biases`, [n], prefix === ''));
            if (batch_normalize) {
                if (prefix) {
                    load_batch_normalize_weights(layer, prefix, n);
                } else {
                    const batchnorm_layer = { weights: [] };
                    load_batch_normalize_weights(batchnorm_layer, prefix, n);
                    layer.chain.push({ type: 'batchnorm', layer: batchnorm_layer });
                }
            }
            layer.weights.push(load_weights(`${prefix}weights`, [Math.floor(c / groups), n, size, size], prefix === ''));
            layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'make_convolutional_layer'));
        };
        const make_deconvolutional_layer = (l, batch, h, w, c, n, size, stride, padding, activation, batch_normalize) => {
            const pad = padding;
            l.out_w = Math.floor((w - 1) * stride + size - 2 * pad);
            l.out_h = Math.floor((h - 1) * stride + size - 2 * pad);
            l.out_c = n;
            l.out = l.out_w * l.out_h * l.out_c;
            l.weights.push(load_weights(`biases`, [n]));
            if (batch_normalize) {
                const batchnorm_layer = { weights: [] };
                load_batch_normalize_weights(batchnorm_layer, '', n);
                l.chain.push({ type: 'batchnorm', layer: batchnorm_layer });
            }
            l.weights.push(load_weights(`weights`, [c, n, size, size]));
            l.outputs[0].type = new darknet.TensorType('float32', make_shape([l.out_w, l.out_h, l.out_c], 'make_convolutional_layer'));
        };
        const make_connected_layer = (layer, prefix, inputs, outputs, batch_normalize) => {
            layer.out_h = 1;
            layer.out_w = 1;
            layer.out_c = outputs;
            layer.out = outputs;
            layer.weights.push(load_weights(`${prefix}biases`, [outputs], prefix === ''));
            if (batch_normalize) {
                if (prefix) {
                    load_batch_normalize_weights(layer, prefix, outputs);
                } else {
                    const batchnorm_layer = { weights: [] };
                    load_batch_normalize_weights(batchnorm_layer, prefix, outputs);
                    layer.chain.push({ type: 'batchnorm', layer: batchnorm_layer });
                }
            }
            layer.weights.push(load_weights(`${prefix}weights`, [inputs, outputs], prefix === ''));
            layer.outputs[0].type = new darknet.TensorType('float32', make_shape([outputs], 'make_connected_layer'));
        };
        if (sections.length === 0) {
            throw new darknet.Error('Config file has no sections.');
        }
        switch (net.type) {
            case 'net':
            case 'network': {
                params.h = option_find_int(net.options, 'height', 0);
                params.w = option_find_int(net.options, 'width', 0);
                params.c = option_find_int(net.options, 'channels', 0);
                params.inputs = option_find_int(net.options, 'inputs', params.h * params.w * params.c);
                for (const key of Object.keys(net.options)) {
                    globals.set(key, net.options[key]);
                }
                break;
            }
            default: {
                throw new darknet.Error(`Unexpected '[${net.type}]' section. First section must be [net] or [network].`);
            }
        }
        const inputType = params.w && params.h && params.c ?
            new darknet.TensorType('float32', make_shape([params.w, params.h, params.c], 'params-if')) :
            new darknet.TensorType('float32', make_shape([params.inputs], ''));
        const inputName = 'input';
        params.value = [new darknet.Value(inputName, inputType, null)];
        this.inputs.push(new darknet.Argument(inputName, params.value));
        for (let i = 0; i < sections.length; i++) {
            const section = sections[i];
            section.name = i.toString();
            section.layer = {
                inputs: [],
                weights: [],
                outputs: [new darknet.Value(section.name, null, null)],
                chain: []
            };
        }
        let infer = true;
        for (let i = 0; i < sections.length; i++) {
            const section = sections[i];
            const options = section.options;
            const layer = section.layer;
            layer.inputs.push(...params.value);
            switch (section.type) {
                case 'shortcut': {
                    let remove = true;
                    const from = options.from ? options.from.split(',').map((item) => Number.parseInt(item.trim(), 10)) : [];
                    for (const route of from) {
                        const index = route < 0 ? i + route : route;
                        const exists = index >= 0 && index < sections.length;
                        remove = exists && remove;
                        if (exists) {
                            const source = sections[index].layer;
                            layer.inputs.push(source.outputs[0]);
                        }
                    }
                    if (remove) {
                        delete options.from;
                    }
                    break;
                }
                case 'sam':
                case 'scale_channels': {
                    const from = option_find_int(options, 'from', 0);
                    const index = from < 0 ? i + from : from;
                    if (index >= 0 && index < sections.length) {
                        const source = sections[index].layer;
                        layer.from = source;
                        layer.inputs.push(source.outputs[0]);
                        delete options.from;
                    }
                    break;
                }
                case 'route': {
                    layer.inputs = [];
                    layer.layers = [];
                    let remove = true;
                    const routes = options.layers ? options.layers.split(',').map((route) => Number.parseInt(route.trim(), 10)) : [];
                    for (const route of routes) {
                        const index = route < 0 ? i + route : route;
                        const exists = index >= 0 && index < sections.length;
                        remove = exists && remove;
                        if (exists) {
                            const source = sections[index].layer;
                            layer.inputs.push(source.outputs[0]);
                            layer.layers.push(source);
                        }
                    }
                    if (remove) {
                        delete options.layers;
                    }
                    break;
                }
                default:
                    break;
            }
            if (infer) {
                switch (section.type) {
                    case 'conv':
                    case 'convolutional': {
                        const shape = layer.inputs[0].type.shape.dimensions;
                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
                            throw new darknet.Error('Layer before convolutional layer must output image.');
                        }
                        const size = option_find_int(options, 'size', 1);
                        const n = option_find_int(options, 'filters', 1);
                        const pad = option_find_int(options, 'pad', 0);
                        const padding = pad ? (size >> 1) : option_find_int(options, 'padding', 0);
                        let stride_x = option_find_int(options, 'stride_x', -1);
                        let stride_y = option_find_int(options, 'stride_y', -1);
                        if (stride_x < 1 || stride_y < 1) {
                            const stride = option_find_int(options, 'stride', 1);
                            stride_x = stride_x < 1 ? stride : stride_x;
                            stride_y = stride_y < 1 ? stride : stride_y;
                        }
                        const groups = option_find_int(options, 'groups', 1);
                        const batch_normalize = option_find_int(options, 'batch_normalize', 0);
                        const activation = option_find_str(options, 'activation', 'logistic');
                        make_convolutional_layer(layer, '', params.w, params.h, params.c, n, groups, size, stride_x, stride_y, padding, batch_normalize);
                        if (activation !== 'logistic' && activation !== 'none') {
                            layer.chain.push({ type: activation });
                        }
                        break;
                    }
                    case 'deconvolutional': {
                        const shape = layer.inputs[0].type.shape.dimensions;
                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
                            throw new darknet.Error('Layer before convolutional layer must output image.');
                        }
                        const n = option_find_int(options, 'filters', 1);
                        const size = option_find_int(options, 'size', 1);
                        const stride = option_find_int(options, 'stride', 1);
                        const activation = option_find_str(options, 'activation', 'logistic');
                        const h = params.h;
                        const w = params.w;
                        const c = params.c;
                        const batch = params.batch;
                        let padding = option_find_int(options, 'padding', 0);
                        const pad = option_find_int(options, 'pad', 0);
                        if (pad) {
                            padding = size / 2;
                        }
                        const batch_normalize = option_find_int(options, 'batch_normalize', 0);
                        make_deconvolutional_layer(layer, batch, h, w, c, n, size, stride, padding, activation, batch_normalize);
                        if (activation !== 'logistic' && activation !== 'none') {
                            layer.chain.push({ type: activation });
                        }
                        break;
                    }
                    case 'connected': {
                        const outputs = option_find_int(options, 'output', 1);
                        const batch_normalize = option_find_int(options, 'batch_normalize', 0);
                        const activation = option_find_str(options, 'activation', 'logistic');
                        make_connected_layer(layer, '', params.inputs, outputs, batch_normalize);
                        if (activation !== 'logistic' && activation !== 'none') {
                            layer.chain.push({ type: activation });
                        }
                        break;
                    }
                    case 'local': {
                        const shape = layer.inputs[0].type.shape.dimensions;
                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
                            throw new darknet.Error('Layer before avgpool layer must output image.');
                        }
                        const n = option_find_int(options, 'filters' , 1);
                        const size = option_find_int(options, 'size', 1);
                        const stride = option_find_int(options, 'stride', 1);
                        const pad = option_find_int(options, 'pad', 0);
                        const activation = option_find_str(options, 'activation', 'logistic');
                        layer.out_h = Math.floor((params.h - (pad ? 1 : size)) / stride) + 1;
                        layer.out_w = Math.floor((params.w - (pad ? 1 : size)) / stride) + 1;
                        layer.out_c = n;
                        layer.out = layer.out_w * layer.out_h * layer.out_c;
                        layer.weights.push(load_weights('weights', [params.c, n, size, size, layer.out_h * layer.out_w]));
                        layer.weights.push(load_weights('biases',[layer.out_w * layer.out_h * layer.out_c]));
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'local'));
                        if (activation !== 'logistic' && activation !== 'none') {
                            layer.chain.push({ type: activation });
                        }
                        break;
                    }
                    case 'batchnorm': {
                        layer.out_h = params.h;
                        layer.out_w = params.w;
                        layer.out_c = params.c;
                        layer.out = layer.in;
                        load_batch_normalize_weights(layer, '', layer.out_c);
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'batchnorm'));
                        break;
                    }
                    case 'activation': {
                        layer.out_h = params.h;
                        layer.out_w = params.w;
                        layer.out_c = params.c;
                        layer.out = layer.in;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'activation'));
                        break;
                    }
                    case 'max':
                    case 'maxpool': {
                        const shape = layer.inputs[0].type.shape.dimensions;
                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
                            throw new darknet.Error('Layer before maxpool layer must output image.');
                        }
                        const antialiasing = option_find_int(options, 'antialiasing', 0);
                        const stride = option_find_int(options, 'stride', 1);
                        const blur_stride_x = option_find_int(options, 'stride_x', stride);
                        const blur_stride_y = option_find_int(options, 'stride_y', stride);
                        const stride_x = antialiasing ? 1 : blur_stride_x;
                        const stride_y = antialiasing ? 1 : blur_stride_y;
                        const size = option_find_int(options, 'size', stride);
                        const padding = option_find_int(options, 'padding', size - 1);
                        const out_channels = option_find_int(options, 'out_channels', 1);
                        const maxpool_depth = option_find_int(options, 'maxpool_depth', 0);
                        if (maxpool_depth) {
                            layer.out_c = out_channels;
                            layer.out_w = params.w;
                            layer.out_h = params.h;
                        } else {
                            layer.out_w = Math.floor((params.w + padding - size) / stride_x) + 1;
                            layer.out_h = Math.floor((params.h + padding - size) / stride_y) + 1;
                            layer.out_c = params.c;
                        }
                        if (antialiasing) {
                            const blur_size = antialiasing === 2 ? 2 : 3;
                            const blur_pad = antialiasing === 2 ? 0 : Math.floor(blur_size / 3);
                            layer.input_layer = { weights: [], outputs: layer.outputs, chain: [] };
                            make_convolutional_layer(layer.input_layer, '', layer.out_h, layer.out_w, layer.out_c, layer.out_c, layer.out_c, blur_size, blur_stride_x, blur_stride_y, blur_pad, 0);
                            layer.out_w = layer.input_layer.out_w;
                            layer.out_h = layer.input_layer.out_h;
                            layer.out_c = layer.input_layer.out_c;
                        } else {
                            layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'maxpool'));
                        }
                        layer.out = layer.out_w * layer.out_h * layer.out_c;
                        break;
                    }
                    case 'avgpool': {
                        const shape = layer.inputs[0].type.shape.dimensions;
                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
                            throw new darknet.Error('Layer before avgpool layer must output image.');
                        }
                        layer.out_w = 1;
                        layer.out_h = 1;
                        layer.out_c = params.c;
                        layer.out = layer.out_c;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'avgpool'));
                        break;
                    }
                    case 'crnn': {
                        const size = option_find_int(options, 'size', 3);
                        const stride = option_find_int(options, 'stride', 1);
                        const output_filters = option_find_int(options, 'output', 1);
                        const hidden_filters = option_find_int(options, 'hidden', 1);
                        const groups = option_find_int(options, 'groups', 1);
                        const pad = option_find_int(options, 'pad', 0);
                        const padding = pad ? (size >> 1) : option_find_int(options, 'padding', 0);
                        const batch_normalize = option_find_int(options, 'batch_normalize', 0);
                        layer.input_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_convolutional_layer(layer.input_layer, 'input_', params.h, params.w, params.c, hidden_filters, groups, size, stride, stride, padding, batch_normalize);
                        layer.self_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_convolutional_layer(layer.self_layer, 'self_', params.h, params.w, hidden_filters, hidden_filters, groups, size, stride, stride, padding, batch_normalize);
                        layer.output_layer = { weights: [], outputs: layer.outputs, chain: [] };
                        make_convolutional_layer(layer.output_layer, 'output_', params.h, params.w, hidden_filters, output_filters, groups, size, stride, stride, padding, batch_normalize);
                        layer.weights = layer.weights.concat(layer.input_layer.weights);
                        layer.weights = layer.weights.concat(layer.self_layer.weights);
                        layer.weights = layer.weights.concat(layer.output_layer.weights);
                        layer.out_h = layer.output_layer.out_h;
                        layer.out_w = layer.output_layer.out_w;
                        layer.out_c = output_filters;
                        layer.out = layer.output_layer.out;
                        break;
                    }
                    case 'rnn': {
                        const outputs = option_find_int(options, 'output', 1);
                        const hidden = option_find_int(options, 'hidden', 1);
                        const batch_normalize = option_find_int(options, 'batch_normalize', 0);
                        const inputs = params.inputs;
                        layer.input_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.input_layer, 'input_', inputs, hidden, batch_normalize);
                        layer.self_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.self_layer, 'self_', hidden, hidden, batch_normalize);
                        layer.output_layer = { weights: [], outputs: layer.outputs, chain: [] };
                        make_connected_layer(layer.output_layer, 'output_', hidden, outputs, batch_normalize);
                        layer.weights = layer.weights.concat(layer.input_layer.weights);
                        layer.weights = layer.weights.concat(layer.self_layer.weights);
                        layer.weights = layer.weights.concat(layer.output_layer.weights);
                        layer.out_w = 1;
                        layer.out_h = 1;
                        layer.out_c = outputs;
                        layer.out = outputs;
                        break;
                    }
                    case 'gru': {
                        const inputs = params.inputs;
                        const outputs = option_find_int(options, 'output', 1);
                        const batch_normalize = option_find_int(options, 'batch_normalize', 0);
                        layer.input_z_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.input_z_layer, 'input_z', inputs, outputs, batch_normalize);
                        layer.state_z_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.state_z_layer, 'state_z', outputs, outputs, batch_normalize);
                        layer.input_r_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.input_r_layer, 'input_r', inputs, outputs, batch_normalize);
                        layer.state_r_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.state_r_layer, 'state_r', outputs, outputs, batch_normalize);
                        layer.input_h_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.input_h_layer, 'input_h', inputs, outputs, batch_normalize);
                        layer.state_h_layer = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.state_h_layer, 'state_h', outputs, outputs, batch_normalize);
                        layer.weights = layer.weights.concat(layer.input_z_layer.weights);
                        layer.weights = layer.weights.concat(layer.state_z_layer.weights);
                        layer.weights = layer.weights.concat(layer.input_r_layer.weights);
                        layer.weights = layer.weights.concat(layer.state_r_layer.weights);
                        layer.weights = layer.weights.concat(layer.input_h_layer.weights);
                        layer.weights = layer.weights.concat(layer.state_h_layer.weights);
                        layer.out = outputs;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([outputs], 'gru'));
                        break;
                    }
                    case 'lstm': {
                        const inputs = params.inputs;
                        const outputs = option_find_int(options, 'output', 1);
                        const batch_normalize = option_find_int(options, 'batch_normalize', 0);
                        layer.uf = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.uf, 'uf_', inputs, outputs, batch_normalize);
                        layer.ui = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.ui, 'ui_', inputs, outputs, batch_normalize);
                        layer.ug = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.ug, 'ug_', inputs, outputs, batch_normalize);
                        layer.uo = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.uo, 'uo_', inputs, outputs, batch_normalize);
                        layer.wf = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.wf, 'wf_', outputs, outputs, batch_normalize);
                        layer.wi = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.wi, 'wi_', outputs, outputs, batch_normalize);
                        layer.wg = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.wg, 'wg_', outputs, outputs, batch_normalize);
                        layer.wo = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_connected_layer(layer.wo, 'wo_', outputs, outputs, batch_normalize);
                        layer.weights = layer.weights.concat(layer.uf.weights);
                        layer.weights = layer.weights.concat(layer.ui.weights);
                        layer.weights = layer.weights.concat(layer.ug.weights);
                        layer.weights = layer.weights.concat(layer.uo.weights);
                        layer.weights = layer.weights.concat(layer.wf.weights);
                        layer.weights = layer.weights.concat(layer.wi.weights);
                        layer.weights = layer.weights.concat(layer.wg.weights);
                        layer.weights = layer.weights.concat(layer.wo.weights);
                        layer.out_w = 1;
                        layer.out_h = 1;
                        layer.out_c = outputs;
                        layer.out = outputs;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([outputs], 'lstm'));
                        weights = null;
                        break;
                    }
                    case 'conv_lstm': {
                        const size = option_find_int(options, "size", 3);
                        const stride = option_find_int(options, "stride", 1);
                        const output_filters = option_find_int(options, "output", 1);
                        const groups = option_find_int(options, "groups", 1);
                        const pad = option_find_int(options, "pad", 0);
                        const padding = pad ? (size >> 1) : option_find_int(options, 'padding', 0);
                        const batch_normalize = option_find_int(options, 'batch_normalize', 0);
                        const bottleneck = option_find_int(options, "bottleneck", 0);
                        const peephole = option_find_int(options, "peephole", 0);
                        layer.uf = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                        make_convolutional_layer(layer.uf, 'uf_', params.h, params.w, params.c, output_filters, groups, size, stride, stride, padding, batch_normalize);
                        layer.ui = { weights: [], outputs: [new darknet.Value('', null, null)], chain: []  };
                        make_convolutional_layer(layer.ui, 'ui_', params.h, params.w, params.c, output_filters, groups, size, stride, stride, padding, batch_normalize);
                        layer.ug = { weights: [], outputs: [new darknet.Value('', null, null)], chain: []  };
                        make_convolutional_layer(layer.ug, 'ug_', params.h, params.w, params.c, output_filters, groups, size, stride, stride, padding, batch_normalize);
                        layer.uo = { weights: [], outputs: [new darknet.Value('', null, null)], chain: []  };
                        make_convolutional_layer(layer.uo, 'uo_', params.h, params.w, params.c, output_filters, groups, size, stride, stride, padding, batch_normalize);
                        layer.weights = layer.weights.concat(layer.uf.weights);
                        layer.weights = layer.weights.concat(layer.ui.weights);
                        layer.weights = layer.weights.concat(layer.ug.weights);
                        layer.weights = layer.weights.concat(layer.uo.weights);
                        if (bottleneck) {
                            layer.wf = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                            make_convolutional_layer(layer.wf, 'wf_', params.h, params.w, output_filters * 2, output_filters, groups, size, stride, stride, padding, batch_normalize);
                            layer.weights = layer.weights.concat(layer.wf.weights);
                        } else {
                            layer.wf = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                            make_convolutional_layer(layer.wf, 'wf_', params.h, params.w, output_filters, output_filters, groups, size, stride, stride, padding, batch_normalize);
                            layer.wi = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                            make_convolutional_layer(layer.wi, 'wi_', params.h, params.w, output_filters, output_filters, groups, size, stride, stride, padding, batch_normalize);
                            layer.wg = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                            make_convolutional_layer(layer.wg, 'wg_', params.h, params.w, output_filters, output_filters, groups, size, stride, stride, padding, batch_normalize);
                            layer.wo = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                            make_convolutional_layer(layer.wo, 'wo_', params.h, params.w, output_filters, output_filters, groups, size, stride, stride, padding, batch_normalize);
                            layer.weights = layer.weights.concat(layer.wf.weights);
                            layer.weights = layer.weights.concat(layer.wi.weights);
                            layer.weights = layer.weights.concat(layer.wg.weights);
                            layer.weights = layer.weights.concat(layer.wo.weights);
                        }
                        if (peephole) {
                            layer.vf = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                            make_convolutional_layer(layer.vf, 'vf_', params.h, params.w, output_filters, output_filters, groups, size, stride, stride, padding, batch_normalize);
                            layer.vi = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                            make_convolutional_layer(layer.vi, 'vi_', params.h, params.w, output_filters, output_filters, groups, size, stride, stride, padding, batch_normalize);
                            layer.vo = { weights: [], outputs: [new darknet.Value('', null, null)], chain: [] };
                            make_convolutional_layer(layer.vo, 'vo_', params.h, params.w, output_filters, output_filters, groups, size, stride, stride, padding, batch_normalize);
                            layer.weights = layer.weights.concat(layer.vf.weights);
                            layer.weights = layer.weights.concat(layer.vi.weights);
                            layer.weights = layer.weights.concat(layer.vo.weights);
                        }
                        layer.out_h = layer.uo.out_h;
                        layer.out_w = layer.uo.out_w;
                        layer.out_c = output_filters;
                        layer.out = layer.out_h * layer.out_w * layer.out_c;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'conv_lstm'));
                        break;
                    }
                    case 'softmax': {
                        layer.out_w = params.w;
                        layer.out_h = params.h;
                        layer.out_c = params.c;
                        layer.out = params.inputs;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out], 'softmax'));
                        break;
                    }
                    case 'dropout': {
                        layer.out_w = params.w;
                        layer.out_h = params.h;
                        layer.out_c = params.c;
                        layer.out = params.inputs;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'dropout'));
                        break;
                    }
                    case 'upsample': {
                        const stride = option_find_int(options, 'stride', 2);
                        layer.out_w = params.w * stride;
                        layer.out_h = params.h * stride;
                        layer.out_c = params.c;
                        layer.out = layer.out_w * layer.out_h * layer.out_c;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'upsample'));
                        break;
                    }
                    case 'crop': {
                        const shape = layer.inputs[0].type.shape.dimensions;
                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
                            throw new darknet.Error('Layer before crop layer must output image.');
                        }
                        const crop_height = option_find_int(options, 'crop_height', 1);
                        const crop_width = option_find_int(options, 'crop_width', 1);
                        layer.out_w = crop_width;
                        layer.out_h = crop_height;
                        layer.out_c = params.c;
                        layer.out = layer.out_w * layer.out_h * layer.out_c;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'crop'));
                        break;
                    }
                    case 'yolo': {
                        const classes = option_find_int(options, 'classes', 20);
                        const n = option_find_int(options, 'num', 1);
                        layer.out_h = params.h;
                        layer.out_w = params.w;
                        layer.out_c = n * (classes + 4 + 1);
                        layer.out = layer.out_h * layer.out_w * layer.out_c;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'yolo'));
                        break;
                    }
                    case 'Gaussian_yolo': {
                        const classes = option_find_int(options, 'classes', 20);
                        const n = option_find_int(options, 'num', 1);
                        layer.out_h = params.h;
                        layer.out_w = params.w;
                        layer.out_c = n * (classes + 8 + 1);
                        layer.out = layer.out_h * layer.out_w * layer.out_c;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'Gaussian_yolo'));
                        break;
                    }
                    case 'region': {
                        const coords = option_find_int(options, 'coords', 4);
                        const classes = option_find_int(options, 'classes', 20);
                        const num = option_find_int(options, 'num', 1);
                        layer.out = params.h * params.w * num * (classes + coords + 1);
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([params.h, params.w, num, (classes + coords + 1)], 'region'));
                        break;
                    }
                    case 'cost': {
                        layer.out = params.inputs;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out], 'cost'));
                        break;
                    }
                    case 'reorg': {
                        const stride = option_find_int(options, 'stride', 1);
                        const reverse = option_find_int(options, 'reverse', 0);
                        const extra = option_find_int(options, 'extra', 0);
                        if (reverse) {
                            layer.out_w = params.w * stride;
                            layer.out_h = params.h * stride;
                            layer.out_c = Math.floor(params.c / (stride * stride));
                            layer.out = layer.out_h * layer.out_w * layer.out_c;
                        } else {
                            layer.out_w = Math.floor(params.w / stride);
                            layer.out_h = Math.floor(params.h / stride);
                            layer.out_c = params.c * (stride * stride);
                            layer.out = layer.out_h * layer.out_w * layer.out_c;
                        }
                        if (extra) {
                            layer.out_w = 0;
                            layer.out_h = 0;
                            layer.out_c = 0;
                            layer.out = (params.h * params.w * params.c) + extra;
                            layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out], 'reorg'));
                        } else {
                            layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'reorg'));
                        }
                        break;
                    }
                    case 'route': {
                        const layers = [].concat(layer.layers);
                        const groups = option_find_int(options, 'groups', 1);
                        layer.out = 0;
                        for (const next of layers) {
                            layer.out += next.outputs / groups;
                        }
                        if (layers.length > 0) {
                            const first = layers.shift();
                            layer.out_w = first.out_w;
                            layer.out_h = first.out_h;
                            layer.out_c = first.out_c / groups;
                            while (layers.length > 0) {
                                const next = layers.shift();
                                if (next.out_w === first.out_w && next.out_h === first.out_h) {
                                    layer.out_c += next.out_c;
                                    continue;
                                }
                                infer = false;
                                break;
                            }
                            if (infer) {
                                layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'route'));
                            }
                        } else {
                            infer = false;
                        }
                        if (!infer) {
                            layer.out_h = 0;
                            layer.out_w = 0;
                            layer.out_c = 0;
                        }
                        break;
                    }
                    case 'sam':
                    case 'scale_channels': {
                        const activation = option_find_str(options, 'activation', 'linear');
                        const from = layer.from;
                        if (from) {
                            layer.out_w = from.out_w;
                            layer.out_h = from.out_h;
                            layer.out_c = from.out_c;
                            layer.out = layer.out_w * layer.out_h * layer.out_c;
                            layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out_w, layer.out_h, layer.out_c], 'shortcut|scale_channels|sam'));
                        }
                        if (activation !== 'linear' && activation !== 'none') {
                            layer.chain.push({ type: activation });
                        }
                        break;
                    }
                    case 'shortcut': {
                        const activation = option_find_str(options, 'activation', 'linear');
                        layer.out_w = params.w;
                        layer.out_h = params.h;
                        layer.out_c = params.c;
                        layer.out = params.w * params.h * params.c;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([params.w, params.h, params.c], 'shortcut|scale_channels|sam'));
                        if (activation !== 'linear' && activation !== 'none') {
                            layer.chain.push({ type: activation });
                        }
                        break;
                    }
                    case 'detection': {
                        layer.out_w = params.w;
                        layer.out_h = params.h;
                        layer.out_c = params.c;
                        layer.out = params.inputs;
                        layer.outputs[0].type = new darknet.TensorType('float32', make_shape([layer.out], 'detection'));
                        break;
                    }
                    default: {
                        infer = false;
                        break;
                    }
                }
                params.h = layer.out_h;
                params.w = layer.out_w;
                params.c = layer.out_c;
                params.inputs = layer.out;
                params.last = section;
            }
            params.value = layer.outputs;
        }

        for (let i = 0; i < sections.length; i++) {
            this.nodes.push(new darknet.Node(metadata, net, sections[i]));
        }

        if (weights) {
            weights.validate();
        }
    }
};

darknet.Argument = class {

    constructor(name, value, type = null, visible = true) {
        this.name = name;
        this.value = value;
        this.type = type;
        this.visible = visible;

    }
};

darknet.Value = class {

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

darknet.Node = class {

    constructor(metadata, net, section) {
        this.name = section.name || '';
        this.identifier = section.line === undefined ? undefined : section.line.toString();
        this.attributes = [];
        this.inputs = [];
        this.outputs = [];
        this.chain = [];
        const type = section.type;
        this.type = metadata.type(type) || { name: type };
        const layer = section.layer;
        if (layer && layer.inputs && layer.inputs.length > 0) {
            this.inputs.push(new darknet.Argument(layer.inputs.length <= 1 ? 'input' : 'inputs', layer.inputs));
        }
        if (layer && layer.weights && layer.weights.length > 0) {
            this.inputs = this.inputs.concat(layer.weights);
        }
        if (layer && layer.outputs && layer.outputs.length > 0) {
            this.outputs.push(new darknet.Argument(layer.outputs.length <= 1 ? 'output' : 'outputs', layer.outputs));
        }
        if (layer && layer.chain) {
            for (const chain of layer.chain) {
                this.chain.push(new darknet.Node(metadata, net, chain, ''));
            }
        }
        const options = section.options;
        if (options) {
            for (const [name, obj] of Object.entries(options)) {
                const schema = metadata.attribute(section.type, name);
                let type = null;
                let value = obj;
                let visible = true;
                if (schema) {
                    type = schema.type || '';
                    switch (type) {
                        case '':
                        case 'string': {
                            break;
                        }
                        case 'int32': {
                            const number = parseInt(value, 10);
                            if (Number.isInteger(number)) {
                                value = number;
                            }
                            break;
                        }
                        case 'float32': {
                            const number = parseFloat(value);
                            if (!isNaN(number)) {
                                value = number;
                            }
                            break;
                        }
                        case 'int32[]': {
                            const numbers = value.split(',').map((item) => parseInt(item.trim(), 10));
                            if (numbers.every((number) => Number.isInteger(number))) {
                                value = numbers;
                            }
                            break;
                        }
                        default: {
                            throw new darknet.Error(`Unsupported attribute type '${type}'.`);
                        }
                    }
                    visible = (schema.visible === false || value === schema.default);
                }
                const attribute = new darknet.Argument(name, value, type, visible);
                this.attributes.push(attribute);
            }
        }
    }
};

darknet.Tensor = class {

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

darknet.TensorType = class {

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

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

darknet.TensorShape = class {

    constructor(dimensions) {
        if (dimensions.some((dimension) => dimension === 0 || dimension === undefined || isNaN(dimension))) {
            throw new darknet.Error(`Invalid tensor shape '${JSON.stringify(dimensions)}'.`);
        }
        this.dimensions = dimensions;
    }

    toString() {
        if (this.dimensions) {
            if (this.dimensions.length === 0) {
                return '';
            }
            return `[${this.dimensions.map((dimension) => dimension.toString()).join(',')}]`;
        }
        return '';
    }
};

darknet.Configuration = class {

    constructor(reader, identifier) {
        this.reader = reader;
        this.identifier = identifier;
    }

    read() {
        // read_cfg
        const sections = [];
        let section = null;
        const reader = this.reader;
        let lineNumber = 0;
        const setup = /^setup.*\.cfg$/.test(this.identifier);
        for (let content = reader.read('\n'); content !== undefined; content = reader.read('\n')) {
            lineNumber++;
            const line = content.replace(/\s/g, '');
            if (line.length > 0) {
                switch (line[0]) {
                    case '#':
                    case ';':
                        break;
                    case '[': {
                        const type = line[line.length - 1] === ']' ? line.substring(1, line.length - 1) : line.substring(1);
                        if (setup) {
                            if (type === 'metadata' || type === 'global' || type === 'wheel' ||
                                type === 'isort' || type === 'flake8' || type === 'build_ext' ||
                                type.startsWith('bdist_') || type.startsWith('tool:') || type.startsWith('coverage:')) {
                                throw new darknet.Error('Invalid file content. File contains Python setup configuration data.');
                            }
                        }
                        section = {
                            line: lineNumber,
                            type,
                            options: {}
                        };
                        sections.push(section);
                        break;
                    }
                    default: {
                        if (!section || line[0] < 0x20 || line[0] > 0x7E) {
                            throw new darknet.Error(`Invalid cfg '${content.replace(/[^\x20-\x7E]+/g, '?').trim()}' at line ${lineNumber}.`);
                        }
                        const index = line.indexOf('=');
                        if (index < 0) {
                            throw new darknet.Error(`Invalid cfg '${content.replace(/[^\x20-\x7E]+/g, '?').trim()}' at line ${lineNumber}.`);
                        }
                        const key = line.substring(0, index);
                        const value = line.substring(index + 1);
                        section.options[key] = value;
                        break;
                    }
                }
            }
        }
        return sections;
    }
};

darknet.Weights = class {

    static async open(context) {
        const reader = await context.read('binary');
        if (reader && reader.length >= 20) {
            const major = reader.int32();
            const minor = reader.int32();
            reader.int32(); // revision
            reader.seek(0);
            const transpose = (major > 1000) || (minor > 1000);
            if (!transpose) {
                const offset = 12 + ((major * 10 + minor) >= 2 ? 8 : 4);
                return new darknet.Weights(reader, offset);
            }
        }
        return null;
    }

    constructor(reader, offset) {
        this._reader = reader;
        this._offset = offset;
    }

    read(size) {
        this._reader.skip(this._offset);
        this._offset = 0;
        return this._reader.read(size);
    }

    validate() {
        if (this._reader.position !== this._reader.length) {
            throw new darknet.Error('Invalid weights size.');
        }
    }
};

darknet.Error = class extends Error {

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

export const ModelFactory = darknet.ModelFactory;