mirror of
https://github.com/wassname/keras-js.git
synced 2026-09-11 12:20:53 +08:00
implement LSTM layer, with tests
This commit is contained in:
Leon Chen
@@ -0,0 +1,237 @@
|
||||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {
|
||||
"collapsed": false
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Using TensorFlow backend.\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import numpy as np\n",
|
||||
"from keras.models import Model\n",
|
||||
"from keras.layers import Input\n",
|
||||
"from keras.layers.recurrent import LSTM\n",
|
||||
"from keras import backend as K"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {
|
||||
"collapsed": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"def format_decimal(arr, places=6):\n",
|
||||
" return [round(x * 10**places) / 10**places for x in arr]"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### LSTM"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"**[recurrent.LSTM.0] output_dim=4, activation='tanh', inner_activation='hard_sigmoid'**\n",
|
||||
"\n",
|
||||
"Note dropout_W and dropout_U are only applied during training phase"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {
|
||||
"collapsed": false
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"W_i shape: (6, 4)\n",
|
||||
"W_i: [0.021587, 0.171208, 0.42254, -0.348744, 0.040626, 0.479164, 0.684427, -0.694989, 0.064855, 0.056754, -0.076463, 0.184194, -0.793392, 0.627222, -0.855503, -0.362412, 0.855828, 0.681804, -0.326362, -0.218483, 0.302837, 0.321287, -0.052374, 0.066035]\n",
|
||||
"U_i shape: (4, 4)\n",
|
||||
"U_i: [-0.486398, -0.924962, 0.924244, -0.099505, 0.697209, -0.988196, 0.312034, -0.82215, -0.450764, -0.702362, 0.86183, -0.505312, 0.224025, -0.210654, -0.378863, 0.270893]\n",
|
||||
"b_i shape: (4,)\n",
|
||||
"b_i: [-0.752966, 0.132946, 0.249025, -0.746897]\n",
|
||||
"W_c shape: (6, 4)\n",
|
||||
"W_c: [-0.321918, 0.410663, -0.775844, -0.651728, 0.078787, 0.493956, -0.207594, -0.280414, -0.172642, -0.673013, -0.577208, -0.485994, 0.501094, 0.138545, 0.010672, 0.470218, -0.560627, 0.188022, 0.383913, 0.692257, -0.360569, -0.077626, -0.009927, 0.454896]\n",
|
||||
"U_c shape: (4, 4)\n",
|
||||
"U_c: [-0.008879, -0.535629, 0.500476, -0.443752, -0.370062, 0.766849, 0.348937, 0.530872, 0.151555, -0.41474, -0.538834, -0.242793, -0.809167, 0.950484, 0.326277, -0.861456]\n",
|
||||
"b_c shape: (4,)\n",
|
||||
"b_c: [-0.852859, 0.946774, 0.323797, -0.109777]\n",
|
||||
"W_f shape: (6, 4)\n",
|
||||
"W_f: [-0.793815, -0.518475, -0.579862, 0.302531, 0.680386, 0.692198, -0.682758, -0.752123, 0.743339, 0.221127, 0.042567, 0.215791, -0.256099, 0.022924, 0.157546, 0.535038, -0.591736, 0.443027, -0.51415, 0.467614, -0.391062, 0.993584, 0.965207, 0.509182]\n",
|
||||
"U_f shape: (4, 4)\n",
|
||||
"U_f: [0.476861, -0.902137, 0.156377, 0.778033, 0.679346, 0.284676, 0.042898, 0.289311, -0.861268, -0.784418, 0.932704, -0.825608, -0.765997, -0.944845, -0.980398, -0.006388]\n",
|
||||
"b_f shape: (4,)\n",
|
||||
"b_f: [0.824388, -0.068399, 0.018677, -0.595055]\n",
|
||||
"W_o shape: (6, 4)\n",
|
||||
"W_o: [0.180806, -0.699554, -0.809467, -0.616054, -0.217633, 0.690901, 0.098324, 0.76948, -0.122458, 0.011661, 0.377493, -0.022008, -0.898668, -0.569468, -0.162894, 0.853575, 0.607888, -0.742334, 0.593786, -0.223577, -0.626006, 0.871527, 0.493996, -0.143212]\n",
|
||||
"U_o shape: (4, 4)\n",
|
||||
"U_o: [0.672312, -0.799748, 0.482037, -0.833993, 0.825852, -0.939344, -0.227248, -0.287931, 0.996954, 0.391395, 0.122945, -0.582373, 0.170427, 0.895754, -0.548469, 0.313066]\n",
|
||||
"b_o shape: (4,)\n",
|
||||
"b_o: [-0.590664, 0.408132, 0.246375, -0.796643]\n",
|
||||
"\n",
|
||||
"in shape: (3, 6)\n",
|
||||
"in: [0.665517, 0.700148, 0.455626, 0.804744, -0.880561, -0.520282, 0.574518, -0.675696, -0.97577, -0.240988, 0.80853, 0.237721, -0.97565, 0.164267, -0.003728, 0.882893, 0.121681, -0.925775]\n",
|
||||
"out shape: (4,)\n",
|
||||
"out: [-0.12453, 0.324853, 0.173839, 0.145323]\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"data_in_shape = (3, 6)\n",
|
||||
"rnn = LSTM(4, activation='tanh', inner_activation='hard_sigmoid')\n",
|
||||
"\n",
|
||||
"layer_0 = Input(shape=data_in_shape)\n",
|
||||
"layer_1 = rnn(layer_0)\n",
|
||||
"model = Model(input=layer_0, output=layer_1)\n",
|
||||
"\n",
|
||||
"# set weights to random (use seed for reproducibility)\n",
|
||||
"weights = []\n",
|
||||
"for i, w in enumerate(model.get_weights()):\n",
|
||||
" np.random.seed(3000 + i)\n",
|
||||
" weights.append(2 * np.random.random(w.shape) - 1)\n",
|
||||
"model.set_weights(weights)\n",
|
||||
"weight_names = ['W_i', 'U_i', 'b_i', 'W_c', 'U_c', 'b_c', 'W_f', 'U_f', 'b_f', 'W_o', 'U_o', 'b_o']\n",
|
||||
"for w_i, w_name in enumerate(weight_names):\n",
|
||||
" print('{} shape:'.format(w_name), weights[w_i].shape)\n",
|
||||
" print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))\n",
|
||||
"\n",
|
||||
"data_in = 2 * np.random.random(data_in_shape) - 1\n",
|
||||
"print('')\n",
|
||||
"print('in shape:', data_in_shape)\n",
|
||||
"print('in:', format_decimal(data_in.ravel().tolist()))\n",
|
||||
"result = model.predict(np.array([data_in]))\n",
|
||||
"print('out shape:', result[0].shape)\n",
|
||||
"print('out:', format_decimal(result[0].ravel().tolist()))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"**[recurrent.LSTM.1] output_dim=5, activation='sigmoid', inner_activation='sigmoid'**\n",
|
||||
"\n",
|
||||
"Note dropout_W and dropout_U are only applied during training phase"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {
|
||||
"collapsed": false
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"W_i shape: (5, 5)\n",
|
||||
"W_i: [0.904105, -0.877164, 0.33871, 0.812244, -0.802633, 0.595832, -0.981528, 0.378119, 0.41493, -0.11458, -0.692809, 0.875448, 0.318268, -0.627292, 0.941065, 0.916899, -0.78002, 0.54744, -0.282447, -0.214613, -0.087847, -0.404508, -0.411046, -0.425703, -0.603508]\n",
|
||||
"U_i shape: (5, 5)\n",
|
||||
"U_i: [-0.708522, -0.706998, 0.233695, 0.906727, -0.986561, 0.797093, -0.065787, 0.680991, 0.839202, 0.009629, -0.39912, -0.935261, 0.187645, 0.505894, -0.292493, 0.909563, -0.869531, -0.741938, -0.617129, -0.750299, 0.78066, -0.489417, 0.306455, -0.684443, -0.097986]\n",
|
||||
"b_i shape: (5,)\n",
|
||||
"b_i: [0.351652, 0.052999, 0.049473, -0.569237, 0.204228]\n",
|
||||
"W_c shape: (5, 5)\n",
|
||||
"W_c: [0.067039, 0.428238, -0.349015, -0.542575, 0.975082, 0.727611, 0.470814, -0.569386, 0.338722, 0.972686, 0.114987, 0.967292, 0.424026, 0.922109, -0.380777, -0.89206, -0.144607, -0.091346, -0.419741, -0.779593, -0.104814, -0.079318, 0.377633, 0.490806, -0.452043]\n",
|
||||
"U_c shape: (5, 5)\n",
|
||||
"U_c: [-0.174106, -0.833394, 0.381703, -0.843934, -0.236694, -0.36737, -0.367741, 0.837889, -0.039786, 0.361505, 0.286408, 0.273419, -0.69914, -0.82985, -0.459621, -0.559184, -0.869297, 0.514745, 0.614001, 0.001062, 0.051854, -0.369003, 0.30649, 0.899397, 0.365186]\n",
|
||||
"b_c shape: (5,)\n",
|
||||
"b_c: [-0.585018, 0.56124, 0.203839, -0.185823, -0.172334]\n",
|
||||
"W_f shape: (5, 5)\n",
|
||||
"W_f: [0.928699, -0.755814, -0.167684, -0.11633, 0.984822, 0.163734, 0.294156, 0.159899, 0.429806, 0.120375, 0.536481, 0.315275, 0.334484, 0.385015, 0.303987, -0.500827, -0.317092, -0.562572, 0.939739, 0.402176, -0.220776, 0.177354, -0.29364, -0.109566, 0.317174]\n",
|
||||
"U_f shape: (5, 5)\n",
|
||||
"U_f: [-0.293165, 0.91974, -0.295233, -0.963946, 0.96014, -0.400338, 0.690912, -0.817071, -0.544351, 0.133677, 0.44247, -0.787779, 0.705348, -0.41608, -0.298659, 0.277014, -0.727674, -0.965871, 0.773519, 0.03901, -0.824034, -0.962203, 0.307807, 0.422613, -0.849913]\n",
|
||||
"b_f shape: (5,)\n",
|
||||
"b_f: [-0.517725, 0.941461, 0.58815, -0.186298, -0.946213]\n",
|
||||
"W_o shape: (5, 5)\n",
|
||||
"W_o: [0.35193, -0.016184, -0.950798, 0.227653, -0.950189, -0.244107, 0.114508, -0.716402, -0.688769, 0.253202, 0.438141, -0.883068, -0.305492, 0.466456, -0.971678, -0.687383, 0.970269, 0.5688, -0.05607, 0.245652, -0.849934, 0.141141, -0.50186, 0.8381, 0.615684]\n",
|
||||
"U_o shape: (5, 5)\n",
|
||||
"U_o: [0.737993, -0.987905, 0.734592, 0.869378, -0.219359, 0.16681, 0.618012, 0.167843, -0.129422, -0.743455, 0.398148, 0.849206, -0.941554, 0.796703, 0.939108, 0.23178, -0.652339, 0.019098, -0.925936, -0.892392, 0.785007, 0.794477, -0.682383, 0.542452, 0.94182]\n",
|
||||
"b_o shape: (5,)\n",
|
||||
"b_o: [-0.80119, -0.093136, -0.781478, 0.664981, 0.81755]\n",
|
||||
"\n",
|
||||
"in shape: (8, 5)\n",
|
||||
"in: [0.218071, -0.494078, 0.765253, 0.73909, -0.631062, 0.886984, -0.606791, -0.153821, -0.772063, 0.758903, -0.855796, -0.307768, 0.947849, 0.654099, 0.100526, -0.662352, -0.969299, 0.120011, -0.356979, 0.062594, 0.727154, -0.336611, -0.870729, -0.949392, -0.881842, 0.731806, 0.361095, -0.168111, 0.835201, 0.10262, -0.76506, -0.680559, 0.56447, 0.52546, 0.503997, -0.947011, -0.578522, 0.289283, 0.649655, 0.453943]\n",
|
||||
"out shape: (5,)\n",
|
||||
"out: [0.143347, 0.400829, 0.363268, 0.449314, 0.438799]\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"data_in_shape = (8, 5)\n",
|
||||
"rnn = LSTM(5, activation='sigmoid', inner_activation='sigmoid')\n",
|
||||
"\n",
|
||||
"layer_0 = Input(shape=data_in_shape)\n",
|
||||
"layer_1 = rnn(layer_0)\n",
|
||||
"model = Model(input=layer_0, output=layer_1)\n",
|
||||
"\n",
|
||||
"# set weights to random (use seed for reproducibility)\n",
|
||||
"weights = []\n",
|
||||
"for i, w in enumerate(model.get_weights()):\n",
|
||||
" np.random.seed(3100 + i)\n",
|
||||
" weights.append(2 * np.random.random(w.shape) - 1)\n",
|
||||
"model.set_weights(weights)\n",
|
||||
"weight_names = ['W_i', 'U_i', 'b_i', 'W_c', 'U_c', 'b_c', 'W_f', 'U_f', 'b_f', 'W_o', 'U_o', 'b_o']\n",
|
||||
"for w_i, w_name in enumerate(weight_names):\n",
|
||||
" print('{} shape:'.format(w_name), weights[w_i].shape)\n",
|
||||
" print('{}:'.format(w_name), format_decimal(weights[w_i].ravel().tolist()))\n",
|
||||
"\n",
|
||||
"data_in = 2 * np.random.random(data_in_shape) - 1\n",
|
||||
"print('')\n",
|
||||
"print('in shape:', data_in_shape)\n",
|
||||
"print('in:', format_decimal(data_in.ravel().tolist()))\n",
|
||||
"result = model.predict(np.array([data_in]))\n",
|
||||
"print('out shape:', result[0].shape)\n",
|
||||
"print('out:', format_decimal(result[0].ravel().tolist()))"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"collapsed": true
|
||||
},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"anaconda-cloud": {},
|
||||
"kernelspec": {
|
||||
"display_name": "Python [default]",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.5.2"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 1
|
||||
}
|
||||
@@ -4,3 +4,4 @@ export * from './convolutional'
|
||||
export * from './pooling'
|
||||
export * from './normalization'
|
||||
export * from './embeddings'
|
||||
export * from './recurrent'
|
||||
|
||||
@@ -0,0 +1,134 @@
|
||||
import * as activations from '../../activations'
|
||||
import Tensor from '../../Tensor'
|
||||
import Layer from '../../Layer'
|
||||
import { gemv } from 'ndarray-blas-level2'
|
||||
import ops from 'ndarray-ops'
|
||||
import cwise from 'cwise'
|
||||
|
||||
/**
|
||||
* LSTM layer class
|
||||
*/
|
||||
export default class LSTM extends Layer {
|
||||
/**
|
||||
* Creates a LSTM layer
|
||||
* @param {number} attrs.outputDim - Number of convolution filters to use.
|
||||
* @param {Object} [attrs] - layer attributes
|
||||
*/
|
||||
constructor (attrs = {}) {
|
||||
super(attrs)
|
||||
this.layerClass = 'LSTM'
|
||||
|
||||
const {
|
||||
outputDim = 1,
|
||||
activation = 'tanh',
|
||||
innerActivation = 'hardSigmoid'
|
||||
} = attrs
|
||||
|
||||
this.outputDim = outputDim
|
||||
|
||||
this.activation = activations[activation]
|
||||
this.innerActivation = activations[innerActivation]
|
||||
|
||||
// Layer weights specification
|
||||
this.params = ['W_i', 'U_i', 'b_i', 'W_c', 'U_c', 'b_c', 'W_f', 'U_f', 'b_f', 'W_o', 'U_o', 'b_o']
|
||||
}
|
||||
|
||||
_combine = cwise({
|
||||
args: ['array', 'array', 'array', 'array'],
|
||||
body: function (_y, _x1, _x2, _b) {
|
||||
_y = _x1 + _x2 + _b
|
||||
}
|
||||
})
|
||||
|
||||
_update = cwise({
|
||||
args: ['array', 'array', 'array', 'array'],
|
||||
body: function (_c, _ctm1, _i, _f) {
|
||||
_c = _c * _i + _ctm1 * _f
|
||||
}
|
||||
})
|
||||
|
||||
/**
|
||||
* Method for layer computational logic
|
||||
* @param {Tensor} x
|
||||
* @returns {Tensor} x
|
||||
*/
|
||||
call (x) {
|
||||
let currentX = new Tensor([], [x.tensor.shape[1]])
|
||||
|
||||
const dimInputGate = this.weights['b_i'].tensor.shape[0]
|
||||
const dimCandidate = this.weights['b_c'].tensor.shape[0]
|
||||
const dimForgetGate = this.weights['b_f'].tensor.shape[0]
|
||||
const dimOutputGate = this.weights['b_o'].tensor.shape[0]
|
||||
|
||||
let currentInputGateState = new Tensor([], [dimInputGate])
|
||||
let tempXI = new Tensor([], [dimInputGate])
|
||||
let tempHI = new Tensor([], [dimInputGate])
|
||||
|
||||
let currentForgetGateState = new Tensor([], [dimForgetGate])
|
||||
let tempXF = new Tensor([], [dimForgetGate])
|
||||
let tempHF = new Tensor([], [dimForgetGate])
|
||||
|
||||
let currentOutputGateState = new Tensor([], [dimOutputGate])
|
||||
let tempXO = new Tensor([], [dimOutputGate])
|
||||
let tempHO = new Tensor([], [dimOutputGate])
|
||||
|
||||
let currentCandidate = new Tensor([], [dimCandidate])
|
||||
let tempXC = new Tensor([], [dimCandidate])
|
||||
let tempHC = new Tensor([], [dimCandidate])
|
||||
let previousCandidate = new Tensor([], [dimCandidate])
|
||||
|
||||
let currentHiddenState = new Tensor([], [dimCandidate])
|
||||
let previousHiddenState = new Tensor([], [dimCandidate])
|
||||
|
||||
const _clearTemp = () => {
|
||||
const tempTensors = [tempXI, tempHI, tempXF, tempHF, tempXO, tempHO, tempXC, tempHC]
|
||||
tempTensors.forEach(temp => ops.assigns(temp.tensor, 0))
|
||||
}
|
||||
|
||||
const _step = () => {
|
||||
ops.assign(previousHiddenState.tensor, currentHiddenState.tensor)
|
||||
|
||||
gemv(1.0, this.weights['W_i'].tensor.transpose(1, 0), currentX.tensor, 1.0, tempXI.tensor)
|
||||
gemv(1.0, this.weights['U_i'].tensor.transpose(1, 0), previousHiddenState.tensor, 1.0, tempHI.tensor)
|
||||
this._combine(currentInputGateState.tensor, tempXI.tensor, tempHI.tensor, this.weights['b_i'].tensor)
|
||||
this.innerActivation(currentInputGateState)
|
||||
|
||||
gemv(1.0, this.weights['W_f'].tensor.transpose(1, 0), currentX.tensor, 1.0, tempXF.tensor)
|
||||
gemv(1.0, this.weights['U_f'].tensor.transpose(1, 0), previousHiddenState.tensor, 1.0, tempHF.tensor)
|
||||
this._combine(currentForgetGateState.tensor, tempXF.tensor, tempHF.tensor, this.weights['b_f'].tensor)
|
||||
this.innerActivation(currentForgetGateState)
|
||||
|
||||
gemv(1.0, this.weights['W_o'].tensor.transpose(1, 0), currentX.tensor, 1.0, tempXO.tensor)
|
||||
gemv(1.0, this.weights['U_o'].tensor.transpose(1, 0), previousHiddenState.tensor, 1.0, tempHO.tensor)
|
||||
this._combine(currentOutputGateState.tensor, tempXO.tensor, tempHO.tensor, this.weights['b_o'].tensor)
|
||||
this.innerActivation(currentOutputGateState)
|
||||
|
||||
gemv(1.0, this.weights['W_c'].tensor.transpose(1, 0), currentX.tensor, 1.0, tempXC.tensor)
|
||||
gemv(1.0, this.weights['U_c'].tensor.transpose(1, 0), previousHiddenState.tensor, 1.0, tempHC.tensor)
|
||||
this._combine(currentCandidate.tensor, tempXC.tensor, tempHC.tensor, this.weights['b_c'].tensor)
|
||||
this.activation(currentCandidate)
|
||||
|
||||
this._update(
|
||||
currentCandidate.tensor,
|
||||
previousCandidate.tensor,
|
||||
currentInputGateState.tensor,
|
||||
currentForgetGateState.tensor
|
||||
)
|
||||
|
||||
ops.assign(previousCandidate.tensor, currentCandidate.tensor)
|
||||
|
||||
this.activation(currentCandidate)
|
||||
ops.mul(currentHiddenState.tensor, currentOutputGateState.tensor, currentCandidate.tensor)
|
||||
}
|
||||
|
||||
for (let i = 0, steps = x.tensor.shape[0]; i < steps; i++) {
|
||||
ops.assign(currentX.tensor, x.tensor.pick(i, null))
|
||||
_clearTemp()
|
||||
_step()
|
||||
}
|
||||
|
||||
x.tensor = currentHiddenState.tensor
|
||||
|
||||
return x
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,5 @@
|
||||
import LSTM from './LSTM'
|
||||
|
||||
export {
|
||||
LSTM
|
||||
}
|
||||
@@ -100,6 +100,9 @@
|
||||
<script src="embeddings/data_Embedding.js"></script>
|
||||
<script src="embeddings/Embedding.js"></script>
|
||||
|
||||
<script src="recurrent/data_LSTM.js"></script>
|
||||
<script src="recurrent/LSTM.js"></script>
|
||||
|
||||
<script>
|
||||
// mocha.checkLeaks();
|
||||
mocha.globals(['jQuery']);
|
||||
|
||||
@@ -0,0 +1,89 @@
|
||||
/* eslint-env browser, mocha */
|
||||
|
||||
describe.only('recurrent layer: LSTM', function () {
|
||||
const assert = chai.assert
|
||||
const styles = testGlobals.styles
|
||||
const logTime = testGlobals.logTime
|
||||
const stringifyCondensed = testGlobals.stringifyCondensed
|
||||
const approxEquals = KerasJS.testUtils.approxEquals
|
||||
const layers = KerasJS.layers
|
||||
|
||||
const testParams = [
|
||||
{
|
||||
inputShape: [3, 6],
|
||||
attrs: { outputDim: 4, activation: 'tanh', innerActivation: 'hardSigmoid' }
|
||||
},
|
||||
{
|
||||
inputShape: [8, 5],
|
||||
attrs: { outputDim: 5, activation: 'sigmoid', innerActivation: 'sigmoid' }
|
||||
}
|
||||
]
|
||||
|
||||
before(function () {
|
||||
console.log('\n%crecurrent layer: LSTM', styles.h1)
|
||||
})
|
||||
|
||||
/*********************************************************
|
||||
* CPU
|
||||
*********************************************************/
|
||||
|
||||
describe('CPU', function () {
|
||||
before(function () {
|
||||
console.log('\n%cCPU', styles.h2)
|
||||
})
|
||||
|
||||
testParams.forEach(({ inputShape, attrs }, i) => {
|
||||
const key = `recurrent.LSTM.${i}`
|
||||
const title = `[${key}] [CPU] test: ${inputShape[0]}x${inputShape[1]} input, activation='${attrs.activation}', innerActivation='${attrs.innerActivation}'`
|
||||
|
||||
it(title, function () {
|
||||
console.log(`\n%c${title}`, styles.h3)
|
||||
let testLayer = new layers.LSTM(attrs)
|
||||
testLayer.setWeights(TEST_DATA[key].weights.map(w => new KerasJS.Tensor(w.data, w.shape)))
|
||||
let t = new KerasJS.Tensor(TEST_DATA[key].input.data, TEST_DATA[key].input.shape)
|
||||
console.log('%cin', styles.h4, stringifyCondensed(t.tensor))
|
||||
const startTime = performance.now()
|
||||
t = testLayer.call(t)
|
||||
const endTime = performance.now()
|
||||
console.log('%cout', styles.h4, stringifyCondensed(t.tensor))
|
||||
logTime(startTime, endTime)
|
||||
const dataExpected = new Float32Array(TEST_DATA[key].expected.data)
|
||||
const shapeExpected = TEST_DATA[key].expected.shape
|
||||
assert.deepEqual(t.tensor.shape, shapeExpected)
|
||||
assert.isTrue(approxEquals(t.tensor, dataExpected))
|
||||
})
|
||||
})
|
||||
})
|
||||
|
||||
/*********************************************************
|
||||
* GPU
|
||||
*********************************************************/
|
||||
|
||||
describe('GPU', function () {
|
||||
before(function () {
|
||||
console.log('\n%cGPU', styles.h2)
|
||||
})
|
||||
|
||||
testParams.forEach(({ inputShape, attrs }, i) => {
|
||||
const key = `recurrent.LSTM.${i}`
|
||||
const title = `[${key}] [GPU] test: ${inputShape[0]}x${inputShape[1]} input, activation='${attrs.activation}', innerActivation='${attrs.innerActivation}'`
|
||||
|
||||
it(title, function () {
|
||||
console.log(`\n%c${title}`, styles.h3)
|
||||
let testLayer = new layers.LSTM(attrs)
|
||||
testLayer.setWeights(TEST_DATA[key].weights.map(w => new KerasJS.Tensor(w.data, w.shape)))
|
||||
let t = new KerasJS.Tensor(TEST_DATA[key].input.data, TEST_DATA[key].input.shape, { gpu: true })
|
||||
console.log('%cin', styles.h4, stringifyCondensed(t.tensor))
|
||||
const startTime = performance.now()
|
||||
t = testLayer.call(t)
|
||||
const endTime = performance.now()
|
||||
console.log('%cout', styles.h4, stringifyCondensed(t.tensor))
|
||||
logTime(startTime, endTime)
|
||||
const dataExpected = new Float32Array(TEST_DATA[key].expected.data)
|
||||
const shapeExpected = TEST_DATA[key].expected.shape
|
||||
assert.deepEqual(t.tensor.shape, shapeExpected)
|
||||
assert.isTrue(approxEquals(t.tensor, dataExpected))
|
||||
})
|
||||
})
|
||||
})
|
||||
})
|
||||
@@ -0,0 +1,130 @@
|
||||
// TEST DATA
|
||||
// Keyed by mocha test ID
|
||||
// Python code for generating test data can be found in the matching jupyter notebook in folder `notebooks/`.
|
||||
|
||||
(function () {
|
||||
var DATA = {
|
||||
'recurrent.LSTM.0': {
|
||||
input: {
|
||||
data: [0.665517, 0.700148, 0.455626, 0.804744, -0.880561, -0.520282, 0.574518, -0.675696, -0.97577, -0.240988, 0.80853, 0.237721, -0.97565, 0.164267, -0.003728, 0.882893, 0.121681, -0.925775],
|
||||
shape: [3, 6]
|
||||
},
|
||||
weights: [
|
||||
{
|
||||
data: [0.021587, 0.171208, 0.42254, -0.348744, 0.040626, 0.479164, 0.684427, -0.694989, 0.064855, 0.056754, -0.076463, 0.184194, -0.793392, 0.627222, -0.855503, -0.362412, 0.855828, 0.681804, -0.326362, -0.218483, 0.302837, 0.321287, -0.052374, 0.066035],
|
||||
shape: [6, 4]
|
||||
},
|
||||
{
|
||||
data: [-0.486398, -0.924962, 0.924244, -0.099505, 0.697209, -0.988196, 0.312034, -0.82215, -0.450764, -0.702362, 0.86183, -0.505312, 0.224025, -0.210654, -0.378863, 0.270893],
|
||||
shape: [4, 4]
|
||||
},
|
||||
{
|
||||
data: [-0.752966, 0.132946, 0.249025, -0.746897],
|
||||
shape: [4]
|
||||
},
|
||||
{
|
||||
data: [-0.321918, 0.410663, -0.775844, -0.651728, 0.078787, 0.493956, -0.207594, -0.280414, -0.172642, -0.673013, -0.577208, -0.485994, 0.501094, 0.138545, 0.010672, 0.470218, -0.560627, 0.188022, 0.383913, 0.692257, -0.360569, -0.077626, -0.009927, 0.454896],
|
||||
shape: [6, 4]
|
||||
},
|
||||
{
|
||||
data: [-0.008879, -0.535629, 0.500476, -0.443752, -0.370062, 0.766849, 0.348937, 0.530872, 0.151555, -0.41474, -0.538834, -0.242793, -0.809167, 0.950484, 0.326277, -0.861456],
|
||||
shape: [4, 4]
|
||||
},
|
||||
{
|
||||
data: [-0.852859, 0.946774, 0.323797, -0.109777],
|
||||
shape: [4]
|
||||
},
|
||||
{
|
||||
data: [-0.793815, -0.518475, -0.579862, 0.302531, 0.680386, 0.692198, -0.682758, -0.752123, 0.743339, 0.221127, 0.042567, 0.215791, -0.256099, 0.022924, 0.157546, 0.535038, -0.591736, 0.443027, -0.51415, 0.467614, -0.391062, 0.993584, 0.965207, 0.509182],
|
||||
shape: [6, 4]
|
||||
},
|
||||
{
|
||||
data: [0.476861, -0.902137, 0.156377, 0.778033, 0.679346, 0.284676, 0.042898, 0.289311, -0.861268, -0.784418, 0.932704, -0.825608, -0.765997, -0.944845, -0.980398, -0.006388],
|
||||
shape: [4, 4]
|
||||
},
|
||||
{
|
||||
data: [0.824388, -0.068399, 0.018677, -0.595055],
|
||||
shape: [4]
|
||||
},
|
||||
{
|
||||
data: [0.180806, -0.699554, -0.809467, -0.616054, -0.217633, 0.690901, 0.098324, 0.76948, -0.122458, 0.011661, 0.377493, -0.022008, -0.898668, -0.569468, -0.162894, 0.853575, 0.607888, -0.742334, 0.593786, -0.223577, -0.626006, 0.871527, 0.493996, -0.143212],
|
||||
shape: [6, 4]
|
||||
},
|
||||
{
|
||||
data: [0.672312, -0.799748, 0.482037, -0.833993, 0.825852, -0.939344, -0.227248, -0.287931, 0.996954, 0.391395, 0.122945, -0.582373, 0.170427, 0.895754, -0.548469, 0.313066],
|
||||
shape: [4, 4]
|
||||
},
|
||||
{
|
||||
data: [-0.590664, 0.408132, 0.246375, -0.796643],
|
||||
shape: [4]
|
||||
}
|
||||
],
|
||||
expected: {
|
||||
data: [-0.12453, 0.324853, 0.173839, 0.145323],
|
||||
shape: [4]
|
||||
}
|
||||
},
|
||||
'recurrent.LSTM.1': {
|
||||
input: {
|
||||
data: [0.218071, -0.494078, 0.765253, 0.73909, -0.631062, 0.886984, -0.606791, -0.153821, -0.772063, 0.758903, -0.855796, -0.307768, 0.947849, 0.654099, 0.100526, -0.662352, -0.969299, 0.120011, -0.356979, 0.062594, 0.727154, -0.336611, -0.870729, -0.949392, -0.881842, 0.731806, 0.361095, -0.168111, 0.835201, 0.10262, -0.76506, -0.680559, 0.56447, 0.52546, 0.503997, -0.947011, -0.578522, 0.289283, 0.649655, 0.453943],
|
||||
shape: [8, 5]
|
||||
},
|
||||
weights: [
|
||||
{
|
||||
data: [0.904105, -0.877164, 0.33871, 0.812244, -0.802633, 0.595832, -0.981528, 0.378119, 0.41493, -0.11458, -0.692809, 0.875448, 0.318268, -0.627292, 0.941065, 0.916899, -0.78002, 0.54744, -0.282447, -0.214613, -0.087847, -0.404508, -0.411046, -0.425703, -0.603508],
|
||||
shape: [5, 5]
|
||||
},
|
||||
{
|
||||
data: [-0.708522, -0.706998, 0.233695, 0.906727, -0.986561, 0.797093, -0.065787, 0.680991, 0.839202, 0.009629, -0.39912, -0.935261, 0.187645, 0.505894, -0.292493, 0.909563, -0.869531, -0.741938, -0.617129, -0.750299, 0.78066, -0.489417, 0.306455, -0.684443, -0.097986],
|
||||
shape: [5, 5]
|
||||
},
|
||||
{
|
||||
data: [0.351652, 0.052999, 0.049473, -0.569237, 0.204228],
|
||||
shape: [5]
|
||||
},
|
||||
{
|
||||
data: [0.067039, 0.428238, -0.349015, -0.542575, 0.975082, 0.727611, 0.470814, -0.569386, 0.338722, 0.972686, 0.114987, 0.967292, 0.424026, 0.922109, -0.380777, -0.89206, -0.144607, -0.091346, -0.419741, -0.779593, -0.104814, -0.079318, 0.377633, 0.490806, -0.452043],
|
||||
shape: [5, 5]
|
||||
},
|
||||
{
|
||||
data: [-0.174106, -0.833394, 0.381703, -0.843934, -0.236694, -0.36737, -0.367741, 0.837889, -0.039786, 0.361505, 0.286408, 0.273419, -0.69914, -0.82985, -0.459621, -0.559184, -0.869297, 0.514745, 0.614001, 0.001062, 0.051854, -0.369003, 0.30649, 0.899397, 0.365186],
|
||||
shape: [5, 5]
|
||||
},
|
||||
{
|
||||
data: [-0.585018, 0.56124, 0.203839, -0.185823, -0.172334],
|
||||
shape: [5]
|
||||
},
|
||||
{
|
||||
data: [0.928699, -0.755814, -0.167684, -0.11633, 0.984822, 0.163734, 0.294156, 0.159899, 0.429806, 0.120375, 0.536481, 0.315275, 0.334484, 0.385015, 0.303987, -0.500827, -0.317092, -0.562572, 0.939739, 0.402176, -0.220776, 0.177354, -0.29364, -0.109566, 0.317174],
|
||||
shape: [5, 5]
|
||||
},
|
||||
{
|
||||
data: [-0.293165, 0.91974, -0.295233, -0.963946, 0.96014, -0.400338, 0.690912, -0.817071, -0.544351, 0.133677, 0.44247, -0.787779, 0.705348, -0.41608, -0.298659, 0.277014, -0.727674, -0.965871, 0.773519, 0.03901, -0.824034, -0.962203, 0.307807, 0.422613, -0.849913],
|
||||
shape: [5, 5]
|
||||
},
|
||||
{
|
||||
data: [-0.517725, 0.941461, 0.58815, -0.186298, -0.946213],
|
||||
shape: [5]
|
||||
},
|
||||
{
|
||||
data: [0.35193, -0.016184, -0.950798, 0.227653, -0.950189, -0.244107, 0.114508, -0.716402, -0.688769, 0.253202, 0.438141, -0.883068, -0.305492, 0.466456, -0.971678, -0.687383, 0.970269, 0.5688, -0.05607, 0.245652, -0.849934, 0.141141, -0.50186, 0.8381, 0.615684],
|
||||
shape: [5, 5]
|
||||
},
|
||||
{
|
||||
data: [0.737993, -0.987905, 0.734592, 0.869378, -0.219359, 0.16681, 0.618012, 0.167843, -0.129422, -0.743455, 0.398148, 0.849206, -0.941554, 0.796703, 0.939108, 0.23178, -0.652339, 0.019098, -0.925936, -0.892392, 0.785007, 0.794477, -0.682383, 0.542452, 0.94182],
|
||||
shape: [5, 5]
|
||||
},
|
||||
{
|
||||
data: [-0.80119, -0.093136, -0.781478, 0.664981, 0.81755],
|
||||
shape: [5]
|
||||
}
|
||||
],
|
||||
expected: {
|
||||
data: [0.143347, 0.400829, 0.363268, 0.449314, 0.438799],
|
||||
shape: [5]
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
window.TEST_DATA = Object.assign({}, window.TEST_DATA, DATA)
|
||||
})()
|
||||
Reference in New Issue
Block a user