[date-conversion-attention] Initial check-in of date-conversion-attention

date-conversion-attention/.babelrc

@@ -0,0 +1,18 @@
+{
+  "presets": [
+    [
+      "env",
+      {
+        "esmodules": false,
+        "targets": {
+          "browsers": [
+            "> 3%"
+          ]
+        }
+      }
+    ]
+  ],
+  "plugins": [
+    "transform-runtime"
+  ]
+}

date-conversion-attention/README.md

@@ -0,0 +1,57 @@
+# TensorFlow.js Example: Date Conversion Through an LSTM-Attention Model
+
+## Overview
+
+This example shows how to use TensorFlow.js to train a model based on
+long short-term memory (LSTM) and the attention mechanism to achieve
+a task of converting various commonly seen date formats (e.g., 01/18/2019,
+18JAN2019, 18-01-2019) to the ISO date format (i.e., 2019-01-18).
+
+We demonstrate the full machine-learning workflow, consisting of
+data engineering, server-side model training, client-side inference,
+model visualization, and unit testing in this example.
+
+The training data is synthesized programmatically.
+
+## Model training in Node.js
+
+For efficiency, the training of the model happens outside the browser
+in Node.js, using tfjs-node or tfjs-node-gpu.
+
+To run the training job, do
+
+```sh
+yarn
+yarn train
+```
+
+By default, the training uses tfjs-node, which runs on the CPU.
+If you have a CUDA-enabled GPU and have the CUDA and CuDNN libraries
+set up properly on your system, you can run the training on the GPU
+by:
+
+```sh
+yarn
+yarn train --gpu
+```
+
+## Using the model in the browser
+
+TODO(cais): Implement it.
+
+### Visualization of the attention mechanism
+
+TODO(cais): Implement it.
+
+## Running unit tests
+
+The data and model code in this example are covered by unit tests.
+To run the unit tests:
+
+```sh
+cd ../
+yarn
+cd date-conversion-attention
+yarn
+yarn test
+```

date-conversion-attention/date_format.js

@@ -0,0 +1,169 @@
+/**
+ * @license
+ * Copyright 2019 Google LLC. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+
+const tf = require('@tensorflow/tfjs');
+
+const MONTH_NAMES_FULL = [
+  'January', 'February', 'March', 'April', 'May', 'June', 'July', 'August',
+  'September', 'October', 'November', 'December'
+];
+const MONTH_NAMES_3LETTER =
+    MONTH_NAMES_FULL.map(name => name.slice(0, 3).toUpperCase());
+
+const MIN_DATE = new Date('1950-01-01').getTime();
+const MAX_DATE = new Date('2050-01-01').getTime();
+
+export const INPUT_LENGTH = 12   // Maximum length of all input formats.
+export const OUTPUT_LENGTH = 10  // Length of 'YYYY-MM-DD'.
+
+// Use "\n" for padding for both input and output. It has to be at the
+// beginning so that `mask_zero=True` can be used in the keras model.
+export const INPUT_VOCAB = '\n0123456789/-, ' +
+    MONTH_NAMES_3LETTER.join('')
+        .split('')
+        .filter(function(item, i, ar) {
+          return ar.indexOf(item) === i;
+        })
+        .join('');
+
+// OUTPUT_VOCAB includes an start-of-sequence (SOS) token, represented as
+// '\t'.
+export const OUTPUT_VOCAB = '\n\t0123456789-';
+
+export const START_CODE = 1;
+
+/**
+ * Generate a random date.
+ *
+ * @return {[number, number, number]} Year as an integer, month as an
+ *   integer >= 1 and <= 12, day as an integer >= 1.
+ */
+export function generateRandomDateTuple() {
+  const date = new Date(Math.random() * (MAX_DATE - MIN_DATE) + MIN_DATE);
+  return [date.getFullYear(), date.getMonth() + 1, date.getDate()];
+}
+
+function toTwoDigitString(num) {
+  return num < 10 ? `0${num}` : `${num}`;
+}
+
+export function dateTupleToDDMMMYYYY(dateTuple) {
+  const monthStr = MONTH_NAMES_3LETTER[dateTuple[1] - 1];
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  return `${dayStr}${monthStr}${dateTuple[0]}`;
+}
+
+export function dateTupleToMMSlashDDSlashYYYY(dateTuple) {
+  const monthStr = toTwoDigitString(dateTuple[1]);
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  return `${monthStr}/${dayStr}/${dateTuple[0]}`;
+}
+
+export function dateTupleToMMSlashDDSlashYY(dateTuple) {
+  const monthStr = toTwoDigitString(dateTuple[1]);
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  const yearStr = `${dateTuple[0]}`.slice(2);
+  return `${monthStr}/${dayStr}/${yearStr}`;
+}
+
+export function dateTupleToMMDDYY(dateTuple) {
+  const monthStr = toTwoDigitString(dateTuple[1]);
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  const yearStr = `${dateTuple[0]}`.slice(2);
+  return `${monthStr}${dayStr}${yearStr}`;
+}
+
+export function dateTupleToMMMSpaceDDSpaceYY(dateTuple) {
+  const monthStr = MONTH_NAMES_3LETTER[dateTuple[1] - 1];
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  const yearStr = `${dateTuple[0]}`.slice(2);
+  return `${monthStr} ${dayStr} ${yearStr}`;
+}
+
+export function dateTupleToMMMSpaceDDSpaceYYYY(dateTuple) {
+  const monthStr = MONTH_NAMES_3LETTER[dateTuple[1] - 1];
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  return `${monthStr} ${dayStr} ${dateTuple[0]}`;
+}
+
+export function dateTupleToMMMSpaceDDCommaSpaceYY(dateTuple) {
+  const monthStr = MONTH_NAMES_3LETTER[dateTuple[1] - 1];
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  const yearStr = `${dateTuple[0]}`.slice(2);
+  return `${monthStr} ${dayStr}, ${yearStr}`;
+}
+
+export function dateTupleToMMMSpaceDDCommaSpaceYYYY(dateTuple) {
+  const monthStr = MONTH_NAMES_3LETTER[dateTuple[1] - 1];
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  return `${monthStr} ${dayStr}, ${dateTuple[0]}`;
+}
+
+export function dateTupleToDDDashMMDashYYYY(dateTuple) {
+  const monthStr = toTwoDigitString(dateTuple[1]);
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  return `${dayStr}-${monthStr}-${dateTuple[0]}`;
+}
+
+export function dateTupleToYYYYDashMMDashDD(dateTuple) {
+  const monthStr = toTwoDigitString(dateTuple[1]);
+  const dayStr = toTwoDigitString(dateTuple[2]);
+  return `${dateTuple[0]}-${monthStr}-${dayStr}`;
+}
+
+export function encodeInputDateStrings(dateStrings) {
+  const n = dateStrings.length;
+  const x = tf.buffer([n, INPUT_LENGTH], 'float32');
+  for (let i = 0; i < n; ++i) {
+    for (let j = 0; j < INPUT_LENGTH; ++j) {
+      if (j < dateStrings[i].length) {
+        const char = dateStrings[i][j];
+        const index = INPUT_VOCAB.indexOf(char);
+        if (index === -1) {
+          throw new Error(`Unknown char: ${char}`);
+        }
+        x.set(index, i, j);
+      }
+    }
+  }
+  return x.toTensor();
+}
+
+export function encodeOutputDateStrings(dateStrings, oneHot = false) {
+  const n = dateStrings.length;
+  const x =
+      oneHot ? tf.buffer([n, OUTPUT_LENGTH, OUTPUT_VOCAB.length], 'float32') :
+      tf.buffer([n, OUTPUT_LENGTH], 'float32');
+  for (let i = 0; i < n; ++i) {
+    tf.util.assert(
+        dateStrings[i].length === OUTPUT_LENGTH,
+        `Date string is not in ISO format: "${dateStrings[i]}"`);
+    for (let j = 0; j < OUTPUT_LENGTH; ++j) {
+      const char = dateStrings[i][j];
+      const index = OUTPUT_VOCAB.indexOf(char);
+      if (index === -1) {
+        throw new Error(`Unknown char: ${char}`);
+      }
+      if (oneHot) {
+        x.set(1, i, j, index);
+      } else {
+        x.set(index, i, j);
+      }
+    }
+  }
+  return x.toTensor();
+}