When creating a tensor with an array of timestamps, the numbers are incorrect - tensorflow.js

Question

Looking for some kind of solution to this issue:
trying to create a tensor from an array of timestamps
[
1612892067115,
],
but here is what happens
tf.tensor([1612892067115]).arraySync()
> [ 1612892078080 ]
as you can see, the result is incorrect.
Somebody pointed out, I may need to use the datatype int64, but this doesn't seem to exist in tfjs 😭
I have also tried to divide my timestamp to a small float, but I get a similar result
tf.tensor([1.612892067115, 1.612892068341]).arraySync()
[ 1.6128920316696167, 1.6128920316696167 ]
If you know a way to work around using timestamps in a tensor, please help :)
:edit:
As an attempted workaround, I tried to remove my year, month, and date from my timestamp
Here are my subsequent input values:
[
56969701,
56969685,
56969669,
56969646,
56969607,
56969602
]
and their outputs:
[
56969700,
56969684,
56969668,
56969648,
56969608,
56969600
]
as you can see, they are still incorrect, and should be well within the acceptable range

Answer 1

found a solution that worked for me:
Since I only require a subset of the timestamp (just the date / hour / minute / second / ms) for my purposes, I simply truncate out the year / month:
export const subts = (ts: number) => {
// a sub timestamp which can be used over the period of a month
const yearMonth = +new Date(new Date().getFullYear(), new Date().getMonth())
return ts - yearMonth
}
then I can use this with:
subTimestamps = timestamps.map(ts => subts(ts))
const x_vals = tf.tensor(subTimestamps, [subTimestamps.length], 'int32')
now all my results work as expected.

Answer 2

Currently only int32 is supported with tensorflow.js, your data has gone out of the range supported by int32.
Until int64 is supported, this can be solved by using a relative timestamp. Currently a timestamp in js uses the number of ms that elapsed since 1 January 1970. A relative timestamp can be used by using another origin and compute the difference of ms that has elapsed since that date. That way, we will have a lower number that can be represented using int32. The best origin to take will be the starting date of the records
const a = Date.now() // computing a tensor out of it will give an accurate result since the number is out of range
const origin = new Date("02/01/2021").now()
const relative = a - origin
const tensor = tf.tensor(relative, undefined, 'int32')
// get back the data
const data = tensor.dataSync()[0]
// get the initial date
const initial date = new Date(data + origin)
In other scenarios, if using the ms is not of interest, using the number of s that has elapsed since the start would be better. It is called the unix time