Add the demo clip: the kit's Forecast example on an iPhone 18 Pro
Browse files
.gitattributes
CHANGED
|
@@ -48,3 +48,5 @@ ios-h19p/timestral_209m_L64_h4_r2_fp16p7-qkv4-tanh-conv.h19p.aimodelc/main-h19p-
|
|
| 48 |
macos/timestral_209m_L256_h32_r2_fp16-attn32.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
|
| 49 |
macos/timestral_209m_L64_h32_r1_fp16p7-qkv4-tanh-conv.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
|
| 50 |
macos/timestral_209m_L64_h4_r2_fp16p7-qkv4-tanh-conv.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
| 48 |
macos/timestral_209m_L256_h32_r2_fp16-attn32.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
|
| 49 |
macos/timestral_209m_L64_h32_r1_fp16p7-qkv4-tanh-conv.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
|
| 50 |
macos/timestral_209m_L64_h4_r2_fp16p7-qkv4-tanh-conv.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text
|
| 51 |
+
demo.gif filter=lfs diff=lfs merge=lfs -text
|
| 52 |
+
demo.mp4 filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -41,6 +41,26 @@ symmetric averaging (the series and its mirror, averaged), the advancing refine
|
|
| 41 |
re-denormalization of the block, `_refine_block`), the cap (the history's range widened by 10⁴ times the running
|
| 42 |
standard deviation at its last step) and the non-negative clip.
|
| 43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
## What this port covers
|
| 45 |
|
| 46 |
- One univariate series without covariates. The released model also forecasts several variates together and takes
|
|
|
|
| 41 |
re-denormalization of the block, `_refine_block`), the cap (the history's range widened by 10⁴ times the running
|
| 42 |
standard deviation at its last step) and the non-negative clip.
|
| 43 |
|
| 44 |
+
## Demo
|
| 45 |
+
|
| 46 |
+

|
| 47 |
+
|
| 48 |
+
The kit's Forecast example ([`Examples/Forecast`](https://github.com/john-rocky/coreai-kit/tree/main/Examples/Forecast))
|
| 49 |
+
on an iPhone 18 Pro (iOS 27.2), 2026-10-24: a year (365 days) of made-up daily order counts goes into the paste box,
|
| 50 |
+
and the next 14 days come back as a median with a 10–90 % band, drawn from the last observed day (the chart shows the
|
| 51 |
+
last 96 days). The app's caption reads 29.5 ms for the forecast, 29.3 ms of it in the graph, on the Neural Engine. The
|
| 52 |
+
app's scripted run puts the numbers in the box, as a paste does, and starts the forecast the button starts. The clip is
|
| 53 |
+
one span of the recording, trimmed at both ends and the status bar cropped; no cut inside, no speed change
|
| 54 |
+
([MP4](https://huggingface.co/mlboydaisuke/Timestral-209M-CoreAI/resolve/main/demo.mp4), 10.8 s).
|
| 55 |
+
|
| 56 |
+
- That forecast is the first one after the app opens. Six launches that day forecast this series on the screen: the
|
| 57 |
+
first forecast took about 21 ms of graph time right at launch and 29–31 ms after the 6 s wait the clip has. The
|
| 58 |
+
[Measured](#measured) table's 8.42 ms for this graph is the median of repeated forecasts within one launch.
|
| 59 |
+
- The clip's launch is a later one in the same install. The first forecast after installing the app downloads the
|
| 60 |
+
three graphs compiled for this phone (about 1.2 GB) and loads them: 66 s in this run.
|
| 61 |
+
- The same app and install, replaying the same 365 values through `-fixture`, matches the released `forecast()`
|
| 62 |
+
within 0.19 % on all 294 numbers (21 quantiles × 14 days).
|
| 63 |
+
|
| 64 |
## What this port covers
|
| 65 |
|
| 66 |
- One univariate series without covariates. The released model also forecasts several variates together and takes
|
demo.gif
ADDED
|
Git LFS Details
|
demo.mp4
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b65dccc90f0cfbd837637c0a39f20779a688eab815ffdfe9cebdf57cf71aefde
|
| 3 |
+
size 259358
|