Upload MTG art CNN

README.md

@@ -21,16 +21,17 @@ with two softmax heads:
 * **Colour identity** (7): White, Blue, Black, Red, Green, Colorless, Multicolor
 * **Card type** (7): Creature, Planeswalker, Land, Artifact, Enchantment, Instant, Sorcery
 
-It is the final project for the Skillspire AI & Machine Learning course (Nathan King, Build IT Systems). The
+It is the final project for the Skillspire AI & Machine Learning course (Nathan King). The
 live demo is the Space [BuildITSystems/mtg-art-demo](https://huggingface.co/spaces/BuildITSystems/mtg-art-demo).
 
 ## Architecture
 
-`TwoHeadCNN(width=32, blocks=[2, 2, 2, 2], dropout=0.2)`: a 3x3 stem, four pre-activation
+`TwoHeadCNN(width=32, blocks=[2, 2, 2, 2], dropout=0.2)`: a 5x5 stride-2 stem, four pre-activation
 residual stages (widths 32/64/128/256, stride 2 each, 8 BasicBlocks),
 global average pooling and two linear heads; **2,800,366 parameters**, 5.6 MB in bf16. Input: the art crop
 resized to 160x224 (HxW), normalised with the training mean/std stored in `config.json`.
-`model.py` in this repo is the complete definition.
+The complete definition is `model.py` in the demo Space
+([BuildITSystems/mtg-art-demo](https://huggingface.co/spaces/BuildITSystems/mtg-art-demo)), byte-identical to the training code.
 
 ## Training
 
@@ -62,8 +63,8 @@ Single-image latency (100 images, batch 1, resize + normalise + forward + softma
 
 | device | mean | p50 | p95 | vs benchmark |
 |---|---:|---:|---:|---:|
-| GPU bf16 (NVIDIA GeForce RTX 3090 Ti) | 3.48 ms | 3.43 ms | 3.80 ms | 0.047x |
-| CPU fp32 (16 threads) | 5.33 ms | 5.19 ms | 5.94 ms | 0.072x |
+| GPU bf16 (NVIDIA GeForce RTX 3090 Ti) | 3.73 ms | 3.68 ms | 4.24 ms | 0.050x |
+| CPU fp32 (16 threads) | 5.20 ms | 5.05 ms | 6.01 ms | 0.070x |
 
 ## How to use
 

config.json

@@ -95,11 +95,11 @@
     }
   },
   "source_checkpoint": "/home/nathank/BITS/dev/MTG Card Classifier/checkpoints/light-aug/best",
-  "exported_at": "2026-10-23 18:39:51",
+  "exported_at": "2026-10-23 20:04:46",
   "metrics_test": {
     "split": "test",
     "n": 3401,
-    "evaluated_at": "2026-10-23 18:39:44",
+    "evaluated_at": "2026-10-23 20:04:39",
     "color": {
       "accuracy": 0.3684210526315789,
       "top2_accuracy": 0.5698324022346368,
@@ -117,20 +117,20 @@
     "samples": 100,
     "gpu_bf16": {
       "n": 100,
-      "mean_s": 0.00348,
-      "p50_s": 0.00343,
-      "p95_s": 0.0038,
-      "max_s": 0.00408,
-      "ratio_to_benchmark": 0.047,
+      "mean_s": 0.00373,
+      "p50_s": 0.00368,
+      "p95_s": 0.00424,
+      "max_s": 0.00475,
+      "ratio_to_benchmark": 0.05,
       "device": "NVIDIA GeForce RTX 3090 Ti"
     },
     "cpu_fp32": {
       "n": 100,
-      "mean_s": 0.00533,
-      "p50_s": 0.00519,
-      "p95_s": 0.00594,
-      "max_s": 0.00885,
-      "ratio_to_benchmark": 0.072,
+      "mean_s": 0.0052,
+      "p50_s": 0.00505,
+      "p95_s": 0.00601,
+      "max_s": 0.00809,
+      "ratio_to_benchmark": 0.07,
       "threads": 16
     },
     "benchmark_s": 0.074

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a081bfd14b72b8e64f3d66cf242da9c72afe65a33b2d7df39a0396b958309354
+oid sha256:62e48f02b52f8b2e670382054ca56088e2c50ee807147f401a4e7c0fd98eea53
 size 5634656