Upload folder using huggingface_hub

best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e369616d25d4620d5389fdd7163100b46f01d6ad02fdd507985f15059c5f34d
+size 1216359894

classification_report.txt

@@ -0,0 +1,11 @@
+               precision    recall  f1-score   support
+
+     carmenta     0.8077    0.7500    0.7778        28
+      healthy     0.9728    0.9670    0.9699       333
+   moniliasis     0.9057    0.8727    0.8889        55
+ phytophthora     0.7797    0.8519    0.8142        54
+witches_broom     0.8000    0.8889    0.8421         9
+
+     accuracy                         0.9290       479
+    macro avg     0.8532    0.8661    0.8586       479
+ weighted avg     0.9304    0.9290    0.9294       479

config.yaml

@@ -0,0 +1,54 @@
+# Cocoa pod 5-class — EVA-02-Large fine-tuning on MI300X
+# Backbone:  EVA-02-L ViT-L/14 pretrained MIM-CLIP, fine-tuned IN22k+IN1k @ 448
+# Dataset :  merged LatAm + Peru YOLO crops (~4.9k crops, 5 classes)
+#
+# Resolution is 448 — quadruples per-image memory vs Track 2's 224×224 DINOv2,
+# so batch_size and num_workers are scaled down. Saturate later if VRAM allows.
+
+seed: 123
+
+data:
+  # splits.json built by prepare_cocoa_data.py — paths inside it are absolute
+  splits_file: splits.json
+  num_workers: 4
+  pin_memory: true
+
+model:
+  # timm name for EVA-02-Large pretrained MIM (LAION) + IN22k + IN1k FT @ 448
+  name: eva02_large_patch14_448.mim_m38m_ft_in22k_in1k
+  num_classes: 5
+  drop_path_rate: 0.1
+  img_size: 448
+
+train:
+  # Single-phase fine-tune (matches the simplified track2/train.py behavior)
+  lr: 1.0e-4
+  epochs: 20             # ~5k crops × 20 epochs ≈ 100k steps total at BS=64
+  batch_size: 64         # safe at 448px on a 192GB MI300X; raise if you have headroom
+  grad_accum_steps: 1
+  weight_decay: 0.05
+  betas: [0.9, 0.999]
+  label_smoothing: 0.1
+  mixup_alpha: 0.1       # matches the reduced setting that worked in Track 2
+  cutmix_alpha: 0.5
+  mixup_prob: 0.5
+  grad_clip: 1.0
+  amp_dtype: bfloat16    # MI300X has native bf16
+  compile: false         # enable after the first stable run if you want speed
+  grad_checkpointing: true   # 448px attn maps are big; checkpoint to free VRAM
+
+augment:
+  rand_augment_n: 2
+  rand_augment_m: 9
+  random_erasing_p: 0.25
+  horizontal_flip: true
+  vertical_flip: true       # cocoa pods don't have a strong up/down orientation
+
+eval:
+  batch_size: 64
+  tta_rounds: 10
+
+log:
+  every_n_steps: 50
+  checkpoint_dir: runs
+  save_best_metric: val_macro_f1   # rare classes (witches_broom, carmenta) → optimize macro F1

confusion_matrix.csv

@@ -0,0 +1,5 @@
+21,4,0,3,0
+3,322,2,4,2
+0,1,48,6,0
+1,4,3,46,0
+1,0,0,0,8

metrics.json

@@ -0,0 +1,12 @@
+{
+  "standard": {
+    "acc": 0.9206680584551148,
+    "macro_f1": 0.8458037279646249
+  },
+  "tta": {
+    "acc": 0.9290187891440501,
+    "macro_f1": 0.8585621479864496,
+    "rounds": 10
+  },
+  "improvement_pp": 0.8350730688935326
+}