sbf-scan-model/Makefile at main · SixByFive/sbf-scan-model · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
# -------- SixByFive Finish Detector — Makefile --------
# Requires: Python 3.12+, pip-installed requirements.
# On Windows, run via Git Bash (or WSL) so rm/find work.

PY        ?= python
TRAIN_IMG ?= 160
EPOCHS    ?= 25

# Thresholds for dataset check (tweak as you like)
MIN_CARDS_PER_CLASS  ?= 20
MIN_IMAGES_PER_CARD  ?= 5
MIN_TOTAL_PER_CLASS  ?= 150

# Paths
RAW_DIR      := datasets/finish_raw
SPLIT_DIR    := datasets/finish_split
PATCH_DIR    := datasets/finish_patches
PATCH_TRAIN  := $(PATCH_DIR)/train
PATCH_VAL    := $(PATCH_DIR)/val
ARTIFACTS    := artifacts/finish_v1

# Convenient split ratios (can override at call time, e.g. make split TRAIN=0.8 VAL=0.2 TEST=0.0)
TRAIN ?= 0.8
VAL   ?= 0.2
TEST  ?= 0.0

.PHONY: help check split patches patches-raw patches-split train eval clean deepclean freeze

help:
	@echo "Targets:"
	@echo "  make check          - Validate dataset balance/leakage on $(RAW_DIR)"
	@echo "  make split          - Create card-aware train/val/test splits in $(SPLIT_DIR)"
	@echo "  make patches        - Build patches (prefers split; falls back to raw)"
	@echo "  make train          - Train model from $(PATCH_TRAIN) and $(PATCH_VAL) -> $(ARTIFACTS)"
	@echo "  make eval           - Evaluate exported TFLite on $(PATCH_VAL)"
	@echo "  make clean          - Remove generated patches and current artifacts"
	@echo "  make deepclean      - Remove ALL generated patches and artifacts"
	@echo "  make freeze         - Write pip freeze to requirements.lock.txt"

# 1) Sanity-check raw dataset before splitting
check:
	$(PY) tools/check_dataset.py \
	  --root $(RAW_DIR) \
	  --min-cards-per-class $(MIN_CARDS_PER_CLASS) \
	  --min-images-per-card $(MIN_IMAGES_PER_CARD) \
	  --min-total-per-class $(MIN_TOTAL_PER_CLASS)

# 2) Create card-aware split (never mixes the same card across splits)
split:
	$(PY) tools/make_split.py \
	  --root $(RAW_DIR) \
	  --out  $(SPLIT_DIR) \
	  --train $(TRAIN) --val $(VAL) --test $(TEST)

# 3) Build patches (auto-crops corners + bottom band)
# Prefer split dirs; if they don't exist, fall back to building from raw (single set).
patches: patches-split

# Build patches from split (train/val must exist)
patches-split:
	@if [ -d "$(SPLIT_DIR)/train" ] && [ -d "$(SPLIT_DIR)/val" ]; then \
	  echo ">> Building patches from split..."; \
	  $(PY) tools/build_patches.py --src $(SPLIT_DIR)/train --dst $(PATCH_TRAIN); \
	  $(PY) tools/build_patches.py --src $(SPLIT_DIR)/val   --dst $(PATCH_VAL); \
	else \
	  echo "!! $(SPLIT_DIR)/train or /val not found. Falling back to raw."; \
	  $(MAKE) patches-raw; \
	fi

# Build patches directly from raw (single directory of patches)
patches-raw:
	$(PY) tools/build_patches.py --src $(RAW_DIR) --dst $(PATCH_DIR)

# 4) Train (explicit dirs). Requires patches in $(PATCH_TRAIN) and $(PATCH_VAL).
train:
	@mkdir -p $(ARTIFACTS)
	$(PY) tools/train_finish_classifier_explicit.py \
	  --train_dir $(PATCH_TRAIN) \
	  --val_dir   $(PATCH_VAL) \
	  --out       $(ARTIFACTS) \
	  --img $(TRAIN_IMG) \
	  --epochs $(EPOCHS)

# 5) Evaluate exported TFLite quickly on val patches
eval:
	$(PY) tools/eval_tflite.py \
	  --data   $(PATCH_VAL) \
	  --tflite $(ARTIFACTS)/model.int8.tflite \
	  --labels $(ARTIFACTS)/labels.txt \
	  --img    $(TRAIN_IMG)

# Clean current outputs
clean:
	@echo ">> Removing generated patches (train/val) and artifacts at $(ARTIFACTS)"
	@rm -rf $(PATCH_TRAIN) $(PATCH_VAL) $(ARTIFACTS)

# Deep clean everything generated
deepclean:
	@echo ">> Removing ALL generated patches and artifacts"
	@rm -rf $(PATCH_DIR) artifacts/*

# Pin current venv state (optional)
freeze:
	$(PY) -m pip freeze > requirements.lock.txt
	@echo "Wrote requirements.lock.txt"