Upload pxg_tiny/config.py with huggingface_hub

pxg_tiny/config.py

@@ -0,0 +1,121 @@
+"""PXG-Tiny shared configuration: master palette, tokenizer, classes, model dims.
+
+Single source of truth imported by both the training pipeline and the
+inference runtime. All sprites are represented as 16x16 grids of *master
+palette indices* (0 = fully transparent alpha, 1..31 = opaque colors).
+This makes reconstruction byte-exact by construction and keeps the
+generation vocabulary tiny (32 visual symbols).
+"""
+
+IMG = 16                 # sprite resolution (16x16, even, inside the 8..256 schema band)
+N_COLORS = 32            # visual vocabulary size: index 0 = transparent, 1..31 colors
+
+# ---------------------------------------------------------------------------
+# Master palette (index -> RGBA). Hand-crafted to span the classic
+# warm_fantasy / nes / cool_sci_fi flavor families while staying <= 31 opaque
+# colors so a single byte encodes any pixel.
+# ---------------------------------------------------------------------------
+PALETTE = {
+    1:  (248, 248, 248),  # white / specular
+    2:  (217, 221, 230),  # pale gray
+    3:  (183, 191, 204),  # silver
+    4:  (120, 129, 142),  # gray
+    5:  (69, 76, 87),     # charcoal
+    6:  (23, 23, 31),     # ink (outline black)
+    7:  (212, 61, 51),    # red
+    8:  (140, 35, 32),    # maroon
+    9:  (232, 137, 46),   # orange / terracotta
+    10: (242, 193, 46),   # gold
+    11: (255, 232, 107),  # lemon (gold highlight)
+    12: (127, 201, 79),   # lime
+    13: (76, 154, 68),    # green mid
+    14: (44, 106, 53),    # forest dark
+    15: (55, 179, 171),   # teal
+    16: (159, 223, 240),  # ice / cyan highlight
+    17: (79, 147, 220),   # sky blue
+    18: (45, 71, 127),    # navy deep
+    19: (124, 80, 173),   # purple
+    20: (180, 140, 224),  # light violet
+    21: (168, 123, 35),   # bronze (dark gold)
+    22: (242, 203, 160),  # peach
+    23: (221, 182, 113),  # sand
+    24: (244, 230, 192),  # cream ivory
+    25: (165, 116, 74),   # tan wood
+    26: (122, 78, 42),    # brown mid
+    27: (76, 48, 26),     # dark wood
+    28: (111, 117, 57),   # moss olive
+    29: (185, 230, 197),  # mint pale (slime)
+    30: (93, 114, 136),   # steel blue-gray
+    31: (238, 173, 174),  # blush soft pink
+}
+
+# ---------------------------------------------------------------------------
+# Text tokenizer (char-level with byte-fallback buckets).
+# id 0 = <pad>, 1 = <bos>, 2 = <sep>, 3..6 punctuation, 7..32 a-z,
+# 33..62 byte-fallback buckets (ord % 30 + 33), 63 spare.
+# Deterministic, lossless round-trip on the supported charset; any other
+# Unicode character folds into its byte-fallback bucket (documented lossy
+# fold for exotic input, mirrors the "byte fallback" philosophy of the big
+# sibling project).
+# ---------------------------------------------------------------------------
+CHARSET = "abcdefghijklmnopqrstuvwxyz ,:-:'"
+PAD_ID, BOS_ID, SEP_ID = 0, 1, 2
+_TEXT_VOCAB = 64         # total text-side embedding width
+CAP_LEN = 32             # fixed padded caption slot count (incl. <pad>)
+PREFIX_LEN = 8           # intent-encoder output prefix vectors fed to decoder
+
+
+def encode_char(c: str) -> int:
+    if c == " ": return 3
+    if c == ",": return 4
+    if c == "-": return 5
+    if c == ":": return 6
+    if "a" <= c <= "z":
+        return 7 + ord(c) - ord("a")
+    if c == "'": return 39
+    # byte fallback bucket for anything else (digits, unicode, etc.)
+    fb = c.encode("utf-8", "replace")[0]        # first UTF-8 byte as int
+    return 33 + (fb % 30)
+
+
+def decode_char(i: int):
+    if i == 3: return " "
+    if i == 4: return ","
+    if i == 5: return "-"
+    if i == 6: return ":"
+    if 7 <= i <= 32: return chr(ord("a") + i - 7)
+    if i == 39: return "'"
+    if 33 <= i <= 38 or 40 <= i <= 62: return "?"   # fallback bucket: unknown glyph
+    return None                                      # control specials
+
+
+def encode_caption(text: str):
+    """text -> list of CAP_LEN ids, left-padded slot layout:
+    [<pad>...]<chars>. No <bos>/<sep> needed: the prefix encoder consumes the
+    whole 32-slot window positionally."""
+    ids = [encode_char(c) for c in text.lower()][:CAP_LEN]
+    return [PAD_ID] * (CAP_LEN - len(ids)) + ids
+
+
+def decode_caption(ids) -> str:
+    out = []
+    for i in ids:
+        c = decode_char(int(i))
+        if c is not None:
+            out.append(c)
+    return "".join(out).strip()
+
+
+# ---------------------------------------------------------------------------
+# Generation model geometry (extremely tiny on purpose).
+# ---------------------------------------------------------------------------
+D_MODEL = 96
+N_LAYERS = 4
+N_HEADS = 4              # 24-dim heads
+FF_HIDDEN = 256          # ffn_mult ~= 2.7 (lean by design)
+SEQ_VIS = IMG * IMG      # 256 autoregressive visual positions
+POS_TOTAL = PREFIX_LEN + SEQ_VIS   # 264 learned positions
+
+SAMPLING = {"temperature": 0.85, "top_k": 10}
+
+APPROX_PARAMS = 483_040  # exact: encoder + 4-block decoder + heads