Spaces:
Runtime error
Runtime error
| import os, struct, math | |
| import numpy as np | |
| from PIL import Image, ImageDraw | |
| PHI = (1.0 + 5.0**0.5) / 2.0 | |
| MAGIC = b"FLC1\x00\x00\x00\x00" | |
| VER_V2 = 2 | |
| # --- Fibonacci & Frequency Logic --- | |
| def fibonacci_sequence(n): | |
| fibs = [1, 2] | |
| while len(fibs) < n: | |
| fibs.append(fibs[-1] + fibs[-2]) | |
| return np.array(fibs[:n], dtype=np.int64) | |
| def fibonacci_sequence_std(n): | |
| fibs = [1, 1] | |
| while len(fibs) < n: | |
| fibs.append(fibs[-1] + fibs[-2]) | |
| return np.array(fibs[:n], dtype=np.int64) | |
| def fibonacci_frequency_boundaries(n_coeffs: int, n_bands: int): | |
| if n_bands < 2: return [0, n_coeffs] | |
| fibs = fibonacci_sequence(n_bands).astype(np.float64) | |
| w = fibs / (fibs.sum() + 1e-12) | |
| cum = np.cumsum(w) | |
| b = [0] | |
| for i in range(n_bands - 1): | |
| b.append(int(round(n_coeffs * cum[i]))) | |
| b.append(n_coeffs) | |
| # Ensure strict monotonicity | |
| for i in range(1, len(b)): | |
| if b[i] <= b[i-1]: b[i] = b[i-1] + 1 | |
| return b | |
| # --- Orthonormal DCT Engine --- | |
| def dct_ortho_1d(x: np.ndarray) -> np.ndarray: | |
| N = x.shape[0] | |
| v = np.concatenate([x, x[::-1]]) | |
| V = np.fft.fft(v) | |
| k = np.arange(N) | |
| X = np.real(V[:N] * np.exp(-1j * np.pi * k / (2 * N))) | |
| X *= 2.0 | |
| X[0] *= (1.0 / math.sqrt(4 * N)) | |
| X[1:] *= (1.0 / math.sqrt(2 * N)) | |
| return X | |
| def idct_ortho_1d(X: np.ndarray) -> np.ndarray: | |
| N = X.shape[0] | |
| x0, xr = X[0] * math.sqrt(4 * N), X[1:] * math.sqrt(2 * N) | |
| c = np.empty(N, dtype=np.complex128) | |
| c[0], c[1:] = x0 / 2.0, xr / 2.0 | |
| k = np.arange(N) | |
| c = c * np.exp(1j * np.pi * k / (2 * N)) | |
| V = np.zeros(2 * N, dtype=np.complex128) | |
| V[:N] = c | |
| V[N+1:] = np.conj(c[1:][::-1]) | |
| return np.fft.ifft(V).real[:N] | |
| def dct_blocks_ortho(x_blocks: np.ndarray) -> np.ndarray: | |
| return np.array([dct_ortho_1d(b) for b in x_blocks]) | |
| def idct_blocks_ortho(X_blocks: np.ndarray) -> np.ndarray: | |
| return np.array([idct_ortho_1d(B) for B in X_blocks]) | |
| # --- Fibonacci Coding (Bit IO) --- | |
| class BitWriter: | |
| def __init__(self): | |
| self.buf, self.acc, self.nbits = bytearray(), 0, 0 | |
| def write_bit(self, b: int): | |
| self.acc = (self.acc << 1) | (b & 1) | |
| self.nbits += 1 | |
| if self.nbits == 8: | |
| self.buf.append(self.acc); self.acc = 0; self.nbits = 0 | |
| def finish(self): | |
| if self.nbits: self.buf.append(self.acc << (8 - self.nbits)) | |
| return bytes(self.buf) | |
| class BitReader: | |
| def __init__(self, data: bytes): | |
| self.data, self.i, self.acc, self.nbits = data, 0, 0, 0 | |
| def read_bit(self): | |
| if self.nbits == 0: | |
| self.acc = self.data[self.i]; self.i += 1; self.nbits = 8 | |
| b = (self.acc >> (self.nbits - 1)) & 1 | |
| self.nbits -= 1 | |
| return b | |
| def fib_encode_nonneg(bw, n): | |
| m = int(n) + 1 | |
| fibs = [1, 2] | |
| while fibs[-1] <= m: fibs.append(fibs[-1] + fibs[-2]) | |
| bits = [0] * (len(fibs) - 1) | |
| for i in reversed(range(len(bits))): | |
| if fibs[i] <= m: bits[i] = 1; m -= fibs[i] | |
| for i in range(max((i for i, b in enumerate(bits) if b), default=0) + 1): | |
| bw.write_bit(bits[i]) | |
| bw.write_bit(1) | |
| def fib_decode_nonneg(br): | |
| fibs, bits, prev = [1, 2], [], 0 | |
| while True: | |
| b = br.read_bit(); bits.append(b) | |
| if prev == 1 and b == 1: break | |
| prev = b | |
| if len(bits) > len(fibs): fibs.append(fibs[-1] + fibs[-2]) | |
| m = sum(fibs[i] for i, bi in enumerate(bits[:-1]) if bi) | |
| return m - 1 | |
| def rle_fib_encode_ints(ints): | |
| bw = BitWriter() | |
| zrun = 0 | |
| for v in ints: | |
| if v == 0: zrun += 1; continue | |
| if zrun: bw.write_bit(0); fib_encode_nonneg(bw, zrun); zrun = 0 | |
| bw.write_bit(1); fib_encode_nonneg(bw, (v << 1) ^ (v >> 63)) | |
| if zrun: bw.write_bit(0); fib_encode_nonneg(bw, zrun) | |
| return bw.finish() | |
| def rle_fib_decode_ints(payload, n_out): | |
| br, out, i = BitReader(payload), np.zeros(n_out, dtype=np.int64), 0 | |
| while i < n_out: | |
| if br.read_bit() == 0: i = min(n_out, i + fib_decode_nonneg(br)) | |
| else: | |
| u = fib_decode_nonneg(br) | |
| out[i] = (u >> 1) ^ (-(u & 1)); i += 1 | |
| return out | |
| # --- Quantization & Spiral Visuals --- | |
| def band_quantize_dct(coeffs, boundaries, base_step): | |
| q = np.zeros_like(coeffs, dtype=np.int32) | |
| for bi in range(len(boundaries) - 1): | |
| a, b = boundaries[bi], boundaries[bi + 1] | |
| step = base_step * (PHI ** bi) | |
| q[:, a:b] = np.round(coeffs[:, a:b] / step) | |
| return q | |
| def band_dequantize_dct(q, boundaries, base_step): | |
| coeffs = np.zeros_like(q, dtype=np.float64) | |
| for bi in range(len(boundaries) - 1): | |
| a, b = boundaries[bi], boundaries[bi + 1] | |
| step = base_step * (PHI ** bi) | |
| coeffs[:, a:b] = q[:, a:b] * step | |
| return coeffs | |
| def hologram_spectrum_image(zints, max_symbols=262144): | |
| z = zints[:max_symbols]; v = np.tanh(z / 32.0) | |
| theta = (2 * math.pi / (PHI**2)) * np.arange(v.size) + 2.0 * math.pi * (v * 0.25) | |
| r = 1.0 + 0.35 * np.abs(v) | |
| syms = r * np.cos(theta) + 1j * r * np.sin(theta) | |
| N = int(2**math.ceil(math.log2(math.sqrt(syms.size or 1)))) | |
| U = np.pad(syms, (0, N*N - syms.size)).reshape(N, N) | |
| mag = np.log1p(np.abs(np.fft.fftshift(np.fft.fft2(U)))) | |
| mag = (mag - mag.min()) / (mag.max() - mag.min() + 1e-12) | |
| return (mag * 255).astype(np.uint8) | |
| def bytes_to_fib_spiral_image(data, max_pixels=262144): | |
| arr = np.frombuffer(data, dtype=np.uint8)[:max_pixels] | |
| fibs = fibonacci_sequence_std(32) | |
| sizes, area = [], 0 | |
| for s in fibs: | |
| sizes.append(int(s)); area += s*s | |
| if area >= arr.size: break | |
| # Simple tile placement logic for demo | |
| tiles, minx, miny, maxx, maxy = [], 0, 0, 0, 0 | |
| curr_x, curr_y = 0, 0 | |
| for i, s in enumerate(sizes): | |
| d = (i-1)%4 | |
| if i>0: | |
| if d==0: curr_x = maxx; curr_y = miny | |
| elif d==1: curr_x = maxx-s; curr_y = maxy | |
| elif d==2: curr_x = minx-s; curr_y = maxy-s | |
| else: curr_x = minx; curr_y = miny-s | |
| tiles.append((curr_x, curr_y, s)) | |
| minx, miny = min(minx, curr_x), min(miny, curr_y) | |
| maxx, maxy = max(maxx, curr_x+s), max(maxy, curr_y+s) | |
| W, H = maxx-minx, maxy-miny | |
| img = np.zeros((H, W), dtype=np.uint8) | |
| idx = 0 | |
| for x, y, s in tiles: | |
| take = min(s*s, arr.size - idx) | |
| if take <= 0: break | |
| block = np.pad(arr[idx:idx+take], (0, s*s-take)).reshape(s, s) | |
| img[H-(y-miny+s):H-(y-miny), x-minx:x-minx+s] = block | |
| idx += take | |
| return img, tiles, (minx, miny, maxx, maxy) | |
| # --- High Level API --- | |
| def flc_encode_file(in_path, out_flc, preview_png=None, unzip_gif=None, block_len=1024, n_bands=10, base_step=0.004, **kwargs): | |
| raw = open(in_path, "rb").read() | |
| x = (np.frombuffer(raw, dtype=np.uint8).astype(np.float64) - 127.5) / 127.5 | |
| pad = (-x.size) % block_len | |
| X = np.pad(x, (0, pad)).reshape(-1, block_len) | |
| C = dct_blocks_ortho(X) | |
| bnds = fibonacci_frequency_boundaries(block_len, n_bands) | |
| Q = band_quantize_dct(C, bnds, base_step) | |
| payload = rle_fib_encode_ints(np.diff(Q.flatten(), prepend=0)) | |
| header = struct.pack("<8sH Q I I H d d H", MAGIC, VER_V2, len(raw), block_len, X.shape[0], n_bands, base_step, 127.5, len(bnds)) | |
| with open(out_flc, "wb") as f: | |
| f.write(header); f.write(struct.pack("<"+ "I"*len(bnds), *bnds)) | |
| f.write(struct.pack("<I", len(payload))); f.write(payload) | |
| if unzip_gif: # Generate the unzip frames | |
| frames = [] | |
| for t in range(1, n_bands + 1): | |
| Q_p = np.zeros_like(Q) | |
| for bi in range(t): Q_p[:, bnds[bi]:bnds[bi+1]] = Q[:, bnds[bi]:bnds[bi+1]] | |
| X_p = idct_blocks_ortho(band_dequantize_dct(Q_p, bnds, base_step)) | |
| recon = np.clip((X_p.flatten()[:len(raw)] * 127.5) + 127.5, 0, 255).astype(np.uint8) | |
| # Create Frame | |
| h_img = Image.fromarray(hologram_spectrum_image(Q_p.flatten())).resize((256, 256)) | |
| s_img_arr, _, _ = bytes_to_fib_spiral_image(recon.tobytes()) | |
| s_img = Image.fromarray(s_img_arr).resize((256, 256)) | |
| frame = Image.new("RGB", (512, 280), (10, 10, 15)) | |
| frame.paste(h_img, (0, 24)); frame.paste(s_img, (256, 24)) | |
| frames.append(frame) | |
| frames[0].save(unzip_gif, save_all=True, append_images=frames[1:], duration=100, loop=0) | |
| return {"n_bytes": len(raw), "payload_len": len(payload), "ratio": len(payload)/len(raw)} | |
| def flc_decode_file(in_flc, out_path): | |
| blob = open(in_flc, "rb").read() | |
| h_sz = struct.calcsize("<8sH Q I I H d d H") | |
| magic, ver, n_bytes, b_len, n_blks, n_bnds, step, mu, bnd_l = struct.unpack_from("<8sH Q I I H d d H", blob) | |
| off = h_sz | |
| bnds = struct.unpack_from("<" + "I"*bnd_l, blob, off); off += 4*bnd_l | |
| p_len = struct.unpack_from("<I", blob, off)[0]; off += 4 | |
| d = rle_fib_decode_ints(blob[off:off+p_len], n_blks * b_len) | |
| Q = np.cumsum(d).reshape(n_blks, b_len) | |
| X = idct_blocks_ortho(band_dequantize_dct(Q, bnds, step)) | |
| res = np.clip((X.flatten()[:n_bytes] * 127.5) + mu, 0, 255).astype(np.uint8) | |
| with open(out_path, "wb") as f: f.write(res.tobytes()) | |
| return {"n_bytes": n_bytes} | |
| def cosine_similarity_bytes(a, b): | |
| x = np.frombuffer(a, dtype=np.uint8).astype(float) | |
| y = np.frombuffer(b, dtype=np.uint8).astype(float) | |
| n = min(len(x), len(y)) | |
| x, y = x[:n]-x[:n].mean(), y[:n]-y[:n].mean() | |
| return np.dot(x, y) / (np.linalg.norm(x) * np.linalg.norm(y) + 1e-12) |