mt/thesis/third_party/PlotNeuralNet/deepsad/arch_lenet_decoder.py

# subter_lenet_arch.py
# Requires running from inside the PlotNeuralNet repo, like: python3 ../subter_lenet_arch.py
import sys, argparse

sys.path.append("../")  # import pycore from repo root

from pycore.tikzeng import *

parser = argparse.ArgumentParser()
parser.add_argument("--rep_dim", type=int, default=1024, help="latent size for FC")
args = parser.parse_args()
REP = int(args.rep_dim)

# Visual scales so the huge width doesn't dominate the figure
H32, H16, H8 = 26, 18, 12
D2048, D1024, D512 = 52, 36, 24
W1, W4, W8 = 1, 2, 4


arch = [
    to_head(".."),
    to_cor(),
    to_begin(),
    to_Conv(
        "latent",
        n_filer="",
        s_filer="latent dim",
        offset="(2,0,0)",
        to="(0,0,0)",
        height=H8 * 1.6,
        depth=1.3,
        width=W1,
        caption=f"Latent Space",
    ),
    # to_connection("fc1", "latent"),
    # --------------------------- DECODER ---------------------------
    # FC back to 16384
    to_fc(
        "fc3",
        n_filer="{{4×512×8}}",
        offset="(2,0,0)",
        to="(latent-east)",
        height=1.3,
        depth=D512,
        width=W1,
        caption=f"FC",
    ),
    # to_connection("latent", "fc3"),
    # Reshape to 4×8×512
    to_UnPool(
        "up1",
        offset="(2,0,0)",
        to="(fc3-east)",
        height=H16,
        depth=D1024,
        width=W4,
        caption="",
    ),
    # Up ×2: 8×512 -> 16×1024 (we just draw a labeled box)
    # DeConv1 (5×5, same): 4->8, 16×1024
    to_Conv(
        "deconv1",
        s_filer="{{1024×16}}",
        n_filer=8,
        offset="(0,0,0)",
        to="(up1-east)",
        height=H16,
        depth=D1024,
        width=W8,
        caption="Deconv1",
    ),
    # to_connection("fc3", "up1"),
    # Up ×2: 16×1024 -> 32×2048
    to_UnPool(
        "up2",
        offset="(2,0,0)",
        to="(deconv1-east)",
        height=H32,
        depth=D2048,
        width=W8,
        caption="",
    ),
    # to_connection("deconv1", "up2"),
    # DeConv2 (5×5, same): 8->1, 32×2048
    to_Conv(
        "deconv2",
        s_filer="{{2048×32}}",
        n_filer=1,
        offset="(0,0,0)",
        to="(up2-east)",
        height=H32,
        depth=D2048,
        width=W1,
        caption="Deconv2",
    ),
    # to_connection("up2", "deconv2"),
    # Output
    to_Conv(
        "out",
        s_filer="{{2048×32}}",
        n_filer=1,
        offset="(2,0,0)",
        to="(deconv2-east)",
        height=H32,
        depth=D2048,
        width=1.0,
        caption="Output",
    ),
    # to_connection("deconv2", "out"),
    to_end(),
]


def main():
    name = "subter_lenet_arch"
    to_generate(arch, name + ".tex")


if __name__ == "__main__":
    main()