mt/thesis/third_party/PlotNeuralNet/deepsad/arch_ef_encoder.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, H1 = 26, 18, 12, 1
D2048, D1024, D512, D256, D128, D1 = 52, 36, 24, 12, 6, 1
W1, W4, W8, W16, W32 = 1, 2, 2, 4, 8


arch = [
    to_head(".."),
    to_cor(),
    to_begin(),
    # --------------------------- ENCODER ---------------------------
    # Input 1×32×2048  (caption carries H×W; s_filer is numeric)
    to_Conv(
        "input",
        zlabeloffset=0.2,
        s_filer="{{2048×32}}",
        n_filer=1,
        offset="(0,0,0)",
        to="(0,0,0)",
        height=H32,
        depth=D2048,
        width=W1,
        caption="Input",
    ),
    # Conv1 (5x5, same): 1->8, 32×2048
    to_Conv(
        "dwconv1",
        s_filer="",
        n_filer=1,
        offset="(2,0,0)",
        to="(input-east)",
        height=H32,
        depth=D2048,
        width=W1,
        caption="",
    ),
    to_Conv(
        "dwconv2",
        s_filer="{{2048×32}}",
        zlabeloffset=0.15,
        n_filer=16,
        offset="(0,0,0)",
        to="(dwconv1-east)",
        height=H32,
        depth=D2048,
        width=W16,
        caption="Conv1",
    ),
    # Pool1 2×2: 32×2048 -> 16×1024
    # to_connection("input", "conv1"),
    to_Pool(
        "pool1",
        offset="(0,0,0)",
        zlabeloffset=0.3,
        s_filer="{{512×32}}",
        to="(dwconv2-east)",
        height=H32,
        depth=D512,
        width=W16,
        caption="",
    ),
    # Conv2 (5x5, same): 8->4, stays 16×1024
    to_Conv(
        "dwconv3",
        s_filer="",
        n_filer=1,
        offset="(2,0,0)",
        to="(pool1-east)",
        height=H32,
        depth=D512,
        width=W1,
        caption="",
    ),
    to_Conv(
        "dwconv4",
        n_filer=32,
        zlabeloffset=0.3,
        s_filer="{{512×32}}",
        offset="(0,0,0)",
        to="(dwconv3-east)",
        height=H32,
        depth=D512,
        width=W32,
        caption="Conv2",
    ),
    # Pool2 2×2: 16×1024 -> 8×512
    # to_connection("pool1", "conv2"),
    to_Pool(
        "pool2",
        offset="(0,0,0)",
        zlabeloffset=0.45,
        s_filer="{{256×16}}",
        to="(dwconv4-east)",
        height=H16,
        depth=D256,
        width=W32,
        caption="",
    ),
    to_Pool(
        "pool3",
        offset="(0,0,0)",
        zlabeloffset=0.45,
        s_filer="{{128×8}}",
        to="(pool2-east)",
        height=H8,
        depth=D128,
        width=W32,
        caption="",
    ),
    to_Conv(
        "squeeze",
        n_filer=8,
        zlabeloffset=0.45,
        s_filer="{{128×8}}",
        offset="(2,0,0)",
        to="(pool3-east)",
        height=H8,
        depth=D128,
        width=W8,
        caption="Squeeze",
    ),
    # FC -> rep_dim (use numeric n_filer)
    to_fc(
        "fc1",
        n_filer="{{8×128×8}}",
        zlabeloffset=0.5,
        offset="(2,0,0)",
        to="(squeeze-east)",
        height=H1,
        depth=D512,
        width=W1,
        caption=f"FC",
    ),
    # to_connection("pool2", "fc1"),
    # --------------------------- LATENT  ---------------------------
    to_Conv(
        "latent",
        n_filer="",
        s_filer="latent dim",
        offset="(2,0,0)",
        to="(fc1-east)",
        height=H8 * 1.6,
        depth=D1,
        width=W1,
        caption=f"Latent Space",
    ),
    to_end(),
]


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


if __name__ == "__main__":
    main()