Virtual-LPU / lpu /benchmarks /models.py
iparaco's picture
Upload 118 files
e59bd9d verified
Raw History Blame Contribute Delete
5.2 kB
"""Benchmark SR models (real PyTorch implementations, seeded deterministically).
- SRResNet (quick tier, ~447K params): the fast A/B/C comparison model.
- EDSR-baseline (extended tier, 16 residual blocks, no BN, ~1.41M params):
the realistic SR training workload.
Both are standard architectures implemented for THIS benchmark so A/B/C run
identical model code - the only variable is the optimization environment.
Scale support is explicit: SRResNet x2/x4 (PixelShuffle chains), EDSR-baseline
x2 (the standard configuration); anything else raises instead of silently
producing wrong upscale factors.
"""
from __future__ import annotations
from ..compat import require_torch
def _conv_block(nn, in_c, out_c, k=3):
return nn.Sequential(
nn.Conv2d(in_c, out_c, k, padding=k // 2),
nn.PReLU(),
)
class SRResNetQuick:
"""Compact SRResNet-style generator (4 residual blocks, upscale x2)."""
name = "srresnet_quick"
@staticmethod
def build(n_res: int = 4, feats: int = 64, scale: int = 2):
torch = require_torch()
nn = torch.nn
if scale not in (2, 4):
raise ValueError(f"SRResNetQuick supports scale 2 or 4, got {scale}")
def res_block():
return nn.Sequential(
nn.Conv2d(feats, feats, 3, padding=1),
nn.PReLU(),
nn.Conv2d(feats, feats, 3, padding=1),
)
class ResChain(nn.Module):
def __init__(self):
super().__init__()
self.blocks = nn.ModuleList([res_block() for _ in range(n_res)])
def forward(self, x):
for b in self.blocks:
x = x + b(x)
return x
class Net(nn.Module):
def __init__(self):
super().__init__()
self.head = nn.Conv2d(3, feats, 3, padding=1)
self.body = ResChain()
up = []
for _ in range(max(1, scale // 2)):
up += [nn.Conv2d(feats, feats * 4, 3, padding=1),
nn.PixelShuffle(2), nn.PReLU()]
self.up = nn.Sequential(*up)
self.tail = nn.Conv2d(feats, 3, 3, padding=1)
def forward(self, x):
x = self.head(x)
x = self.body(x)
x = self.up(x)
return self.tail(x)
return Net()
class EDSRBaseline:
"""EDSR-baseline: 16 residual blocks, 64 features, no BatchNorm, x2.
Matches the canonical EDSR-baseline configuration used across SR research
(Lim et al., CVPRW 2017) closely enough for a realistic training load.
"""
name = "edsr_baseline"
@staticmethod
def build(n_res: int = 16, feats: int = 64, scale: int = 2):
torch = require_torch()
nn = torch.nn
def res():
return nn.Sequential(
nn.Conv2d(feats, feats, 3, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(feats, feats, 3, padding=1),
)
class MeanShift(nn.Module):
# keep the classic EDSR mean-shift at 0 to stay simple+reversible
def forward(self, x):
return x
class Net(nn.Module):
def __init__(self):
super().__init__()
self.shift = MeanShift()
self.head = nn.Conv2d(3, feats, 3, padding=1)
self.body = nn.Sequential(*[res() for _ in range(n_res)])
self.body_conv = nn.Conv2d(feats, feats, 3, padding=1)
up = [nn.Conv2d(feats, feats * 4, 3, padding=1),
nn.PixelShuffle(2), nn.Conv2d(feats, feats, 3, padding=1)] \
if scale == 2 else []
self.up = nn.Sequential(*up)
self.tail = nn.Conv2d(feats, 3, 3, padding=1)
def forward(self, x):
x = self.shift(x)
x = self.head(x)
x = x + self.body_conv(self.body(x))
x = self.up(x)
return self.tail(x)
return Net()
def build_model(tier: str):
if tier == "quick":
return SRResNetQuick.build()
if tier == "extended":
return EDSRBaseline.build()
raise ValueError(f"unknown benchmark tier {tier!r} (quick | extended)")
def model_facts(model) -> dict:
torch = require_torch()
params = list(model.parameters())
n_params = sum(p.numel() for p in params)
mb = sum(p.numel() * p.element_size() for p in params) / (1024 * 1024)
dtype = str(params[0].dtype) if params else "unknown"
return {"params": n_params, "param_mb": round(mb, 2),
"dtype": dtype, "device": str(params[0].device) if params else "?"}
BENCHMARK_TIERS = {
"quick": {
"model": "srresnet_quick",
"patch_lr": 48,
"scale": 2,
"default_batch": 16,
"description": "SRResNet 4RB, LR patch 48x48, x2 upscale",
},
"extended": {
"model": "edsr_baseline",
"patch_lr": 64,
"scale": 2,
"default_batch": 8,
"description": "EDSR-baseline 16RB, LR patch 64x64, x2 upscale",
},
}