From 5873347b901ec9b1b998b2651d7706cd86205363 Mon Sep 17 00:00:00 2001
From: Yifan Zhao <yifanz16@illinois.edu>
Date: Mon, 25 Jan 2021 05:40:43 -0600
Subject: [PATCH] Supports passing param for model saving path from torchapp
---
predtuner/modeledapp.py | 27 ++++++++++++++++++---------
predtuner/torchapp.py | 22 +++++++++++-----------
2 files changed, 29 insertions(+), 20 deletions(-)
diff --git a/predtuner/modeledapp.py b/predtuner/modeledapp.py
index b4eee06..fd224fd 100644
--- a/predtuner/modeledapp.py
+++ b/predtuner/modeledapp.py
@@ -163,9 +163,9 @@ class LinearPerfModel(IPerfModel):
def measure_perf(self, with_approxes: KnobsT) -> float:
"""We implement this using a weighted linear performance model."""
- return float(sum(
- self.cost_df.loc[layer, knob] for layer, knob in with_approxes.items()
- ))
+ return float(
+ sum(self.cost_df.loc[layer, knob] for layer, knob in with_approxes.items())
+ )
class QoSModelP1(IQoSModel):
@@ -212,7 +212,7 @@ class QoSModelP1(IQoSModel):
def _init(self):
dt = self.delta_tensors
- btensor = self.baseline_tensor
+ btensor = self.baseline_tensor
if self.storage and self.storage.is_file():
for op, knob, delta_tensor in self._load(self.storage):
dt[op][knob] = delta_tensor
@@ -239,9 +239,12 @@ class QoSModelP1(IQoSModel):
def _try_append_save(
path: Optional[Path], op_name: str, knob_name: str, tensor: torch.Tensor
):
+ import os
+
if not path:
return
- path.touch(exist_ok=True)
+ if not path.parent.is_dir():
+ os.makedirs(path.parent)
with path.open("ab") as f:
pickle.dump((op_name, knob_name, tensor), f)
@@ -272,9 +275,10 @@ class QoSModelP2(IQoSModel):
def measure_qos(self, with_approxes: KnobsT) -> float:
assert self.baseline_qos is not None and self.qos_df is not None
- delta_qoses = np.array(
- [self.qos_df.loc[kv] for kv in with_approxes.items()]
- ) - self.baseline_qos
+ delta_qoses = (
+ np.array([self.qos_df.loc[kv] for kv in with_approxes.items()])
+ - self.baseline_qos
+ )
ret = delta_qoses.sum() + self.baseline_qos
assert not np.isnan(ret)
return float(ret)
@@ -285,10 +289,11 @@ class QoSModelP2(IQoSModel):
else:
knob_names = [k.name for k in self.app.knobs]
self.qos_df = pd.DataFrame(index=self.app.ops, columns=knob_names)
+ self.qos_df = self.qos_df.where(pd.notnull(self.qos_df), None)
self.baseline_qos = self._empirical_measure_qos({})
df = self.qos_df
for op, knob in barred_ravel_knobs(self.app):
- if not np.isnan(df.loc[op, knob]):
+ if df.loc[op, knob] is not None:
continue
df.loc[op, knob] = self._empirical_measure_qos({op: knob})
if self.storage and not self.storage.is_file():
@@ -312,6 +317,10 @@ class QoSModelP2(IQoSModel):
return df, baseline_qos
def _save(self, path: Path):
+ import os
+
+ if not path.parent.is_dir():
+ os.makedirs(path.parent)
with path.open("w") as f:
json.dump(
{
diff --git a/predtuner/torchapp.py b/predtuner/torchapp.py
index ae62d3b..2e781ac 100644
--- a/predtuner/torchapp.py
+++ b/predtuner/torchapp.py
@@ -1,20 +1,16 @@
import abc
-from typing import Any, Callable, Dict, List, Set, Tuple, Union
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Union
import numpy as np
import torch
from torch.nn import Module
from torch.utils.data.dataloader import DataLoader
+from ._logging import PathLike
from .approxapp import ApproxKnob, KnobsT
-from .modeledapp import (
- IPerfModel,
- IQoSModel,
- LinearPerfModel,
- ModeledApp,
- QoSModelP1,
- QoSModelP2,
-)
+from .modeledapp import (IPerfModel, IQoSModel, LinearPerfModel, ModeledApp,
+ QoSModelP1, QoSModelP2)
from .torchutil import ModuleIndexer, get_summary, move_to_device_recursively
@@ -64,6 +60,7 @@ class TorchApp(ModeledApp, abc.ABC):
tensor_to_qos: Callable[[torch.Tensor, Any], float],
combine_qos: Callable[[np.ndarray], float] = np.mean,
device: Union[torch.device, str] = _default_device,
+ model_storage_folder: Optional[PathLike] = None
) -> None:
self.app_name = app_name
self.module = module
@@ -73,6 +70,7 @@ class TorchApp(ModeledApp, abc.ABC):
self.tensor_to_qos = tensor_to_qos
self.combine_qos = combine_qos
self.device = device
+ self.model_storage = Path(model_storage_folder)
self.module = self.module.to(device)
self.midx = ModuleIndexer(module)
@@ -113,10 +111,12 @@ class TorchApp(ModeledApp, abc.ABC):
qoses.append(qos)
return self.combine_qos(np.array(qoses))
+ p1_storage = self.model_storage / "p1.pkl" if self.model_storage else None
+ p2_storage = self.model_storage / "p2.json" if self.model_storage else None
return [
LinearPerfModel(self._op_costs, self._knob_speedups),
- QoSModelP1(self, self._get_raw_output_valset, batched_valset_qos),
- QoSModelP2(self),
+ QoSModelP1(self, self._get_raw_output_valset, batched_valset_qos, p1_storage),
+ QoSModelP2(self, p2_storage),
]
@torch.no_grad()
--
GitLab