From ade58e30113dcac3edda75c467a2f5a035d77e64 Mon Sep 17 00:00:00 2001
From: Yifan Zhao <yifanz16@illinois.edu>
Date: Thu, 8 Jul 2021 01:21:22 -0500
Subject: [PATCH] Added routine for quantizing miniera model:
added distiller as a submodule
updated env.yaml
auto-install distiller in installer
---
.gitmodules | 3 +
hpvm/env.yaml | 30 ++---
hpvm/projects/distiller | 1 +
hpvm/scripts/hpvm_installer.py | 2 +-
hpvm/test/epoch_dnn/torch_dnn/__init__.py | 1 +
hpvm/test/epoch_dnn/torch_dnn/quantizer.py | 135 +++++++++++++++++++++
6 files changed, 157 insertions(+), 15 deletions(-)
create mode 160000 hpvm/projects/distiller
create mode 100644 hpvm/test/epoch_dnn/torch_dnn/quantizer.py
diff --git a/.gitmodules b/.gitmodules
index c4a0bd02bc..6d6a1a9975 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -5,3 +5,6 @@
[submodule "hpvm/projects/sw"]
path = hpvm/projects/sw
url = https://github.com/nvdla/sw.git
+[submodule "hpvm/projects/distiller"]
+ path = hpvm/projects/distiller
+ url = git@gitlab.engr.illinois.edu:llvm/distiller.git
diff --git a/hpvm/env.yaml b/hpvm/env.yaml
index 612cc956c1..3e1f6394c7 100644
--- a/hpvm/env.yaml
+++ b/hpvm/env.yaml
@@ -1,26 +1,28 @@
name: hpvm
channels:
- pytorch
- - defaults
+ - conda-forge
dependencies:
+ - Cython
- jinja2=2.11
- - jsonpickle=2
- - keras==2.1.6
- matplotlib=3.3
- networkx=2.5
+ - onnx==1.8.0
- pandas=1.1
- - python==3.6.13
- pip
- - pytorch==1.6.0
- - torchvision=0.7
- - tqdm=4.59
- - scipy==1.1.0
- - h5py==2.10.0
+ - pydot=1.4
+ - python=3.7
+ - pytorch==1.5.0
+ - PyYAML
+ - scikit-learn=0.21
+ - scipy=1.3
+ - tabulate=0.8
+ - tensorflow==1.14.0
+ - torchvision==0.6
+ - tqdm=4.33
+ - xlsxwriter=1.2
- pip:
- argparse==1.4
- - onnx==1.8
+ - graphviz==0.10
- onnx-simplifier==0.3
- - opentuner==0.8.3
- - sqlalchemy==1.3.0
- - tensorflow==1.14.0
- - tensorflow-gpu==1.14.0
+ - torchnet==0.0.4
diff --git a/hpvm/projects/distiller b/hpvm/projects/distiller
new file mode 160000
index 0000000000..30c50f368d
--- /dev/null
+++ b/hpvm/projects/distiller
@@ -0,0 +1 @@
+Subproject commit 30c50f368d219efa87d927c0ad0a563bcfa29f0a
diff --git a/hpvm/scripts/hpvm_installer.py b/hpvm/scripts/hpvm_installer.py
index 16e298f90e..1ee808458e 100755
--- a/hpvm/scripts/hpvm_installer.py
+++ b/hpvm/scripts/hpvm_installer.py
@@ -31,7 +31,7 @@ MAKE_TARGETS = ["hpvm-clang"]
MAKE_TEST_TARGETS = ["check-hpvm-dnn", "check-hpvm-pass"]
# Relative to project root which is __file__.parent.parent
-PY_PACKAGES = ["projects/torch2hpvm"]
+PY_PACKAGES = ["projects/torch2hpvm", "projects/distiller"]
PYTHON_REQ = (3, 6) # This means >= 3.6
diff --git a/hpvm/test/epoch_dnn/torch_dnn/__init__.py b/hpvm/test/epoch_dnn/torch_dnn/__init__.py
index 62f29d0d7e..a6fb99465f 100644
--- a/hpvm/test/epoch_dnn/torch_dnn/__init__.py
+++ b/hpvm/test/epoch_dnn/torch_dnn/__init__.py
@@ -1,2 +1,3 @@
from .datasets import CIFAR
from .miniera import MiniERA
+from .quantizer import quantize
diff --git a/hpvm/test/epoch_dnn/torch_dnn/quantizer.py b/hpvm/test/epoch_dnn/torch_dnn/quantizer.py
new file mode 100644
index 0000000000..1c9d5d77ea
--- /dev/null
+++ b/hpvm/test/epoch_dnn/torch_dnn/quantizer.py
@@ -0,0 +1,135 @@
+import os
+from copy import deepcopy
+from pathlib import Path
+from typing import Union
+
+import distiller
+import torch
+import yaml
+from distiller.data_loggers import collect_quant_stats
+from distiller.quantization import PostTrainLinearQuantizer
+from torch import nn
+from torch.utils.data import DataLoader
+
+from .datasets import CIFAR
+from .miniera import MiniERA
+
+PathLike = Union[str, Path]
+STATS_FILENAME = "acts_quantization_stats.yaml"
+QUANT_FILENAME = "layer_quant_params.yaml"
+LAYER_HPVM_NAME = {
+ nn.ReLU: "relu",
+ nn.Linear: "gemm",
+ nn.Conv2d: "conv",
+ nn.MaxPool2d: "pool",
+ nn.Softmax: "softmax",
+ nn.Parameter: "add",
+}
+LAYER_DISTILLER_NAME = {
+ nn.Linear: "fcs",
+ nn.Conv2d: "convs",
+ nn.Softmax: "softmax",
+}
+
+
+def quantize(
+ dataset_path: PathLike,
+ model_chkpt: PathLike,
+ strat: str = "NONE",
+ output: PathLike = "calib.txt",
+):
+ # possible quant strats ['NONE', 'AVG', 'N_STD', 'GAUSS', 'LAPLACE']
+ print("Quantizing...")
+ dataset_path = Path(dataset_path)
+ dataset = CIFAR(dataset_path / "input.bin", dataset_path / "labels.bin")
+ dataloader = DataLoader(dataset, batch_size=1)
+
+ # Load Model
+ model = MiniERA()
+ model.load_state_dict(torch.load(model_chkpt))
+
+ # Collect Pre Quantization Stats
+ distiller.utils.assign_layer_fq_names(model)
+
+ if not os.path.isfile(STATS_FILENAME):
+ # generates STATS_FILENAME
+ collect_quant_stats(
+ model, lambda model: evaluate(model, dataloader), save_dir="."
+ )
+
+ # Generate Quantized Scales
+ quantizer = PostTrainLinearQuantizer(
+ deepcopy(model),
+ model_activation_stats=STATS_FILENAME,
+ mode="SYMMETRIC",
+ bits_activations=8,
+ bits_accum=32,
+ clip_acts=strat,
+ )
+ dummy_input = torch.rand(1, 3, 32, 32)
+ quantizer.prepare_model(dummy_input)
+ quantizer.save_per_layer_parameters()
+
+ print("Quantization process finished.")
+ # converts .yaml file stats to hpvm standard
+ generate_calib_file(model, output)
+
+
+def generate_calib_file(model: MiniERA, output: PathLike):
+ print("Generating calibration file...")
+ with open(QUANT_FILENAME, "r") as stream:
+ scale_data = yaml.safe_load(stream)
+
+ lines = []
+ # add scales for input
+ # fmt: off
+ input_min_max = scale_data["convs.0"]["model_activation_stats"]["convs.0"]["inputs"][0]
+ # fmt: on
+ input_scale = max(abs(input_min_max["min"]), abs(input_min_max["max"])) / 127
+ lines.append(f"input:\t{input_scale}\n")
+
+ # because of definition of miniera
+ layers = [*model.convs, *model.fcs, model.softmax]
+ layer_count = {
+ nn.ReLU: 0,
+ nn.Linear: 0,
+ nn.Conv2d: 0,
+ nn.MaxPool2d: 0,
+ nn.Softmax: 0,
+ nn.Parameter: 0,
+ }
+ # add scales for layers
+ for layer in layers:
+ hpvm_name = LAYER_HPVM_NAME[type(layer)]
+ distiller_typename = LAYER_DISTILLER_NAME[type(layer)]
+ layer_idx = layer_count[type(layer)]
+ layer_count[type(layer)] += 1
+
+ scale_key = f"{distiller_typename}.{layer_idx}.output_scale"
+ layer_scale = 1 / scale_data["linear_quant_params"][scale_key]
+ lines.append(f"{hpvm_name}{layer_idx + 1}:\t{layer_scale}\n")
+
+ if isinstance(layer, nn.Conv2d) or isinstance(layer, nn.Linear):
+ # include 'add' scale
+ add_hpvm_name = LAYER_HPVM_NAME[nn.Parameter]
+ add_idx = layer_count[nn.Parameter]
+ layer_count[nn.Parameter] += 1
+ lines.append(f"{add_hpvm_name}{add_idx + 1}:\t{layer_scale}\n")
+
+ with open(output, "w+") as f:
+ f.writelines(lines)
+ print(f"Calibration file generated to {output}")
+
+
+@torch.no_grad()
+def evaluate(model: MiniERA, dataloader: DataLoader):
+ from torch.nn import functional as F
+
+ # Turn on evaluation mode which disables dropout.
+ model.eval()
+ total_loss = 0
+ for batch in dataloader:
+ data, targets = batch
+ output = model(data)
+ total_loss += len(data) * F.cross_entropy(output, targets)
+ return total_loss / len(dataloader)
--
GitLab