Merge pull request #93 from mjq2020/main

Canonical Code Format

edgelab/engine/apis/mmdet/train.py

@@ -6,7 +6,7 @@
 import torch
 import torch.distributed as dist
 from mmcv.runner import (DistSamplerSeedHook, EpochBasedRunner,
- Fp16OptimizerHook, OptimizerHook, build_runner,builder,
+ Fp16OptimizerHook, OptimizerHook, build_runner,
  get_dist_info)
 
 from mmdet.datasets.utils import replace_ImageToTensor

edgelab/engine/hooks/logger/pavi.py

@@ -3,15 +3,13 @@
 import json
 import warnings
 import os.path as osp
-from functools import partial
 from typing import Optional, Union, Dict
 
 import mmcv
-import torch
 from edgelab.registry import HOOKS
 from mmengine.dist.utils import master_only
 # from mmcv.runner import HOOKS
-from mmengine.dist.utils import master_only
+
 # from mmcv.parallel.scatter_gather import scatter
 # from mmcv.parallel.utils import is_module_wrapper
 
@@ -129,33 +127,33 @@ def _add_ckpt(self, runner, ckpt_path: str, step: int) -> None:
  snapshot_file_path=ckpt_path,
  iteration=step)
 
- def _add_graph(self, runner, step: int) -> None:
- from mmcv.runner.iter_based_runner import IterLoader
- if is_module_wrapper(runner.model):
- _model = runner.model.module
- else:
- _model = runner.model
- device = next(_model.parameters()).device
- # Note that if your sampler indices is generated in init method, your
- # dataset may be one less.
- if isinstance(runner.data_loader, IterLoader):
- data = next(iter(runner.data_loader._dataloader))
- else:
- data = next(iter(runner.data_loader))
- data = scatter(data, [device.index])[0]
- img = data[self.img_key]
- with torch.no_grad():
- origin_forward = _model.forward
- if hasattr(_model, 'forward_dummy'):
- _model.forward = _model.forward_dummy
- if self.dummy_forward_kwargs:
- _model.forward = partial(_model.forward,
- **self.dummy_forward_kwargs)
- self.writer.add_graph(_model,
- img,
- tag=f'{self.run_name}_{step}',
- opset_version=self.opset_version)
- _model.forward = origin_forward
+ # def _add_graph(self, runner, step: int) -> None:
+ #  from mmcv.runner.iter_based_runner import IterLoader
+ #  if is_module_wrapper(runner.model):
+ #  _model = runner.model.module
+ #  else:
+ #  _model = runner.model
+ #  device = next(_model.parameters()).device
+ #  # Note that if your sampler indices is generated in init method, your
+ #  # dataset may be one less.
+ #  if isinstance(runner.data_loader, IterLoader):
+ #  data = next(iter(runner.data_loader._dataloader))
+ #  else:
+ #  data = next(iter(runner.data_loader))
+ #  data = scatter(data, [device.index])[0]
+ #  img = data[self.img_key]
+ #  with torch.no_grad():
+ #  origin_forward = _model.forward
+ #  if hasattr(_model, 'forward_dummy'):
+ #  _model.forward = _model.forward_dummy
+ #  if self.dummy_forward_kwargs:
+ #  _model.forward = partial(_model.forward,
+ #  **self.dummy_forward_kwargs)
+ #  self.writer.add_graph(_model,
+ #  img,
+ #  tag=f'{self.run_name}_{step}',
+ #  opset_version=self.opset_version)
+ #  _model.forward = origin_forward
 
  @master_only
  def log(self, runner) -> None:

edgelab/engine/hooks/logger/tensorboard.py

@@ -1,12 +1,12 @@
-import os.path as osp
+import os
 from typing import Optional, Dict, Union
 
 from edgelab.registry import HOOKS
 from mmengine.dist.utils import master_only
 # from mmcv.runner import HOOKS
 
 # from mmcv.runner.dist_utils import master_only
-# from mmcv.utils import TORCH_VERSION, digit_version
+from mmengine.utils.dl_utils import TORCH_VERSION
 # from mmcv.runner.hooks.logger.text import TextLoggerHook
 
 from .text import TextLoggerHook
@@ -35,8 +35,9 @@ def __init__(self,
  @master_only
  def before_run(self, runner) -> None:
  super().before_run(runner)
- if (TORCH_VERSION == 'parrots'
- or digit_version(TORCH_VERSION) < digit_version('1.1')):
+ if not os.path.exists(self.log_dir):
+ os.makedirs(self.log_dir, exist_ok=True) # type: ignore
+ if TORCH_VERSION == 'parrots':
  try:
  from tensorboardX import SummaryWriter
  except ImportError:
@@ -50,10 +51,7 @@ def before_run(self, runner) -> None:
  'Please run "pip install future tensorboard" to install '
  'the dependencies to use torch.utils.tensorboard '
  '(applicable to PyTorch 1.1 or higher)')
-
- if self.log_dir is None:
- self.log_dir = osp.join(runner.work_dir, 'tf_logs')
- self.writer = SummaryWriter(self.log_dir)
+ self._tensorboard = SummaryWriter(self.log_dir)
 
  @master_only
  def log(self, runner) -> None:

edgelab/engine/hooks/logger/text.py

@@ -1,8 +1,7 @@
 import logging
 import datetime
 from pathlib import Path
-from collections import OrderedDict
-from typing import Optional, Union, Dict
+from typing import Optional, Union
 
 import torch
 import torch.distributed as dist
@@ -154,7 +153,8 @@ def setloglevel(self,
  runner: Runner,
  handler: logging.Handler = logging.StreamHandler,
  level: int = logging.ERROR):
- if handler in self.handltype: return
+ if handler in self.handltype:
+ return
  for i, hand in enumerate(runner.logger.handlers):
  if type(hand) is handler:
  hand.setLevel(level)

edgelab/engine/hooks/visualization_hook.py

@@ -85,7 +85,6 @@ def after_test_iter(self, runner: Runner, batch_idx: int, data_batch: dict,
  ])
  out_file = f'{out_file_name}_{index}.{postfix}'
  out_file = os.path.join(self.out_dir, out_file)
- from mmpose.visualization import PoseLocalVisualizer
  self._visualizer.add_datasample(
  os.path.basename(img_path) if self.show else 'test_img',
  img,

edgelab/engine/runner/loops.py

@@ -1,12 +1,13 @@
 from typing import Dict, Sequence, Union, List
 
+import onnx
+import torch
+from torch.utils.data import DataLoader
 from mmengine.runner import Runner
 from mmengine.evaluator.evaluator import Evaluator
 from mmengine.runner.loops import EpochBasedTrainLoop, BaseLoop
-from torch.utils.data import DataLoader
 
 from edgelab.registry import LOOPS
-from mmengine.registry import RUNNERS
 
 
 @LOOPS.register_module()
@@ -56,29 +57,31 @@ def __init__(
  if isinstance(model, list):
  try:
  import ncnn
- except:
+ except ImportError:
  raise ImportError(
  'You have not installed ncnn yet, please execute the "pip install ncnn" command to install and run again'
  )
  net = ncnn.Net()
  for p in model:
- if p.endswith('param'): param = p
- if p.endswith('bin'): bin = p
+ if p.endswith('param'): 
+ param = p
+ if p.endswith('bin'): 
+ bin = p
  net.load_param(param)
  net.load_model(bin)
  # net.opt.use_vulkan_compute = True
  self.engine = 'ncnn'
  elif model.endswith('onnx'):
  try:
  import onnxruntime
- except:
+ except ImportError:
  raise ImportError(
  'You have not installed onnxruntime yet, please execute the "pip install onnxruntime" command to install and run again'
  )
  try:
  net = onnx.load(model)
  onnx.checker.check_model(net)
- except:
+ except ValueError:
  raise ValueError(
  'onnx file have error,please check your onnx export code!')
  providers = [
@@ -89,7 +92,7 @@ def __init__(
  elif model.endswith('tflite'):
  try:
  import tensorflow as tf
- except:
+ except ImportError:
  raise ImportError(
  'You have not installed tensorflow yet, please execute the "pip install tensorflow" command to install and run again'
  )

edgelab/models/__init__.py

@@ -1,13 +1,6 @@
-from .backbones import *
-from .detectors import *
-from .classifiers import *
-from .heads import *
-from .losses import *
-from .necks import *
-
-__all__ = [
- 'SoundNetRaw', 'Speechcommand', 'PFLD', 'Audio_head', 'Audio_classify',
- 'LabelSmoothCrossEntropyLoss', 'PFLDLoss', 'PFLDhead', 'FastestDet', 'SPP',
- 'NLLLoss', 'BCEWithLogitsLoss', 'Fomo_Head', 'CustomShuffleNetV2',
- 'FomoLoss','Fomo', 'AxesNet', 'AccelerometerClassifier','FPN','MobileNetv2'
-]
+from .backbones import * # noqa
+from .detectors import * # noqa
+from .classifiers import * # noqa
+from .heads import * # noqa
+from .losses import * # noqa
+from .necks import * # noqa

edgelab/models/backbones/AxesNet.py

@@ -1,6 +1,6 @@
-import torch
 import torch.nn as nn
 import torch.nn.functional as F
+
 from edgelab.registry import MODELS
 
 

edgelab/models/backbones/EfficientNet.py

@@ -1,14 +1,14 @@
 from typing import Optional, List, Callable
 from functools import partial
 import copy
-import math
 from torch import nn
 import torch
 from torchvision.ops import StochasticDepth
 from edgelab.registry import BACKBONES
 from mmdet.models.utils.make_divisible import make_divisible
 from mmengine.model import BaseModule
 from edgelab.models.base.general import ConvNormActivation, SqueezeExcitation
+from edgelab.models.layers.rep import RepConv1x1
 
 
 class MBConvConfig:
@@ -36,12 +36,12 @@ def adjust_depth(num_layers: int, depth_mult: float):
 
 class MBConv(nn.Module):
 
- def __init__(
- self,
- cnf: MBConvConfig,
- stochastic_depth_prob: float,
- norm_layer: Callable[..., nn.Module],
- se_layer: Callable[..., nn.Module] = SqueezeExcitation) -> None:
+ def __init__(self,
+  cnf: MBConvConfig,
+  stochastic_depth_prob: float,
+  norm_layer: Callable[..., nn.Module],
+  se_layer: Callable[..., nn.Module] = SqueezeExcitation,
+  rep: bool = False) -> None:
  super().__init__()
 
  if not (1 <= cnf.stride <= 2):
@@ -50,7 +50,7 @@ def __init__(
  self.use_res_connect = cnf.stride == 1 and cnf.input_channels == cnf.out_channels
 
  layers: List[nn.Module] = []
- activation_layer = nn.SiLU
+ activation_layer = nn.ReLU
 
  # expand
  expanded_channels = cnf.adjust_channels(cnf.input_channels,
@@ -64,21 +64,29 @@ def __init__(
  activation_layer=activation_layer))
 
  # depthwise
- layers.append(
- ConvNormActivation(expanded_channels,
- expanded_channels,
- kernel_size=cnf.kernel,
- stride=cnf.stride,
- groups=expanded_channels,
- norm_layer=norm_layer,
- activation_layer=activation_layer))
+ if rep:
+ layers.append(
+ RepConv1x1(expanded_channels,
+ expanded_channels,
+ stride=cnf.stride,
+ act_cfg=activation_layer))
 
- # squeeze and excitation
- squeeze_channels = max(1, cnf.input_channels // 4)
- layers.append(
- se_layer(expanded_channels,
- squeeze_channels,
- activation=partial(nn.SiLU, inplace=True)))
+ else:
+ layers.append(
+ ConvNormActivation(expanded_channels,
+ expanded_channels,
+ kernel_size=cnf.kernel,
+ stride=cnf.stride,
+ groups=expanded_channels,
+ norm_layer=norm_layer,
+ activation_layer=activation_layer))
+
+ # squeeze and excitation
+ squeeze_channels = max(1, cnf.input_channels // 4)
+ layers.append(
+ se_layer(expanded_channels,
+ squeeze_channels,
+ activation=partial(nn.ReLU, inplace=True)))
 
  # project
  layers.append(
@@ -114,7 +122,8 @@ class EfficientNet(BaseModule):
  ]
 
  width_depth_mult = {
- 'b0': [0.15, 0.35, 0.2],
+ 'bt': [0.35, 0.35, 0.2],
+ 'b0': [1.0, 1.0, 0.2],
  'b1': [1.0, 1.1, 0.2],
  'b2': [1.1, 1.2, 0.3],
  'b3': [1.2, 1.4, 0.3],
@@ -131,6 +140,7 @@ def __init__(self,
  norm_cfg='BN',
  frozen_stages=-1,
  norm_eval=False,
+ rep=False,
  init_cfg: Optional[dict] = None):
  super().__init__(init_cfg)
 
@@ -156,7 +166,7 @@ def __init__(self,
  3,
  2,
  norm_layer=norm_cfg,
- activation_layer='SiLU')
+ activation_layer='ReLU')
 
  total_stage_blocks = sum([cnf.num_layers for cnf in arch_param])
  stage_block_id = 0
@@ -170,7 +180,8 @@ def __init__(self,
  conf.stride = 1
  sd_prob = stochastic_depth_prob * float(
  stage_block_id) / total_stage_blocks
- layer.append(MBConv(conf, sd_prob, norm_layer=norm_cfg))
+ layer.append(
+ MBConv(conf, sd_prob, norm_layer=norm_cfg, rep=rep))
  stage_block_id += 1
 
  self.add_module(name, nn.Sequential(*layer))

edgelab/models/backbones/MobileNetv2.py

@@ -5,7 +5,7 @@
 from mmengine.model import BaseModule
 from edgelab.registry import BACKBONES, MODELS
 from torchvision.models._utils import _make_divisible
-from edgelab.models.layers.rep import RepBlock
+from edgelab.models.layers.rep import RepBlock,RepConv1x1
 from ..base.general import InvertedResidual, ConvNormActivation
 
 
@@ -27,7 +27,8 @@ def __init__(self,
  if block is None and not rep:
  block = InvertedResidual
  elif rep:
- block = RepBlock
+ # block = RepBlock
+ block = RepConv1x1
  elif isinstance(block, dict):
  block = MODELS.build(rep)
 
@@ -76,6 +77,9 @@ def __init__(self,
  stride=stride,
  groups=in_channels,
  norm_layer=norm_layer)
+
+ elif block is RepConv1x1:
+ layer=block(in_channels,out_channels,stride=stride,depth=6)
  else:
  layer = block(in_channels,
  out_channels,