[Feature] Support efficientnet in mmdetection. (open-mmlab#7514)

* Initial implementation * Add missing import * Add MemoryEfficientSwishImplementation. Add docstrings * Add efficientnet2mmdet tool * Add config folder * Flake8 * Flake8 * Flake8 * Fix config * Requested changes * docformatter * Update train config from https://github.com/google/automl/blob/master/efficientdet * Run pre-commit * Fix schedule * Set by_epoch=False in scheduler * Train 80 epochs * Remove duplicated arg * Update README.md * efficient3 efficient0 * efficientNet imports * efficientNet * config edit path for eff3 and dropout for eff0 * efficientnet review2 * fix model_converter location and drop path * fix model converter and efficientnet import * register memoryefficietnswish * eff0, eff3 * fix flake8 yapf isort * same padding in tensorflow and edit drop path rate * fix init of utils * Align mmdet utils with mmcls * Align mmdet.models.utils with mmcls * Use mmcls efficientnet backbone * Update * Update * Update metafile Co-authored-by: David de la Iglesia Castro <[email protected]> Co-authored-by: David de la Iglesia Castro <[email protected]> Co-authored-by: jiangyitong <[email protected]> Co-authored-by: jiangyitong <[email protected]>
SakiRinn · Mar 17, 2023 · 52a5d6e · 52a5d6e
1 parent a709183
commit 52a5d6e
Show file tree

Hide file tree

Showing 8 changed files with 595 additions and 2 deletions.
diff --git a/configs/efficientnet/README.md b/configs/efficientnet/README.md
@@ -0,0 +1,30 @@
+# EfficientNet
+
+> [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946v5)
+
+<!-- [BACKBONE] -->
+
+## Introduction
+
+Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performance. Based on this observation, we propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a simple yet highly effective compound coefficient. We demonstrate the effectiveness of this method on scaling up MobileNets and ResNet.
+
+To go even further, we use neural architecture search to design a new baseline network and scale it up to obtain a family of models, called EfficientNets, which achieve much better accuracy and efficiency than previous ConvNets. In particular, our EfficientNet-B7 achieves state-of-the-art 84.3% top-1 accuracy on ImageNet, while being 8.4x smaller and 6.1x faster on inference than the best existing ConvNet. Our EfficientNets also transfer well and achieve state-of-the-art accuracy on CIFAR-100 (91.7%), Flowers (98.8%), and 3 other transfer learning datasets, with an order of magnitude fewer parameters.
+
+## Results and Models
+
+### RetinaNet
+
+|    Backbone     |  Style  | Lr schd | Mem (GB) | Inf time (fps) | box AP | Config | Download |
+| :-------------: | :-----: | :-----: | :------: | :------------: | :----: | :------: | :--------: |
+|Efficientnet-b3  | pytorch |   1x    |   -      |   -           |  40.5  |[config](https://github.com/open-mmlab/mmdetection/tree/master/configs/efficientnet/retinanet_effb3_fpn_crop896_8x4_1x_coco.py) | [model]() &#124; [log]() |
+
+## Citation
+
+```latex
+@article{tan2019efficientnet,
+  title={Efficientnet: Rethinking model scaling for convolutional neural networks},
+  author={Tan, Mingxing and Le, Quoc V},
+  journal={arXiv preprint arXiv:1905.11946},
+  year={2019}
+}
+```
diff --git a/configs/efficientnet/metafile.yml b/configs/efficientnet/metafile.yml
@@ -0,0 +1,19 @@
+Models:
+  - Name: retinanet_effb3_fpn_crop896_8x4_1x_coco
+    In Collection: RetinaNet
+    Config: configs/efficientnet/retinanet_effb3_fpn_crop896_8x4_1x_coco.py
+    Metadata:
+      Epochs: 12
+    Results:
+      - Task: Object Detection
+        Dataset: COCO
+        Metrics:
+          box AP: 40.5
+    Weights: url
+    Paper:
+      URL: https://arxiv.org/abs/1905.11946v5
+      Title: 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks'
+    README: configs/efficientnet/README.md
+    Code:
+      URL: https://github.com/open-mmlab/mmdetection/blob/v2.23.0/mmdet/models/backbones/efficientnet.py#L159
+      Version: v2.23.0
diff --git a/configs/efficientnet/retinanet_effb3_fpn_crop896_8x4_1x_coco.py b/configs/efficientnet/retinanet_effb3_fpn_crop896_8x4_1x_coco.py
@@ -0,0 +1,93 @@
+_base_ = [
+    '../_base_/models/retinanet_r50_fpn.py',
+    '../_base_/datasets/coco_detection.py', '../_base_/default_runtime.py'
+]
+
+cudnn_benchmark = True
+norm_cfg = dict(type='BN', requires_grad=True)
+checkpoint = 'https://download.openmmlab.com/mmclassification/v0/efficientnet/efficientnet-b3_3rdparty_8xb32-aa_in1k_20220119-5b4887a0.pth'  # noqa
+model = dict(
+    backbone=dict(
+        _delete_=True,
+        type='EfficientNet',
+        arch='b3',
+        drop_path_rate=0.2,
+        out_indices=(3, 4, 5),
+        frozen_stages=0,
+        norm_cfg=dict(
+            type='SyncBN', requires_grad=True, eps=1e-3, momentum=0.01),
+        norm_eval=False,
+        init_cfg=dict(
+            type='Pretrained', prefix='backbone', checkpoint=checkpoint)),
+    neck=dict(
+        in_channels=[48, 136, 384],
+        start_level=0,
+        out_channels=256,
+        relu_before_extra_convs=True,
+        no_norm_on_lateral=True,
+        norm_cfg=norm_cfg),
+    bbox_head=dict(type='RetinaSepBNHead', num_ins=5, norm_cfg=norm_cfg),
+    # training and testing settings
+    train_cfg=dict(assigner=dict(neg_iou_thr=0.5)))
+
+# dataset settings
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_size = (896, 896)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(
+        type='Resize',
+        img_scale=img_size,
+        ratio_range=(0.8, 1.2),
+        keep_ratio=True),
+    dict(type='RandomCrop', crop_size=img_size),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=img_size),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=img_size,
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size=img_size),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(pipeline=train_pipeline),
+    val=dict(pipeline=test_pipeline),
+    test=dict(pipeline=test_pipeline))
+# optimizer
+optimizer_config = dict(grad_clip=None)
+optimizer = dict(
+    type='SGD',
+    lr=0.04,
+    momentum=0.9,
+    weight_decay=0.0001,
+    paramwise_cfg=dict(norm_decay_mult=0, bypass_duplicate=True))
+# learning policy
+lr_config = dict(
+    policy='step',
+    warmup='linear',
+    warmup_iters=1000,
+    warmup_ratio=0.1,
+    step=[8, 11])
+# runtime settings
+runner = dict(type='EpochBasedRunner', max_epochs=12)
+
+# NOTE: This variable is for automatically scaling LR,
+# USER SHOULD NOT CHANGE THIS VALUE.
+default_batch_size = 32  # (8 GPUs) x (4 samples per GPU)
diff --git a/mmdet/models/backbones/__init__.py b/mmdet/models/backbones/__init__.py
@@ -3,6 +3,7 @@
 from .darknet import Darknet
 from .detectors_resnet import DetectoRS_ResNet
 from .detectors_resnext import DetectoRS_ResNeXt
+from .efficientnet import EfficientNet
 from .hourglass import HourglassNet
 from .hrnet import HRNet
 from .mobilenet_v2 import MobileNetV2
@@ -20,5 +21,6 @@
     'RegNet', 'ResNet', 'ResNetV1d', 'ResNeXt', 'SSDVGG', 'HRNet',
     'MobileNetV2', 'Res2Net', 'HourglassNet', 'DetectoRS_ResNet',
     'DetectoRS_ResNeXt', 'Darknet', 'ResNeSt', 'TridentResNet', 'CSPDarknet',
-    'SwinTransformer', 'PyramidVisionTransformer', 'PyramidVisionTransformerV2'
+    'SwinTransformer', 'PyramidVisionTransformer',
+    'PyramidVisionTransformerV2', 'EfficientNet'
 ]