--- license: mit library_name: transformers pipeline_tag: image-classification --- # SPG: Sequential Policy Gradient for Adaptive Hyperparameter Optimization This repository contains the models described in the paper [Sequential Policy Gradient for Adaptive Hyperparameter Optimization](https://huggingface.co/papers/2506.15051). [Project page](https://huggingface.co/UniversalAlgorithmic/SPG) [Github repository](https://github.com/SafeAILab/EAGLE) > 🚀 If you're using Jupyter or Colab, you can follow the demo and run it on a single GPU: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/#fileId=https%3A//huggingface.co/UniversalAlgorithmic/SPG/blob/main/demo_nas.ipynb) ## Model Zoo: Adaptive Hyperparameter Optimization (HPO) via SPG Algorithm `Table 1: Performance of pre-trained vs. SPG-retrained models on ImageNet-1K` | Model | SPG | # Params | Acc@1 (%) | Acc@5 (%) | Weights | Command to reproduce | |-------|------|----------|-----------|-----------|---------|----------------------| | MobileNet-V2 | ❌ | 3.5 M | 71.878 | 90.286 |

| Recipe | | MobileNet-V2 | ✅ | 3.5 M | 72.104 | 90.316 |

| [examples/image-classification/run.sh](#retrain-model-on-imagenet-1k) | | ResNet-50 | ❌ | 25.6 M | 76.130 | 92.862 |

| Recipe | | ResNet-50 | ✅ | 25.6 M | 77.234 | 93.322 |

| [examples/image-classification/run.sh](#retrain-model-on-imagenet-1k) | | EfficientNet-V2-M | ❌ | 54.1 M | 85.112 | 97.156 |

| Recipe | | EfficientNet-V2-M | ✅ | 54.1 M | 85.218 | 97.208 |

| [examples/image-classification/run.sh](#retrain-model-on-imagenet-1k) | | ViT-B16 | ❌ | 86.6 M | 81.072 | 95.318 |

| Recipe | | ViT-B16 | ✅ | 86.6 M | 81.092 | 95.304 |

| [examples/image-classification/run.sh](#retrain-model-on-imagenet-1k) | `Table 2: Performance of pre-trained vs. SPG-retrained models. All models are evaluated a subset of COCO val2017, on the 21/20 categories that are present in the Pascal VOC dataset.` > ⚠️ `All model reported on TorchVision (with weight COCO_WITH_VOC_LABELS_V1) were benchmarked using only 20 categories. Researchers should first download the pre-trained model from TorchVision and conduct re-evaluation under the 21-categories (including "background") framework.` | Model | SPG | # Params | mIoU (%) | pixelwise Acc (%) | Weights | Command to reproduce | |---------------------|-----|----------|------------|---------------------|---------|----------------------| | FCN-ResNet50 | ❌ | 35.3 M | 58.9/60.5 | 90.9/91.4 |

| Recipe | | FCN-ResNet50 | ✅ | 35.3 M | 59.4/60.9 | 90.9/91.6 |

| [examples/semantic-segmentation/run.sh](#retrain-model-on-ms-coco-2017) | | FCN-ResNet101 | ❌ | 54.3 M | 62.2/63.7 | 91.1/91.9 |

| Recipe | | FCN-ResNet101 | ✅ | 54.3 M | 62.4/64.3 | 91.1/91.9 |

| [examples/semantic-segmentation/run.sh](#retrain-model-on-ms-coco-2017) | | DeepLabV3-ResNet50 | ❌ | 42.0 M | 63.8/66.4 | 91.5/92.4 |

| Recipe | | DeepLabV3-ResNet50 | ✅ | 42.0 M | 64.2/66.6 | 91.6/92.5 |

| [examples/semantic-segmentation/run.sh](#retrain-model-on-ms-coco-2017) | | DeepLabV3-ResNet101 | ❌ | 61.0 M | 65.3/67.4 | 91.7/92.4 |

| Recipe | | DeepLabV3-ResNet101 | ✅ | 61.0 M | 65.7/67.8 | 91.8/92.5 |

| [examples/semantic-segmentation/run.sh](#retrain-model-on-ms-coco-2017) | `Table 3: Performance comparison of fine-tuned vs. SPG-retrained models across NLP and speech benchmarks.` - GLUE (Text classification: BERT on CoLA, SST-2, MRPC, QQP, QNLI, and RTE task) - SQuAD (Question answering: BERT) - SUPERB (Speech classification: Wav2Vec2 for Audio Classification (AC)) | Task | SPG | Metric Type | Performance (%) | Weights | Command to reproduce | |-------|------|-------------------|-----------------|---------|----------------------| | CoLA | ❌ | Matthews coor | 56.53 |

| [examples/text-classification/run.sh](#transfer-learning-on-glue) | | SST-2 | ❌ | Accuracy | 92.32 |

| Recipe | | SST-2 | ✅ | Accuracy | 92.54 |

| [examples/text-classification/run.sh](#transfer-learning-on-glue) | | QQP | ❌ | F1/Accuracy | 87.49/90.71 |

| Recipe | | QQP | ✅ | F1/Accuracy | 89.72/90.88 |

| [examples/text-classification/run.sh](#transfer-learning-on-glue) | | RTE | ❌ | Accuracy | 65.70 |

| Recipe | | RTE | ✅ | Accuracy | 72.56 |

| [examples/text-classification/run.sh](#transfer-learning-on-glue) | | Q/A* | ❌ | F1/Extra match | 88.52/81.22 |

| Recipe | | Q/A* | ✅ | F1/Extra match | 88.67/81.51 |

| [examples/question-answering/run.sh](#transfer-learning-on-squad) | | AC† | ❌ | Accuracy | 98.26 |

| Recipe | | AC† | ✅ | Accuracy | 98.31 |

| [examples/audio-answering/run.sh](#transfer-learning-on-superb) | ## Model Zoo: Neural Architecture Search (NAS) via SPG Algorithm `Table 4: Performance of pre-trained vs. SPG-retrained models on ImageNet-1K` Depending on the base model, we explore the following architectures: - ResNet-18: ResNet-18, ResNet-27, ResNet-36, ResNet-45 - ResNet-34: ResNet-34, ResNet-40, ResNet-46, ResNet-52 - ResNet-50: ResNet-50, ResNet-53, ResNet-56, ResNet-59 > ⚠️`Our SPG differs from most NAS algorithms, which typically use a gating network for architecture selection. In contrast, we neither employ a gating network nor a proxy network. Instead, after policy optimization, we keep only the base architecture (ResNet-18, ResNet-34, and ResNet-50) and remove all others (ResNet-27/36/45, ResNet-40/46/52, and ResNet-53/56/59).` | Model | SPG | # Params | Acc@1 (%) | Acc@5 (%) | Weights | Command to reproduce | |-------|------|----------|-----------|-----------|---------|----------------------| | ResNet-18 | ❌ | 11.7M | 69.758 | 89.078 |

| Recipe | | ResNet-18 | ✅ | 11.7M | 70.092 | 89.314 |

| [examples/neural-architecture-search/run.sh](#neural-architecture-search-for-resnet-on-imagenet-1k) | | ResNet-34 | ❌ | 21.8M | 73.314 | 91.420 |

| Recipe | | ResNet-34 | ✅ | 21.8M | 73.900 | 93.536 |

| [examples/neural-architecture-search/run.sh](#neural-architecture-search-for-resnet-on-imagenet-1k) | | ResNet-50 | ❌ | 25.6 M | 76.130 | 92.862 |

| Recipe | | ResNet-50 | ✅ | 25.6 M | 77.234 | 93.322 |

| [examples/neural-architecture-search/run.sh](#neural-architecture-search-for-resnet-on-imagenet-1k) | ## Requirements 1. Install `torch>=2.0.0+cu118`. 2. To install other pip packages: ```setup cd examples pip install -r requirements.txt ``` 3. Prepare the [ImageNet](http://image-net.org/) dataset manually and place it in `/path/to/imagenet`. For image classification examples, pass the argument `--data-path=/path/to/imagenet` to the training script. The extracted dataset directory should follow this structure: ```setup /path/to/imagenet/: train/: n01440764: n01440764_18.JPEG ... n01443537: n01443537_2.JPEG ... val/: n01440764: ILSVRC2012_val_00000293.JPEG ... n01443537: ILSVRC2012_val_00000236.JPEG ... ``` 4. Prepare the [MS-COCO 2017](https://cocodataset.org/#home) dataset manually and place it in `/path/to/coco`. For image classification examples, pass the argument `--data-path=/path/to/coco` to the training script. The extracted dataset directory should follow this structure: ```setup /path/to/coco/: annotations: many_json_files.json ... train2017: 000000000009.jpg ... val2017: 000000000139.jpg ... ``` 5. For [🗣️ Keyword Spotting subset](https://huggingface.co/datasets/s3prl/superb#ks), [Common Language](https://huggingface.co/datasets/speechbrain/common_language), [SQuAD](https://huggingface.co/datasets/rajpurkar/squad), [Common Voice](https://huggingface.co/datasets/legacy-datasets/common_voice), [GLUE](https://gluebenchmark.com/) and [WMT](https://huggingface.co/datasets/wmt/wmt17) datasets, manual downloading is not required — they will be automatically loaded via the Hugging Face Datasets library when running our `audio-classification`, `question-answering`, `speech-recognition`, `text-classification`, or `translation` examples. ## Training ### Retrain model on ImageNet-1K We use training recipes similar to those in [PyTorch Vision's classification reference](https://github.com/pytorch/vision/blob/main/references/classification/README.md) to retrain MobileNet-V2, ResNet, EfficientNet-V2, and ViT with our SPG on ImageNet-1K. The following command can be used: ```bash cd ./examples/image-classification # MobileNet-V2 torchrun --nproc_per_node=4 train.py\ --data-path /path/to/imagenet/\ --model mobilenet_v2 --output-dir mobilenet_v2 --weights MobileNet_V2_Weights.IMAGENET1K_V1\ --batch-size 192 --epochs 40 --lr 0.0004 --lr-step-size 10 --lr-gamma 0.5 --wd 0.00004 --apply-trp --trp-depths 1 --trp-p 0.15 --trp-lambdas 0.4 0.2 0.1 # ResNet-50 torchrun --nproc_per_node=4 train.py\ --data-path /path/to/imagenet/\ --model resnet50 --output-dir resnet50 --weights ResNet50_Weights.IMAGENET1K_V1\ --batch-size 64 --epochs 40 --lr 0.0004 --lr-step-size 10 --lr-gamma 0.5 --print-freq 100\ --apply-trp --trp-depths 1 --trp-p 0.2 --trp-lambdas 0.4 0.2 0.1 # EfficientNet-V2 M torchrun --nproc_per_node=4 train.py \ --data-path /path/to/imagenet/\ --model efficientnet_v2_m --output-dir efficientnet_v2_m --weights EfficientNet_V2_M_Weights.IMAGENET1K_V1\ --epochs 10 --batch-size 64 --lr 5e-9 --lr-scheduler cosineannealinglr --weight-decay 0.00002 \ --lr-warmup-method constant --lr-warmup-epochs 8 --lr-warmup-decay 0. \ --auto-augment ta_wide --random-erase 0.1 --label-smoothing 0.1 --mixup-alpha 0.2 --cutmix-alpha 1.0 --norm-weight-decay 0.0 \ --train-crop-size 384 --val-crop-size 480 --val-resize-size 480 --ra-sampler --ra-reps 4 --print-freq 100\ --apply-trp --trp-depths 1 --trp-p 0.2 --trp-lambdas 0.4 0.2 0.1 # ViT-B-16 torchrun --nproc_per_node=4 train.py\ --data-path /path/to/imagenet/\ --model vit_b_16 --output-dir vit_b_16 --weights ViT_B_16_Weights.IMAGENET1K_V1\ --epochs 5 --batch-size 196 --opt adamw --lr 5e-9 --lr-scheduler cosineannealinglr --wd 0.3\ --lr-warmup-method constant --lr-warmup-epochs 3 --lr-warmup-decay 0. \ --amp --label-smoothing 0.11 --mixup-alpha 0.2 --auto-augment ra --clip-grad-norm 1 --cutmix-alpha 1.0\ --apply-trp --trp-depths 1 --trp-p 0.1 --trp-lambdas 0.4 0.2 0.1 --print-freq 100 ``` ### Retrain model on MS-COCO 2017 We use training recipes similar to those in [PyTorch Vision's segmentation reference](https://github.com/pytorch/vision/blob/main/references/segmentation/README.md) to retrain FCN and DeepLab-V3 with our SPG on COCO dataset. The following command can be used: ```bash cd ./examples/semantic-segmentation # FCN-ResNet50 torchrun --nproc_per_node=4 train.py\ --workers 4 --dataset coco --data-path /path/to/coco/\ --model fcn_resnet50 --aux-loss --output-dir fcn_resnet50 --weights FCN_ResNet50