diff options
| -rw-r--r-- | .gitignore | 1 | ||||
| -rw-r--r-- | python-keras-efficientnet-v2.spec | 775 | ||||
| -rw-r--r-- | sources | 1 |
3 files changed, 777 insertions, 0 deletions
@@ -0,0 +1 @@ +/keras-efficientnet-v2-1.2.2.tar.gz diff --git a/python-keras-efficientnet-v2.spec b/python-keras-efficientnet-v2.spec new file mode 100644 index 0000000..965ca58 --- /dev/null +++ b/python-keras-efficientnet-v2.spec @@ -0,0 +1,775 @@ +%global _empty_manifest_terminate_build 0 +Name: python-keras-efficientnet-v2 +Version: 1.2.2 +Release: 1 +Summary: (Unofficial) Tensorflow keras efficientnet v2 with pre-trained +License: Apache 2.0 +URL: https://github.com/leondgarse/keras_efficientnet_v2 +Source0: https://mirrors.nju.edu.cn/pypi/web/packages/b7/b4/c00aca7ee9a0ed592f3a34915a843b29d3b02142f6f33ee1b83d5eea5728/keras-efficientnet-v2-1.2.2.tar.gz +BuildArch: noarch + +Requires: python3-tensorflow + +%description +# ___Keras EfficientNetV2___ +*** +- **As `EfficientNetV2` is included in `keras.application` now, merged this project into [Github leondgarse/keras_cv_attention_models/efficientnet](https://github.com/leondgarse/keras_cv_attention_models/tree/main/keras_cv_attention_models/efficientnet). Package `keras-efficientnet-v2` moved into `stable` status**. + +## Table of Contents +<!-- TOC depthFrom:1 depthTo:6 withLinks:1 updateOnSave:1 orderedList:0 --> + +- [___Keras EfficientNetV2___](#keras-efficientnetv2) + - [Table of Contents](#table-of-contents) + - [Summary](#summary) + - [Usage](#usage) + - [Training detail from article](#training-detail-from-article) + - [Detailed conversion procedure](#detailed-conversion-procedure) + - [Progressive train test on cifar10](#progressive-train-test-on-cifar10) + - [Related Projects](#related-projects) + +<!-- /TOC --> +*** + +## Summary + - My own keras implementation of [Official efficientnetv2](https://github.com/google/automl/tree/master/efficientnetv2). Article [arXiv 2104.00298 EfficientNetV2: Smaller Models and Faster Training](https://arxiv.org/abs/2104.00298) by Mingxing Tan, Quoc V. Le. + - `h5` model weights converted from official publication. + - `effv2-t-imagenet.h5` model weights converted from [Github rwightman/pytorch-image-models](https://github.com/rwightman/pytorch-image-models#july-5-9-2021). which claimed both faster and better accuracy than `b3`. Please notice that `PyTorch` using different `bn_epsilon` and `padding` strategy. + + | V2 Model | Params | Top1 | Input | ImageNet21K | Imagenet21k-ft1k | Imagenet | + | ----------- | ------ | ---- | ----- | ----------- | ---------------- | -------- | + | EffV2B0 | 7.1M | 78.7 | 224 | [v2b0-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-21k.h5)|[v2b0-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-21k-ft1k.h5)|[v2b0-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-imagenet.h5)| + | EffV2B1 | 8.1M | 79.8 | 240 | [v2b1-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-21k.h5)|[v2b1-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-21k-ft1k.h5)|[v2b1-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-imagenet.h5)| + | EffV2B2 | 10.1M | 80.5 | 260 | [v2b2-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-21k.h5)|[v2b2-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-21k-ft1k.h5)|[v2b2-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-imagenet.h5)| + | EffV2B3 | 14.4M | 82.1 | 300 | [v2b3-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-21k.h5)|[v2b3-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-21k-ft1k.h5)|[v2b3-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-imagenet.h5)| + | EffV2T | 13.6M | 82.5 | 320 | | |[v2t-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-t-imagenet.h5)| + | EffV2S | 21.5M | 84.9 | 384 | [v2s-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-21k.h5) |[v2s-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-21k-ft1k.h5)|[v2s-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-imagenet.h5)| + | EffV2M | 54.1M | 86.2 | 480 | [v2m-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-21k.h5) |[v2m-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-21k-ft1k.h5)|[v2m-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-imagenet.h5)| + | EffV2L | 119.5M| 86.9 | 480 | [v2l-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-21k.h5) |[v2l-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-21k-ft1k.h5)|[v2l-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-imagenet.h5)| + | EffV2XL | 206.8M| 87.2 | 512 | [v2xl-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-xl-21k.h5)|[v2xl-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-xl-21k-ft1k.h5)| | + + - **EfficientNetV1 noisy_student models** from [Github tensorflow/tpu/efficientnet](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet). Paper [PDF 1911.04252 Self-training with Noisy Student improves ImageNet classification](https://arxiv.org/pdf/1911.04252.pdf). Parameter `include_preprocessing=False` is added, and the default `False` value makes expecting input value in range `[-1, 1]`, same with `EfficientNetV2`. Default `pretrained` is `noisy_student`. + + | V1 Model | Params | Top1 | Input | noisy_student | Imagenet | + | -------------- | ------- | ---- | ----- | ------------- | -------- | + | EffV1B0 | 5.3M | 78.8 | 224 | [v1-b0-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b0-noisy_student.h5) | [v1-b0-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b0-imagenet.h5) | + | EffV1B1 | 7.8M | 81.5 | 240 | [v1-b1-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b1-noisy_student.h5) | [v1-b1-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b1-imagenet.h5) | + | EffV1B2 | 9.1M | 82.4 | 260 | [v1-b2-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b2-noisy_student.h5) | [v1-b2-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b2-imagenet.h5) | + | EffV1B3 | 12.2M | 84.1 | 300 | [v1-b3-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b3-noisy_student.h5) | [v1-b3-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b3-imagenet.h5) | + | EffV1B4 | 19.3M | 85.3 | 380 | [v1-b4-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b4-noisy_student.h5) | [v1-b4-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b4-imagenet.h5) | + | EffV1B5 | 30.4M | 86.1 | 456 | [v1-b5-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b5-noisy_student.h5) | [v1-b5-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b5-imagenet.h5) | + | EffV1B6 | 43.0M | 86.4 | 528 | [v1-b6-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b6-noisy_student.h5) | [v1-b6-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b6-imagenet.h5) | + | EffV1B7 | 66.3M | 86.9 | 600 | [v1-b7-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b7-noisy_student.h5) | [v1-b7-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b7-imagenet.h5) | + | EffV1L2 | 480.3M | 88.4 | 800 | [v1-l2-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-l2-noisy_student.h5) | | + + - **Self tested imagenet accuracy** + - `rescale_mode` `torch` means `(image - [0.485, 0.456, 0.406]) / [[0.229, 0.224, 0.225]]`, `tf` means `(image - 0.5) / 0.5` + - All `resize_method` is `bicubic`. + - Some testing detail is not clear, so not exactly matching official reported results. + - Testing Detail is [EfficientNetV2 self tested imagenet accuracy](https://github.com/leondgarse/keras_cv_attention_models/discussions/19). + + | model | input | rescale_mode | central_crop | top 1 | top 5 | Reported top1 | + | ------------ | ----- | ------------ | ------------ | ------- | ------- | ----------------- | + | EffV2B0 | 224 | torch | 0.875 | 0.78748 | 0.94386 | 0.787 | + | EffV2B1 | 240 | torch | 0.95 | 0.7987 | 0.94936 | 0.798 | + | EffV2B2 | 260 | torch | 0.95 | 0.80642 | 0.95262 | 0.805 | + | EffV2B3 | 300 | torch | 0.95 | 0.82098 | 0.95896 | 0.821 | + | EffV2T | 320 | torch | 0.99 | 0.82506 | 0.96228 | 0.823 (input 288) | + | EffV2S | 384 | tf | 0.99 | 0.8386 | 0.967 | 0.839 | + | EffV2M | 480 | tf | 0.99 | 0.8509 | 0.973 | 0.852 | + | EffV2L | 480 | tf | 0.99 | 0.855 | 0.97324 | 0.857 | + | EffV2S ft1k | 384 | tf | 0.99 | 0.84328 | 0.97254 | 0.849 | + | EffV2M ft1k | 480 | tf | 0.99 | 0.85606 | 0.9775 | 0.862 | + | EffV2L ft1k | 480 | tf | 0.99 | 0.86294 | 0.9799 | 0.869 | + | EffV2XL ft1k | 512 | tf | 0.99 | 0.86532 | 0.97866 | 0.872 | +## Usage + - This repo can be installed as a pip package, or just `git clone` it. + ```py + pip install -U keras-efficientnet-v2 + # Or + pip install -U git+https://github.com/leondgarse/keras_efficientnet_v2 + ``` + - **Define model and load pretrained weights** Parameter `pretrained` is added in value `[None, "imagenet", "imagenet21k", "imagenet21k-ft1k"]`, default is `imagenet`. Model input value should be in range `[-1, 1]`. + ```py + # Will download and load `imagenet` pretrained weights. + # Model weight is loaded with `by_name=True, skip_mismatch=True`. + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2S(pretrained="imagenet") + + # Run prediction + import tensorflow as tf + from tensorflow import keras + from skimage.data import chelsea + imm = tf.image.resize(chelsea(), model.input_shape[1:3]) # Chelsea the cat + pred = model(tf.expand_dims(imm / 128. - 1., 0)).numpy() + print(keras.applications.imagenet_utils.decode_predictions(pred)[0]) + # [('n02124075', 'Egyptian_cat', 0.8642886), ('n02123159', 'tiger_cat', 0.030793495), ...] + ``` + Or download `h5` model and load directly + ```py + mm = keras.models.load_model('efficientnetv2-b3-21k-ft1k.h5') + ``` + For `"imagenet21k"` pre-trained model, actual `num_classes` is `21843`. + - **Exclude model top layers** by set `num_classes=0`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2B0(dropout=1e-6, num_classes=0, pretrained="imagenet21k") + print(model.output_shape) + # (None, 7, 7, 1280) + + model.save('efficientnetv2-b0-21k-notop.h5') + ``` + - **Use dynamic input resolution** by set `input_shape=(None, None, 3)`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2M(input_shape=(None, None, 3), drop_connect_rate=0.2, num_classes=0, pretrained="imagenet21k-ft1k") + + print(model(np.ones([1, 224, 224, 3])).shape) + # (1, 7, 7, 1280) + print(model(np.ones([1, 512, 512, 3])).shape) + # (1, 16, 16, 1280) + ``` + - **`include_preprocessing`** set `True` will add pre-processing `Rescale + Normalization` after `Input`. Means using input value in range `[0, 255]`. Default value `False` means in range `[-1, 1]`. Works both for `EfficientNetV2` and `EfficientNetV1`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV1B4(pretrained="noisy_student", include_preprocessing=True) + + from skimage.data import chelsea + imm = tf.image.resize(chelsea(), model.input_shape[1:3]) # Chelsea the cat + pred = model(tf.expand_dims(imm, 0)).numpy() # value in range [0, 255] + print(keras.applications.imagenet_utils.decode_predictions(pred)[0]) + # [('n02124075', 'Egyptian_cat', 0.68414235), ('n02123159', 'tiger_cat', 0.04486668), ...] + ``` +## Training detail from article + - **Training configures**, `Eval size` is used as the default `input_shape` for each model type. + + | Model | Train size | Eval size | Dropout | Randaug | Mixup | + | ------- | ---------- | --------- | ------- | ------- | ----- | + | EffV2B0 | 192 | 224 | 0.2 | 0 | 0 | + | EffV2B1 | 192 | 240 | 0.2 | 0 | 0 | + | EffV2B2 | 208 | 260 | 0.3 | 0 | 0 | + | EffV2B3 | 240 | 300 | 0.3 | 0 | 0 | + | EffV2S | 300 | 384 | 0.2 | 10 | 0 | + | EffV2M | 384 | 480 | 0.3 | 15 | 0.2 | + | EffV2L | 384 | 480 | 0.4 | 20 | 0.5 | + | EffV2XL | 384 | 512 | 0.4 | 20 | 0.5 | + + - EfficientNetV2-S architecture + + | Stage | Operator | Stride | #Channels | #Layers | + | ----- | ---------------------- | ------ | --------- | ------- | + | 0 | Conv3x3 | 2 | 24 | 1 | + | 1 | Fused-MBConv1, k3x3 | 1 | 24 | 2 | + | 2 | Fused-MBConv4, k3x3 | 2 | 48 | 4 | + | 3 | Fused-MBConv4, k3x3 | 2 | 64 | 4 | + | 4 | MBConv4, k3x3, SE0.25 | 2 | 128 | 6 | + | 5 | MBConv6, k3x3, SE0.25 | 1 | 160 | 9 | + | 6 | MBConv6, k3x3, SE0.25 | 2 | 256 | 15 | + | 7 | Conv1x1 & Pooling & FC | - | 1280 | 1 | + + - Progressive training settings for EfficientNetV2 + + | | S min | S max | M min | M max | L min | M max | + | ------------ | ----- | ----- | ----- | ----- | ----- | ----- | + | Image Size | 128 | 300 | 128 | 380 | 128 | 380 | + | RandAugment | 5 | 15 | 5 | 20 | 5 | 25 | + | Mixup alpha | 0 | 0 | 0 | 0.2 | 0 | 0.4 | + | Dropout rate | 0.1 | 0.3 | 0.1 | 0.4 | 0.1 | 0.5 | + + - Imagenet training detail + - RMSProp optimizer with decay 0.9 and momentum 0.9 + - batch norm momentum 0.99; weight decay 1e-5 + - Each model is trained for 350 epochs with total batch size 4096 + - Learning rate is first warmed up from 0 to 0.256, and then decayed by 0.97 every 2.4 epochs + - We use exponential moving average with 0.9999 decay rate + - RandAugment (Cubuk et al., 2020) + - Mixup (Zhang et al., 2018) + - Dropout (Srivastava et al., 2014) + - and stochastic depth (Huang et al., 2016) with 0.8 survival probability +## Detailed conversion procedure + - [convert_effnetv2_model.py](convert_effnetv2_model.py) is a modified version of [the orignal effnetv2_model.py](https://github.com/google/automl/blob/master/efficientnetv2/effnetv2_model.py). Check detail by `vimdiff convert_effnetv2_model.py ../automl/efficientnetv2/effnetv2_model.py` + - Delete some `names`, as they may cause confliction in keras. + - Use `.call` directly calling `se` modules and other blocks, so they will not be `blocks` in `model.summary()` + - Just use `Add` layer instead of `utils.drop_connect`, as when `is_training=False`, `utils.drop_connect` functions like `Add`. + - Add a `num_classes` parameter outside of `mconfig`. + - Add `__main__` part, which makes this can be run as a script. Refer to it for converting detail. + - Depends on official repo + ```sh + ../ + ├── automl # Official repo + ├── keras_efficientnet_v2 # This one + ``` + - **Procedure** + ```py + # See help info + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -h + + # Convert by specific model_type and dataset type + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m b0 -d imagenet21k + + # Convert by specific model_type and all its datasets ['imagenet', 'imagenet21k', 'imagenetft'] + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m s -d all + + # Convert all model_types and and all datasets + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m all -d all + ``` +## Progressive train test on cifar10 + - [Colab efficientnetV2_basic_test.ipynb](https://colab.research.google.com/drive/1vmAEfF9tUgK2gkrS5qVftadTyUcX343D?usp=sharing) + ```py + import keras_efficientnet_v2 + from tensorflow import keras + from keras_efficientnet_v2 import progressive_train_test + + model = keras_efficientnet_v2.EfficientNetV2S(input_shape=(None, None, 3), num_classes=10, classifier_activation='softmax', dropout=0.1) + model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"]) + + hhs = progressive_train_test.progressive_with_dropout_randaug( + model, + data_name="cifar10", + lr_scheduler=None, + total_epochs=36, + batch_size=64, + dropout_layer=-2, + target_shapes=[128, 160, 192, 224], # [128, 185, 242, 300] for final shape (300, 300) + dropouts=[0.1, 0.2, 0.3, 0.4], + magnitudes=[5, 8, 12, 15], + ) + + with open("history_ev2s_imagenet_progressive_224.json", "w") as ff: + json.dump(hhs, ff) + ``` +  +## Related Projects + - [tfa.layers.StochasticDepth](https://www.tensorflow.org/addons/api_docs/python/tfa/layers/StochasticDepth) + - [Official efficientnetv2](https://github.com/google/automl/tree/master/efficientnetv2) + - [Github rwightman/pytorch-image-models](https://github.com/rwightman/pytorch-image-models#july-5-9-2021) +*** + + + + +%package -n python3-keras-efficientnet-v2 +Summary: (Unofficial) Tensorflow keras efficientnet v2 with pre-trained +Provides: python-keras-efficientnet-v2 +BuildRequires: python3-devel +BuildRequires: python3-setuptools +BuildRequires: python3-pip +%description -n python3-keras-efficientnet-v2 +# ___Keras EfficientNetV2___ +*** +- **As `EfficientNetV2` is included in `keras.application` now, merged this project into [Github leondgarse/keras_cv_attention_models/efficientnet](https://github.com/leondgarse/keras_cv_attention_models/tree/main/keras_cv_attention_models/efficientnet). Package `keras-efficientnet-v2` moved into `stable` status**. + +## Table of Contents +<!-- TOC depthFrom:1 depthTo:6 withLinks:1 updateOnSave:1 orderedList:0 --> + +- [___Keras EfficientNetV2___](#keras-efficientnetv2) + - [Table of Contents](#table-of-contents) + - [Summary](#summary) + - [Usage](#usage) + - [Training detail from article](#training-detail-from-article) + - [Detailed conversion procedure](#detailed-conversion-procedure) + - [Progressive train test on cifar10](#progressive-train-test-on-cifar10) + - [Related Projects](#related-projects) + +<!-- /TOC --> +*** + +## Summary + - My own keras implementation of [Official efficientnetv2](https://github.com/google/automl/tree/master/efficientnetv2). Article [arXiv 2104.00298 EfficientNetV2: Smaller Models and Faster Training](https://arxiv.org/abs/2104.00298) by Mingxing Tan, Quoc V. Le. + - `h5` model weights converted from official publication. + - `effv2-t-imagenet.h5` model weights converted from [Github rwightman/pytorch-image-models](https://github.com/rwightman/pytorch-image-models#july-5-9-2021). which claimed both faster and better accuracy than `b3`. Please notice that `PyTorch` using different `bn_epsilon` and `padding` strategy. + + | V2 Model | Params | Top1 | Input | ImageNet21K | Imagenet21k-ft1k | Imagenet | + | ----------- | ------ | ---- | ----- | ----------- | ---------------- | -------- | + | EffV2B0 | 7.1M | 78.7 | 224 | [v2b0-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-21k.h5)|[v2b0-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-21k-ft1k.h5)|[v2b0-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-imagenet.h5)| + | EffV2B1 | 8.1M | 79.8 | 240 | [v2b1-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-21k.h5)|[v2b1-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-21k-ft1k.h5)|[v2b1-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-imagenet.h5)| + | EffV2B2 | 10.1M | 80.5 | 260 | [v2b2-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-21k.h5)|[v2b2-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-21k-ft1k.h5)|[v2b2-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-imagenet.h5)| + | EffV2B3 | 14.4M | 82.1 | 300 | [v2b3-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-21k.h5)|[v2b3-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-21k-ft1k.h5)|[v2b3-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-imagenet.h5)| + | EffV2T | 13.6M | 82.5 | 320 | | |[v2t-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-t-imagenet.h5)| + | EffV2S | 21.5M | 84.9 | 384 | [v2s-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-21k.h5) |[v2s-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-21k-ft1k.h5)|[v2s-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-imagenet.h5)| + | EffV2M | 54.1M | 86.2 | 480 | [v2m-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-21k.h5) |[v2m-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-21k-ft1k.h5)|[v2m-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-imagenet.h5)| + | EffV2L | 119.5M| 86.9 | 480 | [v2l-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-21k.h5) |[v2l-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-21k-ft1k.h5)|[v2l-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-imagenet.h5)| + | EffV2XL | 206.8M| 87.2 | 512 | [v2xl-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-xl-21k.h5)|[v2xl-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-xl-21k-ft1k.h5)| | + + - **EfficientNetV1 noisy_student models** from [Github tensorflow/tpu/efficientnet](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet). Paper [PDF 1911.04252 Self-training with Noisy Student improves ImageNet classification](https://arxiv.org/pdf/1911.04252.pdf). Parameter `include_preprocessing=False` is added, and the default `False` value makes expecting input value in range `[-1, 1]`, same with `EfficientNetV2`. Default `pretrained` is `noisy_student`. + + | V1 Model | Params | Top1 | Input | noisy_student | Imagenet | + | -------------- | ------- | ---- | ----- | ------------- | -------- | + | EffV1B0 | 5.3M | 78.8 | 224 | [v1-b0-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b0-noisy_student.h5) | [v1-b0-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b0-imagenet.h5) | + | EffV1B1 | 7.8M | 81.5 | 240 | [v1-b1-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b1-noisy_student.h5) | [v1-b1-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b1-imagenet.h5) | + | EffV1B2 | 9.1M | 82.4 | 260 | [v1-b2-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b2-noisy_student.h5) | [v1-b2-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b2-imagenet.h5) | + | EffV1B3 | 12.2M | 84.1 | 300 | [v1-b3-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b3-noisy_student.h5) | [v1-b3-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b3-imagenet.h5) | + | EffV1B4 | 19.3M | 85.3 | 380 | [v1-b4-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b4-noisy_student.h5) | [v1-b4-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b4-imagenet.h5) | + | EffV1B5 | 30.4M | 86.1 | 456 | [v1-b5-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b5-noisy_student.h5) | [v1-b5-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b5-imagenet.h5) | + | EffV1B6 | 43.0M | 86.4 | 528 | [v1-b6-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b6-noisy_student.h5) | [v1-b6-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b6-imagenet.h5) | + | EffV1B7 | 66.3M | 86.9 | 600 | [v1-b7-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b7-noisy_student.h5) | [v1-b7-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b7-imagenet.h5) | + | EffV1L2 | 480.3M | 88.4 | 800 | [v1-l2-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-l2-noisy_student.h5) | | + + - **Self tested imagenet accuracy** + - `rescale_mode` `torch` means `(image - [0.485, 0.456, 0.406]) / [[0.229, 0.224, 0.225]]`, `tf` means `(image - 0.5) / 0.5` + - All `resize_method` is `bicubic`. + - Some testing detail is not clear, so not exactly matching official reported results. + - Testing Detail is [EfficientNetV2 self tested imagenet accuracy](https://github.com/leondgarse/keras_cv_attention_models/discussions/19). + + | model | input | rescale_mode | central_crop | top 1 | top 5 | Reported top1 | + | ------------ | ----- | ------------ | ------------ | ------- | ------- | ----------------- | + | EffV2B0 | 224 | torch | 0.875 | 0.78748 | 0.94386 | 0.787 | + | EffV2B1 | 240 | torch | 0.95 | 0.7987 | 0.94936 | 0.798 | + | EffV2B2 | 260 | torch | 0.95 | 0.80642 | 0.95262 | 0.805 | + | EffV2B3 | 300 | torch | 0.95 | 0.82098 | 0.95896 | 0.821 | + | EffV2T | 320 | torch | 0.99 | 0.82506 | 0.96228 | 0.823 (input 288) | + | EffV2S | 384 | tf | 0.99 | 0.8386 | 0.967 | 0.839 | + | EffV2M | 480 | tf | 0.99 | 0.8509 | 0.973 | 0.852 | + | EffV2L | 480 | tf | 0.99 | 0.855 | 0.97324 | 0.857 | + | EffV2S ft1k | 384 | tf | 0.99 | 0.84328 | 0.97254 | 0.849 | + | EffV2M ft1k | 480 | tf | 0.99 | 0.85606 | 0.9775 | 0.862 | + | EffV2L ft1k | 480 | tf | 0.99 | 0.86294 | 0.9799 | 0.869 | + | EffV2XL ft1k | 512 | tf | 0.99 | 0.86532 | 0.97866 | 0.872 | +## Usage + - This repo can be installed as a pip package, or just `git clone` it. + ```py + pip install -U keras-efficientnet-v2 + # Or + pip install -U git+https://github.com/leondgarse/keras_efficientnet_v2 + ``` + - **Define model and load pretrained weights** Parameter `pretrained` is added in value `[None, "imagenet", "imagenet21k", "imagenet21k-ft1k"]`, default is `imagenet`. Model input value should be in range `[-1, 1]`. + ```py + # Will download and load `imagenet` pretrained weights. + # Model weight is loaded with `by_name=True, skip_mismatch=True`. + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2S(pretrained="imagenet") + + # Run prediction + import tensorflow as tf + from tensorflow import keras + from skimage.data import chelsea + imm = tf.image.resize(chelsea(), model.input_shape[1:3]) # Chelsea the cat + pred = model(tf.expand_dims(imm / 128. - 1., 0)).numpy() + print(keras.applications.imagenet_utils.decode_predictions(pred)[0]) + # [('n02124075', 'Egyptian_cat', 0.8642886), ('n02123159', 'tiger_cat', 0.030793495), ...] + ``` + Or download `h5` model and load directly + ```py + mm = keras.models.load_model('efficientnetv2-b3-21k-ft1k.h5') + ``` + For `"imagenet21k"` pre-trained model, actual `num_classes` is `21843`. + - **Exclude model top layers** by set `num_classes=0`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2B0(dropout=1e-6, num_classes=0, pretrained="imagenet21k") + print(model.output_shape) + # (None, 7, 7, 1280) + + model.save('efficientnetv2-b0-21k-notop.h5') + ``` + - **Use dynamic input resolution** by set `input_shape=(None, None, 3)`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2M(input_shape=(None, None, 3), drop_connect_rate=0.2, num_classes=0, pretrained="imagenet21k-ft1k") + + print(model(np.ones([1, 224, 224, 3])).shape) + # (1, 7, 7, 1280) + print(model(np.ones([1, 512, 512, 3])).shape) + # (1, 16, 16, 1280) + ``` + - **`include_preprocessing`** set `True` will add pre-processing `Rescale + Normalization` after `Input`. Means using input value in range `[0, 255]`. Default value `False` means in range `[-1, 1]`. Works both for `EfficientNetV2` and `EfficientNetV1`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV1B4(pretrained="noisy_student", include_preprocessing=True) + + from skimage.data import chelsea + imm = tf.image.resize(chelsea(), model.input_shape[1:3]) # Chelsea the cat + pred = model(tf.expand_dims(imm, 0)).numpy() # value in range [0, 255] + print(keras.applications.imagenet_utils.decode_predictions(pred)[0]) + # [('n02124075', 'Egyptian_cat', 0.68414235), ('n02123159', 'tiger_cat', 0.04486668), ...] + ``` +## Training detail from article + - **Training configures**, `Eval size` is used as the default `input_shape` for each model type. + + | Model | Train size | Eval size | Dropout | Randaug | Mixup | + | ------- | ---------- | --------- | ------- | ------- | ----- | + | EffV2B0 | 192 | 224 | 0.2 | 0 | 0 | + | EffV2B1 | 192 | 240 | 0.2 | 0 | 0 | + | EffV2B2 | 208 | 260 | 0.3 | 0 | 0 | + | EffV2B3 | 240 | 300 | 0.3 | 0 | 0 | + | EffV2S | 300 | 384 | 0.2 | 10 | 0 | + | EffV2M | 384 | 480 | 0.3 | 15 | 0.2 | + | EffV2L | 384 | 480 | 0.4 | 20 | 0.5 | + | EffV2XL | 384 | 512 | 0.4 | 20 | 0.5 | + + - EfficientNetV2-S architecture + + | Stage | Operator | Stride | #Channels | #Layers | + | ----- | ---------------------- | ------ | --------- | ------- | + | 0 | Conv3x3 | 2 | 24 | 1 | + | 1 | Fused-MBConv1, k3x3 | 1 | 24 | 2 | + | 2 | Fused-MBConv4, k3x3 | 2 | 48 | 4 | + | 3 | Fused-MBConv4, k3x3 | 2 | 64 | 4 | + | 4 | MBConv4, k3x3, SE0.25 | 2 | 128 | 6 | + | 5 | MBConv6, k3x3, SE0.25 | 1 | 160 | 9 | + | 6 | MBConv6, k3x3, SE0.25 | 2 | 256 | 15 | + | 7 | Conv1x1 & Pooling & FC | - | 1280 | 1 | + + - Progressive training settings for EfficientNetV2 + + | | S min | S max | M min | M max | L min | M max | + | ------------ | ----- | ----- | ----- | ----- | ----- | ----- | + | Image Size | 128 | 300 | 128 | 380 | 128 | 380 | + | RandAugment | 5 | 15 | 5 | 20 | 5 | 25 | + | Mixup alpha | 0 | 0 | 0 | 0.2 | 0 | 0.4 | + | Dropout rate | 0.1 | 0.3 | 0.1 | 0.4 | 0.1 | 0.5 | + + - Imagenet training detail + - RMSProp optimizer with decay 0.9 and momentum 0.9 + - batch norm momentum 0.99; weight decay 1e-5 + - Each model is trained for 350 epochs with total batch size 4096 + - Learning rate is first warmed up from 0 to 0.256, and then decayed by 0.97 every 2.4 epochs + - We use exponential moving average with 0.9999 decay rate + - RandAugment (Cubuk et al., 2020) + - Mixup (Zhang et al., 2018) + - Dropout (Srivastava et al., 2014) + - and stochastic depth (Huang et al., 2016) with 0.8 survival probability +## Detailed conversion procedure + - [convert_effnetv2_model.py](convert_effnetv2_model.py) is a modified version of [the orignal effnetv2_model.py](https://github.com/google/automl/blob/master/efficientnetv2/effnetv2_model.py). Check detail by `vimdiff convert_effnetv2_model.py ../automl/efficientnetv2/effnetv2_model.py` + - Delete some `names`, as they may cause confliction in keras. + - Use `.call` directly calling `se` modules and other blocks, so they will not be `blocks` in `model.summary()` + - Just use `Add` layer instead of `utils.drop_connect`, as when `is_training=False`, `utils.drop_connect` functions like `Add`. + - Add a `num_classes` parameter outside of `mconfig`. + - Add `__main__` part, which makes this can be run as a script. Refer to it for converting detail. + - Depends on official repo + ```sh + ../ + ├── automl # Official repo + ├── keras_efficientnet_v2 # This one + ``` + - **Procedure** + ```py + # See help info + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -h + + # Convert by specific model_type and dataset type + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m b0 -d imagenet21k + + # Convert by specific model_type and all its datasets ['imagenet', 'imagenet21k', 'imagenetft'] + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m s -d all + + # Convert all model_types and and all datasets + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m all -d all + ``` +## Progressive train test on cifar10 + - [Colab efficientnetV2_basic_test.ipynb](https://colab.research.google.com/drive/1vmAEfF9tUgK2gkrS5qVftadTyUcX343D?usp=sharing) + ```py + import keras_efficientnet_v2 + from tensorflow import keras + from keras_efficientnet_v2 import progressive_train_test + + model = keras_efficientnet_v2.EfficientNetV2S(input_shape=(None, None, 3), num_classes=10, classifier_activation='softmax', dropout=0.1) + model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"]) + + hhs = progressive_train_test.progressive_with_dropout_randaug( + model, + data_name="cifar10", + lr_scheduler=None, + total_epochs=36, + batch_size=64, + dropout_layer=-2, + target_shapes=[128, 160, 192, 224], # [128, 185, 242, 300] for final shape (300, 300) + dropouts=[0.1, 0.2, 0.3, 0.4], + magnitudes=[5, 8, 12, 15], + ) + + with open("history_ev2s_imagenet_progressive_224.json", "w") as ff: + json.dump(hhs, ff) + ``` +  +## Related Projects + - [tfa.layers.StochasticDepth](https://www.tensorflow.org/addons/api_docs/python/tfa/layers/StochasticDepth) + - [Official efficientnetv2](https://github.com/google/automl/tree/master/efficientnetv2) + - [Github rwightman/pytorch-image-models](https://github.com/rwightman/pytorch-image-models#july-5-9-2021) +*** + + + + +%package help +Summary: Development documents and examples for keras-efficientnet-v2 +Provides: python3-keras-efficientnet-v2-doc +%description help +# ___Keras EfficientNetV2___ +*** +- **As `EfficientNetV2` is included in `keras.application` now, merged this project into [Github leondgarse/keras_cv_attention_models/efficientnet](https://github.com/leondgarse/keras_cv_attention_models/tree/main/keras_cv_attention_models/efficientnet). Package `keras-efficientnet-v2` moved into `stable` status**. + +## Table of Contents +<!-- TOC depthFrom:1 depthTo:6 withLinks:1 updateOnSave:1 orderedList:0 --> + +- [___Keras EfficientNetV2___](#keras-efficientnetv2) + - [Table of Contents](#table-of-contents) + - [Summary](#summary) + - [Usage](#usage) + - [Training detail from article](#training-detail-from-article) + - [Detailed conversion procedure](#detailed-conversion-procedure) + - [Progressive train test on cifar10](#progressive-train-test-on-cifar10) + - [Related Projects](#related-projects) + +<!-- /TOC --> +*** + +## Summary + - My own keras implementation of [Official efficientnetv2](https://github.com/google/automl/tree/master/efficientnetv2). Article [arXiv 2104.00298 EfficientNetV2: Smaller Models and Faster Training](https://arxiv.org/abs/2104.00298) by Mingxing Tan, Quoc V. Le. + - `h5` model weights converted from official publication. + - `effv2-t-imagenet.h5` model weights converted from [Github rwightman/pytorch-image-models](https://github.com/rwightman/pytorch-image-models#july-5-9-2021). which claimed both faster and better accuracy than `b3`. Please notice that `PyTorch` using different `bn_epsilon` and `padding` strategy. + + | V2 Model | Params | Top1 | Input | ImageNet21K | Imagenet21k-ft1k | Imagenet | + | ----------- | ------ | ---- | ----- | ----------- | ---------------- | -------- | + | EffV2B0 | 7.1M | 78.7 | 224 | [v2b0-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-21k.h5)|[v2b0-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-21k-ft1k.h5)|[v2b0-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b0-imagenet.h5)| + | EffV2B1 | 8.1M | 79.8 | 240 | [v2b1-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-21k.h5)|[v2b1-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-21k-ft1k.h5)|[v2b1-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b1-imagenet.h5)| + | EffV2B2 | 10.1M | 80.5 | 260 | [v2b2-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-21k.h5)|[v2b2-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-21k-ft1k.h5)|[v2b2-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b2-imagenet.h5)| + | EffV2B3 | 14.4M | 82.1 | 300 | [v2b3-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-21k.h5)|[v2b3-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-21k-ft1k.h5)|[v2b3-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-b3-imagenet.h5)| + | EffV2T | 13.6M | 82.5 | 320 | | |[v2t-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-t-imagenet.h5)| + | EffV2S | 21.5M | 84.9 | 384 | [v2s-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-21k.h5) |[v2s-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-21k-ft1k.h5)|[v2s-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-s-imagenet.h5)| + | EffV2M | 54.1M | 86.2 | 480 | [v2m-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-21k.h5) |[v2m-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-21k-ft1k.h5)|[v2m-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-m-imagenet.h5)| + | EffV2L | 119.5M| 86.9 | 480 | [v2l-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-21k.h5) |[v2l-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-21k-ft1k.h5)|[v2l-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-l-imagenet.h5)| + | EffV2XL | 206.8M| 87.2 | 512 | [v2xl-21k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-xl-21k.h5)|[v2xl-21k-ft1k.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv2_pretrained/efficientnetv2-xl-21k-ft1k.h5)| | + + - **EfficientNetV1 noisy_student models** from [Github tensorflow/tpu/efficientnet](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet). Paper [PDF 1911.04252 Self-training with Noisy Student improves ImageNet classification](https://arxiv.org/pdf/1911.04252.pdf). Parameter `include_preprocessing=False` is added, and the default `False` value makes expecting input value in range `[-1, 1]`, same with `EfficientNetV2`. Default `pretrained` is `noisy_student`. + + | V1 Model | Params | Top1 | Input | noisy_student | Imagenet | + | -------------- | ------- | ---- | ----- | ------------- | -------- | + | EffV1B0 | 5.3M | 78.8 | 224 | [v1-b0-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b0-noisy_student.h5) | [v1-b0-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b0-imagenet.h5) | + | EffV1B1 | 7.8M | 81.5 | 240 | [v1-b1-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b1-noisy_student.h5) | [v1-b1-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b1-imagenet.h5) | + | EffV1B2 | 9.1M | 82.4 | 260 | [v1-b2-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b2-noisy_student.h5) | [v1-b2-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b2-imagenet.h5) | + | EffV1B3 | 12.2M | 84.1 | 300 | [v1-b3-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b3-noisy_student.h5) | [v1-b3-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b3-imagenet.h5) | + | EffV1B4 | 19.3M | 85.3 | 380 | [v1-b4-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b4-noisy_student.h5) | [v1-b4-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b4-imagenet.h5) | + | EffV1B5 | 30.4M | 86.1 | 456 | [v1-b5-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b5-noisy_student.h5) | [v1-b5-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b5-imagenet.h5) | + | EffV1B6 | 43.0M | 86.4 | 528 | [v1-b6-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b6-noisy_student.h5) | [v1-b6-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b6-imagenet.h5) | + | EffV1B7 | 66.3M | 86.9 | 600 | [v1-b7-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b7-noisy_student.h5) | [v1-b7-imagenet.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-b7-imagenet.h5) | + | EffV1L2 | 480.3M | 88.4 | 800 | [v1-l2-noisy_student.h5](https://github.com/leondgarse/keras_efficientnet_v2/releases/download/effnetv1_pretrained/efficientnetv1-l2-noisy_student.h5) | | + + - **Self tested imagenet accuracy** + - `rescale_mode` `torch` means `(image - [0.485, 0.456, 0.406]) / [[0.229, 0.224, 0.225]]`, `tf` means `(image - 0.5) / 0.5` + - All `resize_method` is `bicubic`. + - Some testing detail is not clear, so not exactly matching official reported results. + - Testing Detail is [EfficientNetV2 self tested imagenet accuracy](https://github.com/leondgarse/keras_cv_attention_models/discussions/19). + + | model | input | rescale_mode | central_crop | top 1 | top 5 | Reported top1 | + | ------------ | ----- | ------------ | ------------ | ------- | ------- | ----------------- | + | EffV2B0 | 224 | torch | 0.875 | 0.78748 | 0.94386 | 0.787 | + | EffV2B1 | 240 | torch | 0.95 | 0.7987 | 0.94936 | 0.798 | + | EffV2B2 | 260 | torch | 0.95 | 0.80642 | 0.95262 | 0.805 | + | EffV2B3 | 300 | torch | 0.95 | 0.82098 | 0.95896 | 0.821 | + | EffV2T | 320 | torch | 0.99 | 0.82506 | 0.96228 | 0.823 (input 288) | + | EffV2S | 384 | tf | 0.99 | 0.8386 | 0.967 | 0.839 | + | EffV2M | 480 | tf | 0.99 | 0.8509 | 0.973 | 0.852 | + | EffV2L | 480 | tf | 0.99 | 0.855 | 0.97324 | 0.857 | + | EffV2S ft1k | 384 | tf | 0.99 | 0.84328 | 0.97254 | 0.849 | + | EffV2M ft1k | 480 | tf | 0.99 | 0.85606 | 0.9775 | 0.862 | + | EffV2L ft1k | 480 | tf | 0.99 | 0.86294 | 0.9799 | 0.869 | + | EffV2XL ft1k | 512 | tf | 0.99 | 0.86532 | 0.97866 | 0.872 | +## Usage + - This repo can be installed as a pip package, or just `git clone` it. + ```py + pip install -U keras-efficientnet-v2 + # Or + pip install -U git+https://github.com/leondgarse/keras_efficientnet_v2 + ``` + - **Define model and load pretrained weights** Parameter `pretrained` is added in value `[None, "imagenet", "imagenet21k", "imagenet21k-ft1k"]`, default is `imagenet`. Model input value should be in range `[-1, 1]`. + ```py + # Will download and load `imagenet` pretrained weights. + # Model weight is loaded with `by_name=True, skip_mismatch=True`. + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2S(pretrained="imagenet") + + # Run prediction + import tensorflow as tf + from tensorflow import keras + from skimage.data import chelsea + imm = tf.image.resize(chelsea(), model.input_shape[1:3]) # Chelsea the cat + pred = model(tf.expand_dims(imm / 128. - 1., 0)).numpy() + print(keras.applications.imagenet_utils.decode_predictions(pred)[0]) + # [('n02124075', 'Egyptian_cat', 0.8642886), ('n02123159', 'tiger_cat', 0.030793495), ...] + ``` + Or download `h5` model and load directly + ```py + mm = keras.models.load_model('efficientnetv2-b3-21k-ft1k.h5') + ``` + For `"imagenet21k"` pre-trained model, actual `num_classes` is `21843`. + - **Exclude model top layers** by set `num_classes=0`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2B0(dropout=1e-6, num_classes=0, pretrained="imagenet21k") + print(model.output_shape) + # (None, 7, 7, 1280) + + model.save('efficientnetv2-b0-21k-notop.h5') + ``` + - **Use dynamic input resolution** by set `input_shape=(None, None, 3)`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV2M(input_shape=(None, None, 3), drop_connect_rate=0.2, num_classes=0, pretrained="imagenet21k-ft1k") + + print(model(np.ones([1, 224, 224, 3])).shape) + # (1, 7, 7, 1280) + print(model(np.ones([1, 512, 512, 3])).shape) + # (1, 16, 16, 1280) + ``` + - **`include_preprocessing`** set `True` will add pre-processing `Rescale + Normalization` after `Input`. Means using input value in range `[0, 255]`. Default value `False` means in range `[-1, 1]`. Works both for `EfficientNetV2` and `EfficientNetV1`. + ```py + import keras_efficientnet_v2 + model = keras_efficientnet_v2.EfficientNetV1B4(pretrained="noisy_student", include_preprocessing=True) + + from skimage.data import chelsea + imm = tf.image.resize(chelsea(), model.input_shape[1:3]) # Chelsea the cat + pred = model(tf.expand_dims(imm, 0)).numpy() # value in range [0, 255] + print(keras.applications.imagenet_utils.decode_predictions(pred)[0]) + # [('n02124075', 'Egyptian_cat', 0.68414235), ('n02123159', 'tiger_cat', 0.04486668), ...] + ``` +## Training detail from article + - **Training configures**, `Eval size` is used as the default `input_shape` for each model type. + + | Model | Train size | Eval size | Dropout | Randaug | Mixup | + | ------- | ---------- | --------- | ------- | ------- | ----- | + | EffV2B0 | 192 | 224 | 0.2 | 0 | 0 | + | EffV2B1 | 192 | 240 | 0.2 | 0 | 0 | + | EffV2B2 | 208 | 260 | 0.3 | 0 | 0 | + | EffV2B3 | 240 | 300 | 0.3 | 0 | 0 | + | EffV2S | 300 | 384 | 0.2 | 10 | 0 | + | EffV2M | 384 | 480 | 0.3 | 15 | 0.2 | + | EffV2L | 384 | 480 | 0.4 | 20 | 0.5 | + | EffV2XL | 384 | 512 | 0.4 | 20 | 0.5 | + + - EfficientNetV2-S architecture + + | Stage | Operator | Stride | #Channels | #Layers | + | ----- | ---------------------- | ------ | --------- | ------- | + | 0 | Conv3x3 | 2 | 24 | 1 | + | 1 | Fused-MBConv1, k3x3 | 1 | 24 | 2 | + | 2 | Fused-MBConv4, k3x3 | 2 | 48 | 4 | + | 3 | Fused-MBConv4, k3x3 | 2 | 64 | 4 | + | 4 | MBConv4, k3x3, SE0.25 | 2 | 128 | 6 | + | 5 | MBConv6, k3x3, SE0.25 | 1 | 160 | 9 | + | 6 | MBConv6, k3x3, SE0.25 | 2 | 256 | 15 | + | 7 | Conv1x1 & Pooling & FC | - | 1280 | 1 | + + - Progressive training settings for EfficientNetV2 + + | | S min | S max | M min | M max | L min | M max | + | ------------ | ----- | ----- | ----- | ----- | ----- | ----- | + | Image Size | 128 | 300 | 128 | 380 | 128 | 380 | + | RandAugment | 5 | 15 | 5 | 20 | 5 | 25 | + | Mixup alpha | 0 | 0 | 0 | 0.2 | 0 | 0.4 | + | Dropout rate | 0.1 | 0.3 | 0.1 | 0.4 | 0.1 | 0.5 | + + - Imagenet training detail + - RMSProp optimizer with decay 0.9 and momentum 0.9 + - batch norm momentum 0.99; weight decay 1e-5 + - Each model is trained for 350 epochs with total batch size 4096 + - Learning rate is first warmed up from 0 to 0.256, and then decayed by 0.97 every 2.4 epochs + - We use exponential moving average with 0.9999 decay rate + - RandAugment (Cubuk et al., 2020) + - Mixup (Zhang et al., 2018) + - Dropout (Srivastava et al., 2014) + - and stochastic depth (Huang et al., 2016) with 0.8 survival probability +## Detailed conversion procedure + - [convert_effnetv2_model.py](convert_effnetv2_model.py) is a modified version of [the orignal effnetv2_model.py](https://github.com/google/automl/blob/master/efficientnetv2/effnetv2_model.py). Check detail by `vimdiff convert_effnetv2_model.py ../automl/efficientnetv2/effnetv2_model.py` + - Delete some `names`, as they may cause confliction in keras. + - Use `.call` directly calling `se` modules and other blocks, so they will not be `blocks` in `model.summary()` + - Just use `Add` layer instead of `utils.drop_connect`, as when `is_training=False`, `utils.drop_connect` functions like `Add`. + - Add a `num_classes` parameter outside of `mconfig`. + - Add `__main__` part, which makes this can be run as a script. Refer to it for converting detail. + - Depends on official repo + ```sh + ../ + ├── automl # Official repo + ├── keras_efficientnet_v2 # This one + ``` + - **Procedure** + ```py + # See help info + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -h + + # Convert by specific model_type and dataset type + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m b0 -d imagenet21k + + # Convert by specific model_type and all its datasets ['imagenet', 'imagenet21k', 'imagenetft'] + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m s -d all + + # Convert all model_types and and all datasets + CUDA_VISIBLE_DEVICES='-1' python convert_effnetv2_model.py -m all -d all + ``` +## Progressive train test on cifar10 + - [Colab efficientnetV2_basic_test.ipynb](https://colab.research.google.com/drive/1vmAEfF9tUgK2gkrS5qVftadTyUcX343D?usp=sharing) + ```py + import keras_efficientnet_v2 + from tensorflow import keras + from keras_efficientnet_v2 import progressive_train_test + + model = keras_efficientnet_v2.EfficientNetV2S(input_shape=(None, None, 3), num_classes=10, classifier_activation='softmax', dropout=0.1) + model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"]) + + hhs = progressive_train_test.progressive_with_dropout_randaug( + model, + data_name="cifar10", + lr_scheduler=None, + total_epochs=36, + batch_size=64, + dropout_layer=-2, + target_shapes=[128, 160, 192, 224], # [128, 185, 242, 300] for final shape (300, 300) + dropouts=[0.1, 0.2, 0.3, 0.4], + magnitudes=[5, 8, 12, 15], + ) + + with open("history_ev2s_imagenet_progressive_224.json", "w") as ff: + json.dump(hhs, ff) + ``` +  +## Related Projects + - [tfa.layers.StochasticDepth](https://www.tensorflow.org/addons/api_docs/python/tfa/layers/StochasticDepth) + - [Official efficientnetv2](https://github.com/google/automl/tree/master/efficientnetv2) + - [Github rwightman/pytorch-image-models](https://github.com/rwightman/pytorch-image-models#july-5-9-2021) +*** + + + + +%prep +%autosetup -n keras-efficientnet-v2-1.2.2 + +%build +%py3_build + +%install +%py3_install +install -d -m755 %{buildroot}/%{_pkgdocdir} +if [ -d doc ]; then cp -arf doc %{buildroot}/%{_pkgdocdir}; fi +if [ -d docs ]; then cp -arf docs %{buildroot}/%{_pkgdocdir}; fi +if [ -d example ]; then cp -arf example %{buildroot}/%{_pkgdocdir}; fi +if [ -d examples ]; then cp -arf examples %{buildroot}/%{_pkgdocdir}; fi +pushd %{buildroot} +if [ -d usr/lib ]; then + find usr/lib -type f -printf "/%h/%f\n" >> filelist.lst +fi +if [ -d usr/lib64 ]; then + find usr/lib64 -type f -printf "/%h/%f\n" >> filelist.lst +fi +if [ -d usr/bin ]; then + find usr/bin -type f -printf "/%h/%f\n" >> filelist.lst +fi +if [ -d usr/sbin ]; then + find usr/sbin -type f -printf "/%h/%f\n" >> filelist.lst +fi +touch doclist.lst +if [ -d usr/share/man ]; then + find usr/share/man -type f -printf "/%h/%f.gz\n" >> doclist.lst +fi +popd +mv %{buildroot}/filelist.lst . +mv %{buildroot}/doclist.lst . + +%files -n python3-keras-efficientnet-v2 -f filelist.lst +%dir %{python3_sitelib}/* + +%files help -f doclist.lst +%{_docdir}/* + +%changelog +* Thu May 18 2023 Python_Bot <Python_Bot@openeuler.org> - 1.2.2-1 +- Package Spec generated @@ -0,0 +1 @@ +f7b2c1147a486916a3c9627f73b6e306 keras-efficientnet-v2-1.2.2.tar.gz |