AutoKeras: Keras 官方出品基于 Keras 的 AutoML 系统
Keras 官方出品基于 Keras 的 AutoML 系统。支持 CPU 和 GPU 训练，傻瓜式 API，3 行代码就能训练一个模型。目前支持的任务：图像分类、图像回归、文本分类、结构化数据分类等。将人从手工选择超参数中解放出来，快速开发原型 
- Compatible with TensorFlow 2.4.
- Support specify search space for
num_units,num_layers, anddropoutof DenseBlock. Code Example
Assets
2
- Reduces
batch_sizeby 2 when running out of memory. - Add pretrained EfficientNet to the search space.
- Support load data from disk. For more details, read our tutorials on the official website.
- Put data type casting and reshaping into the exported Keras Model.
- Fixed the bug of breaking when validation_data is a tf.data.Dataset instance.
Assets
2
- Improved text data performance by adding pretrained BERT model to the search space.
- Added Adam optimizer with weight decay to the search space.
Assets
2
- Performance improvements for structured data classification and regression tasks.
- Bug fix for not using the best number of epochs for the final model training when validation data is not provided.
Assets
2
- Bug fix for using multiple GPUs.
- Bug fix for KeyError when searching for large number of trials.
- Grammar fixes for error messages.
Assets
2
- Reduced GPU memory consumption.
- Optimized greedy tuning algorithm.
- Support using multiple GPUs.
Assets
2
- Updated TF dependency version to 2.3.0.
- Bug fix for checkpoint not found during evaluate and predict.
Assets
2
Bug Fixes:
- Fixed fit the final model only one epoch after search all the trials.
- Fixed the checkpoint not found issue during fit.
New Features:
- Add pretrained XceptionNet and ResNet with ImageNet weights to the search space of image tasks.
- Enlarge the search space for more optimizers and laerning rates.
- Transformer model included in the search space.
API Changes:
- ResNetBlock and XceptionBlock
poolingargument removed. - All task APIs use
overwrite=Falseby default. - Change all
dropout_ratearguments todropout.
Assets
2
Bump dependency TensorFlow version to 2.2.0.
User now can use custom metrics and loss.
User now can specify the tuner to use for Task APIs like ImageClassifier.
Use Keras preprocessing layers for ImageAugmentation.
If not epochs is not specified, and validation_data is not provided,
it will use the best models best number of epochs for retraining the best model with the entire training set.
Bug fixes:
All bugs in the tutorials are fixed.
All the tutorials can run smoothly.
Breaking changes:
ImageAugmentation args updated.
Assets
2
Fixed the bug for low performance in the final model training.
Fixed the bug for does not support tf.data.Dataset for TextClassifier.
Improved performance for ImageClassifier and TextClassifier.
Issues:
It cannot save the preprocessing layers' weights.
The exported model has to be adapted manually if contains any preprocessing layer.
It should be fixed with TF 2.2.
We will have another release afterward.
Assets
2
| Watchers：301 |
| Star：7804 |
| Fork：1267 |
| 创建时间： 2017-11-20 07:18:20 |
| 最后Commits： 4天前 |
| 许可协议：MIT |
dbe0783Compare
Verified
num_units,num_layers, anddropoutof DenseBlock. Code Examplefilters,num_blocks, andnum_layersof ConvBlock.AutoModel.predict(...).