Releases | Open Source Science

Updates TensorFlow from 2.15 to 2.16, which resulted in some non-minor changes.
Dropped support for Bidirectional, GRU, LSTM, and stateful models. They might be re-added again in the future though.

- C++
Published by Dobiasd almost 2 years ago

frugally-deep - v0.15.31

Simplified the internal implementation of sum_tensors
Removed some redundant internal stuff
Auto-formatted all code

- C++
Published by Dobiasd about 2 years ago

frugally-deep - v0.15.30

Added support for MultiHeadAttention layers

- C++
Published by Dobiasd about 2 years ago

frugally-deep - v0.15.29

Removed the -p0 suffix from our version numbers

- C++
Published by Dobiasd about 2 years ago

frugally-deep - v0.15.28-p0

Improved the install process using CMake on Fedora Linux

- C++
Published by Dobiasd about 2 years ago

frugally-deep - v0.15.27-p0

Added support for UnitNormalization layers.

- C++
Published by Dobiasd about 2 years ago

frugally-deep - v0.15.26-p0

Added support for LayerNormalization layer.
Some improvements on BatchNormalization.

- C++
Published by Dobiasd about 2 years ago

frugally-deep - v0.15.25-p0

Improved performance of softmax calculation (https://github.com/Dobiasd/frugally-deep/issues/405)

- C++
Published by Dobiasd over 2 years ago

frugally-deep - v0.15.24-p0

Updated dependency FunctionalPlus to version v0.2.20-p0

- C++
Published by Dobiasd over 2 years ago

frugally-deep - v0.15.23-p0

Fix C++ syntax for or, relevant for Visual Studio

- C++
Published by Dobiasd over 2 years ago

frugally-deep - v0.15.22-p0

Added support for AdditiveAttention layer
Added support for use_scale and score_mode concat in Attention layer
Fixed softmax implementation
Fixed compiler errors with Visual Studio 2022 (C++20)
Avoid unnecessarily compiling the TensorFlow model when loading it for conversion

- C++
Published by Dobiasd over 2 years ago

frugally-deep - v0.15.21-p0

Added support for CategoryEncoding layer with outputmode onehot
Added support for Attention layer
Fixed zero variance in a Normalization layer
Added support for keepdims in global pooling layers

- C++
Published by Dobiasd over 2 years ago

frugally-deep - v0.15.20-p0

Improved the performance of the Dense layer
Fixed variance-size assertion in the Normalization layer
Added support for new layer types:
- ActivityRegularization
- CategoryEncoding
- CenterCrop
- Cropping3D
- Dot
- Minimum
- Resizing
- ZeroPadding3D

- C++
Published by Dobiasd almost 3 years ago

frugally-deep - v0.15.19-p0

Fixes interpretation of axis value in Normalization layer (issue 357)

- C++
Published by Dobiasd over 3 years ago

frugally-deep - v0.15.18-p0

Improved performance (with strides==(1, 1)) of Conv2D, DepthwiseConv2D, and SeparableConv2D.
JSON export: Reduced memory usage and output.file size.

- C++
Published by Dobiasd over 3 years ago

frugally-deep - v0.15.17-p0

Added support for the Rescaling layer.

- C++
Published by Dobiasd almost 4 years ago

frugally-deep - v0.15.16-p0

Fixed missing inline keyword.

- C++
Published by Dobiasd almost 4 years ago

frugally-deep - v0.15.15-p0

Improved performance of SeparableConv2D and DepthwiseConv2D layers

- C++
Published by Dobiasd almost 4 years ago

frugally-deep - v0.15.14-p0

improved checks and docs
added support for FixedDropout (noop)
added support for tensor expansion in Multiply layer

- C++
Published by Dobiasd almost 4 years ago

frugally-deep - v0.15.13-p0

Added support for Normalization layer.

- C++
Published by Dobiasd about 4 years ago

frugally-deep - v0.15.12-p0

Added support for negative_slope and threshold in ReLU layers

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.11-p0

Added support for relu6 activation

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.10-p0

Added support for layers RandomContrast, RandomFlip, RandomHeight, RandomTranslation, RandomWidth, and RandomZoom (all no-ops during prediction)

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.9-p0

Added support for activations exponential, gelu, and softsign

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.8-p0

Added support for RepeatVector layer

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.7-p0

Added support for swish activation.
Added support for full models inside TimeDistributed layers.

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.6-p0

Removed the broken workaround (for training=true) introduced in release v0.15.6-p0

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.5-p0

Added compatibility for models including layers with training=True by ignoring the flag.

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.4-p0

Added support for using BatchNormalization layers as the inner layer in TimeDistributed layers. :zany_face:

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.3-p0

Added support for shape inference in Reshape layers.
Improved some docs.

- C++
Published by Dobiasd over 4 years ago

frugally-deep - v0.15.2-p0

Fix average_pooling_2d_layer and max_pooling_2d_layer for channels_first
Fix MSVC Compiler Warning C4701 "potentially uninitialized local variable" in time_distributed_layer.hpp
Add support for RandomRotation layer (no-op in prediction)
Improved documentation

- C++
Published by Dobiasd almost 5 years ago

frugally-deep - v0.15.1-p0

Adds support for duplicate layer names in nested models (see issue #237)

- C++
Published by Dobiasd over 5 years ago

frugally-deep - v0.15.0-p0

Update TensorFlow to version 2.3

- C++
Published by Dobiasd over 5 years ago

frugally-deep - v0.14.4-p0

Use the latest version of Use FunctionalPlus (0.2.8).
Some CMake improvements.

- C++
Published by Dobiasd over 5 years ago

frugally-deep - v0.14.3-p0

Further performance improvements for 2d convolutions with big input tensors (see issue 226 and issue 227), while also reducing memory usage.

- C++
Published by Dobiasd over 5 years ago

frugally-deep - v0.14.2-p0

Fix some edge case for very small convolutions.

- C++
Published by Dobiasd over 5 years ago

frugally-deep - v0.14.1-p0

Improved performance of convolution (Conv2D) on large input tensors, while also reducing memory usage.

- C++
Published by Dobiasd over 5 years ago

frugally-deep - v0.14.0-p0

Tensors are now stored in aligned memory blocks according to Eigen::aligned_allocator<T> for performance.

- C++
Published by Dobiasd almost 6 years ago

frugally-deep - v0.13.1-p0

Improve performance of LSTM and GRU

- C++
Published by Dobiasd almost 6 years ago

frugally-deep - v0.13.0-p0

Update Tensorflow to version 2.1.
Subsequent adjustments of RNN layers.

- C++
Published by Dobiasd almost 6 years ago

frugally-deep - v0.12.1-p0

Improved performance for LTSM and GRU.

- C++
Published by Dobiasd almost 6 years ago

frugally-deep - v0.12.0-p0

Tensor shapes and positions now explicitly track the tensor's rank.

breaking changes: - fdeep::tensor5 has been renamed to fdeep::tensor - fdeep::tensor5_pos has been renamed to fdeep::tensor_pos - fdeep::shape5 has been renamed to fdeep::tensor_shape - dropped support for shape inference in reshape layers

deprecated functions (will likely be removed from the API soon) - float_type fdeep::tensor5::get(std::size_t, std::size_t, std::size_t, std::size_t, std::size_t) const: Please use float_type fdeep::tensor5::get(const tensor_pos&) const or float_type fdeep::tensor5::get_ignore_rank(const tensor_pos&) const instead. - void fdeep::tensor5::set(std::size_t, std::size_t, std::size_t, std::size_t, std::size_t, float_type): Please use float_type fdeep::tensor5::set(const tensor_pos, float_type) or float_type fdeep::tensor5::set_ignore_rank(const tensor_pos&, float_type) instead.

- C++
Published by Dobiasd almost 6 years ago

frugally-deep - v0.11.1-p0

Support for batch normalization on arbitrary axes
Improved error messages

- C++
Published by Dobiasd about 6 years ago

frugally-deep - v0.11.0-p0

Switch from keras to tf.keras.

With the release of version 2.3.0, team Keras announced the following:

This is also the last major release of multi-backend Keras. Going forward, we recommend that users consider switching their Keras code to tf.keras in TensorFlow 2.0. It implements the same Keras 2.3.0 API (so switching should be as easy as changing the Keras import statements), but it has many advantages for TensorFlow users, such as support for eager execution, distribution, TPU training, and generally far better integration between low-level TensorFlow and high-level concepts like Layer and Model. It is also better maintained.

Development will focus on tf.keras going forward. We will keep maintaining multi-backend Keras over the next 6 months, but we will only be merging bug fixes. API changes will not be ported.

So frugally-deep follows this direction.

Thanks to @keithchugg for doing the needed adjustments of the stateful implementations.