v0.7.4-p0

Support for Permute layer added.

v0.7.3-p0

Fix invalid shape of bias matrix in GRU for reset_after=False

v0.7.2-p0

Calculate hash from model weights and store to JSON

v0.7.1-p0

• Add support for Embedding layers
• Add support for GRU layers
• Fix missing inlines resulting in duplicate function definitions on some compilers

v0.7.0-p0

Adds support for the following layer types:

• Bidirectional
• LSTM
• TimeDistributed

v0.6.0-p0

API-breaking change: Increase dimensionality of tensors from 3 to 5 (needed for future development, like LSTMs and TimeDistributed):

• tensor3 -> tensor5
• shape_hwc -> shape5
• tensor3::get_yxz -> tensor5::get
• tensor3::set_yxz -> tensor5::set

Feature:

• Enable dense layers for non-flattened input tensors.

Fix:

• Safeguard softmax implementation against NaNs.

v0.5.4-p0

• Slight performance improvements

v0.5.3-p0

• Fix PReLU layer with shared axes and variable shapes

v0.5.2-p0

• minor cleanup

v0.5.1-p0

• Add PReLU layer with shared_axes.
• Add compatibility to Eigen 3.2.x.
• Add parametrized ReLU layers, drop relu6.
• Switch from channel_first to channel_last order. This is an API-breaking change:
  • Models converted from .h5 to .json with older versions have to be re-converted.
  • shape3 was dropped, shape_hwc was added instead.
  • tensor3_pos was dropped, tensor3_pos_yxz was added instead.
  • tensor3::get was dropped, tensor3::get_yxz was added instead.
  • tensor3::set was dropped, tensor3::set_yxz was added instead.