v0.47.1

Fix

• Remove wrongly committed files (4fdea0c)

v0.47.0

Feature

• Simplify project structure (81cbcb3)

v0.46.1

Fix

• Fix wrong port definitions (9a4c8af)
• Fix some syntax errors (3997bbd)

v0.46.0

Feature

• Use conv1d arithmetics function to implement conv1d module (69778be)
• Add conv1d function to arithmetics (1cab190)
• Test that conv1d uses different arithmetics (7eb01db)
• Add the ability to sum over dimension (c45c0e6)

Fix

• Quantize weights before inference (61153e6)

v0.45.0

Feature

• Simplify usage for the elasticai.creator.nn.vhdl package by adding layers to init (2c7c968)

Fix

• Fix broken import in base template generator and move it with its template to own folder (9eb1f70)

v0.44.0

Feature

• Check for autowiring protocol violation (3f17e00)
• Add AutoWirer (f4159c8)
• Add intermediate symbols to rule definitions (624b310)
• Support parsing partial files (f2c2eb6)
• Support parsing partial files (8170012)
• Add standalone parser module (5a9b141)
• Add basic vhdl parsing (5df2a3f)
• Port expansion/template based on autowiring protocol (0d14618)
• template: Make precomputed scalar functions bufferless (89986fa)

Fix

• Use new Sequential constructor (6bb111b)
• Port def and impl of monotonous function design (2d423d4)
• Children of sequential layer determine signal widths (3dd5c0c)
• Remove obsolete parsing functionality (7f85d05)
• Adjust tests to follow previous change (c328bd5)
• Correct tuple type annotation (f0e7da0)

v0.43.0

Feature

• Add tests for the FPMonotonouslyIncreasingModule (9ba64ae)
• Introduce FPMonotonouslyIncreasingModule to easily add new activations (b78c922)

Fix

• Set correct signal names for x and y address (5354a2a)
• Use elsif in lookup table (f375ba3)
• Increase default sampling intervall (07620d3)

v0.42.0

Feature

• Reimplement hard tanh activation function (9b86f9d)
• Add working hardsigmoid implementation (db03ff0)
• Make sure that inplace parameter is fixed defined (79b7a1e)

v0.41.0

Feature

• Add fixed point ReLU module (62c1555)

v0.40.0

Feature

• Simplify the use of the sequential layer (same as in torch) (9fad15d)
• Improve performance of the identity step autograd function (46f036c)
• Add quantized tanh implementation with lookup tables (3a1fb10)
• Implement bufferless component interface for precomputed scalar function (f701a57)
• Pass step lut to identity step function and improve readablility (c1b6747)
• Rename autograd function and pass step lut to autograd function (d607e98)
• Implement autograd fn to map inputs to a subset of inputs (26c6ec7)
• Add a function to easily compare tensors with pytest (24e737e)
• Add experimental precomputed tanh in fixed point (0e76d03)

Fix

• Fix that last io pair was dropped when calling save_to function (2bc46ac)
• Fix missing creation of a subpath in the save_to function (2a4dbdf)