1. 20 9月, 2025 8 次提交
  2. 14 9月, 2025 1 次提交
  3. 06 9月, 2025 1 次提交
  4. 05 9月, 2025 3 次提交
  5. 03 9月, 2025 1 次提交
  6. 01 9月, 2025 1 次提交
  7. 31 8月, 2025 3 次提交
  8. 29 8月, 2025 2 次提交
  9. 19 8月, 2025 1 次提交
  10. 07 8月, 2025 6 次提交
  11. 01 8月, 2025 2 次提交
  12. 31 7月, 2025 2 次提交
  13. 29 7月, 2025 2 次提交
  14. 28 7月, 2025 1 次提交
  15. 27 7月, 2025 3 次提交
  16. 26 7月, 2025 1 次提交
  17. 25 7月, 2025 2 次提交
    • Luca Citi's avatar
      Use stricter numerical tolerance in rewrites and allow casting in `PatternNodeRewriter` (#1526) · d4e8f736
      Luca Citi 提交于
      * Implemented allow_cast in PatternNodeRewriter
      to allow rewrites that would otherwise fail when the new and old dtype differ.
      Example:
      `np.array(1., "float64") - sigmoid(x)` cannot be rewritten as
      `sigmoid(-x)` (where x is an fmatrix) because the type would change.
      This commit allows an automatic cast to be added so the expression
      is rewritten as `cast(sigmoid(-x), "float64")`.
      Relevant tests added.
      
      * Added test cases for which issue #1497 fails
      
      * Changed PatternNodeRewriter::transform to allow types that do not contain dtype
      like MyType in the tests
      
      * Address #1497 by changing instances of np.isclose to a function isclose, which uses 10 ULPs by default
      
      * Addressed failed tests (with older python/numpy versions)
      
      * Addressed feedback by ricardoV94
      
      * Test PatternNodeRewriter doesn't support multi-output nodes in pattern
      
      But it's fine if they're just root inputs
      
      ---------
      Co-authored-by: 's avatarLuca Citi <lciti@ieee.org>
      Co-authored-by: 's avatarRicardo Vieira <ricardo.vieira1994@gmail.com>
      d4e8f736
    • Ricardo Vieira's avatar
      Make Dot only accept matrix inputs · 0c138495
      Ricardo Vieira 提交于
      0c138495