Commits · 070c7e964c43f9e5c24f5208ad8e2a1cd41dcd39 · OpenDAS / apex

26 Apr, 2019 2 commits
- Tested on 1x8x1 · 070c7e96
  Michael Carilli authored Apr 26, 2019
  
  070c7e96
- Fixed bounds checking · 3b32c401
  Michael Carilli authored Apr 26, 2019
  
  3b32c401
25 Apr, 2019 1 commit
- let's see · 75139ca3
  Michael Carilli authored Apr 25, 2019
  
  75139ca3
22 Apr, 2019 1 commit
- Updating TensorList->TensorListMetadata · 16a3bdf3
  Michael Carilli authored Apr 22, 2019
  
  16a3bdf3
18 Apr, 2019 1 commit
- cleanup · 651150cb
  Michael Carilli authored Apr 18, 2019
  
  651150cb
11 Mar, 2019 2 commits
- Fix momentum initialization with weight decay · 724672d7
  Simon Layton authored Mar 11, 2019
  
  724672d7
- Fix dispatch, add wd after momentum option · b265b0b5
  Simon Layton authored Mar 11, 2019
```
Fix dispatch where we have a parameter group with multiple
combinations of types
Optionally apply weight decay after momentum
```
  b265b0b5
09 Mar, 2019 1 commit
- Fix momentum in non-nesterov case · ac74f345
  Simon Layton authored Mar 08, 2019
  
  ac74f345
08 Mar, 2019 5 commits
- Simplify noop exit condition · 6d6f0bc2
  Simon Layton authored Mar 08, 2019
  
  6d6f0bc2
- Handle fp16 weights case without forcing fp16 math · a2799893
  Simon Layton authored Mar 08, 2019
```
Incorrect types used in a few places
```
  a2799893
- Simplify C++-side logic · 75c8a97a
  Simon Layton authored Mar 08, 2019
```
Only support the 4 specific cases we care about
Remove more general set of switch statements
```
  75c8a97a
- Code cleanup, add fused fp16 read / write · cac061a1
  Simon Layton authored Mar 08, 2019
```
Fuse in fp16 gradient -> fp32 convert
Additional option fp16 weight copy written out
```
  cac061a1
- Fused multi-tensor SGD · cadad920
  Simon Layton authored Mar 08, 2019
```
Initial implementation, all fp32
Tested against torch.optim.sgd
```
  cadad920