Open jamesavery opened 5 years ago
awesome seems to be working! @jamesavery should we do our own 2D convolution in bohrium?
Comparing pytest -k mackey_2d
for bohrium and numpy:
NUMPY:
5.631 predict
4.853 forward
0.141 tanh
1.996 state_map
2.700 input_map
0.053 gradient
0.062 dct
2.439 conv
0.122 pixels
20.603 optimize
0.408 IMED matmul on wout
18.815 pseudo_inverse_svd
17.812 svd
0.737 IMED matmul on inputs and labels
0.052 IMED prepare
0.504 IMED initialize
23.916 forward_states_only
23.792 forward
0.631 tanh
10.249 state_map
12.839 input_map
0.268 gradient
0.302 dct
11.631 conv
0.530 pixels
BOHRIUM
4.772 predict
3.680 forward
0.132 tanh
0.470 state_map
2.415 input_map
0.675 gradient
0.056 dct
1.579 conv
0.082 pixels
23.538 optimize
0.316 IMED matmul on wout
21.874 pseudo_inverse_svd
17.426 svd
0.764 IMED matmul on inputs and labels
0.052 IMED prepare
0.527 IMED initialize
35.976 forward_states_only
35.506 forward
1.309 tanh
3.617 state_map
24.054 input_map
6.702 gradient
0.549 dct
15.795 conv
0.784 pixels
Something funny is causing massive slowdown for Bohrium - need to talk to Mads when he's back from holiday.
TODO:
imed.py
(turnimed_loss
back on intest_run_kuro.py
andtest_run_mackey_2d.py
params.backen = 'bohrium'