mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-19 05:54:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

Saving version I am trying to debug

Browse Source
This commit is contained in:
Daniel Povey 2022-07-24 11:00:40 +08:00
parent 962e95f119
commit ba96439c76
1 changed files with 31 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
egs/librispeech/ASR/pruned_transducer_stateless7/optim.py
View File

@ -419,7 +419,7 @@ param_rms_smooth1: Smoothing proportion for parameter matrix, if assumed rank of
# rate matrices at most every other time we reach here, and
# less frequently than that later in training.
#self._update_param_scales(group, p, state, P_proj)
self._update_param_scales_simple(group, p, state, P_proj)
#self._update_param_scales_simple(group, p, state, P_proj)
# We won't be doing this any more.
#self._diagonalize_grad_cov(group, p, state)
@ -608,6 +608,7 @@ param_rms_smooth1: Smoothing proportion for parameter matrix, if assumed rank of
logging.info(f"step={step}, dim={dim}, size={size}, scale min,max,mean={scale_min,scale_max,scale_mean}")
Q *= this_P_proj.sqrt()
logging.info(f"Q rms = {(Q**2).mean().sqrt()}, abs-rms = {Q.abs().mean()}")
def _update_param_scales(self,
group: dict,
@ -950,13 +951,42 @@ param_rms_smooth1: Smoothing proportion for parameter matrix, if assumed rank of
# this_P_proj shape: (batch_size, num_blocks, block_size)
this_P_proj = _diag(torch.matmul(U_prod, torch.matmul(P_prime, U_prod.transpose(2, 3))))
P_proj[dim] = this_P_proj.clone().reshape(batch_size, size)
simple_update = True
if simple_update:
# normalize the scales in a way that preserves the Frobenius norm of the
# projected parameter deltas
P_rms = this_P_proj.clone()
P_rms = P_rms / _mean(P_rms, exclude_dims=[0], keepdim=True)
P_rms_compare = this_P_proj.clone()
P_rms_compare /= _mean(P_rms_compare, exclude_dims=[0], keepdim=True)
mean_diff = (P_rms - P_rms_compare)
if True:
ratio = mean_diff.abs().sum() / P_rms.abs().sum()
logging.info(f"ratio for division is {ratio}, shapes are {P_rms.shape}, {P_rms_compare.shape}")
if ratio > 1.0e-10:
logging.warn(f"P_rms={P_rms}, P_rms_compare={P_rms_compare}")
scale = P_rms.unsqueeze(-1).sqrt()
Q *= scale
logging.info(f"Q rms = {(Q**2).mean().sqrt()}, abs-rms = {Q.abs().mean()}")
# no iterative stuff, just use sqrt(P_proj) as scale on Q. If this is False, we need to
# call self._update_param_scales(...) from the calling function.
if True:
# debug output
step = state["step"]
scale_min, scale_max, scale_mean = scale.min().item(), scale.max().item(), scale.mean().item()
logging.info(f"step={step}, dim={dim}, size={size}, scale min,max,mean={scale_min,scale_max,scale_mean}")
if True:
# debug output
this_P_proj_unsmoothed = _diag(torch.matmul(U_prod, torch.matmul(P_prime_unsmoothed,
U_prod.transpose(2, 3))))
this_P_proj_unsmoothed = this_P_proj_unsmoothed.clone().reshape(batch_size, size)
skip = 10 if P_proj[dim].shape[-1] > 40 else 1
logging.info(f"dim={dim}, diag of P_proj is: {P_proj[dim][0,::skip]}, diag of unsmoothed P_proj is: {this_P_proj_unsmoothed[0,::skip]}")
# P_proj won't be needed if simple_update == True.
return P_proj