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

Reduce some print statements, convert some to info

Browse Source
This commit is contained in:
Daniel Povey 2022-06-23 10:41:08 +08:00
parent 7f756b2910
commit 09282ca28c
1 changed files with 3 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
egs/librispeech/ASR/pruned_transducer_stateless7/optim.py
View File

@ -418,7 +418,7 @@ class NeutralGradient(Optimizer):
cur_grad_rms_corrected = cur_grad_rms * ((exp_avg_sq/bias_correction2).mean().sqrt() / cur_grad_rms_corrected = cur_grad_rms * ((exp_avg_sq/bias_correction2).mean().sqrt() /
(grad**2).mean().sqrt()) (grad**2).mean().sqrt())
param_rms = (p**2).mean().sqrt() param_rms = (p**2).mean().sqrt()
print(f"cur_grad_rms={cur_grad_rms.item():.3e}, corrected_grad_rms={cur_grad_rms_corrected.item():.3e}, param_rms={param_rms.item():.3e}") logging.info(f"cur_grad_rms={cur_grad_rms.item():.3e}, corrected_grad_rms={cur_grad_rms_corrected.item():.3e}, param_rms={param_rms.item():.3e}")
if random.random() < 0.0005: if random.random() < 0.0005:
# check the cosine angle between cur_grad and grad, to see how different this update # check the cosine angle between cur_grad and grad, to see how different this update
@ -426,7 +426,7 @@ class NeutralGradient(Optimizer):
prod = (grad*cur_grad).mean() prod = (grad*cur_grad).mean()
cos_angle = prod / ((grad**2).mean() * (cur_grad**2).mean()).sqrt() cos_angle = prod / ((grad**2).mean() * (cur_grad**2).mean()).sqrt()
if random.random() < 0.04 or cos_angle < 0.01: if random.random() < 0.04 or cos_angle < 0.01:
print(f"cos_angle = {cos_angle}, shape={grad.shape}") logging.info(f"cos_angle = {cos_angle}, shape={grad.shape}")
alpha = -lr * (1-beta1) alpha = -lr * (1-beta1)
if param_pow != 1.0 or grad_pow != 1.0: if param_pow != 1.0 or grad_pow != 1.0:
@ -450,7 +450,7 @@ class NeutralGradient(Optimizer):
delta.add_(this_delta, alpha=alpha) delta.add_(this_delta, alpha=alpha)
if random.random() < 0.0001: if random.random() < 0.0001:
print(f"Delta rms = {(delta**2).mean().item()}, shape = {delta.shape}") logging.info(f"Delta rms = {(delta**2).mean().item()}, shape = {delta.shape}")
p.add_(delta) p.add_(delta)
state["step"] += 1 state["step"] += 1
@ -622,7 +622,6 @@ class NeutralGradient(Optimizer):
this_scale = (this_var ** (param_pow * 0.5)).reshape(size) this_scale = (this_var ** (param_pow * 0.5)).reshape(size)
proj = state[f"proj_{dim}"] proj = state[f"proj_{dim}"]
#print(f"iter={_}, dim={dim}, this_scale = {this_scale}")
if proj.ndim == 1: if proj.ndim == 1:
proj *= this_scale proj *= this_scale
else: else:
@ -633,7 +632,6 @@ class NeutralGradient(Optimizer):
if param_pow != 1.0: if param_pow != 1.0:
# need to get the overall scale correct, as if we had param_pow == 1.0 # need to get the overall scale correct, as if we had param_pow == 1.0
scale = (params_sq_partnorm.mean() ** 0.5) scale = (params_sq_partnorm.mean() ** 0.5)
print("scale = ", scale)
for dim in range(p.ndim): for dim in range(p.ndim):
size = p.shape[dim] size = p.shape[dim]
if size == 1: if size == 1:
@ -1177,8 +1175,6 @@ class Cain(Optimizer):
this_delta = grad / denom this_delta = grad / denom
alpha = -lr*(1-beta1)*(bias_correction2 ** 0.5) alpha = -lr*(1-beta1)*(bias_correction2 ** 0.5)
delta.add_(this_delta, alpha=alpha) delta.add_(this_delta, alpha=alpha)
if random.random() < 0.0001:
print(f"Delta rms = {(delta**2).mean().item()}, shape = {delta.shape}")
p.add_(delta) p.add_(delta)
if step % 10 == 0: if step % 10 == 0: