wassname/TTS
1
0
Fork 0
mirror of https://github.com/wassname/TTS.git synced 2026-09-12 12:11:47 +08:00
Code Issues Packages Projects Releases Wiki Activity

Decoder shape comments for Tacotron2, decoupled grad clip for stopnet and the rest of the network. Some variable renaming and bug fix for alignment score logging

Browse Source
This commit is contained in:
Eren Golge
2019-11-12 11:20:53 +01:00
parent 2966e3f2d1
commit 015f7780f4
4 changed files with 56 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files
utils/generic_utils.py
+5 -2
View File
@@ -150,10 +150,13 @@ def save_best_model(model, optimizer, model_loss, best_loss, out_path,
return best_loss
def check_update(model, grad_clip):
def check_update(model, grad_clip, ignore_stopnet=False):
r'''Check model gradient against unexpected jumps and failures'''
skip_flag = False
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
if ignore_stopnet:
grad_norm = torch.nn.utils.clip_grad_norm_([param for name, param in model.named_parameters() if 'stopnet' not in name], grad_clip)
else:
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
if np.isinf(grad_norm):
print(" | > Gradient is INF !!")
skip_flag = True
utils/logger.py
+20 -15
View File
@@ -11,22 +11,27 @@ class Logger(object):
def tb_model_weights(self, model, step):
layer_num = 1
for name, param in model.named_parameters():
self.writer.add_scalar(
"layer{}-{}/max".format(layer_num, name),
if param.numel() == 1:
self.writer.add_scalar(
"layer{}-{}/value".format(layer_num, name),
param.max(), step)
self.writer.add_scalar(
"layer{}-{}/min".format(layer_num, name),
param.min(), step)
self.writer.add_scalar(
"layer{}-{}/mean".format(layer_num, name),
param.mean(), step)
self.writer.add_scalar(
"layer{}-{}/std".format(layer_num, name),
param.std(), step)
self.writer.add_histogram(
"layer{}-{}/param".format(layer_num, name), param, step)
self.writer.add_histogram(
"layer{}-{}/grad".format(layer_num, name), param.grad, step)
else:
self.writer.add_scalar(
"layer{}-{}/max".format(layer_num, name),
param.max(), step)
self.writer.add_scalar(
"layer{}-{}/min".format(layer_num, name),
param.min(), step)
self.writer.add_scalar(
"layer{}-{}/mean".format(layer_num, name),
param.mean(), step)
self.writer.add_scalar(
"layer{}-{}/std".format(layer_num, name),
param.std(), step)
self.writer.add_histogram(
"layer{}-{}/param".format(layer_num, name), param, step)
self.writer.add_histogram(
"layer{}-{}/grad".format(layer_num, name), param.grad, step)
layer_num += 1
def dict_to_tb_scalar(self, scope_name, stats, step):