Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

user-defined training_reuse_old_ratio is mandatory #343

Merged
merged 2 commits into from
Jan 4, 2021
Merged

user-defined training_reuse_old_ratio is mandatory #343

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ad4d07f
user-defined training_reuse_old_ratio is mandatory
Vibsteamer Dec 31, 2020
f80883b
Update run.py
Vibsteamer Dec 31, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 9 additions & 1 deletion dpgen/generator/run.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,13 +197,21 @@ def make_train (iter_index,
training_iter0_model = jdata.get('training_iter0_model_path', [])
training_init_model = jdata.get('training_init_model', False)
training_reuse_iter = jdata.get('training_reuse_iter')
training_reuse_old_ratio = jdata.get('training_reuse_old_ratio', 0.2)
training_reuse_old_ratio = jdata.get('training_reuse_old_ratio', None)
training_reuse_stop_batch = jdata.get('training_reuse_stop_batch', 400000)
training_reuse_start_lr = jdata.get('training_reuse_start_lr', 1e-4)
training_reuse_start_pref_e = jdata.get('training_reuse_start_pref_e', 0.1)
training_reuse_start_pref_f = jdata.get('training_reuse_start_pref_f', 100)
model_devi_activation_func = jdata.get('model_devi_activation_func', None)

if training_reuse_iter is not None and training_reuse_old_ratio is None:
raise RuntimeError("training_reuse_old_ratio not found but is mandatory when using init-model (training_reuse_iter is detected in param).\n" \
"It defines the ratio of the old-data picking probability to the all-data(old-data plus new-data) picking probability in training after training_reuse_iter.\n" \
"Denoting the index of the current iter as N (N >= training_reuse_iter ), old-data refers to those existed before the N-1 iter, and new-data refers to that obtained by the N-1 iter.\n" \
"A recommended strategy is making the old-to-new ratio close to 10 times of the default value, to reasonably increase the sensitivity of the model to the new-data.\n" \
"By default, the picking probability of data from one system or one iter is proportional to the number of batches (the number of frames divided by batch_size) of that systems or iter.\n" \
"Detailed discussion about init-model (in Chinese) please see https://mp.weixin.qq.com/s/qsKMZ0j270YhQKvwXUiFvQ")

if iter_index > 0 and _check_empty_iter(iter_index-1, fp_task_min) :
log_task('prev data is empty, copy prev model')
copy_model(numb_models, iter_index-1, iter_index)
Expand Down