record-keeper
Installation
pip install record-keeper
The Problem:
When running machine-learning experiments, having more logged data is usually better than less. But adding new series of data to log can often require changes to your training code. When you want to log dozens of different series of data, your code starts to look awful.
The Solution:
Use RecordKeeper, and easily add loggable information when you write a new class. The example below is taken from the pytorch-metric-learning library.
First, create a list that contains the names of the attributes you want to record (self.record_these in the example below).
class ContrastiveLoss(GenericPairLoss):
def __init__(self, pos_margin=0, neg_margin=1, use_similarity=False, power=1, avg_non_zero_only=True, **kwargs):
...
self.num_non_zero_pos_pairs = 0
self.num_non_zero_neg_pairs = 0
self.record_these = ["num_non_zero_pos_pairs", "num_non_zero_neg_pairs"]
...
...
Then tell RecordKeeper the name of the list to read. RecordKeeper will log and save all the attributes described in the list.
from torch.utils.tensorboard import SummaryWriter
import record_keeper as record_keeper_package
from pytorch_metric_learning import losses
record_writer = record_keeper_package.RecordWriter(your_folder_for_logs)
tensorboard_writer = SummaryWriter(log_dir=your_tensorboard_folder)
record_keeper = record_keeper_package.RecordKeeper(tensorboard_writer, record_writer, ["record_these"])
your_loss_dictionary = {"metric_loss": losses.ContrastiveLoss()}
# Then at each iteration of training:
record_keeper.update_records(your_loss_dictionary, current_iteration)
Now the attributes described in record_these, (specifically, num_non_zero_pos_pairs and num_non_zero_neg_pairs) can be viewed on Tensorboard.
These data series are also saved in sqlite and CSV format. If you only want to use Tensorboard, then pass in only a SummaryWriter, and vice versa.
The dictionary that you pass into record_keeper.update_records can contain any number of objects, and for each one, RecordKeeper will check if the object has a "record_these" attribute. As long as you're making your dictionaries programmatically, it's possible to add large amounts of loggable data without clogging up your training code. See pytorch-metric-learning and powerful-benchmarker to see RecordKeeper in action.
1 maintainer
Release files for record-keeper 0.9.26
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| record-keeper-0.9.26.tar.gz | 6.7 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| record_keeper-0.9.26-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 14.5 kB
Release files / record-keeper-0.9.26.tar.gz
| Download URL | record-keeper-0.9.26.tar.gz |
|---|---|
| Size | 6.7 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
29951e85b8b82239b2cb2102cd97919013861ea9c21eda53d2c41f091a2c655e
|
|
BLAKE2b-256 checksum How to use checksums |
8a89cf8e2034b15fadf0e3c8def03348a732e630ec99a40c914a072b905efd00
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/3.1.1 pkginfo/1.5.0.1 requests/2.22.0 setuptools/41.6.0.post20191030 requests-toolbelt/0.9.1 tqdm/4.38.0 CPython/3.7.5
|
Release files / record_keeper-0.9.26-py3-none-any.whl
| Download URL | record_keeper-0.9.26-py3-none-any.whl |
|---|---|
| Size | 7.8 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
78c411799fdb463fa11688235140befbb7ae88be5399fa793a6ba6b097670d61
|
|
BLAKE2b-256 checksum How to use checksums |
8d560e646aad413c5b41937104dc07ec28212dd6200510df4911ad4a03936902
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/3.1.1 pkginfo/1.5.0.1 requests/2.22.0 setuptools/41.6.0.post20191030 requests-toolbelt/0.9.1 tqdm/4.38.0 CPython/3.7.5
|
