FlorDB: Nimble Experiment Management for Iterative ML
Flor (for "fast low-overhead recovery") is a record-replay system for deep learning, and other forms of machine learning that train models on GPUs. Flor was developed to speed-up hindsight logging: a cyclic-debugging practice that involves adding logging statements after encountering a surprise, and efficiently re-training with more logging. Flor takes low-overhead checkpoints during training, or the record phase, and uses those checkpoints for replay speedups based on memoization and parallelism.
FlorDB integrates Flor, git and sqlite3 to manage model developer's logs, execution data, versions of code, and training checkpoints. In addition to serving as an experiment management solution for ML Engineers, FlorDB extends hindsight logging across model trainging versions for the retroactive evaluation of iterative ML.
Flor and FlorDB are software developed at UC Berkeley's RISE Lab.
Installation
pip install flordbDemo
First run
Run the examples/rnn.py script to test your installation.
This script will train a small linear model on MNIST.
FLOR shadow branches permit us to commit your work
automatically on every run, without interfering with your
other commits. You can later review and merge the flor shadow branch as you would any other git branch.
git checkout -b flor.shadow
python examples/rnn.py --flor readmeWhen finished, you will have committed to the shadow branch and written execution metadata into a .flor directory in your current directory. Additionally, flor created a directory tree in your HOME to organize your experiments. You can find our running experiment as follows:
ls ~/.flor/Confirm that Flor saved checkpoints of the examples/rnn.py execution on your home directory.
Flor will access and interpret contents of ~/.flor automatically.
You should routinely clear this stash or spool it to the cloud to clear up disk space.
View your experiment history
From the same directory you ran the examples above, open an iPython terminal, then load and pivot the log records.
In [1]: from flor import full_pivot, log_records
In [2]: full_pivot(log_records())
Out[2]:
runid tstamp epoch step device learning_rate loss
0 readme 2023-03-12 12:23:53 1 100 cpu 0.01 0.5304957032203674
1 readme 2023-03-12 12:23:53 1 200 cpu 0.01 0.21829535067081451
2 readme 2023-03-12 12:23:53 1 300 cpu 0.01 0.15856705605983734
3 readme 2023-03-12 12:23:53 1 400 cpu 0.01 0.11441942304372787
4 readme 2023-03-12 12:23:53 1 500 cpu 0.01 0.06835074722766876
5 readme 2023-03-12 12:23:53 1 600 cpu 0.01 0.13750575482845306
6 readme 2023-03-12 12:23:53 2 100 cpu 0.01 0.11708579957485199
7 readme 2023-03-12 12:23:53 2 200 cpu 0.01 0.08852845430374146
8 readme 2023-03-12 12:23:53 2 300 cpu 0.01 0.16527307033538818
9 readme 2023-03-12 12:23:53 2 400 cpu 0.01 0.11036019027233124
10 readme 2023-03-12 12:23:53 2 500 cpu 0.01 0.05740281194448471
11 readme 2023-03-12 12:23:53 2 600 cpu 0.01 0.07785198092460632
Model Training Kit (MTK)
The MTK includes utilities for serializing and checkpointing PyTorch state,
and utilities for resuming, auto-parallelizing, and memoizing executions from checkpoint.
The model developer passes objects for checkpointing to flor,
and gives it control over loop iterators by calling MTK.checkpoints
and MTK.loop as follows:
from flor import MTK as Flor
import torch
trainloader: torch.utils.data.DataLoader
testloader: torch.utils.data.DataLoader
optimizer: torch.optim.Optimizer
net: torch.nn.Module
criterion: torch.nn._Loss
Flor.checkpoints(net, optimizer)
for epoch in Flor.loop(range(...)):
for data in Flor.loop(trainloader):
inputs, labels = data
optimizer.zero_grad()
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
eval(net, testloader)As shown, we pass the neural network and optimizer to Flor
for checkpointing with Flor.checkpoints(net, optimizer).
We wrap both the nested training loop and main loop with
Flor.loop. This lets Flor jump to an arbitrary epoch
using checkpointed state,
and skip the nested training loop when intermediate
state isn't probed.
Hindsight Logging
from flor import MTK as Flor
import torch
trainloader: torch.utils.data.DataLoader
testloader: torch.utils.data.DataLoader
optimizer: torch.optim.Optimizer
net: torch.nn.Module
criterion: torch.nn._Loss
for epoch in Flor.loop(range(...)):
for batch in Flor.loop(trainloader):
...
eval(net, testloader)
log_confusion_matrix(net, testloader)Suppose you want to view a confusion matrix as it changes throughout training. Add the code to generate the confusion matrix, as sugared above.
python3 mytrain.py --replay_flor PID/NGPUS [your_flags]As before, you tell FLOR to run in replay mode by setting --replay_flor.
You'll also tell FLOR how many GPUs from the pool to use for parallelism,
and you'll dispatch this script simultaneously, varying the pid:<int>
to span all the GPUs. To run segment 3 out of 5 segments, you would write: --replay_flor 3/5.
If instead of replaying all of training you wish to re-execute only a fraction of the epochs
you can do this by setting the value of ngpus and pid respectively.
Suppose you want to run the tenth epoch of a training job that ran for 200 epochs. You would set
pid:9and ngpus:200.
Publications
To cite this work, please refer to the Hindsight Logging paper (VLDB '21).
FLOR is open source software developed at UC Berkeley. Joe Hellerstein (databases), Joey Gonzalez (machine learning), and Koushik Sen (programming languages) are the primary faculty members leading this work.
This work is released as part of Rolando Garcia's doctoral dissertation at UC Berkeley, and has been the subject of study by Eric Liu and Anusha Dandamudi, both of whom completed their master's theses on FLOR. Our list of publications are reproduced below. Finally, we thank Vikram Sreekanti, Dan Crankshaw, and Neeraja Yadwadkar for guidance, comments, and advice. Bobby Yan was instrumental in the development of FLOR and its corresponding experimental evaluation.
- Hindsight Logging for Model Training. R Garcia, E Liu, V Sreekanti, B Yan, A Dandamudi, JE Gonzalez, JM Hellerstein, K Sen. The VLDB Journal, 2021.
- Fast Low-Overhead Logging Extending Time. A Dandamudi. EECS Department, UC Berkeley Technical Report, 2021.
- Low Overhead Materialization with FLOR. E Liu. EECS Department, UC Berkeley Technical Report, 2020.
License
FLOR is licensed under the Apache v2 License.