Comments (7)
@AS-researcher6 merged to master. should be correct now. pls verify
from lightning.
All good. Thanks for the fixes!
Working on a SLURM cluster myself so may submit pull requests for expanded Trainer functionality and bring up more things in the future.
from lightning.
@AS-researcher6 good find, i see the bug. the division is done once the step is applied, but the division should be on line 926.
Current order:
- clip
- step
- zero_grad
- divide accumulated loss by nb accumulated batches
Correct order:
- divide accumulated loss by nb accumulated batches
- clip
- step
- zero_grad
Submitting the change.
Mind sanity checking PR #88
from lightning.
Gladly! Not sure if you're done fixing things but I don't see the commit where the loss is divided by self.accumulate_grad_batches before the loss.backward(). Otherwise the original version where self.batch_loss_value += loss.item() is good as long as the self.batch_loss_value is not averaged after.
from lightning.
look at the PR #88
line 928
from lightning.
Sorry, I don't think it's quite right yet. Accounting for multiple batches needs to happen in the actual loss itself before backpropagation. Otherwise, its like N accumulated additions of loss.backward() rather than N accumulated additions of (1/N * loss).backward()
After line 898 and before loss.backward() on the 903 if statement, there needs to be a line that's loss = loss / self.accumulate_grad_batches. Line 922 should still be self.batch_loss_value += loss.item(). Line 928 can be removed entirely, since the averaging has already been accounted for.
from lightning.
@AS-researcher6 i see. updated. Was following the wrong approach before.
How about now?
from lightning.
Related Issues (20)
- Issue in Manual optimisation, during self.manual_backward call HOT 1
- Existing metric keys not moved to device after LearningRateFinder
- Checkpoint every_n_steps reruns epoch on restore HOT 3
- Metrics logged by self.log and metric.compute() are different HOT 1
- Multi-node Training with DDP stuck at "Initialize distributed..." on SLURM cluster HOT 3
- Full validation after first microbatch when training after LearningRateFinder
- Add a warning when some of the modules are in eval mode before the training stage
- why pytorch-lightning doc say "Model-parallel training (FSDP and DeepSpeed)". I think there is something wrong. HOT 2
- AWS Trainium fails number of device validation when using more than 1 accelerator on the instances
- OnExceptionCheckpoint: training resumes if ckpt found, even if no ckpt_path provided
- TensorBoardLogger has the wrong epoch numbers much more than the fact HOT 1
- How to incorporate vLLM in Lightning for LLM inference?
- WandbLogger `save_dir` and `dir` parameters do not work as expected.
- Loading large models with fabric, FSDP and empty_init=True does not work
- Unable to extract confusion matrix as a metric from trainer HOT 1
- Torchmetrics Accuracy issue when dont shuffle test data. HOT 1
- ModelCheckpoint: Using save_top_k, only the first k models are stored, not the best k models HOT 2
- trainer.fit from checkpoint without performance improvement will break 'last' link to checkpoint on window11
- Exception in RecordFunction callback: state_ptr INTERNAL ASSERT FAILED at "../torch/csrc/profiler/standalone/nvtx_observer.cpp":115
- `ckpt_path` in `Trainer` accepts URIs to automatically load checkpoints from remote paths
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from lightning.