Comments (1)
Secondary confirmation.
P100, Driver Version: 375.26, cuDNN: 5, CUDA: 8.0
-----------------------------------------------------------------------------------------
| end of epoch 6 | time: 18.46s | valid loss 4.89 | valid ppl 133.20
-----------------------------------------------------------------------------------------
=========================================================================================
| End of training | test loss 4.86 | test ppl 128.70
=========================================================================================
Turning off gradient clipping:
-----------------------------------------------------------------------------------------
| end of epoch 6 | time: 20.51s | valid loss 4.77 | valid ppl 118.37
-----------------------------------------------------------------------------------------
=========================================================================================
| End of training | test loss 4.74 | test ppl 114.18
=========================================================================================
The issue appears caused by the clip_gradient
function - though I've not yet found the exact source of the issue in that function. Replacing it with the recently added clip_grad_norm
(source) brings us back towards the reported perplexity of 113 while using gradient clipping.
-----------------------------------------------------------------------------------------
| end of epoch 6 | time: 16.87s | valid loss 4.78 | valid ppl 119.09
-----------------------------------------------------------------------------------------
=========================================================================================
| End of training | test loss 4.75 | test ppl 115.49
=========================================================================================
With a different random seed (42), the perplexity is more in line with the reported 113, so the difference is likely superficial.
-----------------------------------------------------------------------------------------
| end of epoch 6 | time: 17.03s | valid loss 4.77 | valid ppl 118.16
-----------------------------------------------------------------------------------------
=========================================================================================
| End of training | test loss 4.74 | test ppl 114.46
=========================================================================================
from examples.
Related Issues (20)
- word_language_model/data.py - remove '<eos>'
- If I am training on a SINGLE GPU, should this "--dist-backend 'gloo'" argument be added to the command? HOT 10
- SSL Error When downloading dataset HOT 3
- Testing a C++ case with MPI failed.
- Long training time for ResNet50 on ImageNet-1k HOT 1
- Segmentation fault (core dumped) at `model(images)` for examples/imagenet/main.py HOT 1
- RuntimeError in Partialconv-master HOT 1
- Pytorch is insufficiently opinionated
- Documentation Mismatch and AssertionError in language_translation
- RuntimeError: HIP error when running ResNet-50 on PRO W7900 with PyTorch HOT 1
- Drawbacks of making the C++ API look like Python HOT 10
- Build error on cpp/custom-dataset
- multi-node Tensor Parallel
- `local_rank` or `rank` for multi-node FSDP
- reference of weight initialization for llama2 model HOT 1
- [DOC] Update mnist.py example HOT 4
- SequenceParallel sharding seems wrong HOT 1
- word_language_model, is it a Transformer, Encoder-only or Decoder only? HOT 1
- Larger image size for DCGAN code with Celeba dataset
- Multinode.py example fails HOT 2
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 examples.