Comments (3)
Yes. It's important that we have either A or B set to zero so the first forward pass doesn't deviate from the original model. It's also important that A and B are not both zeros because the gradient would be forever zero in that case. So either set A to zero and B something else or vice versa.
Setting A to zeros we're essentially discarding all the knowledge encoded in the pretrained weights
Could you explain why?
If we set B to zeros and initialize A with Kaimin uniform distribution, or normal distribution, we might still be stuck in local minimum if A is filled with zeros?
We will get stuck if both A and B are zeros, but not if one of them is not all zeros.
from lora.
I have the same question, however, see Edward's comment here
To my understanding, Edward resets the parameters of the weight matrix the same way as pytorch's Embedding.reset_parameters()
# the code taken from torch.nn.modules.sparse.Embedding
def reset_parameters(self) -> None:
init.normal_(self.weight)
self._fill_padding_idx_with_zero()
According to Edward's comment, it doesn't really matter which adaption matrix is set to 0, if we set both of the matrices to zeros.
However, I have a couple questions about resetting parameters in Linear
and Embedding
@edwardjhu:
-
Setting A to zeros we're essentially discarding all the knowledge encoded in the pretrained weights, doesn't it defeat the purpose of leveraging transfer learning?
-
If we set B to zeros and initialize A with Kaimin uniform distribution, or normal distribution, we might still be stuck in local minimum if A is filled with zeros? Did those initialized values come from experimental results?
from lora.
Yes. It's important that we have either A or B set to zero so the first forward pass doesn't deviate from the original model. It's also important that A and B are not both zeros because the gradient would be forever zero in that case. So either set A to zero and B something else or vice versa.
Setting A to zeros we're essentially discarding all the knowledge encoded in the pretrained weights
Could you explain why?
If we set B to zeros and initialize A with Kaimin uniform distribution, or normal distribution, we might still be stuck in local minimum if A is filled with zeros?
We will get stuck if both A and B are zeros, but not if one of them is not all zeros.
Thanks for your quick response. It sounds to me that setting one of them to be zeros is critical to "preserve" the pretrained weights for the first pass. However, if A is initialized with zeros, the results of B @ A would be zeros, and the output from the previous layer multiples the B @ A would be zeros, too? Or did I misunderstand the implementation? (I admit I haven't run the code so I didn't check whether it's the case at runtime)
from lora.
Related Issues (20)
- Code samples for "UNDERSTANDING THE LOW-RANK UPDATES" chapter (chapter 7). HOT 2
- why use alpha/r in stead of alpha? HOT 2
- Error in MergedLinear HOT 5
- RuntimeError: CUDA error: CUBLAS_STATUS_INVALID_VALUE when calling `cublasSgemm( handle, opa, opb, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc)` HOT 4
- Pre-trained conv weight is not same as that self.conv.weight HOT 3
- matmul ordering in MergedLinear HOT 1
- How does this paper search the hyperparameters on GLUE datasets with Roberta? HOT 1
- PyPi version and GitHub version give different results HOT 1
- Conv1d and Conv3d are not working HOT 4
- fine tuning RoBERTa-base with LoRA (ValueError: Classification metrics can't handle a mix of binary and multilabel-indicator targets)
- The content on pypi does not seem to be updated HOT 3
- Marvelous! Great work! HOT 1
- Bug after the latest commit #63 HOT 1
- MergedLinear bug? HOT 2
- Is it possible to fine-tuning model to extend token limit length? HOT 3
- Replicating Result on WebNLG HOT 1
- About GPU utilization HOT 1
- there's some bug in layer.py
- AB matrix initialization in layers.py does not conform to the description of the paper 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 lora.