Comments (2)
Good catch! I had Conv2d in mind when I first wrote it.
Looks like we just need to instantiate lora_A and lora_B differently depending on the kind of convolution.
Happy to review and merge it if someone wants to implement and test it. Otherwise, I'll do it in the near future.
from lora.
I have this idea but it will change the way current Conv2d LoRA works. We can treat convolution as a matmul with the input as a flattened "window". For example, for Conv2d, the input is a window with (kernel_size, kernel_size)
spatial dimensions, and the flattened input dim is in_channels * kernel_size * kernel_size
. This will naturally extend to Conv1d and Conv3d
- Conv1d: B @ A = (out_channels, rank) @ (rank, in_channels * kernel_size) = (out_channels, in_channels * kernel_size)
- Conv2d: B @ A = (out_channels, rank) @ (rank, in_channels * kernel_size * kernel_size) = (out_channels, in_channels * kernel_size * kernel_size)
- Conv3d: B @ A = (out_channels, rank) @ (rank, in_channels * kernel_size * kernel_size * kernel_size) = (out_channels, in_channels * kernel_size * kernel_size * kernel_size)
There are two benefits with the above implementation. (1) kernel size doesn't need to be the same in all spatial dimensions, and (2) we can use convolution in the LoRA branch in the forward pass instead of merging weights, similar to Linear implementation (relevant issue - #54). The first convolution (with lora_A) is normal convolution, with the same kernel size, but the second convolution (with lora_B) will be point-wise (aka 1x1) convolution. I haven't tested it but from what I understand, it should work.
The situation becomes slightly complicated when grouped convolution is involved (groups > 1
). I'm thinking of accounting for groups in the input channels of lora_A (so lora_A becomes (rank, in_channels / groups * kernel_size * kernel_size)). We can still implement forward pass of LoRA branch as two convolutions with lora_A and lora_B, where we will use grouped convolution for lora_A, similar to the original convolution branch. A problem might arise when we try to merge weights though. Due to how groped convolution works, I think the merged weights might not be lora_B @ lora_A (I will need to test this). If that's the case, we need to use a different calculation to merge weights.
Another way of using groups > 1
is to follow your current implementation, which puts groups in the output of lora_B (out_channels / groups, rank). However, this would sacrifice the ability to use convolution for forward pass in LoRA branch, but maintains the ability to merge weights with simple lora_B @ lora_A.
Let me know what you think @edwardjhu. Thank you!
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
- Embedding reset_parameters() implement wrong HOT 3
- 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
- Using gradient checkpoint with LoRA
- 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.