Comments (6)
I have to revise my issue from above a bit.
I found a little issue in my evaluation that messed up my results.
The numbers I get now are closer to the ones reported in the paper, but not exactly the same:
Val: 86.55 (81.146) -> in paper: 86.5 (81.1)
Real: 89.705 (86.649) -> in paper: 89.6 (86.6)
v2: 78.41 (71.14) -> in paper 78.4 (70.9)
Inet-C: 27.286 (53.422) -> in paper 28.2 (54.4)
Inet-A: 76.08 (34.373) -> in paper 75.9 (33.5)
Inet-R: 79.253 (55.08) -> in paper 78.8 (53.7)
Inet-Sketch: 62.794 (42.166) -> in paper 62.5 (41.2)
Inet-Style: 34.578 (13.51) -> no evaluation in paper
I also verified my evaluation with other models. I was able to reproduce the results reported here for the ResNet50 (for ImageNet-Real I don't have a number to compare to).
I don't think it would be beneficial if I share my entire code here, but I'll try to put together a minimal example once I have the time.
from dinov2.
Sure. I am not an author of the paper. So it is fine. I was just trying to see if there was still an open issue.
from dinov2.
Would you be able to share your full code for this evaluation ?
from dinov2.
The results I get for ViT-g/14 (ViT-S/14) are:
Val: 86.55 (81.146) -> in paper: 86.5 (81.1) Real: 89.705 (86.649) -> in paper: 89.6 (86.6) v2: 78.41 (71.14) -> in paper 78.4 (70.9)
The results I get for ViT-g/14 (ViT-S/14) are:
Inet-A: 76.08 (34.373) -> in paper 75.9 (33.5) Inet-R: 79.253 (55.08) -> in paper 78.8 (53.7) Inet-C: 27.286 (53.422) -> in paper 28.2 (54.4) Inet-Sketch: 62.794 (42.166) -> in paper 62.5 (41.2)
So it is fine, the issue can be closed, right?
You get the same results (or better), except for Inet-C
, but you are close.
from dinov2.
I'm still wondering why I can't reproduce the exact values you report in the paper (especially the OOD evaluation).
These are the parameters that were used to generate the tables in the paper, right?
Deviations of this magnitude are a bit weird, so I'm trying to understand what's going on.
from dinov2.
These are the parameters that were used to generate the tables in the paper, right?
Not always, typically the average of multiple evaluations using different seeds for training is reported.
Also maybe is a random process in augmenting data for evaluation. double check.
from dinov2.
Related Issues (20)
- Run vanilla DinoV2 training with unlabelled dataset to fit specific field data HOT 1
- why do Self-supervised image retrieval? HOT 1
- Can you release the weight of dino head and ibot head๏ผ HOT 2
- Resume training from intermediate checkpoint? HOT 1
- Why is total_iters unrelated to batch size and number of data? HOT 2
- ResNet50 architecture HOT 1
- Train depth / segmentation head HOT 1
- Implementation using ytorch lightning
- Drop path implementation
- Do black borders affect model performance? HOT 4
- run code at about [ 3170/125000], [W CudaIPCTypes.cpp:16] Producer process has been terminated before all shared CUDA tensors released. See Note [Sharing CUDA tensors]
- Memory-efficient attention forward operator error in Docker container with FastAPI and DINOv2 HOT 9
- torchrun --nproc_per_node=4 /data/zxb/DINO_FL/train/train.py HOT 2
- Fine-Tuning ViT backbone for Depth Estimation
- Upload 3D Ct Scans into Dinov2
- Patch to support torch version > 2.1 and running using torchrun
- Docker multi-architecture build issues for DINOv2 API on AWS and Mac
- Get visualization of the three first principal components HOT 1
- Loss doesn't go down.
- Is it possible to Use DINOv2 for Facial Recognition/Search HOT 1
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
๐ 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. ๐๐๐
-
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 โค๏ธ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from dinov2.