Comments (8)
Thank you for your feedback.
In our paper we train the pixel-wise model with batch_size=5
and without DataParallel
. Please refer to Section 4.2 of our paper for more details.
from nae4ps.
I am sorry my setting is different from the paper. I will test it again.
from nae4ps.
@DeanChan
Hi, I tried the identical experimental setting as the paper given.
- Train an NAE model
CUDA_VISIBLE_DEVICES=0 python scripts/train_NAE.py --debug --lr_warm_up -p ./logs/ --batch_size 5 --nw 5 --w_RCNN_loss_bbox 10.0 --epochs 22 --lr 0.003
The trained model achieved 91.74%
mAP.
- Train a pixel-wise version initialized with trained NAE weights.
CUDA_VISIBLE_DEVICES=1 python scripts/train_NAE.py --debug --lr_warm_up -p ./logs/ --batch_size 5 --nw 5 --w_RCNN_loss_bbox 10.0 --epochs 11 --lr 0.003 --pixel_wise --NAE_pretrain --embedding_feat_fuse --lr_decay_step 9
But the performance of the model is not so good (mAP should be around 92.1%).
[~] Evaluating detections:
all detection:
recall = 92.32%
ap = 86.27%
[~] Evaluating search:
search ranking:
mAP = 90.26%
top- 1 = 90.48%
top- 5 = 97.07%
top-10 = 97.97%
Did I miss something? Could you give me some suggestions? I will be very grateful.
from nae4ps.
Your training command is correct. The result is wired. Could you provide your environment information? i.e. graphics card, nvidia driver version, cuda version, python package info, etc.?
from nae4ps.
System
Ubuntu16.04
Graphics and Nvidia driver
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.40.04 Driver Version: 418.40.04 CUDA Version: 10.1 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla V100-PCIE... Off | 00000000:18:00.0 Off | 0 |
| N/A 71C P0 207W / 250W | 23208MiB / 32480MiB | 86% Default |
+-------------------------------+----------------------+----------------------+
| 1 Tesla V100-PCIE... Off | 00000000:86:00.0 Off | 0 |
| N/A 45C P0 38W / 250W | 0MiB / 32480MiB | 3% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 8419 C python 23197MiB |
+-----------------------------------------------------------------------------+
CUDA version
❯ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Sat_Aug_25_21:08:01_CDT_2018
Cuda compilation tools, release 10.0, V10.0.130
from nae4ps.
@DeanChan
Hi, I tried so many times, and I still can not get the same performance as the paper said.
Could you give me some suggestions?
from nae4ps.
Hi there~ Many thanks for your feedback!
Several factors could affect the final result such as the nvidia-driver version, cuda verson and GPU model.
The reported result is tested on NVIDIA Tesla V100 16GB with driver version 418.43 and cuda 10.1 on Debian 9.12;
On P40 with the same driver and cuda version I got higher results;
On P40 with driver version 440.82 and cuda 10.2 I got lower results.
The performance variation is usually within (-2, +2). Technically it shouldn't be too much.
I haven't figured out the exact reason for this phenomenon, but maybe you could try changing the random seed and see how the performance changes.
And one more thing, it seems that you are using python 3 instead of python 2, right? Did you make large changes to the code? Which pytorch version are you using?
Thanks.
from nae4ps.
The conda environment shown above is base env. It was my fault, and I have deleted these text. I haven't modified the code.
Thank you, I will test it on correct NVIDIA driver version and cuda version.
from nae4ps.
Related Issues (20)
- Class FastRCNNPredictorBN is not defined. HOT 2
- Only get mAP=42.82% using the trained model in PRW dataset HOT 1
- about Class Weighted Similarity (CWS) HOT 1
- number of identity for CUHK-SYSU HOT 2
- Cuda out of memory. HOT 1
- the problem of label_ids HOT 1
- mAP=90.41% for NAE+ on CUHK. HOT 3
- Reproducing OIM-base results
- train question!! HOT 7
- Target sizes: [3, 896, 1125]. Tensor sizes: [3, 900, 1125]
- About evaluation
- KeyError: 'feat_res4' HOT 1
- need help
- link to pretrained model and datasets HOT 2
- Cannot create conda env HOT 4
- Question about the logic of evaluator.py HOT 2
- TypeError: resnet_backbone() got an unexpected keyword argument 'return_res4' HOT 7
- Can not reproduce the same performance in PRW dataset HOT 11
- Can one GeForce GTX 1080Ti is enough for model training? 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 nae4ps.