pytorch/examples

Low training accuracy using pre-trained model

Open

#889 opened on Mar 4, 2021

View on GitHub
 (2 comments) (0 reactions) (0 assignees)Python (21,634 stars) (9,429 forks)batch import
help wantedvision

Description

Hello, I am trying to evaluate a pre-trained mobilenetv2 model from torchvision on the ImageNet training dataset using this script. To do so, I modify lines 235-237 to perform validation on the train loader instead of the val loader:

    if args.evaluate:
        validate(train_loader, model, criterion, args)
        return

Everything else is left untouched. The command I use to run is: python imagenet_train_example.py -a mobilenet_v2 -j 16 -b 1024 -e --pretrained /data/ImageNet However, the results are lower than expected: Acc@1 2.926 Acc@5 15.079 Loss 11.795791

Contributor guide

Tech stack
python
Domain
machine learning
Issue type
bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
under 1 hour
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
needs maintainer response
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
PythonPyTorchImageNet dataset
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30
Research direction
The issue is in imagenet train example.py, lines 235-237, where the user modified the script to evaluate on the training set instead of the validation set. The low accuracy suggests a mismatch between training preprocessing (including data augmentation) and the expected input for the pre trained model. Investigate the data transforms for the training loader (which typically include random crops and horizontal flips) versus the validation loader (center crop only). The fix likely involves using the validation loader for evaluation, or applying validation style transforms if evaluating on the training set. No linked PRs or comments provide further direction.