Reference implementation of PyTorch ImageNet Inferencing

FuriouslyCurious · August 18, 2017, 10:16pm

This is a seemingly stupid request for all post-post-doc AI researchers on here, but hear me out.

First thing that DL Newbies look for is an image inferencing example. Inferencing is computationally cheap and has a high WOW factor, which makes it the perfect “Hello World” for DL.

We have the reference PyTorch ImageNet training code on Github, but we don’t have a reference ImageNet inferencing example. If someone has an elegant, memory efficient CPU + GPU inferencing code sitting around please submit a PR and add it to the examples repo!

github.com

pytorch/examples/blob/master/imagenet/main.py

import argparse
import os
import shutil
import time

import torch
import torch.nn as nn
import torch.nn.parallel
import torch.backends.cudnn as cudnn
import torch.distributed as dist
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import torchvision.models as models

model_names = sorted(name for name in models.__dict__
    if name.islower() and not name.startswith("__")
    and callable(models.__dict__[name]))

This file has been truncated. show original

Thanks!

rwightman · August 28, 2017, 8:50pm

I have a pretty lean inference script included with a DPN conversion I recently posted at https://github.com/rwightman/pytorch-dpn-pretrained

It utilizes a Dataset that’s fine with images in arbitrary folder hierarchies, from flat to many levels. The script dumps top-5 prediction (by class id) to a csv file.

SpandanMadan · August 29, 2017, 8:43am

By inferencing you mean making a prediction for new sample? Sure, I’ll add it to the tutorial I wrote for fine tuning. The tutorial currently only does training. This is a good idea. Here’s the repo for now -

FuriouslyCurious · September 13, 2017, 8:28pm

Your inferencing example looks epic Mr Wightman! You should submit it to PyTorch Examples repo below. (Perhaps inside “imagenet” folder, since your code / arguments match training script inside the folder?)

By the way, I missed the DPN paper this summer - thanks for the implementation! I have starred it and will check it out more.