'ResNetV1' object has no attribute 'classes'

aakova · May 20, 2019, 7:10pm

Trained with ‘sh train_resnet18_v1.sh’ from
https://raw.githubusercontent.com/dmlc/web-data/master/gluoncv/logs/classification/imagenet/resnet18_v1.sh

Training completes but I can’t seem to use the generated model, but it
works with the ‘built-in’ params:

~/scratch/Gluon$ python3 demo_imagenet.py --model resnet18_v1 --input-pic ~/Pictures/sebastian-piton-goralski-gaudi3.jpg
The input picture is classified to be
[church], with probability 0.369.
[bell cote], with probability 0.351.
[castle], with probability 0.123.
[monastery], with probability 0.048.
[tile roof], with probability 0.029.
~/scratch/Gluon$ python3 demo_imagenet.py --model resnet18_v1 --input-pic ~/Pictures/sebastian-piton-goralski-gaudi3.jpg --saved-params params_resnet18_v1/0.2930-imagenet-resnet18_v1-118-best.params
The input picture is classified to be
Traceback (most recent call last):
File “demo_imagenet.py”, line 40, in
(net.classes[ind[i].asscalar()], nd.softmax(pred)[0][ind[i]].asscalar()))
AttributeError: ‘ResNetV1’ object has no attribute ‘classes’
~/scratch/Gluon$ python3 demo_imagenet.py --model resnet18_v1 --input-pic ~/Pictures/sebastian-piton-goralski-gaudi3.jpg --saved-params params_resnet18_v1/imagenet-resnet18_v1-119.params
The input picture is classified to be
Traceback (most recent call last):
File “demo_imagenet.py”, line 40, in
(net.classes[ind[i].asscalar()], nd.softmax(pred)[0][ind[i]].asscalar()))
AttributeError: ‘ResNetV1’ object has no attribute ‘classes’

What am I doing wrong ?

lgo · May 20, 2019, 9:32pm

Hi,

you downloaded the demo_imagenet.py from this page right?
https://gluon-cv.mxnet.io/build/examples_classification/demo_imagenet.html#sphx-glr-build-examples-classification-demo-imagenet-py

It points to an old version of that script apparently, the most recent version looks not to have this bug anymore:
https://raw.githubusercontent.com/dmlc/gluon-cv/25b1717635803712193b28ed54ca4aa2c752f975/docs/tutorials/classification/demo_imagenet.py

hth,
Lieven

aakova · May 21, 2019, 6:37pm

Yes as to the origin of the demo_imagenet.py script.

I assume you meant to point to https://github.com/dmlc/gluon-cv/blob/master/scripts/classification/imagenet/demo_imagenet.py as the docs/tutorials link is a prototype web page. My version of demo_imagenet.py does not differ from this file. I have also verified that the training script train_imagenet.py which I used is identical to the git version.

What is imagenet_labels.txt for, is it silently referred to during training ? None of the scripts appear to refer to it, but would absence of this file during training cause this issue ?

lgo · May 21, 2019, 7:31pm

Strange, I had been looking at this version, which looks like the original commit:

github.com

dmlc/gluon-cv/blob/419afe55b6ab663d1f5e7c2f0b8e06bbe0a3ea70/scripts/classification/imagenet/demo_imagenet.py

from __future__ import division

import argparse, time, logging, random, math

import numpy as np
import mxnet as mx
import matplotlib.pyplot as plt

from mxnet import gluon, nd, image
from mxnet.gluon.data.vision import transforms

from gluonvision.model_zoo import get_model

parser = argparse.ArgumentParser(description='Predict ImageNet classes from a given image')
parser.add_argument('--model', type=str, required=True,
                    help='name of the model to use')
parser.add_argument('--saved-params', type=str, default='',
                    help='path to the saved model parameters')
parser.add_argument('--input-pic', type=str, required=True,
                    help='path to the input picture')

This file has been truncated. show original

Here you see that the classnames are loaded from imagenet_labels.txt and then passed as ‘classes’ argument to the model.

Not sure what happened in the meantime, but that demo script certainly isn’t working anymore.
L.

aakova · May 22, 2019, 12:08am

From that version I just pasted in the class_names initialization and replaced the net.classes reference with class_names, and this lets me get results.

Thank you!

lgo · May 22, 2019, 9:28am

FYI, the issue turns out to be that the Imagenet classnames are only loaded when we are using the pretrained parameters.

I have proposed a simple fix to the demo script here:

L.

Topic		Replies	Views
[model_name] object has no attribute 'output' Gluon	0	296	August 3, 2020
Fine-tuning error "gradient has not been updated by backward since last step" Gluon	1	1433	September 1, 2019
GluonCV dem_ufc101 attribute error list object has no shape	2	659	October 23, 2019
Why FashionMNIST-trained resnet returns array of 1000 values? Gluon code-error , gluon-cv	2	420	February 19, 2019
Error when trying to import a trained net: multiple outputs with name	2	814	December 1, 2018

'ResNetV1' object has no attribute 'classes'

Related Topics