ResNets50 fails on "my own image"

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

The pre-trained ResNets50 model in Lab 1 Week 2 has a putative accuracy of 95%, but it does not perform well on my own pictures. For example, this is the output for the image shown

Input image shape: (1, 64, 64, 3) Class prediction vector [p(0), p(1), p(2), p(3), p(4), p(5)] = [[5.5017262e-07 3.0772483e-07 5.2745468e-06 9.9707901e-01 2.8557680e-03 5.9126767e-05]] Class: 3

image
the model also fails with other images. A “one” sign is classified as class 0.
After I trained the model for 12 epochs (reaching accuracy on tests of 97%) the performance on my own images was equally poor.
The elementary logistic model of week 1 does much better. !!>??

2

Yeah same I tried a couple of own images with homogeneous backgrounds, gives terrible results

3

Same, tried a couple of different pictures and gave up on it.

4

Same here. Tried multiple images with all possible rotations of 90 multiples (e.g. 90, 180, 270). All yield poor results.

5

Hi, Same here. Can a mentor please help out with this?

6

Answering my own question. After I progressed to the ResNets50 model (Lab, Week 2), I applied it to a problem of my own (a categorical with 10 classes). The performance was poor, both in training and testing, until I removed a few of the deepest layers in the model, reducing the number of parameters from 25 million to about 10 million. With the smaller (shallower) model, I got to 80% accuracy in training, a bit worse in testing. In short: one of the reasons these models fail is what Prof. Ng defined as “overfitting”. Reducing the number of “neurons”/parameters might be a way to improve performance.