LMDB not found

[elaith9]

I get error that lmdb is missing. How do I create it? Can you please leave instructions how to run this? Thanks

[spMohanty]

@elaith9 : Did you check this script :
https://github.com/spMohanty/PlantVillage-Dataset/blob/master/generate_lmdb.sh

[elaith9]

@spMohanty yes, when I run it I get this: ls: cannot access 'lmdb': No such file or directory

[spMohanty]

Ahh you are right.

Please create an empty folder called "lmdb" in the root of the directory.
I will push a fix soon.
Thanks for reporting.

[elaith9]

@spMohanty so there is something more missing? after creating empty folder and running the script again I get nothing.

[spMohanty]

@elaith9 : Okay, providing support for this might be tricky at the moment for me.

But if you want to replicate the results, I wrote an article about it here : https://www.crowdai.org/articles/plantvillage-image-classification-tutorial

The different versions of the dataset as raw images can be accessed here : https://github.com/spMohanty/PlantVillage-Dataset/tree/master/raw

[elaith9]

@spMohanty the instructions are for alexnet, I wanna do it with googlenet. Anyways, I used raw colored images you have in repository and split them 7:3 for training and validation. The googlenet model I got was not as accurate like described in the research paper. I got map around 87%. I tested it with pictures from the test set I also downloaded, and it works pretty good. But when I test it with images from Internet I get almost every prediction wrong. I'm trying to figure out what is wrong.

I even took some pictures from my cherry and apple tree and predictions were just so wrong. Did you ever tried it with custom data or images from internet?

[spMohanty]

@elaith9 : That is something we reported in the paper as well.
The models we trained also performed with an accuracy of ~55% top-5 for images taken from Bing Search.

[elaith9]

@spMohanty Do you think the problem might be in dataset? Could sth like resnet-50 yield better results?

[spMohanty]

@elaith9 : No the generalisability (to images take from bing) of the trained models doesnt change much even if we try a whole buffet of models.
Images taken in a more real-world context (and not in lab settings) might mitigate that problem.