@pri2si17… Alright… I’ll inform you as soon as I upload it on GitHub. Might take a day or two.
@pri2si17 My institute has been under lockdown since this evening. They are not allowing entry to students from outside the campus. Since I’m a day scholar I will not have acces to my systems where I have the codes and results. I don’t think I’ll be able to share the code. I think I can only work on the proposal
Hi Priyanshu, I have made a POC for the multi-class classification of EMNIST. I have trained nine different CNN architectures in Keras by changing the number of filters, fully-connected layers and no.of neurons in fully-connected layers, and converted them to coreML models for emulation on iOS platforms and then applied quantization to get half-precision ( FP-16 ) and lower-precision ( FP-8 ) models giving me a total of 27 different models to test. I have made a swift playground project and an XCode project to implement the models and recorded the inference-time for various simulators and also recorded the RAM usage. Due to time constraints, I have recorded the results of quantized and Non-quantized models of the first architecture only. I have attached the GitHub link to my POC. I would like to know your thoughts and suggestions on this.
Link to my POC:
Link to my CV:
hii @pri2si17 sorry for the late i have made some project related tonthis topic like…ddico _extracion,object_detection, maleria blood_cell detecion(based on image detection )using COCO-API…is this relevent this to given topic or i have to do something more for this toppic relatd…pls comment…!!!
object-dection…
.dicome extracton…
Please edit this to not use shorthand and to be more coherent. We state not to do that in a post that is prominently featured on our forums in a README post which you should read. Fix typos as well. You are supposed to look somewhat professional. Using shorthand makes you look lazy, and that’s not a first impression you want to make here
Additionally, cool it with the exclamation points.
Hey, sorry for the late reply, took a while to get a grasp of the theory.
https://github.com/Aaron-AB/GSOC—LibreHealth
Used a mobilenet to train my data on. Converted the model to a tflite model. Currently trying to implement the tflite model on an android application. After I do that, I’ll look at using different post-training quantization techniques on the model and test it’s performance and device resource use.
Should get back to you within three days.
1 Like
Yeah sure. I will get back to you after reviewing your code. Give me sometime. But I will suggest you to implement it on radiology images, you can get some from kaggle. It will give more clear picture.
Well this is not radiology images, but you can try on it.
Yeah. Its look good for starting point. Definitely look into quantization techniques 
.
Thank you for your response. I will look into the datasets. I have used EMNIST since it was just a POC.
1 Like
Hi @pri2si17. Here’s my code for network pruning:
I have tried pruning on DenseNet 100 (with bottleneck and compression). My objectives were 1.) Pruning unimportant blocks that are being skipped to the subsequent layers. 2 ) Pruning unimportant channels.
I have tested the techniques on the Cifar10 dataset and have tested the network starting from 0% compression to 30% compression with an interval of 5%. In case 2, I found that there was an increase in the prediction accuracy as compared to that of the original DenseNet at a compression rate greater than 5%. The highest accuracy was achieved with a channel wise pruning compression rate of 25%.
The general pattern was that the accuracy reduces by a very minimal amount even on compression rates as high as 20%. I have just used a linear weight to weight the blocks (in case 1) and the channels (in case 2).
The file titled ‘100bc_prune.py’ in the above mentioned GitHub repository is for case 1, whereas, the file titled ‘100bc_prune_bychannel.py’ is for the case 2.
Apart from the code, I have also gone through the literature on network compression techniques like pruning, quantisation and tensor decomposition. I was thinking of compiling them onto a single report for the proposal. Kindly suggest if anything else is required in the code as well as in the proposal.
1 Like
Hi @Ashraf, good. I will look into the code. In the meantime, can you please look at the dataset I commented above in post Project: Low Powered Models for disease detection and classification for Radiology Images and see the difference.
Thanks for the suggestion @pri2si17. I will look into the dataset but I won’t be able to run the code for the time being as I do not have access to good quality systems due to the lockdown of my university due to the coronavirus crisis. It’s going to be this way for about about 8-10 more days. However, I can do more theoretical work as of now and I’m very comfortable with it.
Hey Hi! New to this chat thread. I’ve read the requirements and want to contribute to LibreHealth. Also I have worked previously in Malarial Retinopathy Detection. I have a work history of working in Health Informatics with Center of Development of Advanced Computing, Government of India, employing sophisticated Deep Learning Models to detect bacterial colonization in a water sample. The project is inline with my research and past works. I look forward to contribute to such a novel cause. Please let me know the further formalities regarding proposal / deliverable task evaluations
Updated my github with findings from quantized and non-quantized models. Used dynamic range quantization and compared it to a normal tflite model. The app is very inaccurate and I’m currently trying to fix that. The demo app, inference times and resource use can be found on the github readme.
Ok. I will look into it. Till then work on your proposal.
Hi @pri2si17… I don’t have access to a PC right now. My laptop stopped working and everything is locked down so I can’t even go to an internet café. I don’t think I’ll be able to complete the proposal by the GSOC deadline. However, I’m still interested in this project. Can I still work on this after I get access to a system ??
Not if we select another student, no.