Gensyn Testnet - Deep Learning Poll

JavaScript isn't enabled in your browser, so this file can't be opened. Enable and reload.

Hello everyone! As the Gensyn core team continues to work towards releasing the Testnet, we're interested in finding out more about the deep learning engineers/researchers/hobbyists in our community.

If this is you - please provide detail below!

Some notes:
* If you don't know the answer to any question please say 'don't know'
* If you are receiving this and are not in the community, join here! -> https://discord.gg/cHNFXCv7Wr

What's your name on the Gensyn discord? *

Do you exclusively train deep learning models? *

Deep learning models only (i.e. Neural Networks)

Deep learning and other models (e.g. XGBoost)

Other models only (e.g. XGBoost)

Neither

Do you ever use non gradient based optimisers (e.g. swarm optimisation)? *

Do you train deep learning models for work? *

Roughly, what % of the time you spend training a model is dedicated to grid searching/hyperparameter optimisation? *

What services do you regularly use to train models? (select as many as you like) *

Personal GPUs

Company's GPUs (e.g. the company you work at owns a rack of GPUs)

University/research center GPUs

Western cloud providers (e.g. AWS, GCP, Azure)

Eastern cloud providers (e.g. Alicloud, Tencent Cloud)

Low cost cloud provider (e.g. Vast AI, OVH)

Decentralised web3 protocols

Other:

Required

Roughly, how much do you pay for compute per hour? (e.g. cost of AWS instance) *

Do you do any form of cost optimisation? (e.g. sending models to train on different virtual machines) *

How large, on average, is the training data you work with? (in GB) *

What is the primary framework you use? (E.g. PyTorch) *

What is the main use case you work? (e.g. ‘computer vision for pose estimation’) *

What is the largest problem you face when performing deep learning work? *

Are you under significant time pressure when training models? (e.g. you would pay significantly more to make them train significantly faster) *

Do you have issues accessing the scale of compute you need? (provide some detail if you can) *

Do you consider the training data secret or valuable? *

Do you consider the model architecture to be secret/valuable? *

Are the models you train easily decomposed into smaller training tasks that can be parallelised? *

In general, are you trying to solve new problems or work out more efficient/effective ways to solve existing problems? *

Very generally speaking, what's the usual batch size you train with? (e.g. for MNIST this might be 1 x 712 x 30 - if you were using 30 samples of flattened image data) *

Do you design your own model architectures from scratch, implement published architectures yourself, or use libraries with existing architectures pre-made (e.g. HuggingFace Transformers for NLP)? *

Do you typically train models on single GPUs or distribute over multiple GPUs / multiple computational nodes? *

How long (wall clock) do you typically train your models for? *

Anything else you'd like to add?

Submit

Clear form

Never submit passwords through Google Forms.

This form was created inside of Gensyn. Report Abuse

Forms