i use it as my single point of knowledge for all my experiments results, including model weights, configs, false analysis etc Review collected by and hosted on G2.com.

many specific use cases, which are not that specific imo, i had to implement myself, Review collected by and hosted on G2.com.

W&B is so user-friendly and useful for any ML practitioner but if you are a serious one, you need to get your hands on this tool. Not only you can monitor the performance of your different architecture changes and hyper-parameters, but you can also debug some of the problems with your training. For example, one time I was pulling my hair understanding why my training is so slow, and just by looking at the system dashboard, I realized that CUDA had failed for some reason and I was training on CPU. The system dashboard is also so helpful to find the right batch size to make use of the last MBs of your VRAM, if you know what I mean ;) . All the different plotting options and model/hyper-parameter comparison capabilities, give you a lot of freedom and power to efficiently train machine learning models.

I also appreciate the fact the product is constantly evolving and adapting in flow with the scene of AI. Their blog posts are also a treasure trove of ML knowledge which shows some top-notch serious ML people are working on the product.

All in all, go try it, it is fun and useful! Review collected by and hosted on G2.com.

The UI has a very small delay in updating the progress of your training which you might find annoying if you are an impatient person. Also, I would have loved it if they could add other features like the estimated time to finish the training or even show the time scales of the training steps on the plots (maybe there is a way to activate it but did not find) Review collected by and hosted on G2.com.

Used Weight & Biases for Natural language processing task where i have to train pre-trained models like Bert and RoBerta for classification models. By using Weight & Biases i don't have to manage the weights, loss and accuracy charts. All i have to login and initilize. You can login with your github account. Review collected by and hosted on G2.com.

Honestly nothing, only i think confusion part by viewing different analytics charts get valuable information. Default names mention are also little bit long and confusing. Review collected by and hosted on G2.com.

Very clean and easy to understand UI. Easy integration with Tensorflow, it's nice to see metrics per epoch Review collected by and hosted on G2.com.

Would be nice if there was model deployment functionality. Also, it would be nice to have a service user option or a team API key. Since our runs are triggered using AWS Sagemaker pipelines, we have had to hardcode one of our team member's user API keys which isn't the nicest solution since he isn't always the person triggering the run yet it's still linked to him. Review collected by and hosted on G2.com.

This is the rare product that both engineers and researchers love, and it has been transformative for our team's ability to work on large, complex problems together. In particular, the Reports feature has become our main medium for collaboration, almost more essential than Github. It makes it easy to keep a shared ground truth for baselines while enabling everyone to fork their own versions. You can pull in as much or as little of the other team members’ work to your own current workspace and Reports – the filtering by tag, time, etc. makes this easy. We primarily use tags to make runs available with a quick semantic hook.

The ability to create custom visualizations (via Vega) and filter across many runs during sweeps has been very useful. We’ve made everything from embedding projections (tsne/umap) to sweep overviews here, and then been able to share them for everyone to use. Review collected by and hosted on G2.com.

I was pretty skeptical initially that it would help to have more collaborative visualization tools beyond Tensorboard, etc – and I was completely wrong! I wish I’d realized this sooner, wandb seems to know the current flaws in our workflow better than we do :) Review collected by and hosted on G2.com.

I like that W&B provides an on-premise solution (in our cloud environment) that allows us to manage our data completely internally. Their python library is easy to use and can integrate quickly with our existing workflows for ML research. Specifically, we're able to automatically collect relevant ML data and display appropriate visualizations to help us find the best models. More generally, W&B lets us better keep track of our models and test various experiments easily.

They have a great reporting feature as well, where we can easily create and share reports relating to our ML experiments. The visualizations also flexible, and we can basically create whatever visuals we want (although with some effort)

I also like that they keep adding more features to help us accelerate and manage all of our ML operations easily. To my knowledge, we haven't made use of all of these features yet (at least Artifacts and Tables), but they will definitely help us with our workflows as we grow and mature our teams. Review collected by and hosted on G2.com.

Since we're running on-premise in our cloud, it takes a little effort to maintain the product in our environment. Review collected by and hosted on G2.com.

Wandb allows my team to collaborate and share information. As soon as we became users of the tool I noticed that we would spend time analyzing the training loss graphs for model runs, and asking each other for help. These runs used to be squirrelled away on people's desktop machines, and were nearly impossible to reconstruct old runs. Now we can look at older runs very easily and our team can collaborate on experiment results. The support from the wandb team has been amazing too. Review collected by and hosted on G2.com.

nothing really missing in my opinion. I would like to be able to administer my account a little easier, like seeing how many seats I have left and which users are dormant would allow me to manage my license pool more effectively. Review collected by and hosted on G2.com.

The quality of the features on wandb is always very high. We use metrics, artifact management, and hyperparameter sweep extensively and find that wanbd fits seamlessly with our training. Review collected by and hosted on G2.com.

wandb is developing their feature set but doesn't yet have a complete solution to all of ML Ops. That means I still have to find other tools to fill the gaps. Usually wandb integrates well with these tools but the integration always requires work. Review collected by and hosted on G2.com.

W&B is an excellent tool, particularly for collaborating on and maintaining a record of machine learning experiments. I think its role is in closing the gap between training and analysis, which it does very well. Review collected by and hosted on G2.com.

* Reports performance can be slow at times for a large number of displayed runs (e.g. 300-600)

* Tools such as Sweeps don't allow for an alternative backend, and available frontend tools are somewhat clunky without specific customization towards the end use (e.g. using reports for analyzing tune results). As it exists now, I think W&B offers more to ML teams that don't have a supporting SW infrastructure team. Review collected by and hosted on G2.com.

Weights and Biases makes our entire workflow easier. It biases new (and sometimes seasoned!) engineers toward better best practices, makes it easier to introspect, improve, store, and serve models - it makes the entire process better. Can't recommend highly enough! Review collected by and hosted on G2.com.

Really can't think of anything. Wish they'd ship even more new features faster; the ones they've added recently are spectacular! Review collected by and hosted on G2.com.