If I train with standard_p_100 GPU's and am running some batch predict jobs with the trained models, is there a way for me to specify or request that the batch predictions be performed on GPU's? For comparison, the batch prediction for an epoch of training data seems to be taking 8-10x than it would during training time which leads me to suspect that it is not taking advantage of GPUs and i'm wondering if I have any control over speeding this up?
Related
Is it possible to do a hyper-parameter optimization on Sagemaker for a flow (e.g., pipeline) of a processing job followed by a training job?
In Sagemaker pipelines, I see I can use the tuner step with any training step. However, I can't see any helpful resource for integrating the processing job into the optimization.
Any ideas on how to do this task without merging two steps into one step?
In this relatively old question, it was asking about optimizing 2 models jointly. Here, I am asking about Processing and Training jobs.
I really appreciate any help you can provide.
There is no way to run a HyperparameterTuner for Processor, it expects an estimator as an input. On the other hand, you can just put your processing logic with hyperparameters into an appropriate Estimator (sklearn for example) and then output processing script with hyperparameters as an artefact from /opt/ml/model/model.joblib into a S3 or as a model artefact. When done tuning, just load it into Processor via model_dir, and you got your Processor with tuned hyperparameters.
I am building a classification model using AutoML and I have some basic usage questions about the GCP.
1 - Data privacy question; if we save behavior data to train our model in BigQuery, does Google have access to that data? Could Google ever use that data to learn more about behavior of individuals we collected data from?
2 - Since training costs are charged by the hour, I would like to understand the relationship between data and training time. Does the time increase linearly with the size of the training data set? For example, we trained a classification using 1.7MB of data and it took 3 hrs. So, would training a model with 17MB of data take 30 hours?
3 - A batch prediction costs 1.16 USD per hour. However, our data is in a csv and it seems that we cannot upload a csv to do a batch prediction. So, we will try using the API. Therefore I have two questions: A) can we do a batch upload using the API and B) what are the associated costs?
4 - What exactly is an online prediction?
5 - When using the cost calculator (for machine learning), what is a node hour?
1- As is mentioned in the Data Usage FAQ, Google does not use any of your content for any purpose except to provide you with the Cloud AutoML service.
2- The time required to train your model depends on the size and complexity of your training data, for detailed explanation take a look at the Vision documentation for example.
3- You need to upload your csv file to Google Cloud Storage and then you can use it in the API or any of the available client libraries. See Natural Language batch prediction, for example. For costs, check the documentation for the desired product. AutoML pricing depends on what feature you are using: Vision, Natural Language, Translation, Video Intelligence.
4- After you have created (trained) a model, you can deploy the model and request online (single, low-latency and real-time) predictions. Online predictions accept one row of data and provide a predicted result based on your model for that data. You use online predictions when you need a prediction as input for your business logic flow.
5- You can think of node as a single Virtual Machine, which resources are used for computing purposes. Machine types are different depending the product and purpose for which they are used. For example in image classification, the cost for AutoML Vision Image Classification model training is $3.15 per node hour, each node is equivalent to a n1-standard-8 machine with an attached NVIDIA Tesla V100.GPU. Then, node hour are the resources of such node used by one hour.
I'm trying to design how my system uses Google ML for predictions.
I have a largish set of images I need to use for prediction. I can either:
send one big dataset to Google ML, and have the model break the data up into smaller iterative datasets, or
break the dataset up before-hand and send multiple Google ML prediction requests with the smaller datasets. This will ensure parallelism, but I'll end up sending about 3X the data.
If I allocate more resources for Google ML, will Google ML automatically parallelize the iterative datasets if I do option 1? Or am I better off handling this myself?
Thanks in advance!
I have a csv file of 500GB and a mysql database of 1.5 TB of data and I want to run aws sagemaker classification and regression algorithm and random forest on it.
Can aws sagemaker support it? can model be read and trained in batches or chunks? any example for it
Amazon SageMaker is designed for such scales and it is possible to use it to train on very large datasets. To take advantage of the scalability of the service you should consider a few modifications to your current practices, mainly around distributed training.
If you want to use distributed training to allow much faster training (“100 hours of a single instance cost exactly the same as 1 hour of 100 instances, just 100 times faster”), more scalable (“if you have 10 times more data, you just add 10 times more instances and everything just works”) and more reliable, as each instance is only handling a small part of the datasets or the model, and doesn’t go out of disk or memory space.
It is not obvious how to implement the ML algorithm in a distributed way that is still efficient and accurate. Amazon SageMaker has modern implementations of classic ML algorithms such as Linear Learner, K-means, PCA, XGBoost etc. that are supporting distributed training, that can scale to such dataset sizes. From some benchmarking these implementations can be 10 times faster compared to other distributed training implementations such as Spark MLLib. You can see some examples in this notebook: https://github.com/awslabs/amazon-sagemaker-workshop/blob/master/notebooks/video-game-sales-xgboost.ipynb
The other aspect of the scale is the data file(s). The data shouldn’t be in a single file as it limits the ability to distribute the data across the cluster that you are using for your distributed training. With SageMaker you can decide how to use the data files from Amazon S3. It can be in a fully replicated mode, where all the data is copied to all the workers, but it can also be sharded by key, that distributed the data across the workers, and can speed up the training even further. You can see some examples in this notebook: https://github.com/awslabs/amazon-sagemaker-examples/tree/master/advanced_functionality/data_distribution_types
Amazon Sagemaker is built to help you scale your training activities. With large datasets, you might consider two main aspects:
The way data are stored and accessed,
The actual training parallelism.
Data storage: S3 is the most cost-effective way to store your data for training. To get faster startup and training times, you can consider the followings:
If your data is are already stored on Amazon S3, you might want first to consider leveraging the Pipe mode with built-in algorithms or bringing your own. But Pipe mode is not suitable all the time, for example, if your algorithm needs to backtrack or skip ahead within an epoch (the underlying FIFO cannot support lseek() operations) or if it is not easy to parse your training dataset from a streaming source.
In those cases, you may want to leverage Amazon FSx for Lustre and Amazon EFS file systems. If your training data is already in an Amazon EFS, I recommend using it as a data source; otherwise, choose Amazon FSx for Lustre.
Training Parallelism: With large datasets, it is likely you'll want to train on different GPUs. In that case, consider the followings:
If your training is already Horovod ready, you can do it with Amazon SageMaker (notebook).
In December, AWS has released managed data parallelism, which simplifies parallel training over multiple GPUs. As of today, it is available for TensorFlow and PyTorch.
(bonus) Cost Optimisation: Do not forget to leverage Managed Spot training to save up to 90% of the compute costs.
You will find other examples on the Amazon SageMaker Distributed Training documentation page
You can use SageMaker for large scale Machine Learning tasks! It's designed for that. I developed this open source project https://github.com/Kenza-AI/sagify (sagify), it's a CLI tool that can help you train and deploy your Machine Learning/Deep Learning models on SageMaker in a very easy way. I managed to train and deploy all of my ML models whatever library I was using (Keras, Tensorflow, scikit-learn, LightFM, etc)
I am using Google Cloud ML to for training jobs. I observe a peculiar behavior in which I observe varying time taken for the training job to complete for the same data. I analyzed the CPU and Memory utilization in the cloud ML console and see very similar utilization in both the cases(7min and 14mins).
Can anyone let me know what would be the reason for the service to take inconsistent time for the job to complete.
I have the same parameters and data in both the cases and also verified that the time spent in the PREPARING phase is pretty much the same in both cases.
Also would it matter that I schedule simultaneous multiple independent training job on the same project, if so then would like to know the rationale behind it.
Any help would be greatly appreciated.
The easiest way is to add more logging to inspect where the time was spent. You can also inspect training progress using TensorBoard. There's no VM sharing between multiple jobs, so it's unlikely caused by simultaneous jobs.
Also, the running time should be measured from the point when the job enters RUNNING state. Job startup latency varies depending on it's cold or warm start (i.e., we keep the VMs from previous job running for a while).