Azure ML: From Zero to Hero

Azure ML is a complex Microsoft offering for Data Scientists, and sometimes it seems too complex to just start using it. In this talk, I will demonstrate a simple way to start with Azure ML using Visual Studio Code, and then proceed to show more complex examples, such as training GANs to produce artistic paintings and using pre-trained BERT with DeepPavlov library for answering questions on COVID-19. I will also give a general overview of how and when Machine Learning and AI can be effectively used, and cover some practical cases.

Transcript

1. Azure Machine Learning: from zero to hero Dmitry Soshnikov, Cloud Developer Advocate, Microsoft http://soshnikov.com – @shwars
2. Data is more important than algorithms… AI
3. AI / Machine Learning on Azure Domain specific pretrained models To reduce time to market Azure Databricks Machine Learning VMs Popular frameworks To build advanced deep learning solutions TensorFlow Pytorch Onnx Azure Machine Learning Language Speech … Search Vision Productive services To empower data science and development teams Powerful infrastructure To accelerate deep learning Scikit-Learn PyCharm Jupyter Familiar Data Science tools To simplify model development Visual Studio Code Command line CPU GPU FPGA From the Intelligent Cloud to the Intelligent Edge
4. Where does Azure ML fit? Data Science VM Azure ML Cognitive Services Infrastructure services Platform Services Pre-trained Models For beginners Advanced Data Science Experimentation
5. How to Start with Azure ML: Read my blog series: • The best way to start with Azure ML using VS Code • Using Azure ML for Hyperparameter Optimization • Training GAN to Produce Art • Training BERT Question Answering with DeepPavlov ❶ ❷ Try it out: http://github.com/CloudAdvocacy/AzureMLStarter
6. Simplest Way to Start with Azure ML Azure ML for VS Code Portal http://portal.azure.com
7. Azure ML Workspace: A container for Everything Azure ML Workspace encapsulates it all: 1. Storage 2. Datasets 3. Compute 4. Notebooks 5. Experiment Results 6. Models 7. Deployments ❶ az extension add -n azure-cli-ml az group create -n ml -l westus2 az ml workspace create -w AzML -g ml az ml folder attach -w AzML -g ml Create Workspace using Azure CLI: az ml computetarget create amlcompute -n cpu --min-nodes 0 --max-nodes 2 -s STANDARD_D3_V2 Create Cluster using Azure CLI: MS Docs: HERE
8. Azure ML service Workspace Taxonomy
9. Submit and Track Experiments Experiment is represented by a Python Script + Environment that run on Compute (Local Compute, Azure ML Cluster or Databricks) 1. Auto-package code 2. Keep track of results 3. Store models 4. Queue runs 5. Programmatically spawn many runs with different parameters ❷ az ml run submit-script -c sklearn –e MyExp train.py Submit Experiment using CLI: Log Metrics in the script: from azureml.core.run import Run run = Run.get_submitted_run() run.log('accuracy', acc)
10. Toy Problem: MNIST Digit Recognition http://yann.lecun.com/exdb/mnist/ mnist=fetch_openml('mnist_784’) X = mnist[‘data’] y = mnist[‘target’] fig,ax=plt.subplots(1,15) for i in range(15): z=X_train[i].reshape(28,28) ax[i].imshow(z) plt.show()
11. My Computer Data Store Azure ML Workspace Compute Target Docker Image How Azure ML Experimentation Works Experiment
12. What happens after you submit the job? Post-Processing The ./outputs directory of the run is copied over to the run history in your workspace so you can access these results. Running In this stage, the necessary scripts and files are sent to the compute target, then data stores are mounted/copied, then the entry_script is run. While the job is running, stdout and the ./logs directory are streamed to the run history. You can monitor the run's progress using these logs. Image creation A Docker image is created matching the Python environment specified by the estimator. The image is uploaded to the workspace. Image creation and uploading takes about 5 minutes. This happens once for each Python environment since the container is cached for subsequent runs. During image creation, logs are streamed to the run history. You can monitor the image creation progress using these logs. Scaling If the remote cluster requires more nodes to execute the run than currently available, additional nodes are added automatically. Scaling typically takes about 5 minutes.
13. Azure ML Currently Supported Compute Targets Compute target GPU acceleration Hyperdrive Automated model selection Can be used in pipelines Local computer Maybe ✓ Data Science Virtual Machine (DSVM) ✓ ✓ ✓ ✓ Azure ML compute ✓ ✓ ✓ ✓ Azure Databricks ✓ ✓ ✓ Azure Data Lake Analytics ✓ Azure HDInsight ✓ https://docs.microsoft.com/en-us/azure/machine-learning/service/how-to-set-up-training-targets#supported-compute-targets
14. Run Notebooks and Create Datasets When you do a lot of training, it makes sense to store data inside the workspace. To run Python code inside the workspace – use Notebooks! You need to create separate compute (not cluster) to do that! ❸
15. Submitting Experiments / Training Script Use args to pass different parameters, including data path parser = argparse.ArgumentParser(description='MNIST Train') parser.add_argument(‘--data_folder', type=str, dest='data_folder', help='data folder mount point') parser.add_argument('--epochs', type=int, default=3) parser.add_argument('--batch_size', type=int, default=128) parser.add_argument('--hidden', type=int, default=100) parser.add_argument('--dropout', type=float) Store model into outputs directory os.makedirs('outputs',exist_ok=True) model.save('outputs/mnist_model.hdf5') Load data as files fn = os.path.join(args.data_folder, 'mnist_data/mnist.pkl') with open(fn,'rb') as f: X,y = pickle.load(f) run = Run.get_context() run.log('Test Loss', score[0]) run.log('Accuracy', score[1]) Log Result ❹
16. Hyperparameter Optimization Define Parameter Sampling Strategy Submit the Experiment ❺ param_sampling = RandomParameterSampling({ '--hidden': choice([50,100,200,300]), '--batch_size': choice([64,128]), '--epochs': choice([5,10,50]), '--dropout': choice([0.5,0.8,1]) }) Define Hyperdrive Configuration hd_config = HyperDriveConfig(estimator=est, experiment = Experiment(workspace=ws, name='keras-hyperdrive') hyperdrive_run = experiment.submit(hd_config) Strategies: Grid, Random, Bayesian Distribution: choice, uniform, normal hd_config = HyperDriveConfig(estimator=est, hyperparameter_sampling=param_sampling, policy=early_termination_policy, primary_metric_name='Accuracy', primary_metric_goal=MAXIMIZE, max_total_runs=16, max_concurrent_runs=4)
17. Tools for Simplified ML: AutoML, Designer Automatic ML ❻ Designer Run the experiment to automatically try different models and select the one that performs best Can do some feature optimization (data balancing, irrelevant feature elimination) Similar to Azure ML Studio Classic Perform ML experiments without coding, by composing pre-build blocks Defines pipelines in a graphical way
18. Model Deployment ❼
19. Case Study: Generative Adversarial Networks on AML http://aka.ms/azml_gan
20. How Generative Adversarial Networks Work Random Vector Generator (Neural Net) Discriminator (Neural Net) ✓ ✗
21. GAN Library: keragan http://github.com/shwars/keragan Generator Discriminator Random Noise (dim=100) Conv Matrix Conv Matrix Reshape DeConv DeConv Conv Matrix Conv Matrix Feature Vector Classifier
22. GAN Library: keragan http://github.com/shwars/keragan Generator Discriminator discriminator = Sequential() for x in [16,32,64]: # number of filters on next layer discriminator.add(Conv2D(x, (3,3), strides=1, padding="same")) discriminator.add(AveragePooling2D()) discriminator.addBatchNormalization(momentum=0.8)) discriminator.add(LeakyReLU(alpha=0.2)) discriminator.add(Dropout(0.3)) discriminator.add(Flatten()) discriminator.add(Dense(1, activation='sigmoid')) generator = Sequential() generator.add(Dense(8 * 8 * 2 * size, activation="relu", input_dim=latent_dim)) generator.add(Reshape((8, 8, 2 * size))) for x in [64;32;16]: generator.add(UpSampling2D()) generator.add(Conv2D(x, kernel_size=(3,3),strides=1, padding="same")) generator.add(BatchNormalization()) generator.add(Activation("relu")) generator.add(Conv2D(3, kernel_size=3, padding="same")) generator.add(Activation("tanh"))
23. Model Training discriminator.trainable = False noise = np.random.normal(0, 1, (batch_size, latent_dim)) gen_imgs = generator.predict(noise) imgs = get_batch(batch_size) d_loss_r = discriminator.train_on_batch(imgs, ones) d_loss_f = discriminator.train_on_batch(gen_imgs, zeros) d_loss = np.add(d_loss_r , d_loss_f)*0.5 g_loss = combined.train_on_batch(noise, ones) res = generator.predict(np.random.normal(3,latent_dim)) fig,ax = plt.subplots(1,len(res)) for i,v in enumerate(res): ax[i].imshow(v[0]) run.log_image("Sample",plot=plt) # Generate Noise Vector & Images # Train Discriminator # Train Generator (by training combined model) # Log Sample Images through Azure ML Code Sample on GitHub Code Sample on GitHub
24. Getting and Using the Model fnames = list(filter(lambda x : x.startswith('outputs/models/gen_’), run.get_file_names())) no = max(map(lambda x: int(x[19:x.find('.')]), fnames)) fname = 'outputs/models/gen_{}.h5'.format(no) run.download_file(fname) # Get the Latest Model File model = keras.models.load_model(fname) latent_dim=model.layers[0].input.shape[1].value vec = np.random.normal(0,1,(10,latent_dim)) res = model.predict(vec)) res = (res+1.0)/2 # Predict 10 images Code Sample on GitHub
25. Case Study: Open Domain Question Answering Common tasks for NLP: • Intent Classification • Named Entity Recognition (NER) • Keyword Extraction • Text Summarization • Question Answering Open Domain Question Answering – a task, when a model is able to give specific answers contained in a large volume of text (e.g. Wikipedia) - Where did guinea pigs originate? - Andes of South America - When did the Lynmouth floods happen? - 1804 Neural Language Models: • Recurrent Neural Network (RNN) • LSTM, GRU • Transformers • GPT-2 • BERT • Microsoft Turing-NLG
26. How BERT Works (Simplified) Masked Language Model + Next Sentence Prediction During holidays, I like to ______ with my dog. It is so cute. 0.85 Play 0.05 Sleep 0.09 Fight 0.80 YES 0.20 NO BERT contains 345 million parameters => very difficult to train from scratch! In most of the cases it makes sense to use pre-trained language model.
27. Text Processing Pipelines BERT for Classification Input Text BERT Features Classifier BERT for Entity Extraction Input Text BERT Features Mask Generator Class Prob Vector Entity Masks BERT for Question Answering Input Text BERT Features Bounds Generator Answer Bounds 0.85 Insult 0.15 Neutral I live in France My age is 21 LOC
28. DeepPavlov: “Keras” for NLP http://deeppavlov.ai $ pip install deeppavlov python -m deeppavlov install config.json python -m deeppavlov download config.json python -m deeppavlov train config .json Text processing pipeline is defined in JSON config: • Processing steps, their inputs and outputs • Weight location for pre-trained models • Data shape and location • Training parameters
29. Example: Open Domain Question Answering for COVID http://aka.ms/deeppavlov
30. Using Azure ML to Train the ODQA Model We will use the following features of Azure ML: • Define file dataset that points to data location • Create cheap non-GPU compute for data exploration and preparation • Use GPU-enabled compute on the same data to train the model • All code would be in the form of Jupyter Notebooks We do not use training on Azure ML Cluster in this case to have better control on the environment. DeepPavlov downloads large amounts of pre-trained data from the network, and for simple cases it is better to use single node. Link to the non-commercial CORD-19 dataset: here (.tar.gz)
31. Getting Wikipedia ODQA Up and Running import sys !{sys.executable} -m pip install deeppavlov !{sys.executable} -m deeppavlov install en_odqa_infer_wiki !{sys.executable} -m deeppavlov download en_odqa_infer_wiki from deeppavlov import configs from deeppavlov.core.commands.infer import build_model odqa = build_model(configs.odqa.en_odqa_infer_wiki) answers = odqa([ "Where did guinea pigs originate?", "When did the Lynmouth floods happen?" ]) # Get the Library and Required Models # Build Model from Config and Run Inference ['Andes of South America', '1804']
32. ODQA Configs Ranker en_ranker_tdifd_wiki SQuAD multi_squad_noans_infer (R-NET) Config on GitHub en_odqa_infer_wiki question question document answer TRAIN Replace with BERT
33. Train the Ranker from deeppavlov.core.common.file import read_json model_config = read_json(configs.doc_retrieval.en_ranker_tfidf_wiki) model_config["dataset_reader"]["data_path"] = os.path.join(os.getcwd(),"text") model_config["dataset_reader"]["dataset_format"] = "txt" model_config["train"]["batch_size"] = 1000 # Specify Data Path & Format doc_retrieval = train_model(model_config) doc_retrieval(['hydroxychloroquine']) # Train the Model and See the Results "dataset_reader": { "class_name": "odqa_reader", "data_path": "{DOWNLOADS_PATH}/odqa/enwiki", "save_path": "{DOWNLOADS_PATH}/odqa/enwiki.db", "dataset_format": "wiki" } Part of en_ranker_tfidf_wiki config
34. Results with R-NET Question Answering # Download R-NET SQuAD model squad = build_model(configs.squad.multi_squad_noans_infer, download = True) # Do not download the ranker model, we've just trained it odqa = build_model(configs.odqa.en_odqa_infer_wiki, download = False) odqa(["what is coronavirus?","is hydroxychloroquine suitable?"]) ['an imperfect gold standard for identifying King County influenza admissions', 'viral hepatitis']
35. Use BERT for QA # Download Pre-trained BERT Q&A Model # Replace Q&A Model in the Master Config Part of en_odqa_infer_wiki config !{sys.executable} -m deeppavlov install squad_bert_infer bsquad = build_model(configs.squad.squad_bert_infer, download = True) odqa_config = read_json(configs.odqa.en_odqa_infer_wiki) odqa_config['chainer']['pipe'][-1]['squad_model']['config_path'] = '{CONFIGS_PATH}/squad/squad_bert_infer.json' odqa = build_model(odqa_config, download = False) odqa(["what is coronavirus?", "is hydroxychloroquine suitable?", "which drugs should be used?"]) # Build and Use Model { "class_name": "logit_ranker", "squad_model": {"config_path": ".../multi_squad_noans_infer.json"} "in": ["chunks","questions"], "out": ["best_answer","best_answer_score"] }
36. Question Answer what is coronavirus? respiratory tract infection is hydroxychloroquine suitable? well tolerated which drugs should be used? antibiotics, lactulose, probiotics what is incubation period? 3-5 days how to contaminate virus? helper-cell-based rescue system cells what is coronavirus type? enveloped single stranded RNA viruses what are covid symptoms? insomnia, poor appetite, fatigue, and attention deficit what is reproductive number? 5.2 what is the lethality? 10% where did covid-19 originate? uveal melanocytes is antibiotics therapy effective? less effective what are effective drugs? M2, neuraminidase, polymerase, attachment and signal-transduction inhibitors what is effective against covid? Neuraminidase inhibitors is covid similar to sars? All coronaviruses share a very similar organization in their functional and structural genes what is covid similar to? thrombogenesis Results
37. Conclusions Azure ML enhances your ML experience by: • Grouping everything together in workspace • Journaling all experiment results automatically • Helping with hyperparameter optimization and scalable compute • Supporting distributed training ❶ ❷ You should try it out: • http://github.com/CloudAdvocacy/AzureMLStarter • http://aka.ms/azmlstarter - Blog Post
38. Further Reading  How to train your own neural network to generate paintings http://aka.ms/azml_gan  Can AI be creative http://aka.ms/creative_ai  Creating interactive exhibit based on cognitive portraits http://aka.ms/cognitive_portrait_exhibit  Training COVID ODQA on Azure ML: http://aka.ms/deeppavlov
39. © Copyright Microsoft Corporation. All rights reserved. @art_of_artificial