BeautifulSoup quick notes

While preparing the driving test, have though to use BeautifulSoup to get all questions for my notes

Amazing AI deployment journey

We have shifted our Voice assistant AI modules on Openshift.
Nice experience building modules using Jenkins and deploying on Openshift

Voice bot with Splunk

I got the opportunity to work with Splunk and Elastic Search for NLP projects.

Requirement:

We have to track the user utterance and responses by Splunk. In my Flask API , have used RASA for a portion of NLU tasks.

Rasa RESTful API calls

One of my requirement is Parse a message with the trained RASA nlu model.
1) Train a Rasa model using rasa train
2) Start a Rasa server with rasa run --enable-api
3) Parse a message by running
curl -X POST http://52d9691d.ngrok.io/model/parse -H 'Content-Type: application/json' -d '{"text":"Your text"}'

Used ngrok and Postman

error failed building wheel for httptools

While installing latest version of RASA have faced the following issue

Solution: conda install -c conda-forge httptools

Generate Text using OpenAIGPT2 in Python

!git clone https://github.com/graykode/gpt-2-Pytorch
import os
os.chdir('gpt-2-Pytorch')
!curl --output gpt2-pytorch_model.bin https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-pytorch_model.bin
!pip install -r requirements.txt
!python main.py --text "Chandrayaan-2 is India's second lunar exploration mission after Chandrayaan-1. Developed by the Indian Space Research Organisation, the mission was launched from the second launch pad at Satish Dhawan Space Centre on 22 July 2019 at 2.43 PM IST to the Moon by a Geosynchronous Satellite Launch Vehicle Mark III."

Free GPU, TPU on Google Colab

Happy learning and experiments with Google Colab

Note: Data might be lost after 12hrs
We can run the tensorboard on colab too.
Reference links
Google SEEDBANK
https://research.google.com/seedbank/seeds?keyword=text
https://colab.research.google.com/

Question-Answering system with cdQA-suite

This is Transfer Learning era on NLP tasks.
Have been working on how to use my data set on XLNet and BERT.
Found excellent references with community help

https://towardsdatascience.com/how-to-create-your-own-question-answering-system-easily-with-python-2ef8abc8eb5
https://github.com/cdqa-suite/cdQA

XLNet - A SOTA model

While working on the Q&A system, have found pretrained model on NLP. XLNET is state of the art model for Q&A applications.
XLNet is the combination of Autoregressive  and autoencoding

Generalized Autoregressive Pretraining for Language Understanding.

Generalized -- Pretrain without data correption( masking) by using permutation LM.

Auto regressive – Autoregressive language model but also utilizes bidirectional context

This is better than BERT

References:
https://arxiv.org/abs/1906.08237
https://mlexplained.com/2019/06/30/paper-dissected-xlnet-generalized-autoregressive-pretraining-for-language-understanding-explained/

AWS IoT Greengrass supports Python 3.7

AWS IoT Greengrass Core 1.9.0 is now available.
https://aws.amazon.com/about-aws/whats-new/2019/04/aws-iot-greengrass-adds-support-python-3-7-node-v-8-10-0-and-expands-support-elliptic-curve-cryptography/
https://docs.aws.amazon.com/greengrass/latest/developerguide/what-is-gg.html#gg-platforms

12 Greatest Success Tips in Life

1. SKETCH THE OPPORTUNITIES
2. SETTING THE GOAL
3. STUDY PLANNING
4. SELF CONFIDENCE 
5. SELF ESTEEM 
6. SUCCESS STORIES 
7. STUDY TIME 
8. STUDY ENVIRONMENT 
9. STUDY NOTES 
10. STRATEGIES 
11. STRESS MANAGEMENT 
12. SOCIAL RESPONSIBILITY

VMware Workstation and Device/Credential Guard are not compatible. VMware Workstation can be run after disabling Device/Credential Guard.

Follow below link for solution
https://www.youtube.com/watch?v=CGpv2Dvzyeg

Journey with AWS SageMaker

SageMaker is a fully managed machine learning service offered by AWS.Build, train and deploy machine learning models on the AWS cloud.
As part of ML inference at Edge demo first have prepared my ML model using AWS SageMaker.
Have faced few interesting challenges like policy issues and instances availability etc

References:
https://aws.amazon.com/sagemaker/pricing/
https://console.aws.amazon.com/support/home

AWS IOT Greengrass ML inference

Right now am working one of the use case using AWS Greengrass.

AWS Greengrass is a service that allows you to take a lot of the capabilities provided by the AWS IoT service and run that at the edge closer to your devices. AWS Greengrass ensures your IoT devices can respond quickly to local events, use Lambda functions running on Greengrass Core to interact with local resources, operate with irregular connections, stay updated with over the air updates, and minimize the cost of transmitting IoT data to the cloud.

Deep Learning challenges at the Edge

Resource-constrained devices

CPU, memory, storage, power consumption.

Network connectivity

Latency, bandwidth, availability.

On-device prediction may be the only option.

Deployment

Updating code and models on a fleet of devices is not easy.

Value of ML inference at the Edge

·       Latency

·       Bandwidth

·       Availability

·       Privacy

https://aws.amazon.com/solutions/case-studies/iot/
https://aws.amazon.com/greengrass/faqs/

ONNX

With ONNX, AI developers can more easily move models between state-of-the-art tools and choose the combination that is best for them.
Enabling interoperability makes it possible to get great ideas into production faster.
Ex: PyTorch > ONNX > Tensorflow

Edge Computing

As a Automotive industry Machine Learning Engineer, today i thrilled about Edge Computing technology.
Edge computing to be bigger than cloud computing
Edge computing brings memory and computing power closer to the location where it is needed.
As part of my R&D have been deploying my tiny TF model on Android Apps and Embedded devices

More exciting to know Edge computing technology
https://www.sparkfun.com/products/15170

Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX

Am running some model on my brand new 32bit RAM laptop and faced the below error msg

Soulution: os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

Reference: https://stackoverflow.com/questions/47068709/your-cpu-supports-instructions-that-this-tensorflow-binary-was-not-compiled-to-u

Tensorflow Lite Demo

As part of my requirements, have started deploying tensorflow models on Android using Tensorflow Lite
Amazing experience

import tensorflow as tf
# Convert to TensorFlow Lite model.
converter = tf.lite.TFLiteConverter.from_keras_model_file('sunroof_model.h5')
tflite_model = converter.convert()
open("sunroof_model.tflite", "wb").write(tflite_model)

Below are few errors have experienced while developing android code

"Attempt to invoke virtual method 'void org.tensorflow.lite.Interpreter.run(java.lang.Object, java.lang.Object)' on a null object reference"
Cannot copy between a TensorFlowLite tensor with shape [1, 1] and a Java object with shape [1].
Cannot copy between a TensorFlowLite tensor with shape [1, 6] and a Java object with shape [6].

What are the contraints for tensorflow scope names?

While developing TF model, have faced the below issue

Solution: remove the spaces in the feature names

The kernel appears to have died. It will restart automatically

Got strange issue on my new setup
No module named 'numpy.core._multiarray_umath'

Solution: Just upgrade numpy

Success formula in our life

1. Insatiable appetite for knowledge
2. Know the basics 
3. Identify the gaps in your knowledge 
4. Constantly update 
5. Expose the linkages 
6. Expose diverse views 
7. Write, argue and debate

Reference: https://www.youtube.com/watch?v=_svETMOY9zY

Hyper parameters with GridSearch

Building a better model involves iteration and tuning these hyperparameters.
Grid search uses cross validation, splitting the training data up into folds for training and testing of each hyperparameter combination. After all hyperparameter variants are trained, the original test data is used to validate the final model.

GridSearchCV takes a dictionary that describes the parameters that should be tried and a model to train. The grid of parameters is defined as a dictionary, where the keys are the parameters and the values are the settings to be tested.

from sklearn.model_selection import GridSearchCV
from sklearn.metrics import classification_report
from keras.wrappers.scikit_learn import KerasClassifier
from keras.models import Sequential
import time

def dense_model(units, dropout):
    model = Sequential()
    model.add(Dense(units, activation='relu', input_shape=(28, 28,)))
    model.add(Dropout(dropout))
    model.add(Dense(units, activation='relu'))
    model.add(Dropout(dropout))
    model.add(Flatten())
    model.add(Dense(10, activation='softmax'))
    model.compile(loss='categorical_crossentropy',
              optimizer='adam',
              metrics=['accuracy'])
    return model

hyperparameters = {
    'epochs': [1],
    'batch_size': [64],
    'units': [32, 64, 128],
    'dropout': [0.1, 0.2, 0.4]
}

model = KerasClassifier(build_fn=dense_model, verbose=0)
start = time.clock()
grid = GridSearchCV(estimator=model, param_grid=hyperparameters, cv=6, verbose=4)
grid_result = grid.fit(x_train, y_train)
print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_))

y_true, y_pred = np.argmax(y_test, axis=1), grid.predict(x_test)
print()
print(classification_report(y_true, y_pred))
print()
print(time.clock() - start)

How to get current available GPUs in tensorflow?

Working with a GPU can be 20-50x faster!
Reference Link:
https://stackoverflow.com/questions/38559755/how-to-get-current-available-gpus-in-tensorflow

Another Computer Vision application on AWS

URL: http://iamsam.tech/

Technical areas: S3 Bucket, Amazon ML+ Rekognition, boto3 python package