thushv89 / manning_tf2_in_action Goto Github PK

View Code? Open in Web Editor NEW

83.0 5.0 44.0 55.63 MB

The official code repository for "TensorFlow in Action" by Manning.

Home Page: https://www.manning.com/books/tensorflow-in-action

Jupyter Notebook 99.95% Python 0.04% Dockerfile 0.01% Shell 0.01%

tensorflow python machine-learning deep-learning tensorflow2 nlp computer-vision tf tf2 notebook

manning_tf2_in_action's Introduction

TensorFlow in Action

This project is the code repository for TensorFlow in Action.

IMPORTANT: You will need to have Visual Studio installed. Make sure your installation has Windows SDK

Video Guides for Installation

Prerequisites (If you want GPU support)

Install the latest NVIDIA driver for your GPU from this page
Install CUDA 11.2 (TensorFlow 2.9.1)
Setup CuDNN (v8.2 for CUDA 11.2)
Make sure your $PATH variable contains the path to the bin folder of cuda (e.g. On Windows - C:\CUDA\v11.2\bin)
- On UNIX - Set LD_LIBRARY_PATH to lib64 folder (e.g. /usr/local/cuda-11.2/lib64)

Creating a Virtual Environment (Anaconda) (Recommended)

Install Anaconda
Open up Anaconda Prompt
Setup a conda virtual environment with conda create -n manning.tf2 python=3.9
Activate the environment with conda activate manning.tf2
Install the required libraries using pip install -r requirements.txt

Creating a Virtual Environment (virtualenv)

Install Python 3.9 by following instructions on the webpage. Make sure it is added to the PATH variable.
To verify the installation of Python3.9, go to the command line terminal and type python and press enter. You should see a message similar to the following.
```
 Python 3.9.0 (...) [...] on ...
 Type "help", "copyright", "credits" or "license" for more information.
```
Type exit() to exit the Python interpreter.
Run pip3 install virtualenv to install Python virtual environment package
Go to where you would like the virtual env created (e.g. cd C:\Users\<user>\Documents\code\python_venvs) and type python -m venv manning.tf2. You should see a directory called manning.tf2 created in the directory you are in.
Now to activate the virtual environment, from the directory you were originally in (e.g. C:\Users\<user>\Documents\code\python_venvs),
- On Windows - Run manning.tf2\Scripts\activate.bat
- On UNIX - Run source manning.tf2\bin\activate
If successfully activated, you should see a (manning.tf2) in front of the normal prompt you get in the command line interface (CLI).
Now run pip install -r requirements.txt to install all the necessary packages

Tutorial on virtual environments: Here

Important notes

For some plotting capability provided in TensorFlow/Keras, you have installed a Python package called graphviz (Installation instructions). Make sure you add it to your $PATH variable.

Getting Jupyter Notebook server up

Now you are ready to run the Jupyter notebook server, allowing you to run the notebooks provided in the code repository.

Open up the command line terminal and activate the virtual environment manning.tf2 if you haven't already
Go in to the directory you downloaded code to using cd in the CLI (e.g. cd C:\Users\<user>\Documents\code\manning_tf2_in_action)
Run jupyter notebook in the CLI
This should open up the jupyter notebook server's landing page on your default browser
Now you can navigate the folder structure within that directory, open any notebook and run it.

manning_tf2_in_action's People

Contributors

Stargazers

Watchers

manning_tf2_in_action's Issues

Reusable utils for all chapters

After all checks copy util functions to separate chapter specific util

JSON file is missing for chapter 9

Hello, It seems the file for the video game review has been removed, but I found it on Kaggle, here is a link:
https://www.kaggle.com/code/idrimadrid/projet-text-mining-amazon-reviews/output

Ch07 Minception `inception_resnet_a` function is wrong

This should not be there

out2_3 = Conv2D(n_filters[1][2], (1,1), strides=(1,1), activation=None, 
                        kernel_initializer=initializer, padding='same')(out2_2)
if bn:
    out2_3 = BatchNormalization()(out2_3)
out2_3 = Activation(activation)(out2_3)

and

out4_1 = Concatenate(axis=-1)([out1_1, out2_2, out3_4])

Then when putting the network together it should be,

inc_a = inception_resnet_a(stem_out, [(32,),(32,32), (32, 48, 64, 384),(384,)], initializer=init)

Files to be added/removed after renaming

Files to be removed

Ch05-Transformers/Untitled.ipynb
Ch08-Image-Segmentation-with-CNNs/ch08_utils.ipynb

Files to be added

Ch11-Ch12-Sequence-to-Sequence-Learning-with-TF2/attention.py -- is that needed?
Ch14-Ch15-TFX-for-MLOps-in-TF2/tfx/*

Final check (chapters 6, 7 , 8, 9, 10, 12) - Merge `dev` to `master`

Ch06/6.1 - Move Bonus: Metrics of the model section below training - say the model needs to be fit before calling this
Ch06 checked
Ch07/7.1 - Move Metrics that we'll be looking at under the training
Ch07 checked
Ch08/8.1 - Move/Delte first "Plot some train images and their predictions" section under Unet
Ch09 checked
Ch10 - Under Let's look at some word ID sequences, in the for loop rename ngrams to something else.
Ch10 checked
Commit Ch12 / 12.2 after running / 12.1 after running
Ch12 checked

Add tensorflow_addons to requirements.txt
Add tf-models-official to requirements.txt
After all checks copy util functions to separate chapter specific util

Some extra things to talk about

https://www.tensorflow.org/api_docs/python/tf/keras/layers/experimental/preprocessing/TextVectorization

transformers library `return_dict` in graph mode issue

An error was thrown because the return_dict argument cannot be set in graph mode

import transformers
from transformers import TFDistilBertForQuestionAnswering

config = DistilBertConfig.from_pretrained("distilbert-base-uncased", return_dict=False)
model = TFDistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased", config=config)

input_ids = tf.keras.layers.Input([None,], dtype=tf.int32, name="input_ids")
attention_mask = tf.keras.layers.Input([None,], dtype=tf.int32, name="attention_mask")
out = model([input_ids, attention_mask])
model_v2 = tf.keras.models.Model([input_ids, attention_mask], outputs=(out.start_logits, out.end_logits))

loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.SparseCategoricalAccuracy()
#model.distilbert.return_dict = False # if using Transformers >3.02, make sure outputs are tuples

optimizer = tf.keras.optimizers.Adam(learning_rate=5e-5)
model_v2.compile(optimizer=optimizer, loss=loss) # can also use any keras loss fn

model_v2.fit(train_dataset, epochs=3)

Add tensorflow addons to requirements

Ch09 fixes

Use random seed for everything in train/valid/test split
Use percentile 0.9 for sequence length analysis
Add no to stop words
Rename data to inputs (also the save name)

add input to compute_multihead_output (the multi-head list)
adjust h1 and h2 accordingly.

TextVectorization for handling [SOS] and [EOS] with a special standardize callable

Ch14 - Viewing Validation results

import tensorflow_model_analysis as tfma

res = tfma.load_validation_result(validation_path)
print(res)

Input of upsampling layer in ASPP

In code listing 8.11 in the Python notebook the upsampling layer is defined as:

outb_1_up = layers.UpSampling2D((24,24), interpolation='bilinear')(outb_1_avg)

From what I understand, the input to the upsampling layer here (outb_1_up) should be the output of the earlier convolution (outb_1_conv), not the output of the averaging layer (outb_1_avg).

Would be nice to confirm this.