# project-sheeko/sheeko-vagrant

No description, website, or topics provided.
Python Other
1. Python 99.2%
2. Other 0.8%
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
data_preparation
evaluation
inference
library
train
Vagrantfile
installation.sh
sheeko_demo_1.jpg
sheeko_demo_2.jpg
sheeko_demo_3.jpg
sheeko_demo_4.jpg
sheeko_demo_5.jpg
show_and_tell.jpg

# Sheeko Project

Machine learning implementation package to generate descriptive metadata for digitized historical images. The project is funded through the LYRASIS Catalyst grant.

## PROJECT OVERVIEW

This project provides deployable machine learning environment including pre-trained models and code packages for training, inference and evaluation for generating caption metadata. Inference codes for label generating (classification, object detect) are also included in this project.

### Introduction

Inference Example for given image

Caption:

a bird sitting on top of a wooden bench . (p=0.001655)


Label:

jay
indigo bunting, indigo finch, indigo bird, Passerina cyanea
magpie
pill bottle
water ouzel, dipper
Bird


### Architecture

This code package is the customization of Im2txt. Please see https://github.com/tensorflow/models/tree/master/research/im2txt for more details about im2txt

Architecture of Caption generating model in this project is "Show and Tell model" which is an encoder-decoder neural network.

## GETTING STARTED

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes. See deployment for notes on how to deploy the project on a live system.

### Prerequisites

There're two ways deploying this project on your platform.

1. Deploy locally. Windows and Linux OS are supported.
1. Deploy in VirtualBox

Difference between these two ways:

• Deploy locally could maximize your GPU acceleration. This is useful when performing training and evaluation
• Deploy in VirtualBox could run the script without doing additional installation and for sure it's cross-platform.

You could deploy the project in both ways so you could get best performance in all functionalities.

#### To deploy in Virtualbox

• Virtualbox Installed

• vt-x enabled

• Vagrant Installed

### Installation

#### To Install Locally

It's recommened to run the code in Ubuntu 16.04.6 LTS OS or Windows OS environment. Make sure you have tensorflow-gpu and CUDA tookit installed to enable GPU accleration. Installation for tensorflow-gpu is not covered in this document.

• Run the script to see python version. Make sure python is installed properly.
python -V

• For Windows OS

Install the required packages listed above.

To install punkt, run python script:

import nltk


• For Linux OS

if you have the permission for sudo, run script under project root directory:

bash installation.sh


Otherwise install the required packages listed same as Windows OS.

#### Use VirtualBox

This code package is supported using vagrant. Virualbox and vt-x option need to be enabled in BIOS mode.

Please be aware that there is limitation for GPU allocation in VirtualBox environment. So It's not recommended to run either training or evaluation script in VirtualBox.

Alternatively, if you plan to run training or evalutaion, please deploy the code package locally.

• Download code package and run the script under project root directory through command line

vagrant up


it will take few minutes if you running this script at the first time. It will create the VM and install all pakcages on the VM.

After the script is done, the vagrant box is up, and you may notice that VM "sheeko_project" is created. OS inside the VM is Ubuntu 16.04.6 LTS OS

### Test Installation

If you deploy the code project in VM, you need to check installation inside the VM. Otherwise you can skip this step.

• 1.Start up and enter vagrant box
vagrant up
vagrant ssh

• 2.Run following scripts to test packages installation
python -c 'import tensorflow as tf; print(tf.__version__)'


Version of the tensorflow will be printed as result if successfully installed

python -c 'import nltk; print(nltk.__version__)'


Version of the nltk will be printed as result if successfully installed

## HOW TO USE

### Execute script

This code package is executed through Command Line. To run the script, you need to navigate to the directory of the script then run the script using "python script_name.py".

Here's an example:

cd /path/to/project/dir/data_preparation
python validate_data_run.py


### Configure File

To make the script work, you need to configure the scripts with the information needed.

Basically you need to configure the fields within the block in each script.

Here's an example:

-------------------configuration start here-----------------------------------------------------------------------
# Field name in annotation file containing metadata
FIELD_NM = "caption"
# Field name in the annotation file containing the image file name
FIELD_ID = "image_id"

# List of paths (relative or absolute) of directories containing the annotation files
CAPTION_DIR_LIST = ['clean/demo_1', 'clean/demo_2']

# Path to output directory
OUTPUT_PATH = "build"

# Segment method: seg_by_image, seg_by_dir
SEG_METHOD ='seg_by_image'

# Training set percent in int
TRAIN_PERCENT = 80
-------------------configuration end here-----------------------------------------------------------------------


### OS Compatibility

If you plan to run the script under Windows OS, path format with "\".

e.g. CHECKPOINT_PATH = "path\to\dir\pretrained_model\graph.pb"

If Linux OS, path format with "/".

e.g. CHECKPOINT_PATH = "path/to/dir/pretrained_model/graph.pb"

You could easily switch between Linux OS and Windows OS by replacing "\" with "/" or replacing "/" with "\" .

When using a model, check the checkpoint file: Here's an example: model_checkpoint_path: "/vagrant/models/caption/train/model.ckpt-20"

If you're planning to migrate this model, make sure you adjust the path correctly to prevent the "Not Found" error. If you're running the code under Windows OS, you need to adjust the path by changing / with \.

When you run build_data_run.py: check the annotation file generated in annotation.json. Here's an example in : "file_name": "/vagrant/data_preparation/data/demo_2/128269.jpg" If you plan to run build_TF_run.py with this data in Windows OS or if data has been moved somewhere else. You have to run run build_data_run.py under Windows OS with the updated configuration and use the new dataset for your purpose to prevent "Not Found" error.

## WALKTHROUGH

### What's included in the package

• Data Preparation
• Training
• Evaluation
• Inference

In this section we're going to demo how to perform the inference and how to get your own model

To make the script work, we need to configure the code file.

Please see configuration for details.

### Inference Walkthrough

#### 1. First we need to get the pretrained models, you could skip this if you have trained your own model already.

See our pretrained models website to download our pretrained models.

In our walkthrough For caption inference we use the ptm-im2txt-incv3-mlib-cleaned-3m which is trained on MSCOCO dataset (http://cocodataset.org) with 3 million steps.

For classification: Here's the resource you could download the models and labels https://github.com/googlecodelabs/nest-tensorflow/tree/master/tmp/imagenet

For object detection: Here's the resource you could download the models and labels https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md https://github.com/tensorflow/models/tree/master/research/object_detection/data

In our demo we use faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28 and oid_bbox_trainable_label_map.pbtxt

#### 2. Next let's get the test images for inference.

In our code package we provide a few demo images under test_image directory. Here we're using the demo images for inference.

#### 3. Caption Inference

Let's start with turn on vagrant box

vagrant up

vagrant ssh


Now let's do the the caption inference.

the project directory is also the shared folder with the Virutalbox and that means you could edit the file either inside the VM or locally. In VM, the path the project is under /vagrant/.

For example: in vagrant box, Open /vagrant/inference/caption/caption_inference_test.py and configure the following fields:

1. CHECKPOINT_PATH. Path to the directory containing the checkpoint file (pre-trained model)
1. VOCAB_FILE. Path to the Vocabulary Dictionary file
1. IMAGE_FILE. Path to the JPEG image file to generate caption

You could also locate the file under the project_directory/inference/caption/caption_inference_test.py

Here's an example of configuration under linux OS (vagrant).

'''Path to the directory containing the checkpoint file (pre-trained model)'''
CHECKPOINT_PATH="../../models/caption/mscoco2017_3M/train"

# Vocabulary Dictionary file generated by the preprocessing script.
VOCAB_FILE = "../../models/caption/mscoco2017_3M/word_counts_mscoco.txt"

# Path to the JPEG image file to generate caption
IMAGE_FILE = "../../test_image/sheeko_demo_4.jpg"


Then we need to navigate to caption_inference_test located directory and run the script

cd /vagrant/inference/caption

python caption_inference_test.py


So we get the following results printed

  0) a cow standing in a field of grass . (p=0.000430)
1) a cow is standing in a field of grass . (p=0.000424)
2) a cow standing in a field of grass (p=0.000283)


caption_inference_run.py has the similar fields to configure. There's slight diference that it loop over all images under the each directory in the list and generate the captions. Also instead of printing out the results, it creates a JSON file containing of the results generated.

Here's an example of configuration under linux OS (vagrant).

# List of paths to directories containing JPEG image files to caption.
# The script will grab all JPEG in specified directories.
# No need to mention files individually.
# This will only grab images place directly in the directories. It will not go into child dirs.
IMAGE_DIR_LIST = ["../../test_image"]

# Path to output json file
OUTPUT_PATH = "output/captions.json"


After running the script, the results will be stored in JSON file located at output/captions.json.

Feel free to do inference with your own images. However, this model is not perfect for doing inference for all images.

e.g. the sheeko_demo_1.jpg the result generated says it's a bird standing on the top of the wooden branch.

#### 4. Classification Inference

Now let's do the the classification labels inference.

Open /vagrant/inference/classification/classify_image_test.py and configure the following fields:

1. CHECKPOINT_PATH. Path to the directory containing the checkpoint file (pre-trained model)
1. VOCAB_DIR. Path to the directory containing Vocabulary Dictionary file
1. IMAGE_FILE. Path to the JPEG image file to generate caption

You could also locate the file under the project_directory/inference/classification/classify_image_test.py

Here's an example of configuration under linux OS (vagrant).

# Path to pretrained model graph.pb file
CHECKPOINT_PATH = "../../models/labels/classification/classify_image_graph_def.pb"

# Path to vocabulary directory that containing pbtxt and txt dictionary file.
VOCAB_DIR = "../../models/labels/classification"

# Path to the JPEG image file to generate label
IMAGE_FILE = "../../test_image/sheeko_demo_4.jpg"


Then we need to navigate to classify_image_test located directory and run the script

cd /vagrant/inference/classification

python classify_image_test.py


So we get the following results printed

ox (score = 0.89056)
oxcart (score = 0.03067)
water buffalo, water ox, Asiatic buffalo, Bubalus bubalis (score = 0.00574)
hog, pig, grunter, squealer, Sus scrofa (score = 0.00301)
lakeside, lakeshore (score = 0.00288)


similar to caption_inference_run.py, classify_image_run.py also stored generated results in JSON File.

Here's an example of configuration under linux OS (vagrant).

# Path to pretrained model graph.pb file
CHECKPOINT_PATH = "../../models/labels/classification/classify_image_graph_def.pb"

# Path to vocabulary directory that containing pbtxt and txt dictionary file.
VOCAB_DIR = "../../models/labels/classification"

# List of paths to JPEG image files to labels
# The script will grab all JPEG in specified directories.
# No need to mention files individually.
# This will only grab images place directly in the directories. It will not go into child dirs.
IMAGE_DIR_LIST = ["../../test_image"]

# Path to output json file
OUTPUT_PATH = "output/classifications.json"


After running the script, the results will be stored in JSON file located at output/classifications.json.

#### 5. Object Detection Inference

Now let's do the the object detect labels inference.

Open /vagrant/inference/object_detect/object_detect_test.py and configure the following fields:

1. CHECKPOINT_PATH. Path to the directory containing the checkpoint file (pre-trained model)
1. VOCAB_FILE. Path to Vocabulary Dictionary file
1. IMAGE_FILE. Path to the JPEG image file to generate caption

You could also locate the file under the project_directory/inference/object_detect/object_detect_test.py

Here's an example of configuration under linux OS (vagrant).

# Path to pretrained model graph.pb file
CHECKPOINT_PATH = "../../models/labels/object_detect/models/faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28/frozen_inference_graph.pb"

# Path to label mapping pbtxt file
VOCAB_FILE = "../../models/labels/object_detect/vocab/oid_bbox_trainable_label_map.pbtxt"

# Path to the JPEG image file to generate label
IMAGE_FILE = "../../test_image/sheeko_demo_4.jpg"


Then we need to navigate to object_detect_test located directory and run the script

cd /vagrant/inference/object_detect

python object_detect_test.py


Since object detect is GPU usage consuming, it's slower to run it under vagrant box. If you plan to do inference for multiple images, it's recommended to run the script locally.

Notice in vagrant box, gpu acceleration is limited, so for training purpose we use local deployed code pakcage as demo

In this case we need to exit vagrant box, run code in command line

exit

logout
Connection to 127.0.0.1 closed.


then you run

cd /path/to/project/dir/inference/object_detect

python object_detect_test.py


Here's the result:

[{'score': '0.72273505', 'label_text': u'Cattle'}, {'score': '0.3564907', 'label_text': u'Animal'}, {'score': '0.34420493', 'label_text': u'Furniture'}]



similar to caption_inference_run.py, object_detect_run.py also stored generated results in JSON File.

Here's an example of configuration under linux OS (vagrant).

# Path to pretrained model graph.pb file
CHECKPOINT_PATH = "path/to/dir/pretrained_model/graph.pb"

# Path to label mapping pbtxt file
VOCAB_FILE = "path/to/dir/pretrained_model/label_map.pbtxt"

# List of paths to JPEG image files to caption. It's no longer needed since image_dir is used to grab all images inside to generate the captions, for legacy testing, please go to research/im2tx test_model.py
IMAGE_DIR_LIST = ["path/to/dir/"]

# Path to output json file
OUTPUT_PATH = "path/to/output/object_detect.json"


After running the script, the results will be stored in JSON file located at OUTPUT_PATH.

### Train your own model Walkthrough

#### 1. The first thing for your training is always about the data.

Please see Data Prepartion Section for details. Here to make the demo faster, we use the provided data under data_prepartion/data. Under data directory both annotation files and JPEG images are provided.

#### 2. Clean and Structure your data

Most time the data you get is always not clean, so cleaning actually takes most of the time for your training. Here we provide the code to do the trick. It cleans up your data by removing proper-noun (See Clean Data for how it works) and also structures your data into training and testing sets. (See Build Data for how it works)

You could skip data cleaning if you already have the clean data.

Open /vagrant/data_preparation/clean_data_run.py and configure the following fields:

1. FIELD_NM. Field name in annotation file containing metadata
1. DATA_DIR_LIST. List of directories containing annotation files that will process data cleaning
1. OUTPUT_PATH. Path to output directory
1. DATA_DIR. This is alternative way for providing list of directories, you could loop over the given root directory

You could also locate the file under the project_directory/data_preparation/clean_data_run.py

Here's an example of configuration under linux OS (vagrant). Here we loop over data directory, so all directories under "data" will be processed.

# Field name in annotation file containing metadata
FIELD_NM = "description_t"

# Example of looping all dirs under given path
DATA_DIR = 'data'

for dir in os.listdir(os.path.abspath(DATA_DIR)):
if os.path.isdir(os.path.abspath(os.path.join(os.path.abspath(DATA_DIR),dir))):
DATA_DIR_LIST.append(DATA_DIR + '/' + dir)

# Path to output directory
OUTPUT_PATH = "clean"


Then we need to navigate to clean_data_run located directory and run the script

cd /vagrant/data_preparation/

python clean_data_run.py


After running the script, the clean annotation files will be stored under "clean".

Once we've got the clean data needed. We could now structure our data into the format for training.

Open build_data_run.py and configure the fields:

1. FIELD_NM. Field name in annotation file containing metadata
1. FIELD_ID. Field name in annotation file containing image id matching image file name.
1. IMAGE_DIR_LIST. List of directories containing image files.
1. CAPTION_DIR_LIST. List of directories containing annotations files.
1. OUTPUT_PATH. Path to output directory
1. DATA_DIR. This is alternative way for providing list of directories, you could loop over the given root directory
1. SEG_METHOD. Segment method: seg_by_image, seg_by_dir.
1. TRAIN_PERCENT. Integer of percent to put into training data set. The rest of data will go the test set.

Here we put the cleaned annotations file along with the image files in the build_data_run.py.

# Field name in annotation file containing metadata
FIELD_NM = "caption"
# Field name in the annotation file containing the image file name
FIELD_ID = "image_id"

# List of paths (relative or absolute) of directories containing the annotation files
CAPTION_DIR_LIST = ['clean/demo_1', 'clean/demo_2']
# List of paths (relative or absolute) of directories containing the images
IMAGE_DIR_LIST = ['data/demo_1', 'data/demo_2']

# Path to output directory
OUTPUT_PATH = "build"

# Segment method: seg_by_image, seg_by_dir
SEG_METHOD ='seg_by_image'

# Training set percent in int
TRAIN_PERCENT = 80


Run the script:

python build_data_run.py


The structured data is stored under "build". Next we need to convert data into TF Record.

#### 3. Convert Data into TF Records

To make data runnable by training script, we need to convert the images and captions into TF records which are serial image-caption pairs.

Open /vagrant/data_preparation/build_TF_run.py and configure the following fields:

1. TRAIN_SET_IMAGE. Path to directory containing training set images
1. TEST_SET_IMAGE.Path to directory containing testing set images
1. TRAIN_SET_CAPTION. Path to training set annotation file
1. TEST_SET_CAPTION. Path to testing set annotation file
1. OUTPUT_PATH. Path to output directory

You could also locate the file under the project_directory/data_preparation/build_TF_run.py

# Path to directory containing training set images
TRAIN_SET_IMAGE = "build/train/images"
# Path to directory containing testing set images
TEST_SET_IMAGE = "build/test/images"
# Path to training set annotation file
TRAIN_SET_CAPTION = "build/train/annotations/annotation.json"
# Path to testing set annotation file
TEST_SET_CAPTION = "build/test/annotations/annotation.json"

# Path to output directory
OUTPUT_PATH = "TF"


Then we need to navigate to build_TF_run located directory and run the script

cd /vagrant/data_preparation/

python build_TF_run.py


After running the script, the TF Records will be stored under "TF".

#### 4. Now, let's run training for caption model

Finally we can train our own model for generating caption. We need to provide an inception checkpoint file. In this walkthrough we use

Open /vagrant/train/caption_train_run.py and configure the following fields:

1. DATA_DIR. Path to directory containing the saved training TF files
1. VOCAB_FILE. Path to dictionary file generated through TF build script
1. INCEPTION_CHECKPOINT. Path to Inception checkpoint file
1. MODEL_DIR. Directory to save or restore the training process of trained model
1. TRAIN_STEPS. Number of Steps to train
1. GPU_DEVICE. Gpu device to train your model, use integer number to refer to the device. Default set to 0

You could also locate the file under the project_directory/train/caption_train_run.py

# Path to directory containing the saved training TF files
DATA_DIR = "../data_preparation/TF"

# Path to dictionary file generated through TF build script
VOCAB_FILE = "../data_preparation/TF/word_count.txt"

# Path to Inception checkpoint file.
INCEPTION_CHECKPOINT = "../models/inception_v3.ckpt"

# Directory to save or restore the training process of trained model.
MODEL_DIR = "../models/caption"

# Number of Steps to train
TRAIN_STEPS = 20

# Select gpu device to train your model, use integer number to refer to the device: e.g. 0 -> gpu_0
GPU_DEVICE = 0


Here we just run 20 steps (which can do nothing for performance) to create a demo model as example. However you could continue the training by running the script (unless you change MODEL_DIR).

Notice in vagrant box, gpu acceleration is limited, so for training purpose we use local deployed code pakcage as demo

In this case we need to exit vagrant box, run code in command line

exit

logout
Connection to 127.0.0.1 closed.


then you run

cd /path/to/project/dir/train

python caption_train_run.py


After running the script, the model file will be stored under "models/caption/train". Under the model directory you will find model files including:

• 1.checkpoint file containing path to the model. If you plan to migrate your model or run the code under different OS environment, see Model Adjust
1. graph.pbtxt
1. model.ckpt-n.data-00000-of-00001
1. model.ckpt-n.index
1. model.ckpt-n.meta

#### 5. Run Inference

Now you could do the same thing as Caption Inference

#### 6. Run Evaluation

For NLG models, there're a few metrics for evaluation.

In our code package we use perplexity to measure the performance.

Here instead of using the model we just trained since we just trained it for 20 steps, we use ptm-im2txt-incv3-mlib-cleaned-3m model instead for evaluation. Download the models under /path/to/project/dir/models/caption

Open /vagrant/evaluation/caption_eval_run.py and configure the following fields:

1. DATA_DIR. Path to directory containing the saved evaluate TF files
1. MODEL_DIR. Directory containing model files to evaluate.
1. GPU_DEVICE. Gpu device to evaluate your model, use integer number to refer to the device. Default set to 0

You could also locate the file under the project_directory/evaluation/caption_eval_run.py

# Path to directory containing the saved evaluate TF files
DATA_DIR = "../data_preparation/TF"

# Directory containing model files to evaluate.
MODEL_DIR = "../models/caption/mscoco2017_3M/train/"

# Select gpu device to evaluate your model, use integer number to refer to the device: e.g. 0 -> gpu_0
GPU_DEVICE = 0


Notice in vagrant box, gpu acceleration is limited, so for evaluation purpose we use local deployed code pakcage as demo

In this case we need to exit vagrant box (if you are vagrant box), run code in command line

exit

logout
Connection to 127.0.0.1 closed.


then you run

cd /path/to/project/dir/evaluation

python caption_eval_run.py


## DATA PREPARATION

Data from any source for caption generating model need following the required pattern.

Both images files in JPEG format and descriptive annotation files in json format are required.

• JPEG image
• JSON annotation file containing metadata of caption associated with JPEG image

#### JSON Annotation File Example


[
images:[{
"file_name": "/vagrant/data/dc_bpc/125853.jpg", "id": 57870
}, ...
]
annotations:[{  id: 1,  image_id: 57870,  file_name: "COCO_train2014_000000000009.jpg"  caption: "a cat and dog on wooden floor next to a stairwell."}, ...
]



The sheeko pretrained models are trained using the data source below:

### Clean Data

This step is not required if your data set is clean and doesn't contain any noise (e.g. proper-noun). In this section we're removing proper-noun type noises using spaCy based NLP process in data set.

#### Toolkit

API: spaCy

Model: en_core_web_sm

#### What does it do?

• Remove proper-noun noise

• Simplify annotation file by leaving image_id and metadata field only

• Save cleaned annotation file in the destination directory following the same directory structure

• Create data report in JSON format under destination directory

• valid metadata: total number of captions that have valid JPEG images found and got non empty result after NLP process
• dir name: /path/to/data/dir,
• none pronoun: total number of captions that have valid JPEG images found and are proper-noun noise free
• pronoun: total number of captions that have valid JPEG images found and have proper-noun noise
• data count: total number of captions that have valid JPEG images found

#### Neuro-linguistic programming (NLP) Steps

##### step 1: PRONOUN cleaning
• 1.1 replace named person with "person"

example: Mary Jane with daughters --> a person with daughters

• 1.2 remove DATE, TIME, ORG, GPE, LOC, PRODUCT, EVENT, WORK ART, LAW, LANGUAGE, FAC, NORP, PERCENT, MONEY, QUANTITY, ORIGINAL
and get their parent node, if there's no parent node, then abort this sentence)

example: A women group is photographed at the Ashley Power Plant up Taylor Mountain road --> A women group is photographed

##### step 2: Parent nodes cleaning
• 2.1 if the parent node is 'DET', 'NOUN', 'NUM', 'PUNCT', 'CCONJ', 'VERB', remove the subtrees nodes of the lefts if it's not in the list
• 2.2 if the parent node is PREP or other types, remove the nodes and their children nodes

example: A woman tends to a fire at a Ute Indian camp near Brush Creek. --> A woman tends to a fire at a camp .

##### step 3: NOUN entities cleaning
• 3.1 replace the chunk of non-named enties chunk with simple entity along with CC, CD, DT and IN
• 3.2 replace CD type number with string value, e.g. 3, three

example: Ute couple with child stand in front of old cars --> couple with child stand in front of cars

##### step 4: Reform
• 4.1 put person entity and chunk replacement in position
• 4.2 replace person entities with understandable words, e.g. ,if multiple people then use num people instead, e.g. "a person" "two people", else say "a group of people"
• 4.3 replace nums of person entities (>=3) replace with "a group of " + noun
• 4.4 remove the tokens with index in indexes list of the sentence and convert the array to sentence

### Build Data

#### Description

Formatted data in structure is required for the package. To create format data we need to go through build data section.

#### What does it do?

• Filter data set by getting data that have both images and associating annotation only
• Segment images and annotation files into training and testing data sets
• Resize image to trainable format into training and testing data set
• Build annotation file for training and testing data set
##### Segment Mode

We provide two segment methods:

1. segment all data through out all directories as one into training and testing data set.

2. segment the list of directories into training and testing sets.

• seg_by_image segment data set in each directory with given percentage of data into training set, the rest into test

• seg_by_dir segment directories with given percentage into training set, the rest of directories into test

#### Steps

• Make clear the purpose: Training, Inference or Evaluation
• Prepare your data according to the purpose
• Make sure you have enough space in the output directory to store the output result ( 6 X of original data set in file size)

#### For particular purpose

##### For Training
• Prepare images file in JPG format (single file size no more than 15 mb) and associating annotation files in JSON format. Structure your data: image file name has to be unique id and annotation file has to has the same image_id of image file name (e.g. 15376.jpg, annotation: "image_id": 15376 )
• Structure your data: annotation files have to use the consistent field for getting metadata (e.g. "description_t")
##### For cleaning up proper-noun and data noise in your annotation file by using clean_data_run.py and configure field name of metadata field, list of paths to annotation files and output directory, for more details, please see Data Cleaning page.
• Configure arguments in build_data_run.py (field names in annotation file, list of paths to annotation files and image files, output directory and data segmentation args, e.g. method and training set percentage ) Run build_data_run.py
• Formatted data will be available in the directory you specified as OUTPUT_PATH in build_data_run.py
• For im2txt captioning model training. Run build_TF_run.py to generating TF Records. Each image object contains "file_name" and "id". Each annotation object contains "id", "image_id" and "caption". Each image object may refers to multiple caption objects
• TF Records will be located under OUTPUT_PATH which is runnable data for the training
##### For Inference
• Prepare images file for inference in JPG format
• Prepare checkpoint file of model and corresponding vocabulary file
• Configure arguments in build_data_run.py (field names in annotation file, list of paths to annotation files and image files and output directory)
• Run build_data_run.py
• Inference results will be stored in OUTPUT_PATH in caption_infer.py

### Build TF Records

To make data runnable by training script, data need to be converted into TF Record format. Each TF record represents serial pairs of encoded image-caption data. Please see Convert Data into TF Records in Walkthrough section to see details.

## TRAINING A MODEL

Our code package provides the code to train your customized model for generating captions.

### Keys to get a good model

Here're the keys for getting a good models: 1). Data 2). Steps to train your model

Data quality and labels in dictionary decide what your model finally outouts.

Total number of training steps will decide whether your model will be underfitted or overfitted. Unfortunately here's no tutorial telling you how many steps you need to configure for getting that right point since that depends on your data. In our experience from im2txt, 3 million steps is getting the good model for MSCOCO data.

### Accelerate your training

GPU with higher memory can accelerate your training. This code package doesn't support multi-GPU training. However, you could do parallel training with multi-GPU. Specify the GPU by assigning the index of GPU. e.g. 1

### Initial Training

To run the training, an Inception model is required. You could download Inception v3 , See training walkthrough for details.

### Fine Tune the im2txt Model

You could transfer your training with new dataset. The output will follow the new dataset's pattern. Convert new dataset into TF Records then replace the DATA_DIR with the TF Records' path.

## INFERENCE

### Image Caption

Caption model translates given JPEG image into natural language text.

### Image Classification

Classification model translates given JPEG image into single label out of 1000 classes

### Object Detection

Object Detection model identifies the objects in the given JPEG image. You could find models here

Please see Inference Walkthrough for code walkthrough

## EVALUATION CAPTION MODEL

Code package provides evaluation script for caption model.

### Metric

Caption model is Natural-language generation model(NLG model). For NLG model, see metrics to measure performance.

In our code package we use perplexity as our metric to measure model's performance

Go to Evaluation Walkthrough for code walkthrough

## SHEEKO PRETRAINED MODELS RESOURCE

This project provides Sheeko Pretrained Models Resource for generating captions. Models with description are available for downloading. It's highly recommended to try the downloaded model in Inference Walkthrough to test the performance.

You can’t perform that action at this time.