#!/usr/bin/env python
# coding: utf-8

# # Sketch Around Transkribus Python API
# 
# [Transkribus](https://transkribus.eu/Transkribus/) is a desktop application fronted tool for transcribing handwritten texts and generating machine learning models that can then be used to automatically transcribe such texts.
# 
# As well as the desktop app, there's also a [simple python API](https://github.com/Transkribus/TranskribusPyClient) for the Transkribus web service [[docs](http://htmlpreview.github.io/?https://github.com/Transkribus/TranskribusPyClient/blob/master/src/TranskribusPyClient/client.html) and [wiki](https://github.com/Transkribus/TranskribusPyClient/wiki)].
# 
# The wiki docs are primarily focused on calling the API from the command line.
# 
# Whilst we can use CLI calls from `%bash` cells in a notebook, and so doing documenting (or helping automate) a workflow, I'm more interested in using the package in a Python scripting context.
# 
# So what can we do with it?

# ## Installation
# 
# At the moment, the package is not pip installable, so we need to download the repo and then import the required module from a file:

# In[ ]:


get_ipython().run_cell_magic('capture', '', '!wget https://github.com/Transkribus/TranskribusPyClient/archive/master.zip\n!unzip master.zip\n')


# In[1]:


get_ipython().run_line_magic('cd', 'TranskribusPyClient-master/src')
from TranskribusPyClient.client import TranskribusClient
get_ipython().run_line_magic('cd', '../..')


# ## The `TranskribusClient`
# 
# To get started with the Python Transkribus API client, we need to get an instance of it:

# In[2]:


t = TranskribusClient()


# The service that client calls is an authenticated one, so we need to supply credentials for it:

# In[5]:


from getpass import getpass 
  
user = input('User: ')
pwd = getpass('Password: ')


# Logging in updates the state of the client:

# In[6]:


t.auth_login(user, pwd)


# In[7]:


#Review the methods available from the client
#dir(t)


# Many of the client calls require a collection ID, but I can see how to request available collection IDs directly?
# 
# We can, however, find client ID values from any recent jobs we've run, such as jobs run via the desktop application:

# In[24]:


#We can find recent collection IDs from jobs...
jobs = t.getJobs()

colIds = list({j[k] for j in jobs for k in j if k=='colId'})
colIds_str = ', '.join([str(c) for c in colIds])

print(f'Recent collectionIDs: {colIds_str}')


# We can then get a list of documents associated with a particular collection:

# In[32]:


#How do we get the collectionId?
docs = t.listDocsByCollectionId(colIds[0])

md=''
for doc in docs:
    md = f"{md}\n### docId: {doc['docId']}\n{doc['title']}, {doc['nrOfPages']} pages\n"
    
print(md)


# ## Downloading Transkribus Data Structure XML
# 
# We can download the XML metadata associated with a document's transcripts either as a parsed `lxml.etree` document (`bParse=True [default]`) or as a text string. 

# In[71]:


col_id = colIds[0]

#doc_id = [d['docId'] for d in docs if d['title']=='English Handwriting 0.1'][0]
doc_id = 268888


# In[72]:


#Get XML for a doc
_xml_str = t.getDocByIdAsXml(col_id, doc_id, bParse=False)


# If we grab the text string, we can use the convenient `xmltodict` Python package to convert the XML text to a `dict`:

# In[73]:


#!pip3 install xmltodict


# In[74]:


import xmltodict

xml_dict = xmltodict.parse(_xml_str)
xml_dict.keys()


# The document comes in three parts:
# 
# - metadata (`md`)
# - page data [`pageList`]
# - collection information [`collection`]

# In[75]:


xml_dict['trpDoc'].keys()


# In[76]:


xml_dict['trpDoc']['md']


# In[82]:


for k in xml_dict['trpDoc']['collection'].keys():
    print(k,xml_dict['trpDoc']['collection'][k])


# The `pageList` structure contains transcript information about each page:

# In[77]:


xml_dict['trpDoc']['pageList']['pages']['tsList']['transcripts'][0]


# ## Downloading Transkribus Documents
# 
# We can download complete Transkribus documents given a collection and document ID.
# 
# The complete document contains the coordinates for segmented text regions as well as the transcript for each region Image downloads for each page are enabaled by default (`bNoImage=False`).
# 
# If the specified download folder does not exist, create it. If it does exist, and `bForce=False [default]` an error is raised; if the argument is true, delete the directory, and recreate an empty one of the same name.

# In[10]:


download_dir = 'testDOwn'


# In[87]:


t.download_document(col_id, doc_id, download_dir)


# In[88]:


get_ipython().system('ls $download_dir')


# The `trp.json` file looks like it's the Transkribus data structure we could download as the XML file.
# 
# *(It would be nice if `getDocByIdAsXml()` were refactored as `getDocById()` with a switch allowing for `xml|json` and the `bParse` flag, when set, returning the `etree` or a Python `dict` correspondingly.)*

# In[81]:


get_ipython().system('cat $download_dir/trp.json')


# The `png` file is an image file for the page:

# In[4]:


from IPython.display import Image
Image('testDOwn/HO-40-2_13_14.png')


# The `pxml` file contains the line segmentationa and transcript information.

# ## Parsing the XML and Generating a Markdown Document With Individual Text Lines and Transcript
# 
# The aim here is to see if we can take the XML from a document that has already been segmented and transcribed and process it in some way. For example, crop out each line from the image file and place it in a markdown document, along with the transcription of that line.

# We could represent the `pxml` file as an ordered Python dict using `xmltodict`, or we could parse it as an XML document. Let's do the latter for now:

# In[25]:


import xml.etree.ElementTree as ET
def transkribus_tree(fn):
    with open(fn) as f:
        _xml = f.read()

        #Clean the XML of namespace cruft
        _xml = _xml.replace('xmlns="http://schema.primaresearch.org/PAGE/gts/pagecontent/2013-07-15" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://schema.primaresearch.org/PAGE/gts/pagecontent/2013-07-15 http://schema.primaresearch.org/PAGE/gts/pagecontent/2013-07-15/pagecontent.xsd"','')

        #Parse the XML
        tree = ET.fromstring(_xml)
    return tree


# In[28]:


tree = transkribus_tree('testDown/HO-40-2_13_14.pxml')


# For each text region, which may contain one or more text lines, each with its own transcript, a complete transcript is also provided:

# In[165]:


get_ipython().system('head -n 200 testDown/HO-40-2_13_14.pxml | tail -n 30')


# We can parse this out as follows:

# In[164]:


for t in tree.findall('Page/TextRegion/TextEquiv/Unicode'):
    print('...\n',t.text)


# We can also pull out the co-ordinates defining lines identified within each text region.
# 
# Here's what the XML looks like:

# In[120]:


get_ipython().system('head -n 45 testDown/HO-40-2_13_14.pxml')


# So let's get some co-ordinates...

# In[181]:


for line in tree.findall('Page/TextRegion')[1].findall('TextLine')[3:6]:
    te = line.find('TextEquiv/Unicode')
    if te is not None:
        print(te.text, ' ::: ', line.find('Coords').attrib['points'],'\n')


# Using these co-ordinates, can we crop a region of the image corresponding to the text?
# 
# How about trying with OpenCV?

# In[9]:


import numpy as np
import cv2

img = cv2.imread("testDOwn/HO-40-2_13_14.png")


# Preview the image:

# In[10]:


from matplotlib import pyplot as plt

plt.imshow(img)
plt.title('Sample Page')
plt.show()


# Extract some co-ordinates:

# In[12]:


_pts =  tree.findall('Page/TextRegion')[1].findall('TextLine')[4].find('Coords').attrib['points']

pts = np.array([[int(_p.split(',')[0]), int(_p.split(',')[1])] for _p in _pts.split()])

pts[:3]


# And then via a [handy little script](https://stackoverflow.com/a/48301735/454773) I found on Stack Overflow we can crop out the specified region:

# In[ ]:


import matplotlib as mpl

def getCroppedArea(img, pts, dpi=300):
    """Get a cropped area from an image.
        Via: https://stackoverflow.com/a/48301735/454773"""
    
    _dpi = mpl.rcParams['figure.dpi']
    mpl.rcParams['figure.dpi'] = dpi
    
    ## (1) Crop the bounding rect
    rect = cv2.boundingRect(pts)
    x,y,w,h = rect
    cropped = img[y:y+h, x:x+w].copy()

    ## (2) make mask
    pts = pts - pts.min(axis=0)

    mask = np.zeros(cropped.shape[:2], np.uint8)
    cv2.drawContours(mask, [pts], -1, (255, 255, 255), -1, cv2.LINE_AA)

    ## (3) do bit-op
    dst = cv2.bitwise_and(cropped, cropped, mask=mask)

    ## (4) add the white background
    bg = np.ones_like(cropped, np.uint8)*255
    cv2.bitwise_not(bg, bg, mask=mask)
    cropped_area = bg + dst
    
    return cropped_area


# Let's see if it works...

# In[13]:


sentence = getCroppedArea(img, pts, 400)

plt.figure(figsize = (15,5))
plt.axis('off')

plt.imshow(sentence)

#Save the image
plt.savefig("test.png", dpi=400)


# We can also embed the saved image, of course...

# In[59]:


Image("test.png")


# This sets up the possible of creating a markdown file, for example, that embeds single line images as well as transcripts. Using Jupytext, such a document can be edited in a notebook editor.

# In[27]:


#!pip3 install --upgrade tqdm

import os.path
import hashlib
import pathlib
from tqdm.notebook import tqdm

def markdownFromRegion(region, imgfile, fn='testout.md', outdir='test'):
    """Generate markdown page for Transkribus text region."""
    
    #Make sure the path to the outdir exists...
    pathlib.Path(outdir).mkdir(parents=True, exist_ok=True) 
    
    md = ''
    
    img = cv2.imread(imgfile)
    
    #Provide a progress bar using tqdm
    for line in tqdm(region.findall('TextLine')):
        _te = line.find('TextEquiv/Unicode')
        _pts =  line.find('Coords').attrib['points']

        pts = np.array([[int(_p.split(',')[0]), int(_p.split(',')[1])] for _p in _pts.split()])

        sentence = getCroppedArea(img, pts, 400)

        #Save the image
        _img_uid = hashlib.md5(str(pts).encode()).hexdigest()
        ifn = f"img_{_img_uid}.png"
        
        cv2.imwrite(os.path.join(outdir, ifn), sentence) 
        
        _txt = line.find('TextEquiv/Unicode')
        _txt = _txt.text if _txt is not None else ''
        
        md = f'{md}![{_img_uid}]({ifn})\n\n\n{_txt}\n\n\n'
        
        with open(os.path.join(outdir, fn), 'w') as f:
            f.write(md)


# Running it seems to be pretty quick...

# In[30]:


tree = transkribus_tree('testDown/HO-40-2_13_14.pxml')
markdownFromRegion(tree.findall('Page/TextRegion')[1], "testDOwn/HO-40-2_13_14.png")    


# Here's an example of what the md raw, and rendered, looks like:

# In[32]:


Image("transkribus_testout_md.png")


# ## Working with a 1 or 2 page doc on Transkribus
# 
# Let's see if we can create a collection conting a couple of one or two page PDFs, segment the text lines and retrieve the XML and JPG, then segment the JPG into a markdown document.

# In[ ]:


# Create a collection



# In[7]:


#Get one or two pages out of a PDF as a PDF




# In[8]:


# Submit the doc



# In[ ]:


# Segment the Lines



# In[9]:


# Get the XML and JPG Back



# In[ ]:


# Crop each line, saving image fragments




# In[ ]:


# Insert image link to each fragment into md along with trancscription text




# ## Submitting Content Back to the Transkribus Server
# 
# If we implemented a transciption client via a Jupyter extension, for example, we'd want to be able to submit the transcript back.
# 
# The `postPageTranscript()` method allows us to do that. See the [docs](http://htmlpreview.github.io/?https://github.com/Transkribus/TranskribusPyClient/blob/master/src/TranskribusPyClient/client.html) for the XML format that we need to return. 

# In[ ]: