I would like to use this code for a Project. Could you please create a LICENSE.md file, stating the type of license you are using?

💡 Tip:
MIT License would be great for this project.

I get a tick before the first character of every line, is there a way to remove this?
e.g.

Hi,
I am working on a project where I need to define text from windows pre-defined font files.
How can styles (strokes and char files) be created for texts generated with .ttf files?

Thank you in advance.

Regards,
Roy.

Hi, in its current form, this repo is unusable without too much hassle.

I have successfully migrated this to tfv2, so you don't need to waste your time

I'm planning to add web GUI and docker image too

https://github.com/otuva/handwriting-synthesis

System:

• Windows 10 1803
• Python 3.6.2
• TensorFlow 1.11.0

Error:

Traceback (most recent call last):
File "test.py", line 1, in <module>
    import demo
File "C:\Projects\ml\handwriting-synthesis\demo.py", line 9, in <module>
    from rnn import rnn
File "C:\Projects\ml\handwriting-synthesis\rnn.py", line 10, in <module>
    from rnn_ops import rnn_free_run
File "C:\Projects\ml\handwriting-synthesis\rnn_ops.py", line 9, in <module>
    from tensorflow.python.ops.rnn_cell_impl import _concat, _like_rnncell
ImportError: cannot import name '_like_rnncell'

Solution:

Hacky, but I simply removed the reference to _like_rnncell on lines 9, 29, and 30 in rnn_ops.py. Seems to run fine afterwards (though this led to encountering Issue #17).

Hello,

I want to create a new style of handwriting using the same pretrained model.

Thanks

When I look at the relevant code for your data preprocessing, I don't quite understand the method used in it. Can you tell me the relevant literature on your method?

no state attrivute is found in call method of the class "tf.keras.layers.LSTMCell"

First of all, thanks for this amazing tool! It looks incredible. I'm having a little trouble getting it to run locally with Python3 on my Macbook, but that may because I'm not super familiar with python.

Here's the error when I run python3 demo.py

  File "/handwriting-synthesis/rnn_ops.py", line 246, in rnn_free_run
    states, outputs, final_state = raw_rnn(cell, loop_fn, scope=scope)
  File "/handwriting-synthesis/rnn_ops.py", line 40, in raw_rnn
    if context.in_graph_mode():
AttributeError: module 'tensorflow.python.eager.context' has no attribute 'in_graph_mode'

I found tensorflow/tensorflow#14069 which suggested trying

pip install --upgrade --force-reinstall tf-nightly

but they're talking mostly about 1.3 and 1.4 there, while I see you have 1.6 in the requirements.txt.

Can you point me in the right direction?

Something strange happens but only with the few of the consecutive letters like:
HHHHHHHHHHH
EEEEEEEEEEEEEEE
AAAAAAAAAAAA
KKKKKKKKKKKKK
ZZZZZZZZZZZZZ
XXXXXXXXXXXXX
BBBBBBBBBBBBB

Result are unexpected, by having either completely weird output or something unexpected like just single parallel lines

Hey

What changes need to be made to enable training for other languages, other than the training data? I can see some style files being used, does that need to change as well? In which case, could you share how the style.npy files were created.

What resource are there to mail merge for a list of letters but with minor differences like name or address?

Any idea about this error? I get it when running python demo.py
18.04 Ubuntu running stock Python 2.7.15rc1
Have installed:
apt install python-tk
pip install numpy svgwrite matplotlib scipy tensorflow pandas sklearn

Full error:
Traceback (most recent call last): File "demo.py", line 153, in <module> hand = Hand() File "demo.py", line 37, in __init__ attention_mixture_components=10 File "/home/john/handwriting-synthesis/rnn.py", line 85, in __init__ super(rnn, self).__init__(**kwargs) File "/home/john/handwriting-synthesis/tf_base_model.py", line 108, in __init__ self.graph = self.build_graph() File "/home/john/handwriting-synthesis/tf_base_model.py", line 399, in build_graph self.loss = self.calculate_loss() File "/home/john/handwriting-synthesis/rnn.py", line 204, in calculate_loss lambda: self.sample(cell) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/util/deprecation.py", line 432, in new_func return func(*args, **kwargs) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 2063, in cond orig_res_t, res_t = context_t.BuildCondBranch(true_fn) File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 1913, in BuildCondBranch original_result = fn() File "/home/john/handwriting-synthesis/rnn.py", line 203, in <lambda> lambda: self.primed_sample(cell), File "/home/john/handwriting-synthesis/rnn.py", line 162, in primed_sample scope='rnn' File "/home/john/handwriting-synthesis/rnn_ops.py", line 246, in rnn_free_run states, outputs, final_state = raw_rnn(cell, loop_fn, scope=scope) File "/home/john/handwriting-synthesis/rnn_ops.py", line 40, in raw_rnn if context.in_graph_mode(): AttributeError: 'module' object has no attribute 'in_graph_mode'

I realize this wasn't originally developed to be released as a project, but it the code has been posted and no license has been declared. This leaves a really weird limbo situation where the worst-faith actors are the most free to take advantage of the situation but honest actors who build systems that stick to licensing terms can't do anything with this at all.

It also inhibits contributions! You state in the readme that you'd be interested in accepting contributions which helped make this into a package and made it more usable. I'd be willing to contribute some towards that, but not until the licensing is clarified.

Issue #16 also broached this topic, but the OP unilaterally decided what they wanted it to be and then closed the issue and ran away. I also think the MIT license would be a good candidate for this project, but the original author(s) need to state their intention in this regard.

I'd be happy to help with advice on what different licenses are good for and how to pick. Once you do pick I'm happy to contribute the proper files to this repo so that it's properly licensed and usable by others. But I can't make the call what that license will be.

is there any pre existing way that anyone has come across to make custom data in the correct format?

Style 2 gets confused when the legibility is set to high.
Tests run on https://www.calligrapher.ai/

• python3 -m pip install -r requirements.txt # after #5 is merged
• python3 demo.py

Traceback (most recent call last):
  File "demo.py", line 153, in <module>
    stroke_widths=stroke_widths
  File "demo.py", line 42, in write
    strokes = self._sample(lines, biases=biases, styles=styles)
  File "demo.py", line 61, in _sample
    c_p = drawing.encode_ascii(c_p)
  File "~/handwriting-synthesis/drawing.py", line 77, in encode_ascii
    return np.array(map(lambda x: alpha_to_num[x], ascii_string) + [0])
TypeError: unsupported operand type(s) for +: 'map' and 'list'

Hello, is it posibble add new language ? How i can do it. thanks

is there a script that allows us to create our own style files from drawing, or a link to the program that was used to generate them? ^_^

Any chance of featuring extreme style like doctor?

Why did you use following code ?
-1*int(point.attrib['y']),
I think it should
int(point.attrib['y']),

I have an issue with the digits. In most of my test cases, digits are not produced very well and they are not readable. is it something that I can fix with re-training?

Hello, are there any plans to support the Chinese?

Output is aligned to the center, how do I align it to the left?
Thanks

The defined alphabet is missing 'Q', 'X' and 'Z' characters. Is there any reason behind this?

EDIT: Answering myself, this is probably due to low number of samples for those characters in the dataset.

I kept getting this error message when I was running this program. If you also run into this issue the resolution is:

Replace

np.ones_like in the rnn_cell.py file with keras.backend.ones_like.

You will also have to import keras as well so add from tensorflow import keras to the top of the rnn_cell.py file as well.

Hi,

I attempted to generate handwritten "@", "+" and "✓" signs, but was unable to.
E.g. the phone number +1234, and the email [email protected] will not generate anything at all.

Within the "write" function of demo.py (line 42) => valid_char_set = set(drawing.alphabet), alphabet is called from drawing.py. Therefore, I modified within drawing.py (see line 10), the list alphabet by adding the new characters. Adding "@" or "+" to the list caused major problems as even regular words such as "John Smith" can no longer be written i.e. no handwriting output at all.

I have identified where within demo.py, the code breaks after adding the new characters to the list. The line strokes = self._sample(lines, biases=biases, styles=styles) (line 61), calls the function _sample. Within _sample, it is from line 94 to line 106, which fails (see below)

    [samples] = self.nn.session.run(
        [self.nn.sampled_sequence],
        feed_dict={
            self.nn.prime: styles is not None,
            self.nn.x_prime: x_prime,
            self.nn.x_prime_len: x_prime_len,
            self.nn.num_samples: num_samples,
            self.nn.sample_tsteps: max_tsteps,
            self.nn.c: chars,
            self.nn.c_len: chars_len,
            self.nn.bias: biases
        }
    ) 

Would anyone be able to provide an explanation and solution for this please? i.e. explanation to why nothing is drawn at all, and why adding an additional character could not solve this. And what the solution would be to generate handwriting for "@", "+" and "✓"?

Many Thanks!

PS: I also attempted on the web demo to create"@", "+" and "✓". Although not drawn properly, at least handwriting can be generated.

How can I create a new dataset from images of handwriting? Are the XML files in the IAM set some sort of SVG format that I could interpret my images into? How many writing samples should it take to decently train the model?

I haven't found the version of this repo to work with latest Python 3.8 and TensorFlow 2.4, so working on updating it.

I used the TensorFlow upgrade scripts recommended by TensorFlow to convert the code from 1.X to 2.X. It can be found here https://www.tensorflow.org/guide/upgrade

Haven't completely converted yet, but working on it. Looks like there are some functions that are deprecated. Is anyone else working on upgrading to work with TF 2? Thanks

Demo 3 is "Rick Astley - Never Gonna Give You Up", not "Leonard Cohen - Hallelujah".

At multiple places int32 is being used for tensorflow computations. As you know, this prevents the code from using GPU. See this StackOverflow.

Could you please elaborate on why using float64 be a bad idea?
How much would the code need to be changed to make this happen?
Is this doable?

Hi @sjvasquez Nice one

Can we do Handwriting Synthesis for other languages.Am interested to implement for Tamil language
Tamil. If possible kindly let us know.

Why not make an API and let Python or other languages can use it?
My Idea:
- Python post a request to it,it has "text","style" and other keys.
- It recive a key named "text" and do something for it.
- It return a picture and it will named "xx.png"
Maybe.If the idea comes true,I will thanks a lot.

I just finished running a re-training session and the results were extremely impressive! However I'm noticing a loss in numbers. What are some ways I could improve the training?

Enter this string "Hawaii lieben leidige Fijii"
Style 6
Speed , legibility and stroke width at default .

After "Fijii" the Ai continues with another ~6 characters of jibber jabber in > 50% of the cases.

That might be an interesting corner case too look at.

Thanks for the great work.

Hello 👋 I would like to know if it possible to train a model on other languages like Hebrew, if u can help me with that I will be very happy 😊 THX🙏

Ran across https://www.calligrapher.ai on HN and someone mentioned that this repo appeared to be the source for that (or at least, the tech used). If so, that might be cool to call out in the readme.

There's technically no license in this repo, so depending on where you live this code is all rights reserved despite what's at the bottom of the readme, so if that's just someone who used your code without informing you, I guess you might now know >_>

I am trying to train the model. I have prepared the data by running prepare_data.py. When I run the rnn.py I am getting the following error:

Traceback (most recent call last):
File "rnn.py", line 235, in
nn.fit()
File "D:\Data\handwriting-synthesis-master\handwriting-synthesis-master\tf_base_model.py", line 163, in fit
fetches=[self.loss] + self.metrics.values(),
TypeError: can only concatenate list (not "dict_values") to list

Waiting for your reply. Thanks...

Is there a non minified version of the js code available?

Hi. where did you get the stroke.npy from for each of the styles?

Hi Everyone,
Is anyone aware of a public forum to tackle questions about this information? I believe this part of the forum is for technical bug's more then Q&A.
I would like to add a slight slant to the end of the sentences as most normal handwriting takes this into account as they slide to the end of the paper.

I would like to generate some words with accentuation marks, like those in Spanish, Portuguese or French. I know there is the RHIMES dataset that includes accents from French, but I am willing to generate spanish-only forms like ñ or ü. Take the next .svg for example:

Is there any way to manually add a - above the n during the generation process? Perhaps splitting the word before getting the samples and apply it just to the n?

Hi,

When I give words with capital Z and Q in them, I get a complaint that the model does not have those letters as its trained letters. However, I noticed that your web demo has not issue with those letter.
Could you please share your model that includes those characters?

I tried to install an older setup and still doesn't work.

Python: 3.5.4
TensorFlow: 1.6.0

Traceback (most recent call last):
  File "R:\handwriting-synthesis-master\demo.py", line 9, in <module>
    from rnn import rnn
  File "R:\handwriting-synthesis-master\rnn.py", line 9, in <module>
    from rnn_cell import LSTMAttentionCell
  File "R:\handwriting-synthesis-master\rnn_cell.py", line 16, in <module>
    class LSTMAttentionCell(tf.compat.v1.nn.rnn_cell.RNNCell):
AttributeError: module 'tensorflow.python.util.compat' has no attribute 'v1'

would you please give us a working environment? which python+tensorflow version would be appropriate?

I am having trouble with my installations, and would love to know from one of you who did manage to install and run what are your versions.
I tried:

Python3.7 with tensorflow 1.8/1.6/1.5 - all leading to missing modules such as:
ModuleNotFoundError: No module named '_pywrap_tensorflow_internal'
I followed this thread about it: tensorflow/tensorflow#20690

Python3.6.8 with tensorflow 1.6.0:
tensorflow couldn't be installed due to
protobuf requires Python '>=3.7' but the running Python is 3.6.8

The results are impressive, perhaps you can upgrade to using tf 2.x OR give instructions on how to install using tf 1.x. OR allow us to use calligrapher.ai programmatically.

I'm trying to create a portable windows library of it but tensorflow.contrib must be removed.