Custom Dataset, train.py not creating 'model.ckpt-0'

[frederickk]

caveat: I'm very noob with all things machine learning as well as python

I've got the sample data working just fine e.g. clone repo and run python sample.py --filename example_name --sample_length 1000

Now, I'm trying to use my own SVG dataset (placed within ./data and when I run python train.py after about 2–3 minutes it generates a strokes_training_data.cpkl in the ./data folder and config.pkl in the ./save folder.

I've ran other examples of neural training and typically it takes much longer to train a dataset and there are also a series of model checkpoint (?) files e.g. model.ckpt-0 generated per epoch (?). Is there a reason why I'm not getting the same results?

[hardmaru]

Hi Ken

I'm not familiar with your dataset. Part of your learning experience is to
understand the limitations of this model and why it works / doesn't work
well with your data. Have fun :)

On 19 June 2016 at 00:19, kenfrederick notifications@github.com wrote:

caveat: I'm very noob with all things machine learning as well as python

I've got the sample data working just fine e.g. clone repo and run python
sample.py --filename example_name --sample_length 1000

Now, I'm trying to use my own SVG dataset (placed within ./data and when
I run python train.py after about 2–3 minutes it generates a
strokes_training_data.cpkl in the ./data folder and config.pkl in the
./save folder.

I've ran other examples of neural training and typically it takes much
longer to train a dataset and there are also a series of model checkpoint
(?) files e.g. model.ckpt-0 generated per epoch (?). Is there a reason
why I'm not getting the same results?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#5, or mute the
thread
https://github.com/notifications/unsubscribe/AGBoHkfIpUam4F5ksTlp4hzkLfV_h_Hpks5qNO12gaJpZM4I5HA9
.

[hardmaru]

I would also investigates the average number of line segments per sample in
your .svg dataset, and compare it to the typical average number of line
segments per typical kanji in the kanji vg dataset, that should provide
some clues for performance and required time for epoch

On 19 June 2016 at 00:22, hard maru hardmaru@gmail.com wrote:

Hi Ken

I'm not familiar with your dataset. Part of your learning experience is
to understand the limitations of this model and why it works / doesn't work
well with your data. Have fun :)

On 19 June 2016 at 00:19, kenfrederick notifications@github.com wrote:

caveat: I'm very noob with all things machine learning as well as python

I've got the sample data working just fine e.g. clone repo and run python
sample.py --filename example_name --sample_length 1000

Now, I'm trying to use my own SVG dataset (placed within ./data and when
I run python train.py after about 2–3 minutes it generates a
strokes_training_data.cpkl in the ./data folder and config.pkl in the
./save folder.

I've ran other examples of neural training and typically it takes much
longer to train a dataset and there are also a series of model checkpoint
(?) files e.g. model.ckpt-0 generated per epoch (?). Is there a reason
why I'm not getting the same results?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#5, or mute the
thread
https://github.com/notifications/unsubscribe/AGBoHkfIpUam4F5ksTlp4hzkLfV_h_Hpks5qNO12gaJpZM4I5HA9
.

[frederickk]

Fair :)

So this type of "error" comes more from the data? For reference I have a collection of SVG's that more or less take on this structure:

<svg version="1.1" id="Layer_1" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" x="0px" y="0px" viewBox="140 -298 1000 1000" style="enable-background:new 140 -298 1000 1000;" xml:space="preserve">
<style type="text/css">
    .st0{fill:none;stroke:#000000;stroke-width:13;stroke-linecap:round;stroke-miterlimit:10;}
    .st1{fill:none;stroke:#FF0000;stroke-width:13;stroke-linecap:round;stroke-miterlimit:10;}
</style>
<circle id="XMLID_647_" cx="319.9" cy="202" r="179.9"/>
<path id="XMLID_646_" class="st0" d="M1140,114.3v175.3"/>
<path id="XMLID_645_" class="st0" d="M1033.3,122.6v158.7"/>
<path id="XMLID_642_" class="st0" d="M926.6,123.9V280"/>
<path id="XMLID_641_" class="st1" d="M819.9,191.4v21.1"/>
<path id="XMLID_640_" class="st0" d="M713.2,114.3v175.3"/>
<path id="XMLID_563_" class="st1" d="M606.5,191.4v21.1"/>
</svg>

I'll keep plugging away at it, but if you have any tips, I'm all ears.

Thanks!
Ken

[impactcolor]

@frederickk did you ever figure out how to do your own dataset and train it? If so could you share the experience?