interpolation

Author: sjvasquez

.gitignore

@@ -1,6 +1,7 @@
 data/raw/ascii
 data/raw/lineStrokes
 data/raw/original
+data/processed
 
 logs
 predictions

demo.py

@@ -159,7 +159,7 @@ def _draw(self, strokes, lines, filename, stroke_colors=None, stroke_widths=None
     styles = [12 for i in lines]
 
     hand.write(
-        filename='img/all_star2.svg',
+        filename='img/all_star.svg',
         lines=lines,
         biases=biases,
         styles=styles,
@@ -171,7 +171,7 @@ def _draw(self, strokes, lines, filename, stroke_colors=None, stroke_widths=None
     styles = np.cumsum(np.array([len(i) for i in lines]) == 0).astype(int)
 
     hand.write(
-        filename='img/downtown2.svg',
+        filename='img/downtown.svg',
         lines=lines,
         biases=biases,
         styles=styles,
@@ -183,7 +183,7 @@ def _draw(self, strokes, lines, filename, stroke_colors=None, stroke_widths=None
     styles = [7 for i in lines]
 
     hand.write(
-        filename='img/give_up2.svg',
+        filename='img/give_up.svg',
         lines=lines,
         biases=biases,
         styles=styles,

drawing.py

@@ -3,6 +3,7 @@
 import matplotlib.pyplot as plt
 import numpy as np
 from scipy.signal import savgol_filter
+from scipy.interpolate import interp1d
 
 
 alphabet = [
@@ -93,6 +94,40 @@ def denoise(coords):
     return coords
 
 
+def interpolate(coords, factor=2):
+    """
+    interpolates strokes using cubic spline
+    """
+    coords = np.split(coords, np.where(coords[:, 2] == 1)[0] + 1, axis=0)
+    new_coords = []
+    for stroke in coords:
+
+        if len(stroke) == 0:
+            continue
+
+        xy_coords = stroke[:, :2]
+
+        if len(stroke) > 3:
+            f_x = interp1d(np.arange(len(stroke)), stroke[:, 0], kind='cubic')
+            f_y = interp1d(np.arange(len(stroke)), stroke[:, 1], kind='cubic')
+
+            xx = np.linspace(0, len(stroke) - 1, factor*(len(stroke)))
+            yy = np.linspace(0, len(stroke) - 1, factor*(len(stroke)))
+
+            x_new = f_x(xx)
+            y_new = f_y(yy)
+
+            xy_coords = np.hstack([x_new.reshape(-1, 1), y_new.reshape(-1, 1)])
+
+        stroke_eos = np.zeros([len(xy_coords), 1])
+        stroke_eos[-1] = 1.0
+        stroke = np.concatenate([xy_coords, stroke_eos], axis=1)
+        new_coords.append(stroke)
+
+    coords = np.vstack(new_coords)
+    return coords
+
+
 def normalize(offsets):
     """
     normalizes strokes to median unit norm
@@ -118,12 +153,22 @@ def offsets_to_coords(offsets):
     return np.concatenate([np.cumsum(offsets[:, :2], axis=0), offsets[:, 2:3]], axis=1)
 
 
-def draw(offsets, ascii_seq=None, align_strokes=True, denoise_strokes=True, save_file=None):
+def draw(
+        offsets,
+        ascii_seq=None,
+        align_strokes=True,
+        denoise_strokes=True,
+        interpolation_factor=None,
+        save_file=None
+):
     strokes = offsets_to_coords(offsets)
 
     if denoise_strokes:
         strokes = denoise(strokes)
 
+    if interpolation_factor is not None:
+        strokes = interpolate(strokes, factor=interpolation_factor)
+
     if align_strokes:
         strokes[:, :2] = align(strokes[:, :2])
 

img/banner.svg

@@ -1,5 +1,6 @@
 # Handwriting Synthesis
-Implementation of the handwriting synthesis experiments in the paper <a href="https://arxiv.org/abs/1308.0850">Generating Sequences with Recurrent Neural Networks</a> by Alex Graves.  The implementation very closely follows the original paper, with a few slight deviations, and the generated samples are of similar quality to those presented in the paper.
+![](img/banner.svg)
+Implementation of the handwriting synthesis experiments in the paper <a href="https://arxiv.org/abs/1308.0850">Generating Sequences with Recurrent Neural Networks</a> by Alex Graves.  The implementation closely follows the original paper, with a few slight deviations, and the generated samples are of similar quality to those presented in the paper.
 
 ## Usage
 ```python
@@ -48,12 +49,11 @@ The following samples were generated with varying style and fixed bias.  Each ve
 ### Demo #3
 The following samples were generated with a fixed style and varying bias.  Each verse has a lower bias than the previous, with the last verse being unbiased.
 
-**Leonard Cohen – Hallelujah (<a href="https://www.youtube.com/watch?v=dQw4w9WgXcQ">lyrics</a>)
-![](img/give_up.svg)**
+**Leonard Cohen – Hallelujah (<a href="https://www.youtube.com/watch?v=dQw4w9WgXcQ">lyrics</a>)**
+![](img/give_up.svg)
 
 ## Contribute
-This project was intended to serve as a reference implementation for a research paper, but since the results are of decent quality, it may worthwile to make the project more broadly usable.  I plan to continue focusing on the machine learning side of things.  That said, I'd welcome contributors who can:
+This project was intended to serve as a reference implementation for a research paper, but since the results are of decent quality, it may be worthwile to make the project more broadly usable.  I plan to continue focusing on the machine learning side of things.  That said, I'd welcome contributors who can:
 
   - Package this, and otherwise make it look more like a usable software project and less like research code.
   - Add support for more sophisticated drawing, animations, or anything else in this direction.  Currently, the project only creates some simple svg files.
-