init

app.py

@@ -0,0 +1,91 @@
+from imgofai import *
+
+import matplotlib.pyplot as plt
+
+import PIL
+
+import numpy as np
+from pathlib import Path
+
+from imgofai.tree import img2df, df2xy
+import pandas as pd 
+from sklearn.tree import DecisionTreeRegressor
+from sklearn.preprocessing import FunctionTransformer
+from sklearn.pipeline import make_pipeline
+
+import streamlit as st
+
+import requests
+
+st.write("# Images of AI Demo")
+
+st.write("This page demonstrates how I created the images I submitted for [Better Images of AI project](https://betterimagesofai.org/images)")
+
+def add_radial_features(X,y=None):
+    assert isinstance(X, pd.DataFrame), "X is not a dataframe"
+    xp = X.copy()
+    xp['dim0'] = np.sqrt(((X - X.mean())**2).sum(axis=1))
+    xp['dim1'] = np.arctan2(X['dim1'],X['dim0'])
+
+    X = pd.concat([
+        X,
+        xp,
+        ],axis=1)
+
+    return X
+
+
+
+def make_tree_approximator(radial = False, max_depth=4):
+    if radial:
+        model = make_pipeline(
+            FunctionTransformer(add_radial_features),
+            DecisionTreeRegressor(max_depth=max_depth),
+        )
+    else:
+        model =  DecisionTreeRegressor(max_depth=max_depth)
+
+
+    model.fit(x_raw, y)
+    pred = PIL.Image.fromarray(
+        model.predict(x_raw).reshape(img_array.shape).round().astype("uint8")
+    )
+    score = model.score(x_raw, y)
+    return pred
+
+
+st.write("## Try it out yourself:")
+
+url = st.text_input("Image url:", "https://images.unsplash.com/reserve/bOvf94dPRxWu0u3QsPjF_tree.jpg?ixlib=rb-1.2.1&ixid=MnwxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8&auto=format&fit=crop&w=1752&q=80")
+img = PIL.Image.open(requests.get(url, stream=True).raw)
+
+img_array = np.array(img)
+df = img2df(img_array)
+x_raw, y = df2xy(df)
+
+
+
+ccol1, ccol2, _ = st.columns(3)
+
+with ccol1:
+    max_depth1 = st.slider("max depth left:",1,12,2)
+    radial1 = st.checkbox("radial features left", value=False)
+
+with ccol2:
+    max_depth2 = st.slider("max depth middle:",1,12,6)
+    radial2 = st.checkbox("radial features middle", value=False)
+
+
+st.write("## Output")
+col1, col2, col3 = st.columns(3)
+
+with col1:
+    left_img = make_tree_approximator(radial1, max_depth=max_depth1)
+    st.image(left_img)
+
+with col2:
+    mid_img = make_tree_approximator(radial2, max_depth=max_depth2)
+    st.image(mid_img)
+
+with col3:
+    st.image(img)

imgofai/__init__.py

@@ -0,0 +1,8 @@
+import datetime
+from .tree import treeify
+
+__version__ = "0.1.0"
+
+
+def timestamp():
+    return datetime.datetime.now().strftime("%Y%m%d%H%M%S")

imgofai/tree.py

@@ -0,0 +1,101 @@
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+from sklearn.tree import DecisionTreeRegressor
+import PIL
+
+
+def img2df(img_array):
+    dim0_arr = (
+        np.arange(img_array.shape[0])
+        .reshape((-1, 1))
+        .repeat(img_array.shape[1], axis=1)
+        .flatten()
+    )
+
+    dim1_arr = (
+        np.arange(img_array.shape[1])
+        .reshape((1, -1))
+        .repeat(img_array.shape[0], axis=0)
+        .flatten()
+    )
+
+    df = pd.DataFrame({"dim0": dim0_arr, "dim1": dim1_arr})
+
+    values = img_array.reshape((img_array.shape[0] * img_array.shape[1], -1))
+    for col in range(values.shape[1]):
+        df[f"value{col}"] = values[:, col]
+
+    return df
+
+
+def normalize(img):
+    return (img - img.min()) / (img.max() - img.min())
+
+
+def df2xy(df):
+    x = df[[c for c in df if c.startswith("dim")]]
+    y = df[[c for c in df if c.startswith("value")]]
+
+    if len(y.columns) == 1:
+        y = y.values.reshape(-1)
+    return x, y
+
+
+def tree_window(
+    img, add_cartesian=True, add_rotation=False, add_polar=False, depths=(2, 6)
+):
+    df = img2df(img)
+    x_raw, y = df2xy(df)
+
+    sets = []
+    if add_cartesian:
+        sets.append(x_raw)
+    if add_rotation > 0:
+        # rotate
+        theta = np.radians(add_rotation)
+        c, s = np.cos(theta), np.sin(theta)
+        R = np.array(((c, -s), (s, c)))
+        xr = x_raw @ R
+        sets.append(xr)
+    if add_polar:
+        # polar
+        xp = x_raw.copy()
+        xp["dim0"] = np.sqrt(((x_raw - x_raw.mean()) ** 2).sum(axis=1))
+        xp["dim1"] = np.arctan2(x_raw["dim1"], x_raw["dim0"])
+        sets.append(xp)
+
+    x = pd.concat(sets, axis=1)
+
+    fig, axes = plt.subplots(ncols=len(depths) + 1, figsize=(36, 36))
+
+    for ax, depth in zip(axes, depths):
+        model = DecisionTreeRegressor(max_depth=depth)
+        model.fit(x, y)
+        pred = model.predict(x).reshape(img.shape)
+        if len(y.shape) == 2:
+            ax.imshow(pred)
+        else:
+            ax.imshow(pred, cmap="gray")
+        ax.set_axis_off()
+
+    if len(y.shape) == 2:
+        axes[-1].imshow(img)
+    else:
+        axes[-1].imshow(img, cmap="gray")
+    axes[-1].set_axis_off()
+
+    return fig
+
+
+def treeify(img, max_depth):
+    img_array = np.array(img)
+    df = img2df(img_array)
+    x, y = df2xy(df)
+    model = DecisionTreeRegressor(max_depth=max_depth)
+    model.fit(x, y)
+    pred = PIL.Image.fromarray(
+        model.predict(x).reshape(img_array.shape).round().astype("uint8")
+    )
+    score = model.score(x, y)
+    return pred, score

requirements.txt

@@ -0,0 +1,7 @@
+matplotlib
+Pillow
+numpy
+pandas
+scikit-learn
+streamlit
+requests