From 1d52c5a78e4b3af9db04b85be751c841f1511c8d Mon Sep 17 00:00:00 2001
From: henine <henine@raptorpond.com>
Date: Tue, 2 May 2023 08:51:41 +0200
Subject: [PATCH] something

---
 aesthetic_loss.py |  66 +++-
 colors.py         |  31 ++
 patch_gen.ipynb   | 398 +++++++++++++++++++++++
 patch_gen.py      | 127 +++++++-
 playground.ipynb  | 794 ++++++++++++++++++++++++++++++++++++++++++++++
 playgrund.ipynb   | 155 ---------
 polymap.py        |  28 ++
 7 files changed, 1439 insertions(+), 160 deletions(-)
 create mode 100644 colors.py
 create mode 100644 patch_gen.ipynb
 create mode 100644 playground.ipynb
 delete mode 100644 playgrund.ipynb
 create mode 100644 polymap.py

diff --git a/aesthetic_loss.py b/aesthetic_loss.py
index b504577..959fbb5 100644
--- a/aesthetic_loss.py
+++ b/aesthetic_loss.py
@@ -1,5 +1,65 @@
 import tensorflow as tf
 
-def my_loss_fn(y_true, y_pred):
-    
-    return tf.reduce_mean(squared_difference, axis=-1)  # Note the `axis=-1`
+from colors import vector_to_index
+
+
+@tf.function
+def aesthetic_loss(y_pics):
+
+    flipx = tf.reverse(y_pics, axis=(-2,))
+    flipy = tf.reverse(y_pics, axis=(-3,))
+    center = tf.reverse(y_pics, axis=(-2, -3,))
+    trans = tf.transpose(y_pics, perm=(0, 2, 1, 3))
+    anti_trans = tf.reverse(tf.transpose(tf.reverse(
+        y_pics, axis=(-2,)), perm=(0, 2, 1, 3)), axis=(-2,))
+
+    flipx_score = compute_score(y_pics, flipx)
+    flipy_score = compute_score(y_pics, flipy)
+    trans_score = compute_score(y_pics, trans)
+    anti_trans_score = compute_score(y_pics, anti_trans)
+    center_score = compute_score(y_pics, center)
+
+    # Extract patches for neighborhood check
+    patches = tf.image.extract_patches(
+        y_pics, (1, 3, 3, 1), (1, 1, 1, 1), (1, 1, 1, 1), padding="SAME")
+    # Dot product with central pixel
+    # neigh = tf.reduce_sum(
+    #     tf.maximum(0.0, -tf.einsum("...k,...k->...", tf.repeat(y_pics, 9, -1), patches)))
+    neigh = tf.reduce_sum(
+        tf.abs(tf.einsum("...k,...k->...", tf.repeat(y_pics, 9, -1), patches)))
+
+    divers = tf.reduce_sum(
+        tf.square(tf.einsum("ijkl,mjkl->imjk", y_pics, y_pics)))
+
+    # print(flipx_score)
+    # print( tf.reduce_max(tf.stack([flipx_score, flipy_score, trans_score, anti_trans_score, center_score]), axis=(0,)))
+
+    # idxs = vector_to_index(y_pics)
+    # ncol = tf.map_fn(lambda p: tf.size(
+    #     tf.unique(tf.reshape(p, [-1]),)[0], out_type=tf.int64), idxs)
+    # print(ncol)
+
+    return tf.reduce_sum([
+        -1.0/(64) * tf.reduce_sum(tf.reduce_max(tf.stack([flipx_score, flipy_score,
+                                                          trans_score, anti_trans_score, center_score]), axis=(0,))),
+        -1.0/(64*9) * 0.1 * neigh,
+        #   -1.0/7 * 10 * tf.cast(ncol, tf.float32),
+        # 1.0/64 * 0.01 * tf.reduce_sum(tf.square(y_pics)),
+        1.0/64 * 10 * divers
+    ], )
+
+
+@tf.function
+def compute_score(pic1, pic2):
+    return tf.reduce_sum(
+        tf.abs(
+            tf.einsum(
+                "...k,...k->...",
+                pic1,
+                pic2)), axis=(-2, -1))
+    # return tf.reduce_sum(
+    #     -tf.math.maximum(0.0,
+    #                     -tf.einsum(
+    #                         "...k,...k->...",
+    #                         pic1,
+    #                         pic2)), axis=(-2, -1))
diff --git a/colors.py b/colors.py
new file mode 100644
index 0000000..9fd2c5a
--- /dev/null
+++ b/colors.py
@@ -0,0 +1,31 @@
+from math import pi
+import tensorflow as tf
+
+squares = [chr(i) for i in range(0x1F7E5, 0x1F7EC)]
+tfsquares = tf.constant(squares)
+colors = tf.constant([10, 219, 38, 48, 80,  282, 20] *
+                     tf.constant(pi/180.0), dtype=tf.float32)
+color_vectors = tf.transpose(
+    tf.stack([
+        tf.math.cos(colors),
+        tf.math.sin(colors)]
+    )
+)
+
+
+def vector_to_index(tensor):
+    return tf.argmax((tf.einsum("...ijk,lk->...ijl", tensor, color_vectors)), axis=-1)
+
+
+def index_to_string(tensor):
+    tfstring = tf.strings.join(
+        tf.map_fn(
+            lambda v:
+            tf.strings.join(
+                tf.gather(tfsquares, tf.cast(v, tf.int64))), tensor, fn_output_signature=tf.string), "\n")
+
+    return tfstring.numpy().decode()
+
+
+def vector_to_string(tensor):
+    return index_to_string(vector_to_index(tensor))
diff --git a/patch_gen.ipynb b/patch_gen.ipynb
new file mode 100644
index 0000000..d6483a3
--- /dev/null
+++ b/patch_gen.ipynb
@@ -0,0 +1,398 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "The autoreload extension is already loaded. To reload it, use:\n",
+      "  %reload_ext autoreload\n"
+     ]
+    }
+   ],
+   "source": [
+    "%load_ext autoreload\n",
+    "%autoreload 2"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import tensorflow as tf\n",
+    "from tensorflow.keras.layers import Dense, Conv2D, Reshape\n",
+    "from polymap import Polymap\n",
+    "from aesthetic_loss import aesthetic_loss"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 46,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "['🟥', '🟦', '🟧', '🟨', '🟩', '🟪', '🟫']\n"
+     ]
+    }
+   ],
+   "source": [
+    "from math import pi\n",
+    "squares = [chr(i) for i in range(0x1F7E5, 0x1F7EC)]\n",
+    "tfsquares = tf.constant(squares)\n",
+    "print(squares)\n",
+    "colors = tf.constant([10, 219, 38, 48, 80,  282, 20] * tf.constant(pi/180.0), dtype=tf.float32)\n",
+    "color_vectors = tf.transpose(\n",
+    "    tf.stack([\n",
+    "    tf.math.cos(colors),\n",
+    "    tf.math.sin(colors)]\n",
+    "    )\n",
+    ")\n",
+    "\n",
+    "def vector_to_index(tensor):\n",
+    "    return tf.argmax((tf.einsum(\"ijk,lk->ijl\", tensor, color_vectors)), axis=-1)\n",
+    "\n",
+    "def index_to_string(tensor):\n",
+    "    # square_select = tf.math.argmax(tensor, 2)\n",
+    "\n",
+    "    tfstring = tf.strings.join(\n",
+    "        tf.map_fn(\n",
+    "            lambda v:\n",
+    "            tf.strings.join(\n",
+    "                tf.gather(tfsquares, tf.cast(v, tf.int64))), tensor, fn_output_signature=tf.string), \"\\n\")\n",
+    "\n",
+    "    return tfstring.numpy().decode()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "batch_size = 16"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 51,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# class PatchGen(tf.keras.Model):\n",
+    "\n",
+    "#     def __init__(self, batch_size):\n",
+    "#         super(PatchGen, self).__init__()\n",
+    "#         self.pm = Polymap(8, 3)(input)\n",
+    "#         self.d1 = Dense(64)\n",
+    "#         self.rs = Reshape([8,8,1])\n",
+    "#         self.dc1 = Conv2D(3,(1,1))\n",
+    "\n",
+    "#     def call(self, inputs):\n",
+    "#         x = self.block_1(inputs)\n",
+    "#         x = self.block_2(x)\n",
+    "#         x = self.global_pool(x)\n",
+    "#         return self.classifier(x)\n",
+    "\n",
+    "input = tf.random.normal(shape=[batch_size,2])\n",
+    "pm = Polymap(8, 3)\n",
+    "d1 = Dense(64, activation=\"relu\")\n",
+    "d2 = Dense(64, activation=\"relu\")\n",
+    "rs = Reshape([8,8,1])\n",
+    "dc1 = Conv2D(2,(1,1))\n",
+    "\n",
+    "def gen():\n",
+    "    input = tf.random.normal(shape=[batch_size,2])\n",
+    "    x = pm(input)\n",
+    "    x = d1(x)\n",
+    "    x = d2(x)\n",
+    "    x = rs(x)\n",
+    "    return dc1(x)\n",
+    "# gen()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 52,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "optimizer = tf.keras.optimizers.Adam()\n",
+    "train_loss = tf.keras.metrics.Mean(name='train_loss')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 53,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "@tf.function\n",
+    "def train_step():\n",
+    "    with tf.GradientTape() as tape:\n",
+    "        pics = gen()\n",
+    "        loss = aesthetic_loss(pics)\n",
+    "    tv = sum([pm.trainable_weights, d1.trainable_weights,\n",
+    "             dc1.trainable_weights], [])\n",
+    "    print(tv)\n",
+    "    gradients = tape.gradient(loss, tv)\n",
+    "    optimizer.apply_gradients(zip(gradients, tv))\n",
+    "\n",
+    "    train_loss(loss)\n",
+    "    # train_accuracy(labels, predictions)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 57,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Epoch 1, Loss: 0.0004605448921211064, \n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "Epoch 2, Loss: 9.45829197007697e-06, \n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "Epoch 3, Loss: 8.17671389086172e-05, \n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "Epoch 4, Loss: 0.0001505890249973163, \n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "Epoch 5, Loss: 0.00018983485642820597, \n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "Epoch 6, Loss: 0.00012902110756840557, \n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "Epoch 7, Loss: 1.342521863989532e-05, \n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "Epoch 8, Loss: 0.00034237385261803865, \n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "Epoch 9, Loss: 8.251221879618242e-05, \n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "Epoch 10, Loss: 1.0330303013006414e-07, \n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "Epoch 11, Loss: 3.0386236176127568e-05, \n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "Epoch 12, Loss: 4.6750748879276216e-05, \n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "Epoch 13, Loss: 7.047886185773677e-08, \n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "Epoch 14, Loss: 9.711433317672463e-14, \n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "🟥🟥🟥🟥🟥🟥🟥🟥\n",
+      "Epoch 15, Loss: 7.075177338602006e-23, \n",
+      "🟪🟪🟪🟪🟪🟪🟪🟪\n",
+      "🟪🟪🟪🟪🟪🟪🟪🟪\n",
+      "🟪🟪🟪🟪🟪🟪🟪🟪\n",
+      "🟪🟪🟪🟪🟪🟪🟪🟪\n",
+      "🟪🟪🟪🟪🟪🟪🟪🟪\n",
+      "🟪🟪🟪🟪🟪🟪🟪🟪\n",
+      "🟪🟪🟪🟪🟪🟪🟪🟪\n",
+      "🟪🟪🟪🟪🟪🟪🟪🟪\n",
+      "Epoch 16, Loss: 0.11943158507347107, \n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "Epoch 17, Loss: 0.002167836995795369, \n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "🟨🟨🟨🟨🟨🟨🟨🟨\n",
+      "Epoch 18, Loss: 6.590542034246027e-05, \n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "🟧🟧🟧🟧🟧🟧🟧🟧\n",
+      "Epoch 19, Loss: 0.0017910723108798265, \n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "🟩🟩🟩🟩🟩🟩🟩🟩\n",
+      "Epoch 20, Loss: 0.004703103564679623, \n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n",
+      "🟦🟦🟦🟦🟦🟦🟦🟦\n"
+     ]
+    }
+   ],
+   "source": [
+    "EPOCHS = 50\n",
+    "\n",
+    "for epoch in range(EPOCHS):\n",
+    "    # Reset the metrics at the start of the next epoch\n",
+    "    train_loss.reset_states()\n",
+    "#   train_accuracy.reset_states()\n",
+    "#   test_loss.reset_states()\n",
+    "#   test_accuracy.reset_states()\n",
+    "\n",
+    "#   for images, labels in train_ds:\n",
+    "    for i in range(50):\n",
+    "        train_step()\n",
+    "\n",
+    "#   for test_images, test_labels in test_ds:\n",
+    "#     test_step(test_images, test_labels)\n",
+    "\n",
+    "    print(\n",
+    "        f'Epoch {epoch + 1}, '\n",
+    "        f'Loss: {train_loss.result()}, '\n",
+    "        # f'Accuracy: {train_accuracy.result() * 100}, '\n",
+    "        # f'Test Loss: {test_loss.result()}, '\n",
+    "        # f'Test Accuracy: {test_accuracy.result() * 100}'\n",
+    "    )\n",
+    "\n",
+    "    ps = gen()\n",
+    "    # print(ps[0])\n",
+    "    print(index_to_string(vector_to_index(ps[0])))"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": ".venv_patch_gen",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.9"
+  },
+  "orig_nbformat": 4
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
diff --git a/patch_gen.py b/patch_gen.py
index 04da175..2dbcff0 100644
--- a/patch_gen.py
+++ b/patch_gen.py
@@ -1,4 +1,127 @@
-from tensorflow import keras
+# autopep8: off
+import tensorflow as tf
 
-input = keras.Input((2,))
+physical_devices = tf.config.experimental.list_physical_devices('GPU')
+print(physical_devices)
+tf.config.experimental.set_memory_growth(physical_devices[0], True)
 
+import colors
+from aesthetic_loss import aesthetic_loss
+from polymap import Polymap
+from tensorflow.keras.layers import Dense, Conv2D, Reshape, ReLU, Conv2DTranspose, LeakyReLU, BatchNormalization, GaussianNoise
+# autopep8: on
+
+batch_size = 256
+
+
+input_small = tf.random.normal(shape=[4, 2])
+pm = Polymap(8, 3)
+rl = ReLU()
+d1 = Dense(256, activation="relu",
+           kernel_regularizer=tf.keras.regularizers.l1_l2())
+d2 = Dense(512, activation="relu",
+           kernel_regularizer=tf.keras.regularizers.l1_l2())
+d3 = Dense(1024, activation="relu",
+           kernel_regularizer=tf.keras.regularizers.l1_l2())
+rs = Reshape([2, 2, 1024//4])
+dc1 = Conv2DTranspose(256, (5, 5),
+                      padding="valid", use_bias=False)
+dc2 = Conv2DTranspose(128, (3, 3),
+                      padding="valid", use_bias=False)
+dc3 = Conv2DTranspose(64, (3, 3),
+                      padding="same", use_bias=False)
+dc4 = Conv2D(2, (1, 1), activation="tanh", use_bias=False)
+bn1 = BatchNormalization()
+bn2 = BatchNormalization()
+bn3 = BatchNormalization()
+
+def gen_model():
+    model = tf.keras.Sequential()
+
+    model.add(pm)
+    # model.add(rl)
+
+    model.add(d1)
+    model.add(d2)
+    model.add(d3)
+
+    model.add(rs)
+
+    model.add(dc1)
+    model.add(bn1)
+    model.add(LeakyReLU())
+
+    model.add(dc2)
+    model.add(bn2)
+    model.add(LeakyReLU())
+
+    model.add(dc3)
+    model.add(bn3)
+    model.add(LeakyReLU())
+
+    model.add(dc4)
+    return model
+
+
+optimizer = tf.keras.optimizers.Adam(learning_rate=0.001, epsilon=0.001)
+train_loss = tf.keras.metrics.Mean(name='train_loss')
+
+generator = gen_model()
+
+rng = tf.random.Generator.from_seed(1)
+
+@tf.function
+def train_step(input):
+    # input = tf.random.normal(shape=[batch_size, 2])
+    # input = rng.normal(shape=[batch_size, 2])
+
+    # print(input)
+
+    with tf.GradientTape() as tape:
+        # pics = gen(input)
+        pics = generator(input, training=True)
+        loss = aesthetic_loss(
+            pics) + tf.reduce_sum([d1.losses, d2.losses, d3.losses])
+        + 0.01 * tf.reduce_mean(tf.abs(pm.trainable_weights))
+    # tv = sum([pm.trainable_weights,
+    #           d1.trainable_weights,
+    #           d2.trainable_weights,
+    #           d3.trainable_weights,
+    #          dc1.trainable_weights,
+    #          dc2.trainable_weights,
+    #          dc3.trainable_weights,], [])
+
+    # print(tv)
+    gradients = tape.gradient(loss, generator.trainable_weights)
+    optimizer.apply_gradients(zip(gradients, generator.trainable_weights))
+
+    train_loss(loss + tf.reduce_sum([d1.losses, d2.losses, d3.losses]))
+
+
+EPOCHS = 200
+
+for epoch in range(EPOCHS):
+    train_loss.reset_states()
+
+    input = rng.normal(shape=[batch_size, 2])
+    # print(input)
+
+    for i in range(100):
+        train_step(input)
+
+    print(
+        f'Epoch {epoch + 1}, '
+        f'Loss: {train_loss.result()}, '
+    )
+
+    input_small = rng.normal(shape=[4, 2])
+    ps = generator(input_small, training=False)
+    # print(ps[0])
+    # print(pm.kernels)
+    s0 = colors.vector_to_string(ps[0])
+    s1 = colors.vector_to_string(ps[1])
+    s2 = colors.vector_to_string(ps[2])
+    s3 = colors.vector_to_string(ps[3])
+
+    print("\n".join(" ".join(r)
+          for r in zip(s0.split("\n"), s1.split("\n"), s2.split("\n"), s3.split("\n"))))
diff --git a/playground.ipynb b/playground.ipynb
new file mode 100644
index 0000000..3ebb6b6
--- /dev/null
+++ b/playground.ipynb
@@ -0,0 +1,794 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%load_ext autoreload\n",
+    "%autoreload 2"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import tensorflow as tf\n",
+    "import random\n",
+    "from math import pi"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "['🟥', '🟦', '🟧', '🟨', '🟩', '🟪', '🟫']\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "<tf.Tensor: shape=(7, 2), dtype=float32, numpy=\n",
+       "array([[ 0.9848077 ,  0.17364818],\n",
+       "       [-0.777146  , -0.62932044],\n",
+       "       [ 0.7880108 ,  0.6156615 ],\n",
+       "       [ 0.6691306 ,  0.74314487],\n",
+       "       [ 0.17364822,  0.9848077 ],\n",
+       "       [ 0.20791148, -0.9781476 ],\n",
+       "       [ 0.9396926 ,  0.34202012]], dtype=float32)>"
+      ]
+     },
+     "execution_count": 5,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "squares = [chr(i) for i in range(0x1F7E5, 0x1F7EC)]\n",
+    "tfsquares = tf.constant(squares)\n",
+    "print(squares)\n",
+    "colors = tf.constant([10, 219, 38, 48, 80,  282, 20] * tf.constant(pi/180.0), dtype=tf.float32)\n",
+    "color_vectors = tf.transpose(\n",
+    "    tf.stack([\n",
+    "    tf.math.cos(colors),\n",
+    "    tf.math.sin(colors)]\n",
+    "    )\n",
+    ")\n",
+    "color_vectors"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def one_hot_to_string(tensor):\n",
+    "    square_select = tf.math.argmax(tensor, 2)\n",
+    "\n",
+    "    tfstring = tf.strings.join(\n",
+    "        tf.map_fn(\n",
+    "            lambda v:\n",
+    "            tf.strings.join(\n",
+    "                tf.gather(tfsquares, tf.cast(v, tf.int64))), square_select, fn_output_signature=tf.string), \"\\n\")\n",
+    "\n",
+    "    return tfstring.numpy().decode()\n",
+    "\n",
+    "def index_to_string(tensor):\n",
+    "    return tf.argmax((tf.einsum(\"ijk,lk->ijl\", tensor, color_vectors)), axis=2)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 57,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<tf.Tensor: shape=(8, 8), dtype=int64, numpy=\n",
+       "array([[5, 0, 1, 3, 3, 2, 4, 0],\n",
+       "       [3, 2, 3, 6, 5, 3, 0, 5],\n",
+       "       [1, 3, 0, 1, 6, 4, 5, 3],\n",
+       "       [3, 5, 0, 6, 5, 5, 0, 3],\n",
+       "       [0, 3, 0, 6, 6, 1, 5, 2],\n",
+       "       [4, 1, 4, 6, 6, 1, 0, 5],\n",
+       "       [6, 0, 2, 5, 0, 1, 5, 4],\n",
+       "       [0, 2, 5, 4, 1, 5, 0, 4]])>"
+      ]
+     },
+     "execution_count": 57,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "vi = tf.map_fn(lambda v: tf.map_fn(lambda s: color_vectors[s],v, fn_output_signature=tf.float32),tf.constant(e, shape=(8, 8)),fn_output_signature=tf.float32)\n",
+    "\n",
+    "index_to_string(vi)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<tf.Tensor: shape=(1, 8, 8), dtype=float32, numpy=\n",
+       "array([[[0.71719116, 0.04602498, 1.0273366 , 2.9402392 , 2.4437628 ,\n",
+       "         1.0988994 , 1.7921778 , 0.6444071 ],\n",
+       "        [1.2672982 , 1.8689629 , 3.8761568 , 2.4126112 , 2.840665  ,\n",
+       "         3.0419385 , 4.8909955 , 3.583321  ],\n",
+       "        [0.4506898 , 0.01335579, 3.258954  , 3.80125   , 6.4259176 ,\n",
+       "         6.3595123 , 6.609476  , 1.6973627 ],\n",
+       "        [1.8337119 , 1.4464988 , 4.6484404 , 4.2732587 , 7.6358337 ,\n",
+       "         5.4232316 , 4.313001  , 1.6071305 ],\n",
+       "        [4.9838686 , 1.3140092 , 1.5253077 , 1.9003649 , 6.782009  ,\n",
+       "         3.5835938 , 4.203249  , 1.5495552 ],\n",
+       "        [5.3104043 , 5.0123706 , 3.3745103 , 4.540594  , 6.6606255 ,\n",
+       "         3.9076035 , 4.9301    , 3.4342353 ],\n",
+       "        [3.3523657 , 4.9394774 , 5.672796  , 4.8756404 , 7.1605525 ,\n",
+       "         8.192585  , 7.06907   , 4.199152  ],\n",
+       "        [0.13434029, 4.039145  , 6.2488484 , 3.2168784 , 5.319997  ,\n",
+       "         5.4898353 , 5.301211  , 1.6475667 ]]], dtype=float32)>"
+      ]
+     },
+     "execution_count": 10,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# def colors_to_one_hot(tensor):\n",
+    "vi = tf.map_fn(lambda v: tf.map_fn(lambda s: color_vectors[s],v, fn_output_signature=tf.float32),tf.constant(e, shape=(8, 8)),fn_output_signature=tf.float32)\n",
+    "ns = tf.image.extract_patches(tf.constant(vi, shape=(1,8,8,2)), (1,3,3,1), (1,1,1,1), (1,1,1,1), padding=\"SAME\")\n",
+    "virs = tf.reshape(vi, (8*8,2))\n",
+    "# tf.reduce_sum(tf.multiply(tf.repeat(vi, 9, -1), ns))\n",
+    "tf.abs(tf.einsum(\"...k,...k->...\", tf.repeat(vi, 9, -1), ns))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 27,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "tf.Tensor([0. 0. 0. 1. 0. 0. 0.], shape=(7,), dtype=float32)\n",
+      "tf.Tensor([[3]], shape=(1, 1), dtype=int64)\n",
+      "🟨🟥🟨🟨🟩🟫🟪🟫\n",
+      "🟩🟪🟧🟧🟦🟧🟦🟫\n",
+      "🟩🟦🟫🟨🟦🟨🟩🟫\n",
+      "🟪🟫🟩🟨🟧🟪🟥🟩\n",
+      "🟥🟫🟨🟫🟧🟦🟦🟧\n",
+      "🟩🟫🟫🟧🟨🟨🟪🟫\n",
+      "🟩🟪🟦🟥🟧🟫🟧🟨\n",
+      "🟨🟪🟨🟧🟪🟦🟨🟧\n",
+      "--------------\n",
+      "🟧🟨🟦🟪🟧🟨🟪🟨\n",
+      "🟨🟧🟫🟧🟥🟦🟪🟩\n",
+      "🟫🟪🟨🟨🟧🟫🟫🟩\n",
+      "🟧🟦🟦🟧🟫🟨🟫🟥\n",
+      "🟩🟥🟪🟧🟨🟩🟫🟪\n",
+      "🟫🟩🟨🟦🟨🟫🟦🟩\n",
+      "🟫🟦🟧🟦🟧🟧🟪🟩\n",
+      "🟫🟪🟫🟩🟨🟨🟥🟨\n",
+      "tf.Tensor(\n",
+      "[[0. 0. 0. 0. 0. 0. 0. 0.]\n",
+      " [0. 0. 1. 0. 0. 1. 0. 0.]\n",
+      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
+      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
+      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
+      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
+      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
+      " [0. 0. 0. 0. 0. 0. 0. 0.]], shape=(8, 8), dtype=float32)\n"
+     ]
+    }
+   ],
+   "source": [
+    "oh = tf.one_hot(3, 7)\n",
+    "print(oh)\n",
+    "print(tf.where(oh))\n",
+    "\n",
+    "# a = tf.constant([[list(tf.one_hot(random.randrange(7), 7)) for i in range(8) ] for i in range(8)], shape=(8,8,len(squares)))\n",
+    "\n",
+    "# print([[list(tf.one_hot(random.randrange(7), 7)) for i in range(8) ] for i in range(8)])\n",
+    "e = [[random.randrange(7) for i in range(8)] for i in range(8)]\n",
+    "# print(e)\n",
+    "\n",
+    "a = tf.one_hot(tf.constant(e, shape=(8, 8)), len(squares))\n",
+    "\n",
+    "print(one_hot_to_string(a))\n",
+    "# print(\"--------------\")\n",
+    "# print(one_hot_to_string(tf.reverse(a, axis=(-2,))))\n",
+    "# print(\"--------------\")\n",
+    "# print(one_hot_to_string(tf.reverse(a, axis=(-3,))))\n",
+    "# print(\"--------------\")\n",
+    "# print(one_hot_to_string(tf.transpose(a, perm=(1,0,2))))\n",
+    "print(\"--------------\")\n",
+    "# print(one_hot_to_string(tf.reverse(tf.transpose(tf.reverse(a, axis=(-2,)), perm=(1,0,2)), axis=(-2,))))\n",
+    "print(one_hot_to_string(tf.reverse(a, axis=(-2,-3,))))\n",
+    "# print(tensor_to_string(tf.transpose(a, [1, 0, 2])))\n",
+    "# print(\"--------------\")\n",
+    "# print(tensor_to_string(tf.linalg.matmul( a , tf.constant(\n",
+    "#     [\n",
+    "#         [0, 0, 0, 0, 0, 0, 0, 1],\n",
+    "#         [0, 0, 0, 0, 0, 0, 1, 0],\n",
+    "#         [0, 0, 0, 0, 0, 1, 0, 0],\n",
+    "#         [0, 0, 0, 0, 1, 0, 0, 0],\n",
+    "#         [0, 0, 0, 1, 0, 0, 0, 0],\n",
+    "#         [0, 0, 1, 0, 0, 0, 0, 0],\n",
+    "#         [0, 1, 0, 0, 0, 0, 0, 0],\n",
+    "#         [1, 0, 0, 0, 0, 0, 0, 0]\n",
+    "#     ], dtype=tf.float32, shape=(8,8,1)\n",
+    "# ), transpose_a=True)))\n",
+    "# print(\"--------------\")\n",
+    "# print(tensor_to_string(tf.reverse(a, (0,))))\n",
+    "# print(\"--------------\")\n",
+    "# print(tensor_to_string(tf.reverse(a, (1,))))\n",
+    "# print(tensor_to_string(\n",
+    "#     tf.linalg.matmul(\n",
+    "#         tf.transpose(a, [0, 2, 1]),\n",
+    "#         tf.reverse(a, (1,)))\n",
+    "# ))\n",
+    "print(\n",
+    "    tf.einsum(\n",
+    "        \"ijk,ijk->ij\",\n",
+    "        a,\n",
+    "        tf.reverse(a, (1,)))\n",
+    ")\n",
+    "\n",
+    "\n",
+    "# print(tf.constant(range(6), shape=(2,3)))\n",
+    "# tf.map_fn(tf.math.reduce_sum, tf.constant(range(6), shape=(2,3)))\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 15,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<tf.Tensor: shape=(), dtype=float32, numpy=-0.08645747>"
+      ]
+     },
+     "execution_count": 15,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import aesthetic_loss\n",
+    "\n",
+    "es = [[[random.randrange(7) for i in range(8)]\n",
+    "      for i in range(8)] for j in range(16)]\n",
+    "\n",
+    "vis = tf.stack(\n",
+    "    [tf.map_fn(lambda v: tf.map_fn(lambda s: color_vectors[s], v, fn_output_signature=tf.float32), tf.constant(\n",
+    "        e, shape=(8, 8)), fn_output_signature=tf.float32) for e in es]\n",
+    ")\n",
+    "\n",
+    "aesthetic_loss.aesthetic_loss(vis)\n",
+    "# aesthetic_loss.compute_score(vis, vis)\n",
+    "# tf.linalg.l2_normalize(vis, axis=-1)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 91,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<tf.Tensor: shape=(), dtype=int32, numpy=2890>"
+      ]
+     },
+     "execution_count": 91,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "x = tf.Variable(2)\n",
+    "y = tf.Variable(3)\n",
+    "\n",
+    "xp = tf.pow(x, [0,1,2,3])\n",
+    "yp = tf.pow(y, [0,1,2,3])\n",
+    "\n",
+    "A = tf.Variable([[i+j for i in range(4)] for j in range(4)], shape=(4,4))\n",
+    "tf.einsum(\"i,ij,j\", xp, A, yp)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 70,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<tf.Tensor: shape=(16, 8), dtype=float32, numpy=\n",
+       "array([[-6.07829041e+02, -1.43090210e+03, -2.25397559e+03,\n",
+       "        -3.07704883e+03, -3.90012207e+03, -4.72319531e+03,\n",
+       "        -5.54626855e+03, -6.36934229e+03],\n",
+       "       [-4.28289948e+01, -8.77924423e+01, -1.32755875e+02,\n",
+       "        -1.77719330e+02, -2.22682800e+02, -2.67646210e+02,\n",
+       "        -3.12609680e+02, -3.57573120e+02],\n",
+       "       [ 9.78072433e+01,  2.85662659e+02,  4.73518066e+02,\n",
+       "         6.61373413e+02,  8.49228882e+02,  1.03708435e+03,\n",
+       "         1.22493982e+03,  1.41279517e+03],\n",
+       "       [ 3.64326048e+00,  2.58858261e+01,  4.81283913e+01,\n",
+       "         7.03709641e+01,  9.26135254e+01,  1.14856094e+02,\n",
+       "         1.37098648e+02,  1.59341217e+02],\n",
+       "       [-1.08920467e+00,  3.87969995e+00,  8.84860516e+00,\n",
+       "         1.38175116e+01,  1.87864113e+01,  2.37553177e+01,\n",
+       "         2.87242279e+01,  3.36931267e+01],\n",
+       "       [ 2.75129032e+01,  9.32729721e+01,  1.59033051e+02,\n",
+       "         2.24793137e+02,  2.90553192e+02,  3.56313324e+02,\n",
+       "         4.22073334e+02,  4.87833435e+02],\n",
+       "       [-5.32951474e-01,  8.84730148e+00,  1.82275543e+01,\n",
+       "         2.76078072e+01,  3.69880600e+01,  4.63683128e+01,\n",
+       "         5.57485619e+01,  6.51288223e+01],\n",
+       "       [-7.10889587e+01, -1.97856873e+02, -3.24624786e+02,\n",
+       "        -4.51392792e+02, -5.78160645e+02, -7.04928589e+02,\n",
+       "        -8.31696533e+02, -9.58464539e+02],\n",
+       "       [-3.13533902e-01,  1.77808132e+01,  3.58751602e+01,\n",
+       "         5.39695053e+01,  7.20638504e+01,  9.01582031e+01,\n",
+       "         1.08252541e+02,  1.26346893e+02],\n",
+       "       [ 7.97924957e+01,  2.76383240e+02,  4.72974030e+02,\n",
+       "         6.69564819e+02,  8.66155518e+02,  1.06274634e+03,\n",
+       "         1.25933704e+03,  1.45592786e+03],\n",
+       "       [-2.12144566e+01, -1.41624451e+02, -2.62034424e+02,\n",
+       "        -3.82444489e+02, -5.02854431e+02, -6.23264404e+02,\n",
+       "        -7.43674438e+02, -8.64084229e+02],\n",
+       "       [-1.87502289e+02, -3.70114594e+02, -5.52726929e+02,\n",
+       "        -7.35339233e+02, -9.17951660e+02, -1.10056396e+03,\n",
+       "        -1.28317627e+03, -1.46578857e+03],\n",
+       "       [-1.07772827e+00, -9.14207935e-01, -7.50688553e-01,\n",
+       "        -5.87172508e-01, -4.23656583e-01, -2.60130525e-01,\n",
+       "        -9.66115594e-02,  6.69060946e-02],\n",
+       "       [ 2.14334583e+01,  1.10324554e+02,  1.99215652e+02,\n",
+       "         2.88106781e+02,  3.76997894e+02,  4.65889008e+02,\n",
+       "         5.54780090e+02,  6.43671143e+02],\n",
+       "       [ 5.15513878e+01,  1.88147568e+02,  3.24743774e+02,\n",
+       "         4.61339966e+02,  5.97936157e+02,  7.34532349e+02,\n",
+       "         8.71128601e+02,  1.00772473e+03],\n",
+       "       [ 2.68221771e+02,  7.04315491e+02,  1.14040918e+03,\n",
+       "         1.57650305e+03,  2.01259656e+03,  2.44869043e+03,\n",
+       "         2.88478394e+03,  3.32087769e+03]], dtype=float32)>"
+      ]
+     },
+     "execution_count": 70,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "r = tf.random.normal((16,2))\n",
+    "rp = tf.map_fn(lambda v: tf.pow(tf.repeat(tf.reshape(v, shape=(2,1)), repeats=4, axis=1), tf.constant([[0,1,2,3], [0,1,2,3]], dtype=tf.float32)), r)\n",
+    "A = tf.Variable(tf.reshape([float(i) for i in range(8*4*4)], shape=(8,4,4)), shape=(8,4,4), dtype=tf.float32)\n",
+    "tf.einsum(\"ik,ij,lkj->il\", rp[:,0], rp[:,1], A)\n",
+    "# rp[:,0]\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 82,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "tf.Tensor(\n",
+      "[[ 4.15302128e-01 -1.17927468e+00  2.49658436e-01  6.19551420e-01\n",
+      "  -1.39033377e-01  2.79209971e-01  3.97601783e-01 -2.17277408e-01]\n",
+      " [ 1.74276757e+00  5.69359064e-02  5.38259745e-04  1.49045396e+00\n",
+      "  -7.87080288e-01 -3.98119450e-01  1.18716180e-01  1.86502957e+00]\n",
+      " [ 1.78081155e+00 -1.29199183e+00 -5.72377980e-01  1.73489356e+00\n",
+      "  -1.81523621e+00 -7.56731749e-01  5.50791979e-01  2.23066378e+00]\n",
+      " [-2.86107287e-02  2.38285348e-01 -2.90394664e-01  2.20198780e-02\n",
+      "   3.43291536e-02  2.77290910e-01 -3.25144343e-02  3.50418538e-01]\n",
+      " [-2.32135087e-01  1.95894271e-01  5.75335659e-02 -1.76543556e-03\n",
+      "  -4.01449911e-02  2.00845733e-01 -9.64274257e-02  2.31855541e-01]\n",
+      " [ 4.60646629e-01  7.75275290e-01 -9.96064782e-01 -5.15016496e-01\n",
+      "   2.03099363e-02  7.87942708e-02 -3.07008505e-01  7.16598153e-01]\n",
+      " [ 3.02803397e-01 -8.83126408e-02  3.72119397e-01  1.14753373e-01\n",
+      "  -8.30153376e-03  6.69163764e-02 -3.60930443e-01  4.32033747e-01]\n",
+      " [-1.55036688e-01  2.96032988e-02  5.69820963e-02 -1.04342289e-01\n",
+      "   2.18345195e-01  1.60513148e-01 -2.30654161e-02  2.92460948e-01]\n",
+      " [-9.37448144e-02  4.92064208e-02  1.25066668e-01 -1.08998209e-01\n",
+      "   1.47912696e-01  7.13704750e-02 -9.69988406e-02  3.16246897e-01]\n",
+      " [ 3.76237392e-01 -8.38992000e-02  3.31230164e-01  2.86825836e-01\n",
+      "   3.67110878e-01 -3.61552715e-01  6.06793404e-01 -1.22976625e+00]\n",
+      " [-1.00743435e-01 -1.46258920e-01 -1.82411104e-01 -3.14753801e-01\n",
+      "   7.92229354e-01  2.98170269e-01 -1.82591528e-01  2.78764069e-01]\n",
+      " [ 3.58289123e-01  1.97965109e+00 -1.18626821e+00 -1.47391105e+00\n",
+      "   6.11245155e-01 -4.61193591e-01 -7.64733970e-01  1.45492101e+00]\n",
+      " [-2.00078160e-01  4.32360053e-01  1.63023725e-01 -3.31694707e-02\n",
+      "  -4.53773856e-01  4.14216101e-01  1.76714227e-01  2.12158859e-01]\n",
+      " [ 4.91659880e-01  1.56133473e+00 -8.19299459e-01  9.75016415e-01\n",
+      "  -2.52336192e+00  1.69403172e+00  2.05944443e+00  1.03490460e+00]\n",
+      " [-2.33240795e+00  6.64168453e+00 -4.04779243e+00 -5.99489832e+00\n",
+      "   1.60179734e+00  2.19676828e+00  4.11457443e+00  2.81916380e-01]\n",
+      " [ 1.73429586e-02  4.88152355e-02  3.41429591e-01 -2.39478126e-02\n",
+      "  -2.88104061e-02  1.31737068e-01 -2.55863100e-01  2.94781595e-01]], shape=(16, 8), dtype=float32)\n"
+     ]
+    }
+   ],
+   "source": [
+    "import polymap\n",
+    "r = tf.random.normal((16,2))\n",
+    "\n",
+    "pm = polymap.Polymap(8,4)\n",
+    "# pm.build([16,2])\n",
+    "print(pm(r))\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 111,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<tf.Tensor: shape=(16, 8, 8, 2), dtype=float32, numpy=\n",
+       "array([[[[ 0.26760346,  0.9635291 ],\n",
+       "         [ 0.92846906,  0.37140968],\n",
+       "         [ 0.43006882,  0.90279603],\n",
+       "         ...,\n",
+       "         [ 0.9174245 ,  0.39790976],\n",
+       "         [ 0.68369484,  0.729768  ],\n",
+       "         [ 0.5905379 , -0.8070098 ]],\n",
+       "\n",
+       "        [[-0.9665002 ,  0.25666556],\n",
+       "         [ 0.6174854 , -0.7865824 ],\n",
+       "         [-0.8356619 , -0.54924417],\n",
+       "         ...,\n",
+       "         [-0.42570287,  0.90486294],\n",
+       "         [ 0.99368197,  0.11223278],\n",
+       "         [ 0.22659355,  0.97398937]],\n",
+       "\n",
+       "        [[ 0.9998629 ,  0.01655844],\n",
+       "         [-0.02171519,  0.9997642 ],\n",
+       "         [ 0.765471  , -0.64347047],\n",
+       "         ...,\n",
+       "         [-0.45096368,  0.8925422 ],\n",
+       "         [ 0.23687297,  0.9715406 ],\n",
+       "         [-0.2564354 ,  0.9665613 ]],\n",
+       "\n",
+       "        ...,\n",
+       "\n",
+       "        [[ 0.99743736, -0.07154454],\n",
+       "         [-0.421027  ,  0.907048  ],\n",
+       "         [-0.06904475,  0.99761355],\n",
+       "         ...,\n",
+       "         [-0.14730097, -0.9890917 ],\n",
+       "         [-0.76936895, -0.63880455],\n",
+       "         [ 0.12321126,  0.99238044]],\n",
+       "\n",
+       "        [[ 0.8592728 ,  0.51151776],\n",
+       "         [ 0.5945256 , -0.80407673],\n",
+       "         [-0.926837  , -0.37546387],\n",
+       "         ...,\n",
+       "         [-0.6461148 ,  0.76324016],\n",
+       "         [-0.30289924,  0.95302254],\n",
+       "         [-0.17756721,  0.9841086 ]],\n",
+       "\n",
+       "        [[-0.9989693 ,  0.04539123],\n",
+       "         [ 0.9668924 ,  0.25518435],\n",
+       "         [-0.70309734,  0.7110935 ],\n",
+       "         ...,\n",
+       "         [ 0.9820237 , -0.1887575 ],\n",
+       "         [-0.45441642, -0.8907894 ],\n",
+       "         [ 0.20170917,  0.9794454 ]]],\n",
+       "\n",
+       "\n",
+       "       [[[-0.9041683 , -0.4271764 ],\n",
+       "         [-0.8145172 ,  0.5801394 ],\n",
+       "         [-0.93892664,  0.34411758],\n",
+       "         ...,\n",
+       "         [ 0.7724613 , -0.63506186],\n",
+       "         [ 0.73973364,  0.6728997 ],\n",
+       "         [-0.7584759 , -0.6517011 ]],\n",
+       "\n",
+       "        [[ 0.40460777,  0.91449016],\n",
+       "         [ 0.9511079 ,  0.30885857],\n",
+       "         [-0.8532944 ,  0.5214295 ],\n",
+       "         ...,\n",
+       "         [ 0.3517043 , -0.9361111 ],\n",
+       "         [-0.81593966, -0.57813704],\n",
+       "         [-0.96846116, -0.24916425]],\n",
+       "\n",
+       "        [[ 0.9971029 , -0.07606225],\n",
+       "         [ 0.8346386 , -0.550798  ],\n",
+       "         [ 0.79441607,  0.60737395],\n",
+       "         ...,\n",
+       "         [-0.17288695, -0.9849416 ],\n",
+       "         [ 0.34486932, -0.93865067],\n",
+       "         [ 0.15837449,  0.9873792 ]],\n",
+       "\n",
+       "        ...,\n",
+       "\n",
+       "        [[-0.9557019 ,  0.2943367 ],\n",
+       "         [-0.1128222 , -0.9936151 ],\n",
+       "         [ 0.8779246 , -0.4787989 ],\n",
+       "         ...,\n",
+       "         [ 0.01654347,  0.9998631 ],\n",
+       "         [-0.31221828,  0.9500104 ],\n",
+       "         [-0.70988506,  0.70431745]],\n",
+       "\n",
+       "        [[ 0.59169656, -0.8061608 ],\n",
+       "         [ 0.29944003,  0.95411515],\n",
+       "         [ 0.12906635,  0.9916359 ],\n",
+       "         ...,\n",
+       "         [-0.7345422 , -0.678563  ],\n",
+       "         [-0.6101673 , -0.79227257],\n",
+       "         [ 0.63129115,  0.7755459 ]],\n",
+       "\n",
+       "        [[ 0.15320726, -0.9881941 ],\n",
+       "         [ 0.8636006 , -0.5041763 ],\n",
+       "         [ 0.5430426 ,  0.83970505],\n",
+       "         ...,\n",
+       "         [ 0.99351764, -0.11367732],\n",
+       "         [ 0.8001608 , -0.5997855 ],\n",
+       "         [-0.8337144 ,  0.55219585]]],\n",
+       "\n",
+       "\n",
+       "       [[[-0.6251835 ,  0.78047776],\n",
+       "         [ 0.6738433 ,  0.7388742 ],\n",
+       "         [-0.9261264 , -0.37721333],\n",
+       "         ...,\n",
+       "         [ 0.77851856, -0.6276216 ],\n",
+       "         [ 0.6481785 ,  0.76148844],\n",
+       "         [ 0.8758924 , -0.4825066 ]],\n",
+       "\n",
+       "        [[-0.9522389 , -0.3053537 ],\n",
+       "         [-0.94333917, -0.33183014],\n",
+       "         [-0.8840897 ,  0.46731716],\n",
+       "         ...,\n",
+       "         [ 0.8191087 ,  0.5736384 ],\n",
+       "         [ 0.6949664 ,  0.7190423 ],\n",
+       "         [-0.2683343 ,  0.96332586]],\n",
+       "\n",
+       "        [[-0.83271   ,  0.5537092 ],\n",
+       "         [-0.72585243,  0.6878505 ],\n",
+       "         [ 0.77075607,  0.63713026],\n",
+       "         ...,\n",
+       "         [-0.9409917 ,  0.33842987],\n",
+       "         [-0.9817458 ,  0.19019753],\n",
+       "         [ 0.9783392 ,  0.20700799]],\n",
+       "\n",
+       "        ...,\n",
+       "\n",
+       "        [[-0.04849374,  0.99882346],\n",
+       "         [ 0.9695309 ,  0.24496901],\n",
+       "         [-0.37880003, -0.9254785 ],\n",
+       "         ...,\n",
+       "         [ 0.6892487 , -0.7245249 ],\n",
+       "         [ 0.81647533,  0.57738036],\n",
+       "         [-0.9972612 , -0.07396041]],\n",
+       "\n",
+       "        [[-0.7055271 ,  0.7086829 ],\n",
+       "         [-0.8819862 ,  0.4712752 ],\n",
+       "         [-0.9642273 , -0.2650766 ],\n",
+       "         ...,\n",
+       "         [ 0.82875234, -0.55961555],\n",
+       "         [-0.06711138,  0.9977455 ],\n",
+       "         [-0.1260777 ,  0.99202037]],\n",
+       "\n",
+       "        [[ 0.99550056, -0.09475542],\n",
+       "         [-0.9923597 ,  0.1233783 ],\n",
+       "         [-0.71074235, -0.70345235],\n",
+       "         ...,\n",
+       "         [ 0.8930231 , -0.4500106 ],\n",
+       "         [ 0.1744522 ,  0.9846656 ],\n",
+       "         [ 0.8210883 , -0.570801  ]]],\n",
+       "\n",
+       "\n",
+       "       ...,\n",
+       "\n",
+       "\n",
+       "       [[[-0.7103158 ,  0.703883  ],\n",
+       "         [ 0.37816712, -0.9257372 ],\n",
+       "         [ 0.9501393 ,  0.31182557],\n",
+       "         ...,\n",
+       "         [ 0.66041577,  0.75090003],\n",
+       "         [ 0.00253691,  0.9999967 ],\n",
+       "         [-0.3108576 , -0.95045644]],\n",
+       "\n",
+       "        [[-0.9948109 ,  0.1017414 ],\n",
+       "         [ 0.8379823 ,  0.54569715],\n",
+       "         [-0.42345405,  0.9059175 ],\n",
+       "         ...,\n",
+       "         [ 0.7350055 ,  0.678061  ],\n",
+       "         [-0.6902005 , -0.7236182 ],\n",
+       "         [ 0.3050526 , -0.9523355 ]],\n",
+       "\n",
+       "        [[-0.9992436 , -0.03888722],\n",
+       "         [-0.9524134 , -0.3048092 ],\n",
+       "         [ 0.17434528,  0.9846846 ],\n",
+       "         ...,\n",
+       "         [-0.48622662, -0.8738327 ],\n",
+       "         [ 0.5717803 ,  0.82040673],\n",
+       "         [-0.7644213 ,  0.64471704]],\n",
+       "\n",
+       "        ...,\n",
+       "\n",
+       "        [[-0.0501435 , -0.998742  ],\n",
+       "         [-0.915064  ,  0.40330857],\n",
+       "         [-0.7114537 , -0.7027328 ],\n",
+       "         ...,\n",
+       "         [ 0.9969975 , -0.07743412],\n",
+       "         [ 0.58687896, -0.80967456],\n",
+       "         [ 0.9493494 ,  0.31422222]],\n",
+       "\n",
+       "        [[ 0.49507466,  0.86885047],\n",
+       "         [ 0.8797779 , -0.47538495],\n",
+       "         [-0.9211454 ,  0.38921857],\n",
+       "         ...,\n",
+       "         [ 0.8058742 , -0.5920867 ],\n",
+       "         [ 0.53095114, -0.84740245],\n",
+       "         [ 0.9835844 , -0.18044874]],\n",
+       "\n",
+       "        [[-0.19601943, -0.98059994],\n",
+       "         [ 0.5338162 , -0.8456005 ],\n",
+       "         [-0.9987141 , -0.05069751],\n",
+       "         ...,\n",
+       "         [ 0.73851573, -0.6742363 ],\n",
+       "         [-0.99623436,  0.08670113],\n",
+       "         [ 0.62578243,  0.7799976 ]]],\n",
+       "\n",
+       "\n",
+       "       [[[ 0.99914247, -0.04140271],\n",
+       "         [-0.45437112,  0.8908124 ],\n",
+       "         [-0.42213327, -0.9065337 ],\n",
+       "         ...,\n",
+       "         [ 0.62287563,  0.78232086],\n",
+       "         [-0.34597075, -0.9382453 ],\n",
+       "         [ 0.79978997, -0.6002798 ]],\n",
+       "\n",
+       "        [[ 0.8313185 , -0.5557963 ],\n",
+       "         [ 0.7764067 , -0.6302323 ],\n",
+       "         [-0.4633804 ,  0.8861594 ],\n",
+       "         ...,\n",
+       "         [-0.06348424, -0.99798286],\n",
+       "         [-0.24774338,  0.9688258 ],\n",
+       "         [ 0.7214034 ,  0.6925151 ]],\n",
+       "\n",
+       "        [[-0.7895615 , -0.6136714 ],\n",
+       "         [ 0.7669125 , -0.64175165],\n",
+       "         [-0.60046875,  0.7996482 ],\n",
+       "         ...,\n",
+       "         [-0.22642732,  0.974028  ],\n",
+       "         [-0.874757  , -0.48456183],\n",
+       "         [-0.75133216, -0.6599242 ]],\n",
+       "\n",
+       "        ...,\n",
+       "\n",
+       "        [[ 0.99788564,  0.06499327],\n",
+       "         [-0.9613605 ,  0.27529252],\n",
+       "         [ 0.8874723 , -0.46086106],\n",
+       "         ...,\n",
+       "         [ 0.24368125,  0.96985537],\n",
+       "         [ 0.9091159 , -0.4165433 ],\n",
+       "         [ 0.39610097,  0.9182069 ]],\n",
+       "\n",
+       "        [[ 0.15766756, -0.98749226],\n",
+       "         [-0.40235043,  0.9154857 ],\n",
+       "         [-0.7679899 , -0.64046216],\n",
+       "         ...,\n",
+       "         [ 0.85627615,  0.5165182 ],\n",
+       "         [ 0.43845195,  0.89875454],\n",
+       "         [-0.995217  ,  0.09768905]],\n",
+       "\n",
+       "        [[ 0.31384754, -0.94947344],\n",
+       "         [-0.335953  ,  0.94187874],\n",
+       "         [ 0.32295498,  0.94641435],\n",
+       "         ...,\n",
+       "         [ 0.56830823,  0.82281566],\n",
+       "         [ 0.5089668 ,  0.860786  ],\n",
+       "         [ 0.43783924,  0.89905316]]],\n",
+       "\n",
+       "\n",
+       "       [[[ 0.00368896, -0.99999326],\n",
+       "         [-0.15407981, -0.98805845],\n",
+       "         [-0.9993973 ,  0.03471346],\n",
+       "         ...,\n",
+       "         [-0.7566905 ,  0.6537732 ],\n",
+       "         [-0.58953756,  0.80774087],\n",
+       "         [-0.68740994,  0.72626966]],\n",
+       "\n",
+       "        [[-0.46347883,  0.8861079 ],\n",
+       "         [ 0.00387764,  0.99999243],\n",
+       "         [ 0.58661395, -0.80986667],\n",
+       "         ...,\n",
+       "         [ 0.9909095 ,  0.13453025],\n",
+       "         [-0.44597933,  0.89504325],\n",
+       "         [ 0.6099662 ,  0.7924275 ]],\n",
+       "\n",
+       "        [[ 0.88540715,  0.46481633],\n",
+       "         [-0.9495712 ,  0.31355175],\n",
+       "         [-0.9835763 , -0.1804929 ],\n",
+       "         ...,\n",
+       "         [ 0.11676199,  0.99315995],\n",
+       "         [ 0.78835183, -0.6152245 ],\n",
+       "         [ 0.6080205 , -0.7939213 ]],\n",
+       "\n",
+       "        ...,\n",
+       "\n",
+       "        [[ 0.8907368 , -0.45451924],\n",
+       "         [-0.61389214, -0.7893899 ],\n",
+       "         [-0.9944057 , -0.10562801],\n",
+       "         ...,\n",
+       "         [ 0.37469187,  0.92714936],\n",
+       "         [ 0.94401246, -0.32990944],\n",
+       "         [ 0.7751419 ,  0.6317871 ]],\n",
+       "\n",
+       "        [[-0.14082797, -0.99003404],\n",
+       "         [ 0.27515417,  0.96140003],\n",
+       "         [-0.40955904,  0.9122836 ],\n",
+       "         ...,\n",
+       "         [ 0.8188169 ,  0.5740549 ],\n",
+       "         [-0.9845477 , -0.17511682],\n",
+       "         [-0.71323305,  0.7009269 ]],\n",
+       "\n",
+       "        [[-0.9327261 , -0.36058545],\n",
+       "         [-0.2761146 , -0.9611246 ],\n",
+       "         [ 0.7902948 , -0.6127267 ],\n",
+       "         ...,\n",
+       "         [ 0.999295  ,  0.03754444],\n",
+       "         [ 0.95306456, -0.30276698],\n",
+       "         [-0.9659615 ,  0.2586859 ]]]], dtype=float32)>"
+      ]
+     },
+     "execution_count": 111,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "r = tf.random.normal(shape = (16, 8, 8, 2))\n",
+    "tf.linalg.l2_normalize(r,axis=-1)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": ".patch_gen_venv",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.9"
+  },
+  "orig_nbformat": 4
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
diff --git a/playgrund.ipynb b/playgrund.ipynb
deleted file mode 100644
index fd72057..0000000
--- a/playgrund.ipynb
+++ /dev/null
@@ -1,155 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import tensorflow as tf\n",
-    "import random"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 29,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "['🟥', '🟦', '🟧', '🟨', '🟩', '🟪', '🟫']\n"
-     ]
-    }
-   ],
-   "source": [
-    "squares = [chr(i) for i in range(0x1F7E5, 0x1F7EC)]\n",
-    "tfsquares = tf.constant(squares)\n",
-    "print(squares)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 51,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def tensor_to_string(tensor):\n",
-    "    square_select = tf.math.argmax(tensor, 2)\n",
-    "\n",
-    "    tfstring = tf.strings.join(\n",
-    "        tf.map_fn(\n",
-    "            lambda v:\n",
-    "            tf.strings.join(\n",
-    "                tf.gather(tfsquares, tf.cast(v, tf.int64))), square_select, fn_output_signature=tf.string), \"\\n\")\n",
-    "\n",
-    "    return tfstring.numpy().decode()\n"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 95,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "tf.Tensor([0. 0. 0. 1. 0. 0. 0.], shape=(7,), dtype=float32)\n",
-      "tf.Tensor([[3]], shape=(1, 1), dtype=int64)\n",
-      "🟦🟪🟦🟫🟦🟧🟨🟨\n",
-      "🟦🟫🟧🟫🟪🟥🟧🟪\n",
-      "🟫🟩🟨🟧🟨🟫🟪🟨\n",
-      "🟧🟦🟨🟪🟦🟨🟥🟨\n",
-      "🟫🟧🟧🟨🟧🟧🟩🟪\n",
-      "🟦🟦🟨🟦🟥🟫🟧🟩\n",
-      "🟫🟩🟧🟫🟨🟫🟥🟩\n",
-      "🟧🟫🟫🟩🟧🟧🟫🟩\n",
-      "--------------\n",
-      "tf.Tensor(\n",
-      "[[0. 0. 0. 0. 0. 0. 0. 0.]\n",
-      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
-      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
-      " [0. 0. 1. 0. 0. 1. 0. 0.]\n",
-      " [0. 0. 1. 0. 0. 1. 0. 0.]\n",
-      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
-      " [0. 0. 0. 0. 0. 0. 0. 0.]\n",
-      " [0. 1. 0. 0. 0. 0. 1. 0.]], shape=(8, 8), dtype=float32)\n"
-     ]
-    }
-   ],
-   "source": [
-    "oh = tf.one_hot(3, 7)\n",
-    "print(oh)\n",
-    "print(tf.where(oh))\n",
-    "\n",
-    "# a = tf.constant([[list(tf.one_hot(random.randrange(7), 7)) for i in range(8) ] for i in range(8)], shape=(8,8,len(squares)))\n",
-    "\n",
-    "# print([[list(tf.one_hot(random.randrange(7), 7)) for i in range(8) ] for i in range(8)])\n",
-    "e = [[random.randrange(7) for i in range(8)] for i in range(8)]\n",
-    "# print(e)\n",
-    "\n",
-    "a = tf.one_hot(tf.constant(e, shape=(8, 8)), len(squares))\n",
-    "\n",
-    "print(tensor_to_string(a))\n",
-    "print(\"--------------\")\n",
-    "# print(tensor_to_string(tf.transpose(a, [1, 0, 2])))\n",
-    "# print(\"--------------\")\n",
-    "# print(tensor_to_string(tf.linalg.matmul( a , tf.constant(\n",
-    "#     [\n",
-    "#         [0, 0, 0, 0, 0, 0, 0, 1],\n",
-    "#         [0, 0, 0, 0, 0, 0, 1, 0],\n",
-    "#         [0, 0, 0, 0, 0, 1, 0, 0],\n",
-    "#         [0, 0, 0, 0, 1, 0, 0, 0],\n",
-    "#         [0, 0, 0, 1, 0, 0, 0, 0],\n",
-    "#         [0, 0, 1, 0, 0, 0, 0, 0],\n",
-    "#         [0, 1, 0, 0, 0, 0, 0, 0],\n",
-    "#         [1, 0, 0, 0, 0, 0, 0, 0]\n",
-    "#     ], dtype=tf.float32, shape=(8,8,1)\n",
-    "# ), transpose_a=True)))\n",
-    "# print(\"--------------\")\n",
-    "# print(tensor_to_string(tf.reverse(a, (0,))))\n",
-    "# print(\"--------------\")\n",
-    "# print(tensor_to_string(tf.reverse(a, (1,))))\n",
-    "# print(tensor_to_string(\n",
-    "#     tf.linalg.matmul(\n",
-    "#         tf.transpose(a, [0, 2, 1]),\n",
-    "#         tf.reverse(a, (1,)))\n",
-    "# ))\n",
-    "print(\n",
-    "    tf.einsum(\n",
-    "        \"ijk,ijk->ij\",\n",
-    "        a,\n",
-    "        tf.reverse(a, (1,)))\n",
-    ")\n",
-    "\n",
-    "\n",
-    "# print(tf.constant(range(6), shape=(2,3)))\n",
-    "# tf.map_fn(tf.math.reduce_sum, tf.constant(range(6), shape=(2,3)))\n"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": ".patch_gen_venv",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.10.0"
-  },
-  "orig_nbformat": 4
- },
- "nbformat": 4,
- "nbformat_minor": 2
-}
diff --git a/polymap.py b/polymap.py
new file mode 100644
index 0000000..aa7ecad
--- /dev/null
+++ b/polymap.py
@@ -0,0 +1,28 @@
+import tensorflow as tf
+# from tensorflow import Layer
+
+
+class Polymap(tf.keras.layers.Layer):
+    def __init__(self, num_poly, poly_deg):
+        super(Polymap, self).__init__()
+        self.num_poly = num_poly
+        self.poly_deg = poly_deg
+        self.exponents = tf.constant(
+            [range(self.poly_deg + 1), range(self.poly_deg + 1)], dtype=tf.float32)
+        self.kernels = self.add_weight("kernels",
+                                       shape=[
+                                           self.num_poly, self.poly_deg + 1, self.poly_deg + 1],
+                                       trainable=True)
+
+    def build(self, input_shape):
+        if input_shape[-1] != 2:
+            raise "Input shape must be Ix2"
+        
+
+    def call(self, input):
+        powers = tf.map_fn(
+            lambda v: tf.pow(
+                tf.repeat(tf.reshape(v, shape=(2, 1)),
+                          repeats=self.poly_deg + 1, axis=1),
+                self.exponents), input)
+        return tf.einsum("ik,ij,lkj->il", powers[:, 0], powers[:, 1], self.kernels)