Orbits/Code/PartialDerivativesEstimand.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "working-peeing",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "decimal-boundary",
   "metadata": {},
   "source": [
    "The purpose of this notebook is to allow me to investigate proper shaping of inputs.\n",
    "\n",
    "Typically pytorch chooses a tensor specification\n",
    "$$\n",
    "(N, .*)\n",
    "$$\n",
    "where $N$ is the batch size.\n",
    "For example a Convolutional NN layer expects\n",
    "$$\n",
    "    NCHW\n",
    "$$\n",
    "for BatchSize,ChannelSize,Height,Width.\n",
    "On the other hand, Linear expects\n",
    "$$\n",
    "    N.*H\n",
    "$$\n",
    "for BatchSize,any number of other dimensions, in_features\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "eligible-isolation",
   "metadata": {},
   "outputs": [],
   "source": [
    "class PartialDerivativesEstimand(torch.nn.Module):\n",
    "    def __init__(self,batch_size, number_constellations, number_states,scale_factor=4, layer_size=12):\n",
    "        \"\"\"\n",
    "        \n",
    "        \"\"\"\n",
    "        super().__init__()\n",
    "        self.batch_size = batch_size\n",
    "        self.number_constellations = number_constellations\n",
    "        self.number_states = number_states\n",
    "        self.scale_factor = scale_factor\n",
    "        self.layer_size = layer_size\n",
    "        \n",
    "        \n",
    "        #preprocess (single linear layer in case there is anything that needs to happen to all states)\n",
    "        self.preprocess = torch.nn.Sequential(\n",
    "            torch.nn.ReLU() #cleanup as states must be positive\n",
    "            ,torch.nn.Linear(in_features = self.number_states, out_features=self.number_states)\n",
    "        )\n",
    "        #upscale to get the basic dimensionality correct. From (batch,State) to (batch, constellation, state). Includes a reshape\n",
    "        self.upsample = lambda x: torch.nn.Upsample(scale_factor=self.number_constellations)(x).view(self.batch_size\n",
    "                                                                                            ,self.number_constellations\n",
    "                                                                                            ,self.number_states)\n",
    "        \n",
    "        #sequential steps\n",
    "        self.sequential = torch.nn.Sequential(\n",
    "            torch.nn.Linear(in_features=number_states, out_features=layer_size)\n",
    "            #who knows if a convolution might help here.\n",
    "            ,torch.nn.Linear(in_features=layer_size, out_features=layer_size)\n",
    "            ,torch.nn.Linear(in_features=layer_size, out_features=layer_size)\n",
    "        )\n",
    "\n",
    "        #reduce axis to match expectation\n",
    "        self.feature_reduction = torch.nn.Linear(in_features=layer_size, out_features=number_states)\n",
    "        \n",
    "    def forward(self, input_values):\n",
    "        #Note that the input values are just going to be the state variables\n",
    "        #TODO:check that input values match the prepared dimension?\n",
    "        \n",
    "        #preprocess\n",
    "        intermediate = self.preprocess(input_values)\n",
    "        \n",
    "        #upscale the input values\n",
    "        intermediate = self.upsample(intermediate)\n",
    "        \n",
    "        #intermediate processing\n",
    "        intermediate = self.sequential(intermediate)\n",
    "        \n",
    "        #reduce feature axis to match the expected number of partials\n",
    "        intermediate = self.feature_reduction(intermediate)\n",
    "        \n",
    "        \n",
    "        return intermediate"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "literary-desktop",
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 2\n",
    "constellations = 2\n",
    "number_states = constellations+1\n",
    "\n",
    "#initialize the NN\n",
    "a = PartialDerivativesEstimand(batch_size,constellations,number_states,scale_factor=2)\n",
    "\n",
    "#example state\n",
    "s = torch.rand(size=(batch_size,1,number_states))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "second-graduation",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[[0.9283, 0.9414, 0.3426]],\n",
       "\n",
       "        [[0.1902, 0.0369, 0.4699]]])"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "s"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "reliable-alberta",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[[-0.1991,  0.1335,  0.2821],\n",
       "         [-0.3549,  0.0213,  0.2322]],\n",
       "\n",
       "        [[-0.1701,  0.1557,  0.2954],\n",
       "         [-0.3017,  0.0690,  0.2419]]], grad_fn=<AddBackward0>)"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a(s)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "horizontal-judges",
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "instant-lindsay",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.8.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}