Pytorch Embedding that supports gradient propagation to inputs

pytorch_hot_embedding.ipynb

pytorch_linear_hot_embedding.ipynb

pytorch_linear_hot_embedding_no_square.ipynb

{
 "metadata": {
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.7-final"
  },
  "orig_nbformat": 2,
  "kernelspec": {
   "name": "python37764bitbaseconda4a4f2856d96346b89bf455c97059c11c",
   "display_name": "Python 3.7.7 64-bit ('base': conda)"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2,
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "class HotEmbedding(torch.nn.Module):\n",
    "    def __init__(self, max_val, embedding_dim, eps=1e-2):\n",
    "        super(HotEmbedding, self).__init__()\n",
    "        A = torch.arange(max_val, requires_grad=False)\n",
    "        self.register_buffer(\"A\", A)\n",
    "        self.B = torch.nn.Linear(max_val, embedding_dim)\n",
    "        self.eps = eps\n",
    "\n",
    "    def forward(self, x):\n",
    "        return self.B(1/((x.unsqueeze(1) - self.A)+self.eps))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "tags": []
   },
   "outputs": [
    {
     "output_type": "stream",
     "name": "stdout",
     "text": "tensor([1., 2., 3., 1., 2., 3.], requires_grad=True)\ntensor([[-27.3928, -27.0610,  18.1991,  16.2886,   2.6034],\n        [ 26.3739,  20.9340, -11.5895, -12.7277, -28.9630],\n        [-17.8610,  18.7509, -18.1535,   0.1145, -25.8138],\n        [-27.3928, -27.0610,  18.1991,  16.2886,   2.6034],\n        [ 26.3739,  20.9340, -11.5895, -12.7277, -28.9630],\n        [-17.8610,  18.7509, -18.1535,   0.1145, -25.8138]],\n       grad_fn=<AddmmBackward>)\n"
    }
   ],
   "source": [
    "layer = HotEmbedding(10, 5)\n",
    "x = torch.tensor([1.,2.,3.,1.,2.,3.], requires_grad=True)\n",
    "y = layer(x)\n",
    "print(x)\n",
    "print(y)"
   ]
  }
 ]
}