{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Tensor Train layers for neural networks\n",
    "\n",
    "In this section, the TT layers are introduced.\n",
    "\n",
    "Imports:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch as tn\n",
    "import torch.nn as nn\n",
    "import datetime\n",
    "try:\n",
    "    import torchtt as tntt\n",
    "except:\n",
    "    print('Installing torchTT...')\n",
    "    %pip install git+https://github.com/ion-g-ion/torchTT\n",
    "    import torchtt as tntt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We consider a linear layer $\\mathcal{LTT}(\\mathsf{x}) = \\mathsf{Wx}+\\mathsf{b}$ acting on a tensor input $\\mathsf{x}$ of shape $n_1 \\times \\cdots \\times n_d$ and returning a tensor of shape $m_1\\times\\cdots\\times m_d$. The corresponding weight matrix $\\mathsf{W}$ would have the shape $(m_1\\times\\cdots\\times m_d) \\times (n_1 \\times \\cdots \\times n_d)$. The goal is to represent the weights tensor operator in TT format and perform the learning with respect tot the cores of the TT decomposition (ranks have to be fixed a priori).\n",
    "Due to the AD functionality of `torchtt`, the gradient with respect tot the cores can be computed for any network structure.\n",
    "TT layers can be added using `torchtt.nn.LinearLayerTT()` class. \n",
    "\n",
    "In the following, a neural netywork with 3 hidden layers and one linear layer is created.\n",
    "The shapes of the individual layers are \n",
    "\n",
    "$\\mathbb{R}^{16} \\times\\mathbb{R}^{16} \\times\\mathbb{R}^{16} \\times\\mathbb{R}^{16} \\underset{}{\\longrightarrow} \\mathbb{R}^8 \\times\\mathbb{R}^8 \\times\\mathbb{R}^8 \\times\\mathbb{R}^8 \\underset{}{\\longrightarrow} \\mathbb{R}^4 \\times\\mathbb{R}^4 \\times\\mathbb{R}^4 \\times\\mathbb{R}^4  \\underset{}{\\longrightarrow}  \\mathbb{R}^2 \\times\\mathbb{R}^4 \\times\\mathbb{R}^2 \\times\\mathbb{R}^4 \\underset{}{\\longrightarrow} \\mathbb{R}^{10}$.\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "class BasicTT(nn.Module):\n",
    "    def __init__(self):\n",
    "        super().__init__()\n",
    "        self.ttl1 = tntt.nn.LinearLayerTT([16,16,16,16], [8,8,8,8], [1,3,3,3,1])\n",
    "        self.ttl2 = tntt.nn.LinearLayerTT([8,8,8,8], [4,4,4,4], [1,2,2,2,1])\n",
    "        self.ttl3 = tntt.nn.LinearLayerTT([4,4,4,4], [2,4,2,4], [1,2,2,2,1])\n",
    "        self.linear = nn.Linear(64, 10, dtype = tn.float32)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.ttl1(x)\n",
    "        x = tn.relu(x)\n",
    "        x = self.ttl2(x)\n",
    "        x = tn.relu(x)\n",
    "        x = self.ttl3(x)\n",
    "        x = tn.relu(x)\n",
    "        x = tn.reshape(x,[-1,64])\n",
    "        return self.linear(x)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Create the model and print the number of trainable parameters as well as the model structure."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of trainable parameters: 17\n",
      "BasicTT(\n",
      "  (ttl1): LinearLayerTT(\n",
      "    (cores): ParameterList(\n",
      "        (0): Parameter containing: [torch.float32 of size 1x8x16x3]\n",
      "        (1): Parameter containing: [torch.float32 of size 3x8x16x3]\n",
      "        (2): Parameter containing: [torch.float32 of size 3x8x16x3]\n",
      "        (3): Parameter containing: [torch.float32 of size 3x8x16x1]\n",
      "    )\n",
      "  )\n",
      "  (ttl2): LinearLayerTT(\n",
      "    (cores): ParameterList(\n",
      "        (0): Parameter containing: [torch.float32 of size 1x4x8x2]\n",
      "        (1): Parameter containing: [torch.float32 of size 2x4x8x2]\n",
      "        (2): Parameter containing: [torch.float32 of size 2x4x8x2]\n",
      "        (3): Parameter containing: [torch.float32 of size 2x4x8x1]\n",
      "    )\n",
      "  )\n",
      "  (ttl3): LinearLayerTT(\n",
      "    (cores): ParameterList(\n",
      "        (0): Parameter containing: [torch.float32 of size 1x2x4x2]\n",
      "        (1): Parameter containing: [torch.float32 of size 2x4x4x2]\n",
      "        (2): Parameter containing: [torch.float32 of size 2x2x4x2]\n",
      "        (3): Parameter containing: [torch.float32 of size 2x4x4x1]\n",
      "    )\n",
      "  )\n",
      "  (linear): Linear(in_features=64, out_features=10, bias=True)\n",
      ")\n"
     ]
    }
   ],
   "source": [
    "model = BasicTT()\n",
    "print('Number of trainable parameters:', len(list(model.parameters())))\n",
    "print(model)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "A random input is created and passed as argument to the model. Batch evaluation is also possible by extending the dimensionality of the input before the leading mode."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 3.61 s, sys: 2.22 s, total: 5.84 s\n",
      "Wall time: 450 ms\n"
     ]
    }
   ],
   "source": [
    "input = tn.rand((16,16,16,16), dtype = tn.float32)\n",
    "pred = model.forward(input)\n",
    "\n",
    "input_batch = tn.rand((1000,16,16,16,16), dtype = tn.float32)\n",
    "label_batch = tn.rand((1000,10), dtype = tn.float32)\n",
    "%time pred = model.forward(input_batch)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The obtained network can be trained similarily to other `torch` models.\n",
    "A loss function together with an optimizer are defined. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "criterion = nn.CrossEntropyLoss()\n",
    "optimizer = tn.optim.Adam(model.parameters(), lr = 0.001)   "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "A training loop is executed to exemplify the training parameters update procedure. An example where a true dataset is used is presented [here](https://github.com/ion-g-ion/torchTT/blob/main/examples/mnist_nn.ipynb).  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1, loss 1.196611e+01\n",
      "Epoch 2, loss 1.182479e+01\n",
      "Epoch 3, loss 1.174133e+01\n",
      "Epoch 4, loss 1.169423e+01\n",
      "Epoch 5, loss 1.166731e+01\n",
      "Finished Training\n",
      "CPU times: user 3.85 s, sys: 1.79 s, total: 5.64 s\n",
      "Wall time: 499 ms\n"
     ]
    }
   ],
   "source": [
    "for epoch in range(5):  \n",
    "\n",
    "    optimizer.zero_grad()\n",
    "\n",
    "    outputs = model(input_batch)\n",
    "    loss = criterion(outputs, label_batch)\n",
    "    loss.backward()\n",
    "    optimizer.step()\n",
    "\n",
    "    # print statistics\n",
    "    print('Epoch %d, loss %e'%(epoch+1,loss.item()))\n",
    "\n",
    "\n",
    "print('Finished Training')\n",
    "%time plm = model(input_batch)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "If the GPU is available, the model can be run on it to get a speedup (should be run 2 times to see the speedup due to CUDA warm-up)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Time on CPU  0:00:00.232971\n",
      "Time on GPU  0:00:00.003855\n"
     ]
    }
   ],
   "source": [
    "if tn.cuda.is_available():\n",
    "    model_gpu = BasicTT().cuda()\n",
    "    input_batch_gpu = tn.rand((400,16,16,16,16)).cuda()\n",
    "\n",
    "    input_batch = tn.rand((400,16,16,16,16))\n",
    "    tme = datetime.datetime.now()\n",
    "    pred = model.forward(input_batch)\n",
    "    tme = datetime.datetime.now() - tme\n",
    "    print('Time on CPU ',tme)\n",
    "\n",
    "    tme = datetime.datetime.now()\n",
    "    pred_gpu = model_gpu.forward(input_batch_gpu).cpu()\n",
    "    tme = datetime.datetime.now() - tme\n",
    "    print('Time on GPU ',tme)\n",
    "\n",
    "    "
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "pytorch",
   "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": "3.10.15"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}