{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements; and to You under the Apache License, Version 2.0. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "# SINGA Core Classes\n",
    "\n",
    "<img src=\"http://singa.apache.org/en/_static/images/singav1-sw.png\" width=\"500px\"/>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "# Device\n",
    "\n",
    "A device instance represents a hardware device with multiple execution units, e.g.,\n",
    "* A GPU which has multile cuda streams\n",
    "* A CPU which has multiple threads\n",
    "\n",
    "All data structures (variables) are allocated on a device instance. Consequently, all operations are executed on the resident device.\n",
    "\n",
    "## Create a device instance"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<singa.singa_wrap.Device; proxy of <Swig Object of type 'std::shared_ptr< singa::Device > *' at 0x7f69a05ff330> >"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from singa import device\n",
    "default_dev = device.get_default_device()\n",
    "gpu = device.create_cuda_gpu()  # the first gpu device\n",
    "gpu"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "** NOTE: currently we can only call the creating function once due to the cnmem restriction.**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "gpu = device.create_cuda_gpu_on(1)  # use the gpu device with the specified GPU ID\n",
    "gpu_list1 = device.create_cuda_gpus(2)  # the first two gpu devices\n",
    "gpu_list2 = device.create_cuda_gpus([0,2]) # create the gpu instances on the given GPU IDs\n",
    "opencl_gpu = device.create_opencl_device()  # valid if SINGA is compiled with USE_OPENCL=ON"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "3"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "device.get_num_gpus()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(0, 1, 2)"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "device.get_gpu_ids()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Tensor\n",
    "\n",
    "A tensor instance represents a multi-dimensional array allocated on a device instance.\n",
    "It provides linear algbra operations, like +, -, *, /, dot, pow ,etc\n",
    "\n",
    "NOTE: class memeber functions are inplace; global functions are out-of-place."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Create tensor instances"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(2, 3)"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from singa import tensor\n",
    "import numpy as np\n",
    "a = tensor.Tensor((2, 3))\n",
    "a.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<singa.singa_wrap.Device; proxy of <Swig Object of type 'std::shared_ptr< singa::Device > *' at 0x7f69a02f8a50> >"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a.device"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "gb = tensor.Tensor((2, 3), gpu)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<singa.singa_wrap.Device; proxy of <Swig Object of type 'std::shared_ptr< singa::Device > *' at 0x7f69a05ff330> >"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gb.device"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Initialize tensor values"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "a.set_value(1.2)\n",
    "gb.gaussian(0, 0.1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### To and from numpy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 1.20000005,  1.20000005,  1.20000005],\n",
       "       [ 1.20000005,  1.20000005,  1.20000005]], dtype=float32)"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tensor.to_numpy(a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 0.24042693, -0.21337385, -0.0969397 ],\n",
       "       [-0.010797  , -0.07642138, -0.09220808]], dtype=float32)"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tensor.to_numpy(gb)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(2,)"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "c = tensor.from_numpy(np.array([1,2], dtype=np.float32))\n",
    "c.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 3.,  4.], dtype=float32)"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "c.copy_from_numpy(np.array([3,4], dtype=np.float32))\n",
    "tensor.to_numpy(c)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Move tensor between devices"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<singa.singa_wrap.Device; proxy of <Swig Object of type 'std::shared_ptr< singa::Device > *' at 0x7f69a05ff330> >"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gc = c.clone()\n",
    "gc.to_device(gpu)\n",
    "gc.device"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<singa.singa_wrap.Device; proxy of <Swig Object of type 'std::shared_ptr< singa::Device > *' at 0x7f69a02f8a50> >"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "b = gb.clone()\n",
    "b.to_host()  # the same as b.to_device(default_dev)\n",
    "b.device"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Operations\n",
    "\n",
    "**NOTE: tensors should be initialized if the operation would read the tensor values**"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Summary"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.12169448286294937"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gb.l1()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.4898979663848877"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a.l2()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "False"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "e = tensor.Tensor((2, 3))\n",
    "e.is_empty()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "6L"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gb.size()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "24L"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gb.memsize()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "False"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# note we can only support matrix multiplication for tranposed tensors; \n",
    "# other operations on transposed tensor would result in errors\n",
    "c.is_transpose()  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "et=e.T()\n",
    "et.is_transpose()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(3L, 2L)"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "et.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2L"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "et.ndim()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Member functions (in-place)\n",
    "\n",
    "These functions would change the content of the tensor"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 1.44042695,  0.98662621,  1.10306036],\n",
       "       [ 1.18920302,  1.12357867,  1.10779202]], dtype=float32)"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a += b\n",
    "tensor.to_numpy(a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 1.20000005,  1.20000005,  1.20000005],\n",
       "       [ 1.20000005,  1.20000005,  1.20000005]], dtype=float32)"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a -= b\n",
    "tensor.to_numpy(a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 2.4000001,  2.4000001,  2.4000001],\n",
       "       [ 2.4000001,  2.4000001,  2.4000001]], dtype=float32)"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a *= 2\n",
    "tensor.to_numpy(a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 0.80000007,  0.80000007,  0.80000007],\n",
       "       [ 0.80000007,  0.80000007,  0.80000007]], dtype=float32)"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a /= 3\n",
    "tensor.to_numpy(a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 0.62944734, -0.72904599,  0.81158388], dtype=float32)"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "d = tensor.Tensor((3,))\n",
    "d.uniform(-1,1)\n",
    "tensor.to_numpy(d)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 1.42944741,  0.07095408,  1.61158395],\n",
       "       [ 1.42944741,  0.07095408,  1.61158395]], dtype=float32)"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a.add_row(d)\n",
    "tensor.to_numpy(a)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Global functions (out of place)\n",
    "\n",
    "These functions would not change the memory of the tensor, instead they return a new tensor\n",
    "\n",
    "**Unary functions**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 1., -1.,  1.], dtype=float32)"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "h = tensor.sign(d)\n",
    "tensor.to_numpy(h)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 0.62944734, -0.72904599,  0.81158388], dtype=float32)"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tensor.to_numpy(d)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 0.62944734,  0.72904599,  0.81158388], dtype=float32)"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "h = tensor.abs(d)\n",
    "tensor.to_numpy(h)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 0.62944734,  0.        ,  0.81158388], dtype=float32)"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "h = tensor.relu(d)\n",
    "tensor.to_numpy(h)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(3L,)"
      ]
     },
     "execution_count": 34,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "g = tensor.sum(a, 0)\n",
    "g.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(2L,)"
      ]
     },
     "execution_count": 35,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "g = tensor.sum(a, 1)\n",
    "g.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 1.,  0.], dtype=float32)"
      ]
     },
     "execution_count": 36,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tensor.bernoulli(0.5, g)\n",
    "tensor.to_numpy(g)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([-0.12226005, -0.05827543], dtype=float32)"
      ]
     },
     "execution_count": 37,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "g.gaussian(0, 0.2)\n",
    "tensor.gaussian(0, 0.2, g)\n",
    "tensor.to_numpy(g)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Binary functions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 1.66987431, -0.14241977,  1.51464427],\n",
       "       [ 1.41865039, -0.0054673 ,  1.51937592]], dtype=float32)"
      ]
     },
     "execution_count": 38,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "f = a + b\n",
    "tensor.to_numpy(f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 0.,  0.,  0.],\n",
       "       [ 0.,  0.,  0.]], dtype=float32)"
      ]
     },
     "execution_count": 39,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "g = a < b\n",
    "tensor.to_numpy(g)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 7.66987419,  5.85758018,  7.51464415],\n",
       "       [ 9.41865063,  7.99453259,  9.5193758 ]], dtype=float32)"
      ]
     },
     "execution_count": 40,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tensor.add_column(2, c, 1, f)   # f = 2 *c + 1* f\n",
    "tensor.to_numpy(f)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### BLAS\n",
    "\n",
    "BLAS function may change the memory of input tensor"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 0.24042693, -0.21337385, -0.0969397 ],\n",
       "       [-0.010797  , -0.07642138, -0.09220808]], dtype=float32)"
      ]
     },
     "execution_count": 41,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tensor.axpy(2, a, f)  # f = 2a + f\n",
    "tensor.to_numpy(b)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 0.17231143, -0.16945721],\n",
       "       [ 0.17231143, -0.16945721]], dtype=float32)"
      ]
     },
     "execution_count": 42,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "f = tensor.mult(a, b.T())\n",
    "tensor.to_numpy(f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 0.51693428, -0.50837165],\n",
       "       [ 0.51693428, -0.50837165]], dtype=float32)"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tensor.mult(a, b.T(), f, 2, 1)  # f = 2a*b.T() + 1f\n",
    "tensor.to_numpy(f)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "## Next: [SINGA model classes](./model.ipynb)"
   ]
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [conda env:pysinga]",
   "language": "python",
   "name": "conda-env-pysinga-py"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}