Skip to content

Instantly share code, notes, and snippets.

@xiangze

xiangze/switch_dimshuffle_theano.ipynb

Created June 14, 2015 16:24
Show Gist options
  • Save xiangze/9385de1d8194160295f9 to your computer and use it in GitHub Desktop.
Save xiangze/9385de1d8194160295f9 to your computer and use it in GitHub Desktop.
Download ZIP
Display the source blob
Display the rendered blob
Raw
switch_dimshuffle_theano.ipynb
{
"metadata": {
"name": "",
"signature": "sha256:3acbd332f21cc0cbb0c9d37e650e384a2e8b4fe4b327e83dd98d0c7255712dd5"
},
"nbformat": 3,
"nbformat_minor": 0,
"worksheets": [
{
"cells": [
{
"cell_type": "code",
"collapsed": false,
"input": [
"import theano\n",
"from theano import function, sandbox, shared\n",
"import theano.tensor as T\n",
"import numpy as np"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "stream",
"stream": "stderr",
"text": [
"Using gpu device 0: GeForce GT 620\n"
]
}
],
"prompt_number": 1
},
{
"cell_type": "heading",
"level": 1,
"metadata": {},
"source": [
"Switch to vector,tensor"
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"def sel(a,b):\n",
" return T.switch(a<b,a,b)\n",
"#a=T.tensor3(\"A\")\n",
"a=T.vector(\"a\")\n",
"b=T.vector(\"b\")\n",
"\n",
"f=function([a,b],sel(a,b))\n",
"\n",
"f([1,2,3],[4,1,13])"
],
"language": "python",
"metadata": {},
"outputs": [
{
"metadata": {},
"output_type": "pyout",
"prompt_number": 66,
"text": [
"array([ 1., 1., 3.], dtype=float32)"
]
}
],
"prompt_number": 66
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"a=T.matrix(\"a\")\n",
"b=T.matrix(\"b\")\n",
"#https://groups.google.com/forum/#!topic/theano-users/YfrhjjOiyAQ\n",
"def reshape1(a):\n",
" return T.reshape(a,(1,T.shape(a)[0]*T.shape(a)[1]))\n",
"\n",
"def reshape1T(a):\n",
" return T.reshape(a,(T.shape(a)[0]*T.shape(a)[1],1))\n",
"#(T.shape(x)[0]*2,T.int_div(T.shape(x)[0],2)) \n",
"\n",
"def switchMin(a,b):\n",
" aa=reshape1(a)\n",
" bb=reshape1(b)\n",
" cc=T.switch(aa<bb,aa,bb)\n",
" return T.reshape(cc,(T.shape(a)[0],T.shape(a)[1]))\n",
"\n",
"def switchMax(a,b):\n",
" aa=reshape1(a)\n",
" bb=reshape1(b)\n",
" cc=T.switch(aa>bb,aa,bb)\n",
" return T.reshape(cc,(T.shape(a)[0],T.shape(a)[1]))\n",
"\n",
"fmin=function([a,b],switchMin(a,b),allow_input_downcast=True)"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 75
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"a=np.array(range(16),dtype=float)\n",
"a=a.reshape(4,4)\n",
"b=a+np.random.uniform(-1,1,size=16).reshape(4,4)\n",
"print a\n",
"print b"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "stream",
"stream": "stdout",
"text": [
"[[ 0. 1. 2. 3.]\n",
" [ 4. 5. 6. 7.]\n",
" [ 8. 9. 10. 11.]\n",
" [ 12. 13. 14. 15.]]\n",
"[[ -0.81269797 1.3718386 1.53156345 2.31660372]\n",
" [ 3.86150118 4.15623634 6.75110587 6.54735707]\n",
" [ 7.92685634 8.614342 10.64430822 10.17661102]\n",
" [ 12.15359756 13.85438625 14.28163421 14.25680402]]\n"
]
}
],
"prompt_number": 73
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"fmin(a,b)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"metadata": {},
"output_type": "pyout",
"prompt_number": 74,
"text": [
"array([[ -0.81269795, 1. , 1.5315634 , 2.31660366],\n",
" [ 3.86150122, 4.15623617, 6. , 6.54735708],\n",
" [ 7.92685652, 8.61434174, 10. , 10.17661095],\n",
" [ 12. , 13. , 14. , 14.25680447]], dtype=float32)"
]
}
],
"prompt_number": 74
},
{
"cell_type": "heading",
"level": 1,
"metadata": {},
"source": [
"Dimshuffle"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"http://deeplearning.net/software/theano/library/tensor/basic.html#theano.tensor._tensor_py_operators.dimshuffle"
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"a=T.tensor3('a')\n",
"f=function([a],a.dimshuffle(1,0,2),allow_input_downcast=True)"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 81
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"k=np.array(range(27),dtype=float)\n",
"\n",
"k=k.reshape(3,3,3)\n",
"print k\n",
"print f(k)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"output_type": "stream",
"stream": "stdout",
"text": [
"[[[ 0. 1. 2.]\n",
" [ 3. 4. 5.]\n",
" [ 6. 7. 8.]]\n",
"\n",
" [[ 9. 10. 11.]\n",
" [ 12. 13. 14.]\n",
" [ 15. 16. 17.]]\n",
"\n",
" [[ 18. 19. 20.]\n",
" [ 21. 22. 23.]\n",
" [ 24. 25. 26.]]]\n",
"[[[ 0. 1. 2.]\n",
" [ 9. 10. 11.]\n",
" [ 18. 19. 20.]]\n",
"\n",
" [[ 3. 4. 5.]\n",
" [ 12. 13. 14.]\n",
" [ 21. 22. 23.]]\n",
"\n",
" [[ 6. 7. 8.]\n",
" [ 15. 16. 17.]\n",
" [ 24. 25. 26.]]]\n"
]
}
],
"prompt_number": 82
},
{
"cell_type": "heading",
"level": 1,
"metadata": {},
"source": [
"Hamilton\u7a4d\u5206\u306eTensor\u4e26\u5217\u5316"
]
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"def dynamics(x0, v0, stepsize, n_steps, logP):\n",
"\n",
" def leapfrog(p, v, step):\n",
" dE = T.grad(logP(p).sum(), p)\n",
" vn=v-step*dE\n",
" xn=p+vn*step\n",
" return [xn,vn], {}\n",
"\n",
" E0 = logP(x0)\n",
" dE = T.grad(E0.sum(), x0)\n",
"\n",
" (xs, vs), scan_updates = theano.scan(leapfrog,\n",
" outputs_info=[dict(initial=x0+stepsize*v_half), dict(initial=v0-stepsize/2*dE),],\n",
" non_sequences=[stepsize],\n",
" n_steps=n_steps - 1)\n",
"\n",
" final_x,final_v = xs[-1],vs[-1]\n",
"\n",
" assert not scan_updates\n",
" E=logP(final_x)\n",
" final_v = final_v - stepsize/2 * T.grad(E.sum(), final_x)\n",
" return final_x, final_v"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 94
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"from scipy import linalg"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 86
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"dim=3\n",
"seed=120\n",
"\n",
"rng = np.random.RandomState(seed)\n",
"mu = np.array(rng.rand(dim) * 10, dtype=theano.config.floatX)\n",
"cov = np.array(rng.rand(dim, dim), dtype=theano.config.floatX)\n",
"cov = (cov + cov.T) / 2.\n",
"cov[np.arange(dim), np.arange(dim)] = 1.0\n",
"cov_inv = linalg.inv(cov)\n",
"\n",
"gaussian_energy=lambda x:(T.dot((x - mu), cov_inv) * (x - mu)).sum(axis=1)/2"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 98
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"din=lambda x,v:dynamics(x,v,1,100,gaussian_energy)"
],
"language": "python",
"metadata": {},
"outputs": [],
"prompt_number": 95
},
{
"cell_type": "code",
"collapsed": false,
"input": [
"v=T.vector('v')\n",
"T.grad(gaussian_energy(v),v)"
],
"language": "python",
"metadata": {},
"outputs": [
{
"ename": "ValueError",
"evalue": "Not enough dimensions on Elemwise{mul,no_inplace}.0 to reduce on axis 1",
"output_type": "pyerr",
"traceback": [
"\u001b[1;31m---------------------------------------------------------------------------\u001b[0m\n\u001b[1;31mValueError\u001b[0m Traceback (most recent call last)",
"\u001b[1;32m<ipython-input-97-461572e5f2ed>\u001b[0m in \u001b[0;36m<module>\u001b[1;34m()\u001b[0m\n\u001b[0;32m 1\u001b[0m \u001b[0mv\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0mT\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mvector\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;34m'v'\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m----> 2\u001b[1;33m \u001b[0mT\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mgrad\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mgaussian_energy\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mv\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m,\u001b[0m\u001b[0mv\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m",
"\u001b[1;32m<ipython-input-87-af0492ba03e6>\u001b[0m in \u001b[0;36m<lambda>\u001b[1;34m(x)\u001b[0m\n\u001b[0;32m 9\u001b[0m \u001b[0mcov_inv\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0mlinalg\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0minv\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mcov\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 10\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m---> 11\u001b[1;33m \u001b[0mgaussian_energy\u001b[0m\u001b[1;33m=\u001b[0m\u001b[1;32mlambda\u001b[0m \u001b[0mx\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mT\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mdot\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mx\u001b[0m \u001b[1;33m-\u001b[0m \u001b[0mmu\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0mcov_inv\u001b[0m\u001b[1;33m)\u001b[0m \u001b[1;33m*\u001b[0m \u001b[1;33m(\u001b[0m\u001b[0mx\u001b[0m \u001b[1;33m-\u001b[0m \u001b[0mmu\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0msum\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0maxis\u001b[0m\u001b[1;33m=\u001b[0m\u001b[1;36m1\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m/\u001b[0m\u001b[1;36m2\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m",
"\u001b[1;32mC:\\Python27\\lib\\site-packages\\theano\\tensor\\var.pyc\u001b[0m in \u001b[0;36msum\u001b[1;34m(self, axis, dtype, keepdims, acc_dtype)\u001b[0m\n\u001b[0;32m 450\u001b[0m return theano.tensor.basic.sum(self, axis=axis,\n\u001b[0;32m 451\u001b[0m \u001b[0mdtype\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0mdtype\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0mkeepdims\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0mkeepdims\u001b[0m\u001b[1;33m,\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m--> 452\u001b[1;33m acc_dtype=acc_dtype)\n\u001b[0m\u001b[0;32m 453\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 454\u001b[0m \u001b[1;32mdef\u001b[0m \u001b[0mprod\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mself\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0maxis\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0mNone\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0mdtype\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0mNone\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0mkeepdims\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0mFalse\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0macc_dtype\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0mNone\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
"\u001b[1;32mC:\\Python27\\lib\\site-packages\\theano\\tensor\\basic.pyc\u001b[0m in \u001b[0;36msum\u001b[1;34m(input, axis, dtype, keepdims, acc_dtype)\u001b[0m\n\u001b[0;32m 2668\u001b[0m \"\"\"\n\u001b[0;32m 2669\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m-> 2670\u001b[1;33m \u001b[0mout\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0melemwise\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mSum\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0maxis\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0maxis\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0mdtype\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0mdtype\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0macc_dtype\u001b[0m\u001b[1;33m=\u001b[0m\u001b[0macc_dtype\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0minput\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m\u001b[0;32m 2671\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 2672\u001b[0m \u001b[1;32mif\u001b[0m \u001b[0mkeepdims\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
"\u001b[1;32mC:\\Python27\\lib\\site-packages\\theano\\gof\\op.pyc\u001b[0m in \u001b[0;36m__call__\u001b[1;34m(self, *inputs, **kwargs)\u001b[0m\n\u001b[0;32m 397\u001b[0m \"\"\"\n\u001b[0;32m 398\u001b[0m \u001b[0mreturn_list\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0mkwargs\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mpop\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;34m'return_list'\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0mFalse\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m--> 399\u001b[1;33m \u001b[0mnode\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0mself\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mmake_node\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;33m*\u001b[0m\u001b[0minputs\u001b[0m\u001b[1;33m,\u001b[0m \u001b[1;33m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m\u001b[0;32m 400\u001b[0m \u001b[1;32mif\u001b[0m \u001b[0mself\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0madd_stack_trace_on_call\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 401\u001b[0m \u001b[0mself\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0madd_tag_trace\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mnode\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
"\u001b[1;32mC:\\Python27\\lib\\site-packages\\theano\\tensor\\elemwise.pyc\u001b[0m in \u001b[0;36mmake_node\u001b[1;34m(self, input)\u001b[0m\n\u001b[0;32m 1795\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 1796\u001b[0m \u001b[1;32massert\u001b[0m \u001b[0mop\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0macc_dtype\u001b[0m \u001b[1;32mis\u001b[0m \u001b[1;32mnot\u001b[0m \u001b[0mNone\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m-> 1797\u001b[1;33m \u001b[1;32mreturn\u001b[0m \u001b[0mCAReduce\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mmake_node\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mop\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0minput\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m\u001b[0;32m 1798\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 1799\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n",
"\u001b[1;32mC:\\Python27\\lib\\site-packages\\theano\\tensor\\elemwise.pyc\u001b[0m in \u001b[0;36mmake_node\u001b[1;34m(self, input)\u001b[0m\n\u001b[0;32m 1273\u001b[0m raise ValueError((\n\u001b[0;32m 1274\u001b[0m \u001b[1;34m'Not enough dimensions on %s to reduce on axis %s'\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[1;32m-> 1275\u001b[1;33m % (input, axis)))\n\u001b[0m\u001b[0;32m 1276\u001b[0m \u001b[0minput\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0mas_tensor_variable\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0minput\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m 1277\u001b[0m \u001b[0maxis\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0mself\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0maxis\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
"\u001b[1;31mValueError\u001b[0m: Not enough dimensions on Elemwise{mul,no_inplace}.0 to reduce on axis 1"
]
}
],
"prompt_number": 97
},
{
"cell_type": "code",
"collapsed": false,
"input": [],
"language": "python",
"metadata": {},
"outputs": []
}
],
"metadata": {}
}
]
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment