def check_affine(f, *nu_inputs): types = ",".join(["{%s,%s}" % (x.dtype, x.ndim) for x in nu_inputs]) cgt.utils.colorprint(cgt.utils.Color.YELLOW, "Testing %s(%s)\n" % (f.__name__, types)) sy_inputs = map(tensor_like, nu_inputs) for (i, sy) in enumerate(sy_inputs): sy.name = "x%i" % i sy_result = f(*sy_inputs) def maybeprint(msg): if DISPLAY: print msg maybeprint("Function:") if DISPLAY: cgt.print_tree([sy_result]) f_cgt = cgt.function(sy_inputs, sy_result) sy_grads = cgt.grad(sy_result, sy_inputs) gradf_cgt = cgt.function(sy_inputs, sy_grads) sy_result_simple = core.simplify([sy_result]) sy_grads_simple = core.simplify(sy_grads) maybeprint("Gradient:") if DISPLAY: cgt.print_tree(sy_grads) maybeprint("Gradient after simplification:") if DISPLAY: cgt.print_tree(sy_grads_simple) out_true = f(*nu_inputs) out_cgt = f_cgt(*nu_inputs) grads_true = gradients_affine(f_cgt, nu_inputs, h=1e-4 if "max" in f.__name__ else 1e-1) grads_cgt = gradf_cgt(*nu_inputs) rtol = {"single": 1e-3, "double": 1e-5}[cgt.get_precision()] np.testing.assert_allclose(out_cgt, out_true, rtol=rtol) for (g_cgt, g_true) in zip(grads_cgt, grads_true): np.testing.assert_allclose(g_cgt, g_true, rtol=rtol) result_count = cgt.count_nodes(sy_result_simple) grad_count = cgt.count_nodes(sy_grads_simple) maybeprint("Result before: %i. after: %i" % (cgt.count_nodes([sy_result]), result_count)) maybeprint("Grad before: %i. after: %i" % (cgt.count_nodes(sy_grads), grad_count)) PROB2RESULT[f.__name__] = {} PROB2RESULT[f.__name__]["fn"] = result_count PROB2RESULT[f.__name__]["grad"] = grad_count
def test_flatvec(): cgt.reset_config cgt.set_precision('double') cgt.core.update_config(backend="python") # XXX N = 10 K = 3 Xval = np.random.randn(N, K) wval = np.random.randn(K) bval = np.random.randn() yval = np.random.randn(N) X_nk = cgt.shared(Xval, "X") y_n = cgt.shared(yval, "y") w_k = cgt.shared(wval, "w") b = cgt.shared(bval, name="b") ypred = cgt.dot(X_nk, w_k) + b err = cgt.sum(cgt.square(ypred - y_n)) g = cgt.grad(err, [w_k, b]) g = core.simplify(g) pars = [w_k, b] flatx = nn.setup_contiguous_storage(pars) f = cgt.function([], [err, cgt.flatcat(g)])
def test_flatvec(): cgt.reset_config cgt.set_precision('double') cgt.core.update_config(backend="python") # XXX N = 10 K = 3 Xval = np.random.randn(N,K) wval = np.random.randn(K) bval = np.random.randn() yval = np.random.randn(N) X_nk = cgt.shared(Xval, "X") y_n = cgt.shared(yval, "y") w_k = cgt.shared(wval, "w") b = cgt.shared(bval, name="b") ypred = cgt.dot(X_nk, w_k) + b err = cgt.sum(cgt.square(ypred - y_n)) g = cgt.grad(err, [w_k, b]) g = core.simplify(g) pars = [w_k, b] flatx = nn.setup_contiguous_storage(pars) f = cgt.function([], [err,cgt.flatcat(g)])
def check_scalar_grads(precision, backend): cgt.reset_config() np.random.seed(0) cgt.set_precision(precision) cgt.core.update_config(backend=backend) x = cgt.scalar('x') y = cgt.scalar('y') z = cgt.scalar('z') vars = [x,y,z] #pylint: disable=W0622 vals = nr.rand(len(vars))+1 PROB2RESULT = {} for ((key,_), cls) in it.chain( it.izip(core.UNARY_INFO.items(),it.repeat(core.ElwiseUnary)), it.izip(core.BINARY_INFO.items(),it.repeat(core.ElwiseBinary)) ): if key == "conj": print "skipping conj" continue utils.colorprint(utils.Color.YELLOW, "Testing %s\n"%key) if cls == core.ElwiseUnary: n_in = 1 op = cls(key) else: n_in = 2 op = cls(key, (True,True)) inputvars = vars[0:n_in] inputvals = vals[0:n_in] out = core.Result(op, inputvars) f = cgt.function(inputvars, out) try: grads = cgt.grad(out, inputvars) except core.NonDifferentiable: print "nondiff" continue if DISPLAY: print "Function:" cgt.print_tree(out) print "Gradient original:" cgt.print_tree(grads) print "Gradient simplified:" grads_simple = core.simplify(grads) if DISPLAY: cgt.print_tree(grads_simple) gradf = cgt.function(inputvars, grads) eps = {"single":1e-4,"double":1e-9}[precision] nugrad = numeric_grad(lambda li: f(*li), inputvals,eps=eps) #pylint: disable=W0640 cgtgrad = gradf(*inputvals) np.testing.assert_almost_equal(nugrad,cgtgrad,decimal={"single":3,"double":6}[precision]) grad_count = core.count_nodes(grads_simple) PROB2RESULT[key] = {} PROB2RESULT[key]["grad"] = grad_count if DISPLAY: from thirdparty.tabulate import tabulate print tabulate([[key,val["grad"]] for (key,val) in PROB2RESULT.iteritems()],headers=["funcname","gradcount"])
def test_scalars(): np.random.seed(0) x = cgt.scalar('x') y = cgt.scalar('y') z = cgt.scalar('z') vars = [x,y,z] #pylint: disable=W0622 vals = nr.rand(len(vars))+1 PROB2RESULT = {} for ((key,_), cls) in it.chain( it.izip(core.UNARY_INFO.items(),it.repeat(core.ElwiseUnary)), it.izip(core.BINARY_INFO.items(),it.repeat(core.ElwiseBinary)) ): if key == "conj": print "skipping conj" continue utils.colorprint(utils.Color.YELLOW, "Testing %s\n"%key) if cls == core.ElwiseUnary: n_in = 1 op = cls(key) else: n_in = 2 op = cls(key, (True,True)) inputvars = vars[0:n_in] inputvals = vals[0:n_in] out = core.Result(op, inputvars) f = cgt.function(inputvars, out) try: grads = cgt.grad(out, inputvars) except core.NonDifferentiable: print "nondiff" continue if DISPLAY: print "Function:" cgt.print_tree(out) print "Gradient original:" cgt.print_tree(grads) print "Gradient simplified:" grads_simple = core.simplify(grads) if DISPLAY: cgt.print_tree(grads_simple) gradf = cgt.function(inputvars, grads) eps = {"single":1e-4,"double":1e-9}[cgt.get_precision()] nugrad = numeric_grad(lambda li: f(*li), inputvals,eps=eps) #pylint: disable=W0640 cgtgrad = gradf(*inputvals) np.testing.assert_almost_equal(nugrad,cgtgrad,decimal={"single":3,"double":6}[cgt.get_precision()]) grad_count = core.count_nodes(grads_simple) PROB2RESULT[key] = {} PROB2RESULT[key]["grad"] = grad_count if DISPLAY: from thirdparty.tabulate import tabulate print tabulate([[key,val["grad"]] for (key,val) in PROB2RESULT.iteritems()],headers=["funcname","gradcount"])