brainstate 0.0.1__py2.py3-none-any.whl

brainstate/transform/_autograd_test.py

@@ -0,0 +1,1183 @@
+# Copyright 2024 BDP Ecosystem Limited. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+# -*- coding: utf-8 -*-
+
+import unittest
+from pprint import pprint
+
+import jax
+import jax.numpy as jnp
+import pytest
+
+import brainstate as bc
+from brainstate.transform._autograd import _jacfwd
+
+
+class TestPureFuncGrad(unittest.TestCase):
+  def test_grad_pure_func_1(self):
+    def call(a, b, c): return jnp.sum(a + b + c)
+
+    bc.random.seed(1)
+    a = jnp.ones(10)
+    b = bc.random.randn(10)
+    c = bc.random.uniform(size=10)
+    f_grad = bc.transform.grad(call, argnums=[0, 1, 2])
+    grads = f_grad(a, b, c)
+
+    for g in grads: assert (g == 1.).all()
+
+  def test_grad_pure_func_2(self):
+    def call(a, b, c): return jnp.sum(a + b + c)
+
+    bc.random.seed(1)
+    a = jnp.ones(10)
+    b = bc.random.randn(10)
+    c = bc.random.uniform(size=10)
+    f_grad = bc.transform.grad(call)
+    assert (f_grad(a, b, c) == 1.).all()
+
+  def test_grad_pure_func_aux1(self):
+    def call(a, b, c):
+      return jnp.sum(a + b + c), (jnp.sin(100), jnp.exp(0.1))
+
+    bc.random.seed(1)
+    f_grad = bc.transform.grad(call, argnums=[0, 1, 2])
+    with pytest.raises(TypeError):
+      f_grad(jnp.ones(10), bc.random.randn(10), bc.random.uniform(size=10))
+
+  def test_grad_pure_func_aux2(self):
+    def call(a, b, c):
+      return jnp.sum(a + b + c), (jnp.sin(100), jnp.exp(0.1))
+
+    bc.random.seed(1)
+    f_grad = bc.transform.grad(call, argnums=[0, 1, 2], has_aux=True)
+    grads, aux = f_grad(jnp.ones(10), bc.random.randn(10), bc.random.uniform(size=10))
+    for g in grads: assert (g == 1.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+  def test_grad_pure_func_return1(self):
+    def call(a, b, c): return jnp.sum(a + b + c)
+
+    bc.random.seed(1)
+    a = jnp.ones(10)
+    b = bc.random.randn(10)
+    c = bc.random.uniform(size=10)
+    f_grad = bc.transform.grad(call, return_value=True)
+    grads, returns = f_grad(a, b, c)
+    assert (grads == 1.).all()
+    assert returns == jnp.sum(a + b + c)
+
+  def test_grad_func_return_aux1(self):
+    def call(a, b, c):
+      return jnp.sum(a + b + c), (jnp.sin(100), jnp.exp(0.1))
+
+    bc.random.seed(1)
+    a = jnp.ones(10)
+    b = bc.random.randn(10)
+    c = bc.random.uniform(size=10)
+    f_grad = bc.transform.grad(call, return_value=True, has_aux=True)
+    grads, returns, aux = f_grad(a, b, c)
+    assert (grads == 1.).all()
+    assert returns == jnp.sum(a + b + c)
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+
+class TestObjectFuncGrad(unittest.TestCase):
+  def test_grad_ob1(self):
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+
+        self.a = bc.ParamState(jnp.ones(10))
+        self.b = bc.ParamState(bc.random.randn(10))
+        self.c = bc.ParamState(bc.random.uniform(size=10))
+
+      def __call__(self):
+        return jnp.sum(self.a.value + self.b.value + self.c.value)
+
+    bc.random.seed(0)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars={'a': t.a, 'b': t.b, 'c': t.c})
+    grads = f_grad()
+    for g in grads.values():
+      assert (g == 1.).all()
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=[t.a, t.b])
+    grads = f_grad()
+    for g in grads: assert (g == 1.).all()
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=t.a)
+    grads = f_grad()
+    assert (grads == 1.).all()
+
+  def test_grad_ob_aux(self):
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+        self.a = bc.ParamState(jnp.ones(10))
+        self.b = bc.ParamState(bc.random.randn(10))
+        self.c = bc.ParamState(bc.random.uniform(size=10))
+
+      def __call__(self):
+        return jnp.sum(self.a.value + self.b.value + self.c.value), (jnp.sin(100), jnp.exp(0.1))
+
+    bc.random.seed(0)
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=[t.a, t.b], has_aux=True)
+    grads, aux = f_grad()
+    for g in grads: assert (g == 1.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=t.a, has_aux=True)
+    grads, aux = f_grad()
+    assert (grads == 1.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+  def test_grad_ob_return(self):
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+        self.a = bc.ParamState(jnp.ones(10))
+        self.b = bc.ParamState(bc.random.randn(10))
+        self.c = bc.ParamState(bc.random.uniform(size=10))
+
+      def __call__(self):
+        return jnp.sum(self.a.value + self.b.value + self.c.value)
+
+    bc.random.seed(0)
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=[t.a, t.b], return_value=True)
+    grads, returns = f_grad()
+    for g in grads: assert (g == 1.).all()
+    assert returns == t()
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=t.a, return_value=True)
+    grads, returns = f_grad()
+    assert (grads == 1.).all()
+    assert returns == t()
+
+  def test_grad_ob_aux_return(self):
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+        self.a = bc.ParamState(jnp.ones(10))
+        self.b = bc.ParamState(bc.random.randn(10))
+        self.c = bc.ParamState(bc.random.uniform(size=10))
+
+      def __call__(self):
+        return jnp.sum(self.a.value + self.b.value + self.c.value), (jnp.sin(100), jnp.exp(0.1))
+
+    bc.random.seed(0)
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=[t.a, t.b], has_aux=True, return_value=True)
+    grads, returns, aux = f_grad()
+    for g in grads: assert (g == 1.).all()
+    assert returns == jnp.sum(t.a.value + t.b.value + t.c.value)
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=t.a, has_aux=True, return_value=True)
+    grads, returns, aux = f_grad()
+    assert (grads == 1.).all()
+    assert returns == jnp.sum(t.a.value + t.b.value + t.c.value)
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+  def test_grad_ob_argnums(self):
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+        bc.random.seed()
+        self.a = bc.ParamState(jnp.ones(10))
+        self.b = bc.ParamState(bc.random.randn(10))
+        self.c = bc.ParamState(bc.random.uniform(size=10))
+
+      def __call__(self, d):
+        return jnp.sum(self.a.value + self.b.value + self.c.value + 2 * d)
+
+    bc.random.seed(0)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, t.states(), argnums=0)
+    var_grads, arg_grads = f_grad(bc.random.random(10))
+    for g in var_grads.values(): assert (g == 1.).all()
+    assert (arg_grads == 2.).all()
+
+    t = Test()
+    f_grad = bc.transform.grad(t, t.states(), argnums=[0])
+    var_grads, arg_grads = f_grad(bc.random.random(10))
+    for g in var_grads.values(): assert (g == 1.).all()
+    assert (arg_grads[0] == 2.).all()
+
+    t = Test()
+    f_grad = bc.transform.grad(t, argnums=0)
+    arg_grads = f_grad(bc.random.random(10))
+    assert (arg_grads == 2.).all()
+
+    t = Test()
+    f_grad = bc.transform.grad(t, argnums=[0])
+    arg_grads = f_grad(bc.random.random(10))
+    assert (arg_grads[0] == 2.).all()
+
+  def test_grad_ob_argnums_aux(self):
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+        self.a = bc.ParamState(jnp.ones(10))
+        self.b = bc.ParamState(bc.random.randn(10))
+        self.c = bc.ParamState(bc.random.uniform(size=10))
+
+      def __call__(self, d):
+        return jnp.sum(self.a.value + self.b.value + self.c.value + 2 * d), (jnp.sin(100), jnp.exp(0.1))
+
+    bc.random.seed(0)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=t.states(), argnums=0, has_aux=True)
+    (var_grads, arg_grads), aux = f_grad(bc.random.random(10))
+    for g in var_grads.values(): assert (g == 1.).all()
+    assert (arg_grads == 2.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=t.states(), argnums=[0], has_aux=True)
+    (var_grads, arg_grads), aux = f_grad(bc.random.random(10))
+    for g in var_grads.values(): assert (g == 1.).all()
+    assert (arg_grads[0] == 2.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, argnums=0, has_aux=True)
+    arg_grads, aux = f_grad(bc.random.random(10))
+    assert (arg_grads == 2.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, argnums=[0], has_aux=True)
+    arg_grads, aux = f_grad(bc.random.random(10))
+    assert (arg_grads[0] == 2.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+
+  def test_grad_ob_argnums_return(self):
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+
+        self.a = bc.ParamState(jnp.ones(10))
+        self.b = bc.ParamState(bc.random.randn(10))
+        self.c = bc.ParamState(bc.random.uniform(size=10))
+
+      def __call__(self, d):
+        return jnp.sum(self.a.value + self.b.value + self.c.value + 2 * d)
+
+    bc.random.seed(0)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, t.states(), argnums=0, return_value=True)
+    d = bc.random.random(10)
+    (var_grads, arg_grads), loss = f_grad(d)
+    for g in var_grads.values():
+      assert (g == 1.).all()
+    assert (arg_grads == 2.).all()
+    assert loss == t(d)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, t.states(), argnums=[0], return_value=True)
+    d = bc.random.random(10)
+    (var_grads, arg_grads), loss = f_grad(d)
+    for g in var_grads.values():
+      assert (g == 1.).all()
+    assert (arg_grads[0] == 2.).all()
+    assert loss == t(d)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, argnums=0, return_value=True)
+    d = bc.random.random(10)
+    arg_grads, loss = f_grad(d)
+    assert (arg_grads == 2.).all()
+    assert loss == t(d)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, argnums=[0], return_value=True)
+    d = bc.random.random(10)
+    arg_grads, loss = f_grad(d)
+    assert (arg_grads[0] == 2.).all()
+    assert loss == t(d)
+
+  def test_grad_ob_argnums_aux_return(self):
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+        self.a = bc.ParamState(jnp.ones(10))
+        self.b = bc.ParamState(bc.random.randn(10))
+        self.c = bc.ParamState(bc.random.uniform(size=10))
+
+      def __call__(self, d):
+        return jnp.sum(self.a.value + self.b.value + self.c.value + 2 * d), (jnp.sin(100), jnp.exp(0.1))
+
+    bc.random.seed(0)
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=t.states(), argnums=0, has_aux=True, return_value=True)
+    d = bc.random.random(10)
+    (var_grads, arg_grads), loss, aux = f_grad(d)
+    for g in var_grads.values(): assert (g == 1.).all()
+    assert (arg_grads == 2.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+    assert loss == t(d)[0]
+
+    t = Test()
+    f_grad = bc.transform.grad(t, grad_vars=t.states(), argnums=[0], has_aux=True, return_value=True)
+    d = bc.random.random(10)
+    (var_grads, arg_grads), loss, aux = f_grad(d)
+    for g in var_grads.values(): assert (g == 1.).all()
+    assert (arg_grads[0] == 2.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+    assert loss == t(d)[0]
+
+    t = Test()
+    f_grad = bc.transform.grad(t, argnums=0, has_aux=True, return_value=True)
+    d = bc.random.random(10)
+    arg_grads, loss, aux = f_grad(d)
+    assert (arg_grads == 2.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+    assert loss == t(d)[0]
+
+    t = Test()
+    f_grad = bc.transform.grad(t, argnums=[0], has_aux=True, return_value=True)
+    d = bc.random.random(10)
+    arg_grads, loss, aux = f_grad(d)
+    assert (arg_grads[0] == 2.).all()
+    assert aux[0] == jnp.sin(100)
+    assert aux[1] == jnp.exp(0.1)
+    assert loss == t(d)[0]
+
+
+class TestPureFuncJacobian(unittest.TestCase):
+  def test1(self):
+    jac, aux = _jacfwd(lambda x: (x ** 3, [x ** 2]), has_aux=True)(3.)
+    self.assertTrue(jax.numpy.allclose(jac, jax.jacfwd(lambda x: x ** 3)(3.)))
+    self.assertTrue(aux[0] == 9.)
+
+  def test_jacfwd_and_aux_nested(self):
+    def f(x):
+      jac, aux = _jacfwd(lambda x: (x ** 3, [x ** 3]), has_aux=True)(x)
+      return aux[0]
+
+    f2 = lambda x: x ** 3
+
+    self.assertEqual(_jacfwd(f)(4.), _jacfwd(f2)(4.))
+    self.assertEqual(jax.jit(_jacfwd(f))(4.), _jacfwd(f2)(4.))
+    self.assertEqual(jax.jit(_jacfwd(jax.jit(f)))(4.), _jacfwd(f2)(4.))
+
+    self.assertEqual(_jacfwd(f)(jnp.asarray(4.)), _jacfwd(f2)(jnp.asarray(4.)))
+    self.assertEqual(jax.jit(_jacfwd(f))(jnp.asarray(4.)), _jacfwd(f2)(jnp.asarray(4.)))
+    self.assertEqual(jax.jit(_jacfwd(jax.jit(f)))(jnp.asarray(4.)), _jacfwd(f2)(jnp.asarray(4.)))
+
+    def f(x):
+      jac, aux = _jacfwd(lambda x: (x ** 3, [x ** 3]), has_aux=True)(x)
+      return aux[0] * jnp.sin(x)
+
+    f2 = lambda x: x ** 3 * jnp.sin(x)
+
+    self.assertEqual(_jacfwd(f)(4.), _jacfwd(f2)(4.))
+    self.assertEqual(jax.jit(_jacfwd(f))(4.), _jacfwd(f2)(4.))
+    self.assertEqual(jax.jit(_jacfwd(jax.jit(f)))(4.), _jacfwd(f2)(4.))
+
+    self.assertEqual(_jacfwd(f)(jnp.asarray(4.)), _jacfwd(f2)(jnp.asarray(4.)))
+    self.assertEqual(jax.jit(_jacfwd(f))(jnp.asarray(4.)), _jacfwd(f2)(jnp.asarray(4.)))
+    self.assertEqual(jax.jit(_jacfwd(jax.jit(f)))(jnp.asarray(4.)), _jacfwd(f2)(jnp.asarray(4.)))
+
+  def test_jacrev1(self):
+    def f1(x, y):
+      r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+      return r
+
+    br = bc.transform.jacrev(f1)(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    jr = jax.jacrev(f1)(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    assert (br == jr).all()
+
+    br = bc.transform.jacrev(f1, argnums=(0, 1))(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    jr = jax.jacrev(f1, argnums=(0, 1))(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    assert (br[0] == jr[0]).all()
+    assert (br[1] == jr[1]).all()
+
+  def test_jacrev2(self):
+    print()
+
+    def f2(x, y):
+      r1 = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1]])
+      r2 = jnp.asarray([4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+      return r1, r2
+
+    jr = jax.jacrev(f2)(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(jr)
+
+    br = bc.transform.jacrev(f2)(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(br)
+    assert jnp.array_equal(br[0], jr[0])
+    assert jnp.array_equal(br[1], jr[1])
+
+    br = bc.transform.jacrev(f2)(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(br)
+    assert jnp.array_equal(br[0], jr[0])
+    assert jnp.array_equal(br[1], jr[1])
+
+    def f2(x, y):
+      r1 = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1]])
+      r2 = jnp.asarray([4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+      return r1, r2
+
+    br = bc.transform.jacrev(f2)(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(br)
+    assert jnp.array_equal(br[0], jr[0])
+    assert jnp.array_equal(br[1], jr[1])
+
+    br = bc.transform.jacrev(f2)(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(br)
+    assert jnp.array_equal(br[0], jr[0])
+    assert jnp.array_equal(br[1], jr[1])
+
+  def test_jacrev3(self):
+    print()
+
+    def f3(x, y):
+      r1 = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1]])
+      r2 = jnp.asarray([4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+      return r1, r2
+
+    jr = jax.jacrev(f3, argnums=(0, 1))(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(jr)
+
+    br = bc.transform.jacrev(f3, argnums=(0, 1))(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(br)
+    assert jnp.array_equal(br[0][0], jr[0][0])
+    assert jnp.array_equal(br[0][1], jr[0][1])
+    assert jnp.array_equal(br[1][0], jr[1][0])
+    assert jnp.array_equal(br[1][1], jr[1][1])
+
+    br = bc.transform.jacrev(f3, argnums=(0, 1))(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(br)
+    assert jnp.array_equal(br[0][0], jr[0][0])
+    assert jnp.array_equal(br[0][1], jr[0][1])
+    assert jnp.array_equal(br[1][0], jr[1][0])
+    assert jnp.array_equal(br[1][1], jr[1][1])
+
+    def f3(x, y):
+      r1 = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1]])
+      r2 = jnp.asarray([4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+      return r1, r2
+
+    br = bc.transform.jacrev(f3, argnums=(0, 1))(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(br)
+    assert jnp.array_equal(br[0][0], jr[0][0])
+    assert jnp.array_equal(br[0][1], jr[0][1])
+    assert jnp.array_equal(br[1][0], jr[1][0])
+    assert jnp.array_equal(br[1][1], jr[1][1])
+
+    br = bc.transform.jacrev(f3, argnums=(0, 1))(jnp.array([1., 2., 3.]), jnp.array([10., 5.]))
+    pprint(br)
+    assert jnp.array_equal(br[0][0], jr[0][0])
+    assert jnp.array_equal(br[0][1], jr[0][1])
+    assert jnp.array_equal(br[1][0], jr[1][0])
+    assert jnp.array_equal(br[1][1], jr[1][1])
+
+  def test_jacrev_aux1(self):
+    x = jnp.array([1., 2., 3.])
+    y = jnp.array([10., 5.])
+
+    def f1(x, y):
+      a = 4 * x[1] ** 2 - 2 * x[2]
+      r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], a, x[2] * jnp.sin(x[0])])
+      return r, a
+
+    f2 = lambda *args: f1(*args)[0]
+    jr = jax.jacrev(f2)(x, y)  # jax jacobian
+    pprint(jr)
+    grads, aux = bc.transform.jacrev(f1, has_aux=True)(x, y)
+    assert (grads == jr).all()
+    assert aux == (4 * x[1] ** 2 - 2 * x[2])
+
+    jr = jax.jacrev(f2, argnums=(0, 1))(x, y)  # jax jacobian
+    pprint(jr)
+    grads, aux = bc.transform.jacrev(f1, argnums=(0, 1), has_aux=True)(x, y)
+    assert (grads[0] == jr[0]).all()
+    assert (grads[1] == jr[1]).all()
+    assert aux == (4 * x[1] ** 2 - 2 * x[2])
+
+  def test_jacrev_return_aux1(self):
+    with bc.environ.context(precision=64):
+
+      def f1(x, y):
+        a = 4 * x[1] ** 2 - 2 * x[2]
+        r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], a, x[2] * jnp.sin(x[0])])
+        return r, a
+
+      _x = jnp.array([1., 2., 3.])
+      _y = jnp.array([10., 5.])
+      _r, _a = f1(_x, _y)
+      f2 = lambda *args: f1(*args)[0]
+      _g1 = jax.jacrev(f2)(_x, _y)  # jax jacobian
+      pprint(_g1)
+      _g2 = jax.jacrev(f2, argnums=(0, 1))(_x, _y)  # jax jacobian
+      pprint(_g2)
+
+      grads, vec, aux = bc.transform.jacrev(f1, return_value=True, has_aux=True)(_x, _y)
+      assert (grads == _g1).all()
+      assert aux == _a
+      assert (vec == _r).all()
+
+      grads, vec, aux = bc.transform.jacrev(f1, return_value=True, argnums=(0, 1), has_aux=True)(_x, _y)
+      assert (grads[0] == _g2[0]).all()
+      assert (grads[1] == _g2[1]).all()
+      assert aux == _a
+      assert (vec == _r).all()
+
+
+
+class TestClassFuncJacobian(unittest.TestCase):
+  def test_jacrev1(self):
+    def f1(x, y):
+      r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+      return r
+
+    _x = jnp.array([1., 2., 3.])
+    _y = jnp.array([10., 5.])
+
+    class Test(bc.Module):
+      def __init__(self):
+        super(Test, self).__init__()
+        self.x = bc.State(jnp.array([1., 2., 3.]))
+        self.y = bc.State(jnp.array([10., 5.]))
+
+      def __call__(self, ):
+        a = self.x.value[0] * self.y.value[0]
+        b = 5 * self.x.value[2] * self.y.value[1]
+        c = 4 * self.x.value[1] ** 2 - 2 * self.x.value[2]
+        d = self.x.value[2] * jnp.sin(self.x.value[0])
+        r = jnp.asarray([a, b, c, d])
+        return r
+
+    _jr = jax.jacrev(f1)(_x, _y)
+    t = Test()
+    br = bc.transform.jacrev(t, grad_vars=t.x)()
+    self.assertTrue((br == _jr).all())
+
+    _jr = jax.jacrev(f1, argnums=(0, 1))(_x, _y)
+    t = Test()
+    br = bc.transform.jacrev(t, grad_vars=[t.x, t.y])()
+    self.assertTrue((br[0] == _jr[0]).all())
+    self.assertTrue((br[1] == _jr[1]).all())
+#
+#   def test_jacfwd1(self):
+#     def f1(x, y):
+#       r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+#       return r
+#
+#     _x = jnp.array([1., 2., 3.])
+#     _y = jnp.array([10., 5.])
+#
+#     class Test(bst.Module):
+#       def __init__(self):
+#         super(Test, self).__init__()
+#         self.x = jnp.Variable(jnp.array([1., 2., 3.]))
+#         self.y = jnp.Variable(jnp.array([10., 5.]))
+#
+#       def __call__(self, ):
+#         a = self.x[0] * self.y[0]
+#         b = 5 * self.x[2] * self.y[1]
+#         c = 4 * self.x[1] ** 2 - 2 * self.x[2]
+#         d = self.x[2] * jnp.sin(self.x[0])
+#         r = jnp.asarray([a, b, c, d])
+#         return r
+#
+#     _jr = jax.jacfwd(f1)(_x, _y)
+#     t = Test()
+#     br = bst.transform.jacfwd(t, grad_vars=t.x)()
+#     self.assertTrue((br == _jr).all())
+#
+#     _jr = jax.jacfwd(f1, argnums=(0, 1))(_x, _y)
+#     t = Test()
+#     br = bst.transform.jacfwd(t, grad_vars=[t.x, t.y])()
+#     self.assertTrue((br[0] == _jr[0]).all())
+#     self.assertTrue((br[1] == _jr[1]).all())
+#
+#   def test_jacrev2(self):
+#     def f1(x, y):
+#       r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+#       return r
+#
+#     _x = jnp.array([1., 2., 3.])
+#     _y = jnp.array([10., 5.])
+#
+#     class Test(bst.Module):
+#       def __init__(self):
+#         super(Test, self).__init__()
+#         self.x = jnp.Variable(jnp.array([1., 2., 3.]))
+#
+#       def __call__(self, y):
+#         a = self.x[0] * y[0]
+#         b = 5 * self.x[2] * y[1]
+#         c = 4 * self.x[1] ** 2 - 2 * self.x[2]
+#         d = self.x[2] * jnp.sin(self.x[0])
+#         r = jnp.asarray([a, b, c, d])
+#         return r
+#
+#     _jr = jax.jacrev(f1)(_x, _y)
+#     t = Test()
+#     br = bst.transform.jacrev(t, grad_vars=t.x)(_y)
+#     self.assertTrue((br == _jr).all())
+#
+#     _jr = jax.jacrev(f1, argnums=(0, 1))(_x, _y)
+#     t = Test()
+#     var_grads, arg_grads = bst.transform.jacrev(t, grad_vars=t.x, argnums=0)(_y)
+#     print(var_grads, )
+#     print(arg_grads, )
+#     self.assertTrue((var_grads == _jr[0]).all())
+#     self.assertTrue((arg_grads == _jr[1]).all())
+#
+#   def test_jacfwd2(self):
+#     def f1(x, y):
+#       r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+#       return r
+#
+#     _x = jnp.array([1., 2., 3.])
+#     _y = jnp.array([10., 5.])
+#
+#     class Test(bst.Module):
+#       def __init__(self):
+#         super(Test, self).__init__()
+#         self.x = jnp.Variable(jnp.array([1., 2., 3.]))
+#
+#       def __call__(self, y):
+#         a = self.x[0] * y[0]
+#         b = 5 * self.x[2] * y[1]
+#         c = 4 * self.x[1] ** 2 - 2 * self.x[2]
+#         d = self.x[2] * jnp.sin(self.x[0])
+#         r = jnp.asarray([a, b, c, d])
+#         return r
+#
+#     _jr = jax.jacfwd(f1)(_x, _y)
+#     t = Test()
+#     br = bst.transform.jacfwd(t, grad_vars=t.x)(_y)
+#     self.assertTrue((br == _jr).all())
+#
+#     _jr = jax.jacfwd(f1, argnums=(0, 1))(_x, _y)
+#     t = Test()
+#     var_grads, arg_grads = bst.transform.jacfwd(t, grad_vars=t.x, argnums=0)(_y)
+#     print(var_grads, )
+#     print(arg_grads, )
+#     self.assertTrue((var_grads == _jr[0]).all())
+#     self.assertTrue((arg_grads == _jr[1]).all())
+#
+#   def test_jacrev_aux1(self):
+#     jnp.enable_x64()
+#
+#     def f1(x, y):
+#       r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+#       return r
+#
+#     _x = jnp.array([1., 2., 3.])
+#     _y = jnp.array([10., 5.])
+#
+#     class Test(bst.Module):
+#       def __init__(self):
+#         super(Test, self).__init__()
+#         self.x = jnp.Variable(jnp.array([1., 2., 3.]))
+#
+#       def __call__(self, y):
+#         a = self.x[0] * y[0]
+#         b = 5 * self.x[2] * y[1]
+#         c = 4 * self.x[1] ** 2 - 2 * self.x[2]
+#         d = self.x[2] * jnp.sin(self.x[0])
+#         r = jnp.asarray([a, b, c, d])
+#         return r, (c, d)
+#
+#     _jr = jax.jacrev(f1)(_x, _y)
+#     t = Test()
+#     br, _ = bst.transform.jacrev(t, grad_vars=t.x, has_aux=True)(_y)
+#     self.assertTrue((br == _jr).all())
+#
+#     t = Test()
+#     _jr = jax.jacrev(f1, argnums=(0, 1))(_x, _y)
+#     _aux = t(_y)[1]
+#     (var_grads, arg_grads), aux = bst.transform.jacrev(t, grad_vars=t.x, argnums=0, has_aux=True)(_y)
+#     print(var_grads, )
+#     print(arg_grads, )
+#     self.assertTrue((var_grads == _jr[0]).all())
+#     self.assertTrue((arg_grads == _jr[1]).all())
+#     self.assertTrue(jnp.array_equal(aux, _aux))
+#
+#     jnp.disable_x64()
+#
+#   def test_jacfwd_aux1(self):
+#     jnp.enable_x64()
+#
+#     def f1(x, y):
+#       r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+#       return r
+#
+#     _x = jnp.array([1., 2., 3.])
+#     _y = jnp.array([10., 5.])
+#
+#     class Test(bst.Module):
+#       def __init__(self):
+#         super(Test, self).__init__()
+#         self.x = jnp.Variable(jnp.array([1., 2., 3.]))
+#
+#       def __call__(self, y):
+#         a = self.x[0] * y[0]
+#         b = 5 * self.x[2] * y[1]
+#         c = 4 * self.x[1] ** 2 - 2 * self.x[2]
+#         d = self.x[2] * jnp.sin(self.x[0])
+#         r = jnp.asarray([a, b, c, d])
+#         return r, (c, d)
+#
+#     _jr = jax.jacfwd(f1)(_x, _y)
+#     t = Test()
+#     br, (c, d) = bst.transform.jacfwd(t, grad_vars=t.x, has_aux=True)(_y)
+#     # print(_jr)
+#     # print(br)
+#     a = (br == _jr)
+#     self.assertTrue(a.all())
+#
+#     t = Test()
+#     _jr = jax.jacfwd(f1, argnums=(0, 1))(_x, _y)
+#     _aux = t(_y)[1]
+#     (var_grads, arg_grads), aux = bst.transform.jacfwd(t, grad_vars=t.x, argnums=0, has_aux=True)(_y)
+#     print(var_grads, )
+#     print(arg_grads, )
+#     self.assertTrue((var_grads == _jr[0]).all())
+#     self.assertTrue((arg_grads == _jr[1]).all())
+#     self.assertTrue(jnp.array_equal(aux, _aux))
+#
+#     jnp.disable_x64()
+#
+#   def test_jacrev_return_aux1(self):
+#     jnp.enable_x64()
+#
+#     def f1(x, y):
+#       r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+#       return r
+#
+#     _x = jnp.array([1., 2., 3.])
+#     _y = jnp.array([10., 5.])
+#
+#     class Test(bst.Module):
+#       def __init__(self):
+#         super(Test, self).__init__()
+#         self.x = jnp.Variable(jnp.array([1., 2., 3.]))
+#
+#       def __call__(self, y):
+#         a = self.x[0] * y[0]
+#         b = 5 * self.x[2] * y[1]
+#         c = 4 * self.x[1] ** 2 - 2 * self.x[2]
+#         d = self.x[2] * jnp.sin(self.x[0])
+#         r = jnp.asarray([a, b, c, d])
+#         return r, (c, d)
+#
+#     _jr = jax.jacrev(f1)(_x, _y)
+#     t = Test()
+#     br, _ = bst.transform.jacrev(t, grad_vars=t.x, has_aux=True)(_y)
+#     self.assertTrue((br == _jr).all())
+#
+#     t = Test()
+#     _jr = jax.jacrev(f1, argnums=(0, 1))(_x, _y)
+#     _val, _aux = t(_y)
+#     (var_grads, arg_grads), value, aux = bst.transform.jacrev(t, grad_vars=t.x, argnums=0, has_aux=True, return_value=True)(_y)
+#     print(var_grads, )
+#     print(arg_grads, )
+#     self.assertTrue((var_grads == _jr[0]).all())
+#     self.assertTrue((arg_grads == _jr[1]).all())
+#     self.assertTrue(jnp.array_equal(aux, _aux))
+#     self.assertTrue(jnp.array_equal(value, _val))
+#
+#     jnp.disable_x64()
+#
+#   def test_jacfwd_return_aux1(self):
+#     jnp.enable_x64()
+#
+#     def f1(x, y):
+#       r = jnp.asarray([x[0] * y[0], 5 * x[2] * y[1], 4 * x[1] ** 2 - 2 * x[2], x[2] * jnp.sin(x[0])])
+#       return r
+#
+#     _x = jnp.array([1., 2., 3.])
+#     _y = jnp.array([10., 5.])
+#
+#     class Test(bst.Module):
+#       def __init__(self):
+#         super(Test, self).__init__()
+#         self.x = jnp.Variable(jnp.array([1., 2., 3.]))
+#
+#       def __call__(self, y):
+#         a = self.x[0] * y[0]
+#         b = 5 * self.x[2] * y[1]
+#         c = 4 * self.x[1] ** 2 - 2 * self.x[2]
+#         d = self.x[2] * jnp.sin(self.x[0])
+#         r = jnp.asarray([a, b, c, d])
+#         return r, (c, d)
+#
+#     _jr = jax.jacfwd(f1)(_x, _y)
+#     t = Test()
+#     br, _ = bst.transform.jacfwd(t, grad_vars=t.x, has_aux=True)(_y)
+#     self.assertTrue((br == _jr).all())
+#
+#     t = Test()
+#     _jr = jax.jacfwd(f1, argnums=(0, 1))(_x, _y)
+#     _val, _aux = t(_y)
+#     (var_grads, arg_grads), value, aux = bst.transform.jacfwd(t, grad_vars=t.x, argnums=0, has_aux=True, return_value=True)(_y)
+#     print(_val, )
+#     print('_aux: ', _aux, 'aux: ', aux)
+#     print(var_grads, )
+#     print(arg_grads, )
+#     self.assertTrue((var_grads == _jr[0]).all())
+#     self.assertTrue((arg_grads == _jr[1]).all())
+#     self.assertTrue(jnp.array_equal(aux, _aux))
+#     self.assertTrue(jnp.array_equal(value, _val))
+#
+#     jnp.disable_x64()
+#
+#
+# class TestPureFuncVectorGrad(unittest.TestCase):
+#   def test1(self):
+#     f = lambda x: 3 * x ** 2
+#     _x = jnp.ones(10)
+#     pprint(bst.transform.vector_grad(f, argnums=0)(_x))
+#
+#   def test2(self):
+#     def f(x, y):
+#       dx = x ** 2 + y ** 2 + 10
+#       return dx
+#
+#     _x = jnp.ones(5)
+#     _y = jnp.ones(5)
+#
+#     g = bst.transform.vector_grad(f, argnums=0)(_x, _y)
+#     pprint(g)
+#     self.assertTrue(jnp.array_equal(g, 2 * _x))
+#
+#     g = bst.transform.vector_grad(f, argnums=(0,))(_x, _y)
+#     self.assertTrue(jnp.array_equal(g[0], 2 * _x))
+#
+#     g = bst.transform.vector_grad(f, argnums=(0, 1))(_x, _y)
+#     pprint(g)
+#     self.assertTrue(jnp.array_equal(g[0], 2 * _x))
+#     self.assertTrue(jnp.array_equal(g[1], 2 * _y))
+#
+#   def test3(self):
+#     def f(x, y):
+#       dx = x ** 2 + y ** 2 + 10
+#       dy = x ** 3 + y ** 3 - 10
+#       return dx, dy
+#
+#     _x = jnp.ones(5)
+#     _y = jnp.ones(5)
+#
+#     g = bst.transform.vector_grad(f, argnums=0)(_x, _y)
+#     # pprint(g)
+#     self.assertTrue(jnp.array_equal(g, 2 * _x + 3 * _x ** 2))
+#
+#     g = bst.transform.vector_grad(f, argnums=(0,))(_x, _y)
+#     self.assertTrue(jnp.array_equal(g[0], 2 * _x + 3 * _x ** 2))
+#
+#     g = bst.transform.vector_grad(f, argnums=(0, 1))(_x, _y)
+#     # pprint(g)
+#     self.assertTrue(jnp.array_equal(g[0], 2 * _x + 3 * _x ** 2))
+#     self.assertTrue(jnp.array_equal(g[1], 2 * _y + 3 * _y ** 2))
+#
+#   def test4_2d(self):
+#     def f(x, y):
+#       dx = x ** 2 + y ** 2 + 10
+#       return dx
+#
+#     _x = jnp.ones((5, 5))
+#     _y = jnp.ones((5, 5))
+#
+#     g = bst.transform.vector_grad(f, argnums=0)(_x, _y)
+#     pprint(g)
+#     self.assertTrue(jnp.array_equal(g, 2 * _x))
+#
+#     g = bst.transform.vector_grad(f, argnums=(0,))(_x, _y)
+#     self.assertTrue(jnp.array_equal(g[0], 2 * _x))
+#
+#     g = bst.transform.vector_grad(f, argnums=(0, 1))(_x, _y)
+#     pprint(g)
+#     self.assertTrue(jnp.array_equal(g[0], 2 * _x))
+#     self.assertTrue(jnp.array_equal(g[1], 2 * _y))
+#
+#   def test_aux1(self):
+#     def f(x, y):
+#       dx = x ** 2 + y ** 2 + 10
+#       dy = x ** 3 + y ** 3 - 10
+#       return dx, dy
+#
+#     _x = jnp.ones(5)
+#     _y = jnp.ones(5)
+#
+#     g, aux = bst.transform.vector_grad(f, has_aux=True)(_x, _y)
+#     pprint(g, )
+#     pprint(aux)
+#     self.assertTrue(jnp.array_equal(g, 2 * _x))
+#     self.assertTrue(jnp.array_equal(aux, _x ** 3 + _y ** 3 - 10))
+#
+#   def test_return1(self):
+#     def f(x, y):
+#       dx = x ** 2 + y ** 2 + 10
+#       return dx
+#
+#     _x = jnp.ones(5)
+#     _y = jnp.ones(5)
+#
+#     g, value = bst.transform.vector_grad(f, return_value=True)(_x, _y)
+#     pprint(g, )
+#     pprint(value)
+#     self.assertTrue(jnp.array_equal(g, 2 * _x))
+#     self.assertTrue(jnp.array_equal(value, _x ** 2 + _y ** 2 + 10))
+#
+#   def test_return_aux1(self):
+#     def f(x, y):
+#       dx = x ** 2 + y ** 2 + 10
+#       dy = x ** 3 + y ** 3 - 10
+#       return dx, dy
+#
+#     _x = jnp.ones(5)
+#     _y = jnp.ones(5)
+#
+#     g, value, aux = bst.transform.vector_grad(f, has_aux=True, return_value=True)(_x, _y)
+#     print('grad', g)
+#     print('value', value)
+#     print('aux', aux)
+#     self.assertTrue(jnp.array_equal(g, 2 * _x))
+#     self.assertTrue(jnp.array_equal(value, _x ** 2 + _y ** 2 + 10))
+#     self.assertTrue(jnp.array_equal(aux, _x ** 3 + _y ** 3 - 10))
+#
+#
+# class TestClassFuncVectorGrad(unittest.TestCase):
+#   def test1(self):
+#     class Test(bst.Module):
+#       def __init__(self):
+#         super(Test, self).__init__()
+#         self.x = jnp.Variable(jnp.ones(5))
+#         self.y = jnp.Variable(jnp.ones(5))
+#
+#       def __call__(self, *args, **kwargs):
+#         return self.x ** 2 + self.y ** 2 + 10
+#
+#     t = Test()
+#
+#     g = bst.transform.vector_grad(t, grad_vars=t.x)()
+#     self.assertTrue(jnp.array_equal(g, 2 * t.x))
+#
+#     g = bst.transform.vector_grad(t, grad_vars=(t.x,))()
+#     self.assertTrue(jnp.array_equal(g[0], 2 * t.x))
+#
+#     g = bst.transform.vector_grad(t, grad_vars=(t.x, t.y))()
+#     self.assertTrue(jnp.array_equal(g[0], 2 * t.x))
+#     self.assertTrue(jnp.array_equal(g[1], 2 * t.y))
+#
+#
+# def vgrad(f, *x):
+#   y, vjp_fn = jax.vjp(f, *x)
+#   return vjp_fn(jnp.ones(y.shape).value)[0]
+#
+#
+# class TestDebug(parameterized.TestCase):
+#   def test_debug1(self):
+#     a = bst.random.RandomState()
+#
+#     def f(b):
+#       print(a.value)
+#       return a + b + a.random()
+#
+#     f = bst.transform.vector_grad(f, argnums=0)
+#     f(1.)
+#
+#     with jax.disable_jit():
+#       f(1.)
+#
+#   @parameterized.product(
+#     grad_fun=[bst.transform.grad, bst.transform.vector_grad]
+#   )
+#   def test_print_info1(self, grad_fun):
+#     file = tempfile.TemporaryFile(mode='w+')
+#
+#     @functools.partial(grad_fun, argnums=0)
+#     def f2(a, b):
+#       print('compiling f2 ...', file=file)
+#       return a + b
+#
+#     @functools.partial(grad_fun, argnums=0)
+#     def f1(a):
+#       print('compiling f1 ...', file=file)
+#       return f2(a, 1.)
+#
+#     expect_res = '''
+# compiling f1 ...
+# compiling f2 ...
+# compiling f1 ...
+# compiling f2 ...
+#     '''
+#
+#     print(f1(1.))
+#     file.seek(0)
+#     self.assertTrue(file.read().strip() == expect_res.strip())
+#
+#     file = tempfile.TemporaryFile(mode='w+')
+#     with jax.disable_jit():
+#       expect_res = '''
+# compiling f1 ...
+# compiling f2 ...
+#       '''
+#       self.assertTrue(f1(1.) == 0.)
+#       file.seek(0)
+#       self.assertTrue(file.read().strip() == expect_res.strip())
+#
+#   @parameterized.product(
+#     grad_fun=[bst.transform.grad, bst.transform.vector_grad]
+#   )
+#   def test_print_info2(self, grad_fun):
+#     file = tempfile.TemporaryFile(mode='w+')
+#
+#     @functools.partial(grad_fun, argnums=0)
+#     def f1(a):
+#       @functools.partial(grad_fun, argnums=0)
+#       def f2(a, b):
+#         print('compiling f2 ...', file=file)
+#         return a + b
+#
+#       print('compiling f1 ...', file=file)
+#       return f2(a, 1.)
+#
+#     expect_res = '''
+# compiling f1 ...
+# compiling f2 ...
+# compiling f1 ...
+# compiling f2 ...
+# compiling f2 ...
+#     '''
+#     self.assertTrue(f1(1.) == 0.)
+#     file.seek(0)
+#     self.assertTrue(file.read().strip() == expect_res.strip())
+#
+#     file = tempfile.TemporaryFile(mode='w+')
+#     with jax.disable_jit():
+#       expect_res = '''
+# compiling f1 ...
+# compiling f2 ...
+#       '''
+#       self.assertTrue(f1(1.) == 0.)
+#       file.seek(0)
+#       # print(file.read().strip())
+#       self.assertTrue(file.read().strip() == expect_res.strip())
+#
+#   def test_debug_correctness1(self):
+#     def test_f():
+#       a = jnp.Variable(jnp.ones(2))
+#       b = jnp.Variable(jnp.zeros(2))
+#
+#       @bst.transform.vector_grad(argnums=0)
+#       def f1(c):
+#         a.value += 1
+#         b.value += 10
+#         return a * b * c
+#
+#       return a, b, f1(1.)
+#
+#     r1 = test_f()
+#     print(r1)
+#
+#     with jax.disable_jit():
+#       r2 = test_f()
+#       print(r2)
+#       self.assertTrue(jnp.allclose(r1[0], r2[0]))
+#       self.assertTrue(jnp.allclose(r1[1], r2[1]))
+#       self.assertTrue(jnp.allclose(r1[2], r2[2]))
+#
+#     def f1(c, a, b):
+#       a += 1
+#       b += 10
+#       return a * b * c
+#
+#     r3 = vgrad(f1, 1., jnp.ones(2).value, jnp.zeros(2).value)
+#     self.assertTrue(jnp.allclose(r1[2], r3))
+#
+#   def _bench_f2(self, dd):
+#     a = jnp.Variable(jnp.ones(2))
+#     b = jnp.Variable(jnp.zeros(2))
+#
+#
+#     def run_fun(d):
+#       @bst.transform.vector_grad(argnums=0)
+#       def f1(c):
+#         a.value += d
+#         b.value += 10
+#         return a * b * c
+#
+#       return a, b, f1(1.)
+#
+#     return run_fun(dd)
+#
+#   def test_debug_correctness2(self):
+#     r1 = self._bench_f2(1.)
+#     print(r1)
+#
+#     with jax.disable_jit():
+#       r2 = self._bench_f2(1.)
+#       print(r2)
+#
+#     self.assertTrue(jnp.allclose(r1[0], r2[0]))
+#     self.assertTrue(jnp.allclose(r1[1], r2[1]))
+#     self.assertTrue(jnp.allclose(r1[2], r2[2]))
+#
+#   def test_cache1(self):
+#       file = tempfile.TemporaryFile(mode='w+')
+#
+#       def f(a, b):
+#         print('compiling f ...', file=file)
+#         return a + b
+#
+#       grad1 = bst.transform.grad(f)(1., 2.)  # call "f" twice, one for Variable finding, one for compiling
+#       grad2 = bst.transform.vector_grad(f)(1., 2.)  # call "f" once for compiling
+#
+#       file.seek(0)
+#       print(file.read().strip())
+#
+#       expect_res = '''
+# compiling f ...
+# compiling f ...
+# compiling f ...
+#       '''
+#       file.seek(0)
+#       self.assertTrue(file.read().strip() == expect_res.strip())
+#
+#