author
int64
658
755k
date
stringlengths
19
19
timezone
int64
-46,800
43.2k
hash
stringlengths
40
40
message
stringlengths
5
490
mods
list
language
stringclasses
20 values
license
stringclasses
3 values
repo
stringlengths
5
68
original_message
stringlengths
12
491
260,411
11.04.2020 13:38:29
-7,200
c4dd1cf3fa76fd4ef6449c23efaa540b08030e20
Relax tolerance for new test on TPU
[ { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -1139,7 +1139,7 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nargs_maker = lambda: [rng(xshape, dtype), rng(yshape, dtype)]\nonp_fun = partial(onp_op, mode=mode)\njnp_fun = partial(jnp_op, mode=mode)\n- tol = {onp.float16: 1e-2}\n+ tol = {onp.float16: 1e-2} if jtu.device_under_test() != \"tpu\" else 0.5\nself._CheckAgainstNumpy(onp_fun, jnp_fun, args_maker, check_dtypes=False, tol=tol)\nself._CompileAndCheck(jnp_fun, args_maker, check_dtypes=True)\n" } ]
Python
Apache License 2.0
google/jax
Relax tolerance for new test on TPU
260,384
11.04.2020 20:54:04
0
98d46d39aa9138fbd8143411b99dc42ef0a36ad3
Implement indexing helpers
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -32,6 +32,7 @@ import itertools\nimport os\nimport re\nimport string\n+import textwrap\nimport types\nfrom typing import Callable\nimport warnings\n@@ -46,6 +47,7 @@ from ..abstract_arrays import UnshapedArray, ShapedArray, ConcreteArray\nfrom ..config import flags\nfrom ..interpreters.xla import DeviceArray\nfrom .. import lax\n+from .. import ops\nfrom ..util import partial, get_module_functions, unzip2, prod as _prod, subvals\nfrom ..lib import pytree\nfrom ..lib import xla_client\n@@ -3688,3 +3690,63 @@ def _unstack(x):\nraise ValueError(\"Argument to _unstack must be non-scalar\")\nreturn [lax.index_in_dim(x, i, keepdims=False) for i in range(x.shape[0])]\nsetattr(DeviceArray, \"_unstack\", _unstack)\n+\n+\n+# Syntactic sugar for scatter operations\n+class IndexUpdateHelper(object):\n+ \"\"\"Helper object to call indexed update functions.\n+\n+ This makes it possible to do :code:`x.update[idx](y)` instead of\n+ :code:`jax.ops.index_update(x, jax.ops.index[idx], y)`.\n+ \"\"\"\n+ __slots__ = (\"index_method\", \"array\")\n+\n+ def __init__(self, index_method, array):\n+ self.index_method = index_method\n+ self.array = array\n+\n+ def __getitem__(self, index):\n+ return partial(self.index_method, self.array, index)\n+\n+ def __repr__(self):\n+ return f\"IndexUpdateHelper({self.index_method}, {self.array})\"\n+\n+\n+def _make_index_update_property(property_name, numpy_equiv, method_fn):\n+ fn = partial(IndexUpdateHelper, method_fn)\n+ fn.__doc__ = textwrap.dedent(f\"\"\"\\\n+ Pure equivalent of :code:`{numpy_equiv}`.\n+\n+ :code:`x.{property_name}[idx](y)` is syntactic sugar for\n+ :code:`jax.ops.{method_fn.__name__}(x, jax.ops.index[idx], y)`, and\n+ returns the value of `x` that would result from the NumPy-style\n+ :mod:`indexed assignment <numpy.doc.indexing>` :code:`{numpy_equiv}`.\n+\n+ See :mod:`jax.ops` for details.\"\"\")\n+ return fn\n+\n+\n+_update_helper = _make_index_update_property(\"update\", \"x[idx] = y\",\n+ ops.index_update)\n+setattr(DeviceArray, \"update\", property(_update_helper))\n+setattr(ShapedArray, \"update\", core.aval_property(_update_helper))\n+\n+_add_helper = _make_index_update_property(\"add\", \"x[idx] += y\", ops.index_add)\n+setattr(DeviceArray, \"add\", property(_add_helper))\n+setattr(ShapedArray, \"add\", core.aval_property(_add_helper))\n+\n+_update_add_helper = _make_index_update_property(\"update_add\", \"x[idx] += y\",\n+ ops.index_add)\n+setattr(DeviceArray, \"update_add\", property(_update_add_helper))\n+setattr(ShapedArray, \"update_add\", core.aval_property(_update_add_helper))\n+\n+_update_min_helper = _make_index_update_property(\n+ \"update_min\", \"x[idx] = minimum(x[idx], y)\", ops.index_min)\n+setattr(DeviceArray, \"update_min\", property(_update_min_helper))\n+setattr(ShapedArray, \"update_min\", core.aval_property(_update_min_helper))\n+\n+_update_max_helper = _make_index_update_property(\n+ \"update_max\", \"x[idx] = maximum(x[idx], y)\", ops.index_max)\n+setattr(DeviceArray, \"update_max\", property(_update_max_helper))\n+setattr(ShapedArray, \"update_max\", core.aval_property(_update_max_helper))\n+\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_indexing_test.py", "new_path": "tests/lax_numpy_indexing_test.py", "diff": "@@ -831,51 +831,68 @@ class UpdateOps(enum.Enum):\nUpdateOps.MAX: ops.index_max,\n}[op](x, indexer, y)\n+ def sugar_fn(op, indexer, x, y):\n+ x = jnp.array(x)\n+ return {\n+ UpdateOps.UPDATE: x.update,\n+ UpdateOps.ADD: x.update_add,\n+ UpdateOps.MIN: x.update_min,\n+ UpdateOps.MAX: x.update_max,\n+ }[op][indexer](y)\n+\nclass IndexedUpdateTest(jtu.JaxTestCase):\n@parameterized.named_parameters(jtu.cases_from_list({\n- \"testcase_name\": \"{}_inshape={}_indexer={}_update={}_op={}\".format(\n+ \"testcase_name\": \"{}_inshape={}_indexer={}_update={}_sugared={}_op={}\".format(\nname, jtu.format_shape_dtype_string(shape, dtype), indexer,\n- jtu.format_shape_dtype_string(update_shape, update_dtype), op.name),\n+ jtu.format_shape_dtype_string(update_shape, update_dtype), sugared, op.name),\n\"shape\": shape, \"dtype\": dtype, \"rng_factory\": rng_factory, \"indexer\": indexer,\n\"update_shape\": update_shape, \"update_dtype\": update_dtype,\n- \"op\": op\n+ \"op\": op, \"sugared\": sugared\n} for name, index_specs in STATIC_INDEXING_TESTS\nfor shape, indexer in index_specs\nfor op in UpdateOps\nfor dtype in (all_dtypes if op == UpdateOps.UPDATE else default_dtypes)\nfor update_shape in _broadcastable_shapes(_update_shape(shape, indexer))\nfor update_dtype in ([dtype] if op == UpdateOps.ADD else all_dtypes)\n+ for sugared in [True, False]\nfor rng_factory in [jtu.rand_default]))\ndef testStaticIndexing(self, shape, dtype, update_shape, update_dtype,\n- rng_factory, indexer, op):\n+ rng_factory, indexer, sugared, op):\nrng = rng_factory()\nargs_maker = lambda: [rng(shape, dtype), rng(update_shape, update_dtype)]\nonp_fn = lambda x, y: UpdateOps.onp_fn(op, indexer, x, y)\n+ if sugared:\n+ jax_fn = lambda x, y: UpdateOps.sugar_fn(op, indexer, x, y)\n+ else:\njax_fn = lambda x, y: UpdateOps.jax_fn(op, indexer, x, y)\nself._CheckAgainstNumpy(onp_fn, jax_fn, args_maker, check_dtypes=True)\nself._CompileAndCheck(jax_fn, args_maker, check_dtypes=True)\n@parameterized.named_parameters(jtu.cases_from_list({\n- \"testcase_name\": \"{}_inshape={}_indexer={}_update={}_op={}\".format(\n+ \"testcase_name\": \"{}_inshape={}_indexer={}_update={}_sugared={}_op={}\".format(\nname, jtu.format_shape_dtype_string(shape, dtype), indexer,\n- jtu.format_shape_dtype_string(update_shape, update_dtype), op.name),\n+ jtu.format_shape_dtype_string(update_shape, update_dtype), sugared, op.name),\n\"shape\": shape, \"dtype\": dtype, \"rng_factory\": rng_factory, \"indexer\": indexer,\n\"update_shape\": update_shape, \"update_dtype\": update_dtype,\n- \"op\": op\n+ \"op\": op, \"sugared\": sugared\n} for name, index_specs in ADVANCED_INDEXING_TESTS_NO_REPEATS\nfor shape, indexer in index_specs\nfor op in UpdateOps\nfor dtype in (all_dtypes if op == UpdateOps.UPDATE else default_dtypes)\nfor update_shape in _broadcastable_shapes(_update_shape(shape, indexer))\nfor update_dtype in ([dtype] if op == UpdateOps.ADD else all_dtypes)\n+ for sugared in [True, False]\nfor rng_factory in [jtu.rand_default]))\ndef testAdvancedIndexing(self, shape, dtype, update_shape, update_dtype,\n- rng_factory, indexer, op):\n+ rng_factory, indexer, sugared, op):\nrng = rng_factory()\nargs_maker = lambda: [rng(shape, dtype), rng(update_shape, update_dtype)]\nonp_fn = lambda x, y: UpdateOps.onp_fn(op, indexer, x, y)\n+ if sugared:\n+ jax_fn = lambda x, y: UpdateOps.sugar_fn(op, indexer, x, y)\n+ else:\njax_fn = lambda x, y: UpdateOps.jax_fn(op, indexer, x, y)\nself._CheckAgainstNumpy(onp_fn, jax_fn, args_maker, check_dtypes=True)\nself._CompileAndCheck(jax_fn, args_maker, check_dtypes=True)\n@@ -886,19 +903,23 @@ class IndexedUpdateTest(jtu.JaxTestCase):\njtu.format_shape_dtype_string(update_shape, update_dtype), op.name),\n\"shape\": shape, \"dtype\": dtype, \"rng_factory\": rng_factory, \"indexer\": indexer,\n\"update_shape\": update_shape, \"update_dtype\": update_dtype,\n- \"op\": op\n+ \"op\": op, \"sugared\": sugared\n} for name, index_specs in MIXED_ADVANCED_INDEXING_TESTS_NO_REPEATS\nfor shape, indexer in index_specs\nfor op in UpdateOps\nfor dtype in (all_dtypes if op == UpdateOps.UPDATE else default_dtypes)\nfor update_shape in _broadcastable_shapes(_update_shape(shape, indexer))\nfor update_dtype in ([dtype] if op == UpdateOps.ADD else all_dtypes)\n+ for sugared in [True, False]\nfor rng_factory in [jtu.rand_default]))\ndef testMixedAdvancedIndexing(self, shape, dtype, update_shape, update_dtype,\n- rng_factory, indexer, op):\n+ rng_factory, indexer, sugared, op):\nrng = rng_factory()\nargs_maker = lambda: [rng(shape, dtype), rng(update_shape, update_dtype)]\nonp_fn = lambda x, y: UpdateOps.onp_fn(op, indexer, x, y)\n+ if sugared:\n+ jax_fn = lambda x, y: UpdateOps.sugar_fn(op, indexer, x, y)\n+ else:\njax_fn = lambda x, y: UpdateOps.jax_fn(op, indexer, x, y)\nself._CheckAgainstNumpy(onp_fn, jax_fn, args_maker, check_dtypes=True)\nself._CompileAndCheck(jax_fn, args_maker, check_dtypes=True)\n" } ]
Python
Apache License 2.0
google/jax
Implement indexing helpers
260,411
12.04.2020 19:05:04
-7,200
453dc5f0857477fa672537124b32140f3610fe21
Relax tolerance for LaxBackedNumpyTests.testConvolutions
[ { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -1139,7 +1139,7 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nargs_maker = lambda: [rng(xshape, dtype), rng(yshape, dtype)]\nonp_fun = partial(onp_op, mode=mode)\njnp_fun = partial(jnp_op, mode=mode)\n- tol = {onp.float16: 1e-2} if jtu.device_under_test() != \"tpu\" else 0.5\n+ tol = 1e-2 if jtu.device_under_test() != \"tpu\" else 0.5\nself._CheckAgainstNumpy(onp_fun, jnp_fun, args_maker, check_dtypes=False, tol=tol)\nself._CompileAndCheck(jnp_fun, args_maker, check_dtypes=True)\n" } ]
Python
Apache License 2.0
google/jax
Relax tolerance for LaxBackedNumpyTests.testConvolutions
260,335
13.04.2020 09:44:13
25,200
2d25773c212589b81fa7d6bbb61a43bf9970f3e3
add custom_jvp for logaddexp / logaddexp2 fixes draws from and thanks !
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/ad.py", "new_path": "jax/interpreters/ad.py", "diff": "@@ -143,8 +143,11 @@ def backward_pass(jaxpr: core.Jaxpr, consts, args, cotangents_in):\ndef write_cotangent(v, ct):\n# assert v not in primal_env\n- if ct is not None and type(v) is not Literal:\n+ if ct is not None and type(v) is not Literal and ct is not zero:\nct_env[v] = add_tangents(ct_env[v], ct) if v in ct_env else ct\n+ if not core.skip_checks:\n+ ct_aval = core.get_aval(ct_env[v])\n+ assert v.aval == core.lattice_join(v.aval, ct_aval)\ndef read_cotangent(v):\nreturn ct_env.get(v, zero)\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/__init__.py", "new_path": "jax/lax/__init__.py", "diff": "@@ -20,7 +20,7 @@ from .lax import (_reduce_sum, _reduce_max, _reduce_min, _reduce_or,\n_const, _eq_meet, _broadcasting_select,\n_check_user_dtype_supported, _one, _const,\n_upcast_fp16_for_computation, _broadcasting_shape_rule,\n- _eye, _tri, _delta)\n+ _eye, _tri, _delta, _ones, _zeros)\nfrom .lax_control_flow import *\nfrom .lax_fft import *\nfrom .lax_parallel import *\n" }, { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -39,7 +39,7 @@ import warnings\nimport numpy as onp\nimport opt_einsum\n-from jax import jit, device_put\n+from jax import jit, device_put, custom_jvp\nfrom .. import core\nfrom .. import dtypes\nfrom ..abstract_arrays import UnshapedArray, ShapedArray, ConcreteArray\n@@ -583,6 +583,7 @@ def power(x1, x2):\nreturn acc\n+@custom_jvp\n@_wraps(onp.logaddexp)\ndef logaddexp(x1, x2):\nx1, x2 = _promote_shapes(\"logaddexp\", *_promote_dtypes_inexact(x1, x2))\n@@ -592,7 +593,21 @@ def logaddexp(x1, x2):\nlax.add(x1, x2), # NaNs or infinities of the same sign.\nlax.add(amax, lax.log1p(lax.exp(-lax.abs(delta)))))\n+@logaddexp.defjvp\n+def _logaddexp_jvp(primals, tangents):\n+ x1, x2 = primals\n+ t1, t2 = tangents\n+ x1, x2, t1, t2 = broadcast_arrays(x1, x2, t1, t2)\n+ primal_out = logaddexp(x1, x2)\n+ tangent_out = (t1 * exp(_replace_inf(x1) - _replace_inf(primal_out)) +\n+ t2 * exp(_replace_inf(x2) - _replace_inf(primal_out)))\n+ return primal_out, tangent_out\n+def _replace_inf(x):\n+ return lax.select(isposinf(x), zeros_like(x), x)\n+\n+\n+@custom_jvp\n@_wraps(onp.logaddexp2)\ndef logaddexp2(x1, x2):\nx1, x2 = _promote_shapes(\"logaddexp2\", *_promote_dtypes_inexact(x1, x2))\n@@ -602,6 +617,15 @@ def logaddexp2(x1, x2):\nlax.add(x1, x2), # NaNs or infinities of the same sign.\nlax.add(amax, lax.div(lax.log1p(exp2(-lax.abs(delta))),\n_constant_like(x1, onp.log(2)))))\n+@logaddexp2.defjvp\n+def _logaddexp2_jvp(primals, tangents):\n+ x1, x2 = primals\n+ t1, t2 = tangents\n+ x1, x2, t1, t2 = broadcast_arrays(x1, x2, t1, t2)\n+ primal_out = logaddexp2(x1, x2)\n+ tangent_out = (t1 * 2 ** (_replace_inf(x1) - _replace_inf(primal_out)) +\n+ t2 * 2 ** (_replace_inf(x2) - _replace_inf(primal_out)))\n+ return primal_out, tangent_out\n@_wraps(onp.log2)\n" }, { "change_type": "MODIFY", "old_path": "jax/test_util.py", "new_path": "jax/test_util.py", "diff": "@@ -696,8 +696,10 @@ def cases_from_gens(*gens):\nclass JaxTestCase(parameterized.TestCase):\n\"\"\"Base class for JAX tests including numerical checks and boilerplate.\"\"\"\n- def tearDown(self) -> None:\n- assert core.reset_trace_state()\n+ # TODO(mattjj): this obscures the error messages from failures, figure out how\n+ # to re-enable it\n+ # def tearDown(self) -> None:\n+ # assert core.reset_trace_state()\ndef assertArraysAllClose(self, x, y, check_dtypes, atol=None, rtol=None):\n\"\"\"Assert that x and y are close (up to numerical tolerances).\"\"\"\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -2966,6 +2966,12 @@ GRAD_TEST_RECORDS = [\ngrad_test_spec(jnp.arctanh, nargs=1, order=2,\nrng_factory=partial(jtu.rand_uniform, -0.9, 0.9),\ndtypes=[onp.float64, onp.complex64], tol=1e-4),\n+ grad_test_spec(jnp.logaddexp, nargs=2, order=1,\n+ rng_factory=partial(jtu.rand_uniform, -0.9, 0.9),\n+ dtypes=[onp.float64], tol=1e-4),\n+ grad_test_spec(jnp.logaddexp2, nargs=2, order=2,\n+ rng_factory=partial(jtu.rand_uniform, -0.9, 0.9),\n+ dtypes=[onp.float64], tol=1e-4),\n]\nGradSpecialValuesTestSpec = collections.namedtuple(\n@@ -2975,7 +2981,7 @@ GRAD_SPECIAL_VALUE_TEST_RECORDS = [\nGradSpecialValuesTestSpec(jnp.arcsinh, [0., 1000.], 2),\nGradSpecialValuesTestSpec(jnp.arccosh, [1000.], 2),\nGradSpecialValuesTestSpec(jnp.arctanh, [0.], 2),\n- GradSpecialValuesTestSpec(jnp.sinc, [0.], 1)\n+ GradSpecialValuesTestSpec(jnp.sinc, [0.], 1),\n]\ndef num_float_bits(dtype):\n" }, { "change_type": "MODIFY", "old_path": "tests/nn_test.py", "new_path": "tests/nn_test.py", "diff": "@@ -39,6 +39,27 @@ class NNFunctionsTest(jtu.JaxTestCase):\ncheck_grads(nn.softplus, (1e-8,), order=4,\nrtol=1e-2 if jtu.device_under_test() == \"tpu\" else None)\n+ def testSoftplusGradZero(self):\n+ check_grads(nn.softplus, (0.,), order=1,\n+ rtol=1e-2 if jtu.device_under_test() == \"tpu\" else None)\n+\n+ def testSoftplusGradInf(self):\n+ self.assertAllClose(\n+ 1., jax.grad(nn.softplus)(float('inf')), check_dtypes=True)\n+\n+ def testSoftplusGradNegInf(self):\n+ check_grads(nn.softplus, (-float('inf'),), order=1,\n+ rtol=1e-2 if jtu.device_under_test() == \"tpu\" else None)\n+\n+ def testSoftplusGradNan(self):\n+ check_grads(nn.softplus, (float('nan'),), order=1,\n+ rtol=1e-2 if jtu.device_under_test() == \"tpu\" else None)\n+\n+ @parameterized.parameters([\n+ int, np.int32, float, np.float64, np.float32, np.float64,])\n+ def testSoftplusZero(self, dtype):\n+ self.assertEqual(np.log(dtype(2)), nn.softplus(dtype(0)))\n+\ndef testReluGrad(self):\nrtol = 1e-2 if jtu.device_under_test() == \"tpu\" else None\ncheck_grads(nn.relu, (1.,), order=3, rtol=rtol)\n" } ]
Python
Apache License 2.0
google/jax
add custom_jvp for logaddexp / logaddexp2 fixes #2107, draws from #2356 and #2357, thanks @yingted ! Co-authored-by: Ted Ying <yingted@gmail.com>
260,467
14.04.2020 03:26:30
-7,200
d6ab70c315067bf9a08b1071486715b515188a7c
Fix minor typo in cell * Fix minor typo in cell One of the arguments to `hvp` wasn't being used, which made the example slightly confusing. * Fix both definitions of hvp in the autodiff cookbook.
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/autodiff_cookbook.ipynb", "new_path": "docs/notebooks/autodiff_cookbook.ipynb", "diff": "\"outputs\": [],\n\"source\": [\n\"def hvp(f, x, v):\\n\",\n- \" return grad(lambda x: np.vdot(grad(f)(x), v))\"\n+ \" return grad(lambda x: np.vdot(grad(f)(x), v))(x)\"\n]\n},\n{\n\"outputs\": [],\n\"source\": [\n\"def hvp(f, x, v):\\n\",\n- \" return grad(lambda x: np.vdot(grad(f)(x), v))\"\n+ \" return grad(lambda x: np.vdot(grad(f)(x), v))(x)\"\n]\n},\n{\n" } ]
Python
Apache License 2.0
google/jax
Fix minor typo in cell (#2692) * Fix minor typo in cell One of the arguments to `hvp` wasn't being used, which made the example slightly confusing. * Fix both definitions of hvp in the autodiff cookbook. Co-authored-by: Peter Hawkins <phawkins@google.com>
260,384
15.04.2020 01:20:30
0
72593f0489315cefe685f36f6c89ae8b81dfc860
Modify syntax to `x.at[idx].set(y)` and similar.
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -3692,61 +3692,100 @@ def _unstack(x):\nsetattr(DeviceArray, \"_unstack\", _unstack)\n-# Syntactic sugar for scatter operations\n+# Syntactic sugar for scatter operations.\nclass IndexUpdateHelper(object):\n- \"\"\"Helper object to call indexed update functions.\n-\n- This makes it possible to do :code:`x.update[idx](y)` instead of\n- :code:`jax.ops.index_update(x, jax.ops.index[idx], y)`.\n+ # Note: this docstring will appear as the docstring for the `at` property.\n+ \"\"\"Indexable helper object to call indexed update functions.\n+\n+ The `at` property is syntactic sugar for calling the indexed update functions\n+ defined in :mod:`jax.ops`, and acts as a pure equivalent of in-place\n+ modificatons.\n+\n+ In particular:\n+ - :code:`x = x.at[idx].set(y)` is a pure equivalent of :code:`x[idx] = y`.\n+ - :code:`x = x.at[idx].add(y)` is a pure equivalent of :code:`x[idx] += y`.\n+ - :code:`x = x.at[idx].min(y)` is a pure equivalent of\n+ :code:`x[idx] = minimum(x[idx], y)`.\n+ - :code:`x = x.at[idx].max(y)` is a pure equivalent of\n+ :code:`x[idx] = maximum(x[idx], y)`.\n\"\"\"\n- __slots__ = (\"index_method\", \"array\")\n+ __slots__ = (\"array\",)\n- def __init__(self, index_method, array):\n- self.index_method = index_method\n+ def __init__(self, array):\nself.array = array\ndef __getitem__(self, index):\n- return partial(self.index_method, self.array, index)\n+ return IndexUpdateRef(self.array, index)\n+\n+ def __repr__(self):\n+ return f\"IndexUpdateHelper({repr(self.array)})\"\n+\n+\n+class IndexUpdateRef(object):\n+ \"\"\"Helper object to call indexed update functions for an (advanced) index.\n+\n+ This object references a source array and a specific indexer into that array.\n+ Methods on this object return copies of the source array that have been\n+ modified at the positions specified by the indexer.\n+ \"\"\"\n+ __slots__ = (\"array\", \"index\")\n+\n+ def __init__(self, array, index):\n+ self.array = array\n+ self.index = index\ndef __repr__(self):\n- return f\"IndexUpdateHelper({self.index_method}, {self.array})\"\n+ return f\"IndexUpdateRef({repr(self.array)}, {repr(self.index)})\"\n+ def set(self, values):\n+ \"\"\"Pure equivalent of :code:`x[idx] = y`.\n-def _make_index_update_property(property_name, numpy_equiv, method_fn):\n- fn = partial(IndexUpdateHelper, method_fn)\n- fn.__doc__ = textwrap.dedent(f\"\"\"\\\n- Pure equivalent of :code:`{numpy_equiv}`.\n+ :code:`x.at[idx].set(y)` is syntactic sugar for\n+ :code:`jax.ops.index_update(x, jax.ops.index[idx], y)`, and\n+ returns the value of `x` that would result from the NumPy-style\n+ :mod:`indexed assignment <numpy.doc.indexing>` :code:`x[idx] = y`.\n+\n+ See :mod:`jax.ops` for details.\n+ \"\"\"\n+ return ops.index_update(self.array, self.index, values)\n- :code:`x.{property_name}[idx](y)` is syntactic sugar for\n- :code:`jax.ops.{method_fn.__name__}(x, jax.ops.index[idx], y)`, and\n+ def add(self, values):\n+ \"\"\"Pure equivalent of :code:`x[idx] += y`.\n+\n+ :code:`x.at[idx].add(y)` is syntactic sugar for\n+ :code:`jax.ops.index_add(x, jax.ops.index[idx], y)`, and\nreturns the value of `x` that would result from the NumPy-style\n- :mod:`indexed assignment <numpy.doc.indexing>` :code:`{numpy_equiv}`.\n+ :mod:`indexed assignment <numpy.doc.indexing>` :code:`x[idx] += y`.\n- See :mod:`jax.ops` for details.\"\"\")\n- return fn\n+ See :mod:`jax.ops` for details.\n+ \"\"\"\n+ return ops.index_add(self.array, self.index, values)\n+ def min(self, values):\n+ \"\"\"Pure equivalent of :code:`x[idx] = minimum(x[idx], y)`.\n-_update_helper = _make_index_update_property(\"update\", \"x[idx] = y\",\n- ops.index_update)\n-setattr(DeviceArray, \"update\", property(_update_helper))\n-setattr(ShapedArray, \"update\", core.aval_property(_update_helper))\n+ :code:`x.at[idx].min(y)` is syntactic sugar for\n+ :code:`jax.ops.index_min(x, jax.ops.index[idx], y)`, and\n+ returns the value of `x` that would result from the NumPy-style\n+ :mod:`indexed assignment <numpy.doc.indexing>`\n+ :code:`x[idx] = minimum(x[idx], y)`.\n-_add_helper = _make_index_update_property(\"add\", \"x[idx] += y\", ops.index_add)\n-setattr(DeviceArray, \"add\", property(_add_helper))\n-setattr(ShapedArray, \"add\", core.aval_property(_add_helper))\n+ See :mod:`jax.ops` for details.\n+ \"\"\"\n+ return ops.index_min(self.array, self.index, values)\n-_update_add_helper = _make_index_update_property(\"update_add\", \"x[idx] += y\",\n- ops.index_add)\n-setattr(DeviceArray, \"update_add\", property(_update_add_helper))\n-setattr(ShapedArray, \"update_add\", core.aval_property(_update_add_helper))\n+ def max(self, values):\n+ \"\"\"Pure equivalent of :code:`x[idx] = maximum(x[idx], y)`.\n-_update_min_helper = _make_index_update_property(\n- \"update_min\", \"x[idx] = minimum(x[idx], y)\", ops.index_min)\n-setattr(DeviceArray, \"update_min\", property(_update_min_helper))\n-setattr(ShapedArray, \"update_min\", core.aval_property(_update_min_helper))\n+ :code:`x.at[idx].max(y)` is syntactic sugar for\n+ :code:`jax.ops.index_max(x, jax.ops.index[idx], y)`, and\n+ returns the value of `x` that would result from the NumPy-style\n+ :mod:`indexed assignment <numpy.doc.indexing>`\n+ :code:`x[idx] = maximum(x[idx], y)`.\n-_update_max_helper = _make_index_update_property(\n- \"update_max\", \"x[idx] = maximum(x[idx], y)\", ops.index_max)\n-setattr(DeviceArray, \"update_max\", property(_update_max_helper))\n-setattr(ShapedArray, \"update_max\", core.aval_property(_update_max_helper))\n+ See :mod:`jax.ops` for details.\n+ \"\"\"\n+ return ops.index_max(self.array, self.index, values)\n+setattr(DeviceArray, \"at\", property(IndexUpdateHelper))\n+setattr(ShapedArray, \"at\", core.aval_property(IndexUpdateHelper))\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_indexing_test.py", "new_path": "tests/lax_numpy_indexing_test.py", "diff": "@@ -834,11 +834,11 @@ class UpdateOps(enum.Enum):\ndef sugar_fn(op, indexer, x, y):\nx = jnp.array(x)\nreturn {\n- UpdateOps.UPDATE: x.update,\n- UpdateOps.ADD: x.update_add,\n- UpdateOps.MIN: x.update_min,\n- UpdateOps.MAX: x.update_max,\n- }[op][indexer](y)\n+ UpdateOps.UPDATE: x.at[indexer].set,\n+ UpdateOps.ADD: x.at[indexer].add,\n+ UpdateOps.MIN: x.at[indexer].min,\n+ UpdateOps.MAX: x.at[indexer].max,\n+ }[op](y)\nclass IndexedUpdateTest(jtu.JaxTestCase):\n" } ]
Python
Apache License 2.0
google/jax
Modify syntax to `x.at[idx].set(y)` and similar.
260,384
15.04.2020 01:28:43
0
9e429907c1236e600473d2ff31008cb81e9d1395
Add support for `mul`
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -3803,6 +3803,7 @@ class IndexUpdateHelper(object):\nIn particular:\n- :code:`x = x.at[idx].set(y)` is a pure equivalent of :code:`x[idx] = y`.\n- :code:`x = x.at[idx].add(y)` is a pure equivalent of :code:`x[idx] += y`.\n+ - :code:`x = x.at[idx].mul(y)` is a pure equivalent of :code:`x[idx] *= y`.\n- :code:`x = x.at[idx].min(y)` is a pure equivalent of\n:code:`x[idx] = minimum(x[idx], y)`.\n- :code:`x = x.at[idx].max(y)` is a pure equivalent of\n@@ -3860,6 +3861,18 @@ class IndexUpdateRef(object):\n\"\"\"\nreturn ops.index_add(self.array, self.index, values)\n+ def mul(self, values):\n+ \"\"\"Pure equivalent of :code:`x[idx] += y`.\n+\n+ :code:`x.at[idx].mul(y)` is syntactic sugar for\n+ :code:`jax.ops.index_mul(x, jax.ops.index[idx], y)`, and\n+ returns the value of `x` that would result from the NumPy-style\n+ :mod:`indexed assignment <numpy.doc.indexing>` :code:`x[idx] *= y`.\n+\n+ See :mod:`jax.ops` for details.\n+ \"\"\"\n+ return ops.index_mul(self.array, self.index, values)\n+\ndef min(self, values):\n\"\"\"Pure equivalent of :code:`x[idx] = minimum(x[idx], y)`.\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_indexing_test.py", "new_path": "tests/lax_numpy_indexing_test.py", "diff": "@@ -846,6 +846,7 @@ class UpdateOps(enum.Enum):\nreturn {\nUpdateOps.UPDATE: x.at[indexer].set,\nUpdateOps.ADD: x.at[indexer].add,\n+ UpdateOps.MUL: x.at[indexer].mul,\nUpdateOps.MIN: x.at[indexer].min,\nUpdateOps.MAX: x.at[indexer].max,\n}[op](y)\n" } ]
Python
Apache License 2.0
google/jax
Add support for `mul`
260,384
17.04.2020 01:30:06
0
a5efe842562af81cae4df7aca95e52981dea2802
Update names and documentation.
[ { "change_type": "MODIFY", "old_path": "docs/jax.ops.rst", "new_path": "docs/jax.ops.rst", "diff": "@@ -24,6 +24,27 @@ pure alternatives, namely :func:`jax.ops.index_update` and its relatives.\nindex_min\nindex_max\n+\n+Syntactic sugar for indexed update operators\n+--------------------------------------------\n+\n+JAX also provides an alternate syntax for these indexed update operators.\n+Specifically, JAX ndarray types have a property ``at``, which can be used as\n+follows (where ``idx`` can be an arbitrary index expression).\n+\n+==================== ===================================================\n+Alternate syntax Equivalent expression\n+==================== ===================================================\n+``x.at[idx].set(y)`` ``jax.ops.index_update(x, jax.ops_index[idx], y)``\n+``x.at[idx].add(y)`` ``jax.ops.index_add(x, jax.ops_index[idx], y)``\n+``x.at[idx].mul(y)`` ``jax.ops.index_mul(x, jax.ops_index[idx], y)``\n+``x.at[idx].min(y)`` ``jax.ops.index_min(x, jax.ops_index[idx], y)``\n+``x.at[idx].max(y)`` ``jax.ops.index_max(x, jax.ops_index[idx], y)``\n+==================== ===================================================\n+\n+Note that none of these expressions modify the original `x`; instead they return\n+a modified copy of `x`.\n+\nOther operators\n---------------\n" }, { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -3791,7 +3791,7 @@ setattr(DeviceArray, \"_unstack\", _unstack)\n# Syntactic sugar for scatter operations.\n-class IndexUpdateHelper(object):\n+class _IndexUpdateHelper:\n# Note: this docstring will appear as the docstring for the `at` property.\n\"\"\"Indexable helper object to call indexed update functions.\n@@ -3800,13 +3800,13 @@ class IndexUpdateHelper(object):\nmodificatons.\nIn particular:\n- - :code:`x = x.at[idx].set(y)` is a pure equivalent of :code:`x[idx] = y`.\n- - :code:`x = x.at[idx].add(y)` is a pure equivalent of :code:`x[idx] += y`.\n- - :code:`x = x.at[idx].mul(y)` is a pure equivalent of :code:`x[idx] *= y`.\n- - :code:`x = x.at[idx].min(y)` is a pure equivalent of\n- :code:`x[idx] = minimum(x[idx], y)`.\n- - :code:`x = x.at[idx].max(y)` is a pure equivalent of\n- :code:`x[idx] = maximum(x[idx], y)`.\n+ - ``x = x.at[idx].set(y)`` is a pure equivalent of ``x[idx] = y``.\n+ - ``x = x.at[idx].add(y)`` is a pure equivalent of ``x[idx] += y``.\n+ - ``x = x.at[idx].mul(y)`` is a pure equivalent of ``x[idx] *= y``.\n+ - ``x = x.at[idx].min(y)`` is a pure equivalent of\n+ ``x[idx] = minimum(x[idx], y)``.\n+ - ``x = x.at[idx].max(y)`` is a pure equivalent of\n+ ``x[idx] = maximum(x[idx], y)``.\n\"\"\"\n__slots__ = (\"array\",)\n@@ -3814,13 +3814,13 @@ class IndexUpdateHelper(object):\nself.array = array\ndef __getitem__(self, index):\n- return IndexUpdateRef(self.array, index)\n+ return _IndexUpdateRef(self.array, index)\ndef __repr__(self):\n- return f\"IndexUpdateHelper({repr(self.array)})\"\n+ return f\"_IndexUpdateHelper({repr(self.array)})\"\n-class IndexUpdateRef(object):\n+class _IndexUpdateRef:\n\"\"\"Helper object to call indexed update functions for an (advanced) index.\nThis object references a source array and a specific indexer into that array.\n@@ -3834,69 +3834,69 @@ class IndexUpdateRef(object):\nself.index = index\ndef __repr__(self):\n- return f\"IndexUpdateRef({repr(self.array)}, {repr(self.index)})\"\n+ return f\"_IndexUpdateRef({repr(self.array)}, {repr(self.index)})\"\ndef set(self, values):\n- \"\"\"Pure equivalent of :code:`x[idx] = y`.\n+ \"\"\"Pure equivalent of ``x[idx] = y``.\n- :code:`x.at[idx].set(y)` is syntactic sugar for\n- :code:`jax.ops.index_update(x, jax.ops.index[idx], y)`, and\n- returns the value of `x` that would result from the NumPy-style\n- :mod:`indexed assignment <numpy.doc.indexing>` :code:`x[idx] = y`.\n+ ``x.at[idx].set(y)`` is syntactic sugar for\n+ ``jax.ops.index_update(x, jax.ops.index[idx], y)``, and\n+ returns the value of ``x`` that would result from the NumPy-style\n+ :mod:indexed assignment <numpy.doc.indexing>` ``x[idx] = y``.\nSee :mod:`jax.ops` for details.\n\"\"\"\nreturn ops.index_update(self.array, self.index, values)\ndef add(self, values):\n- \"\"\"Pure equivalent of :code:`x[idx] += y`.\n+ \"\"\"Pure equivalent of ``x[idx] += y``.\n- :code:`x.at[idx].add(y)` is syntactic sugar for\n- :code:`jax.ops.index_add(x, jax.ops.index[idx], y)`, and\n- returns the value of `x` that would result from the NumPy-style\n- :mod:`indexed assignment <numpy.doc.indexing>` :code:`x[idx] += y`.\n+ ``x.at[idx].add(y)`` is syntactic sugar for\n+ ``jax.ops.index_add(x, jax.ops.index[idx], y)``, and\n+ returns the value of ``x`` that would result from the NumPy-style\n+ :mod:indexed assignment <numpy.doc.indexing>` ``x[idx] += y``.\nSee :mod:`jax.ops` for details.\n\"\"\"\nreturn ops.index_add(self.array, self.index, values)\ndef mul(self, values):\n- \"\"\"Pure equivalent of :code:`x[idx] += y`.\n+ \"\"\"Pure equivalent of ``x[idx] += y``.\n- :code:`x.at[idx].mul(y)` is syntactic sugar for\n- :code:`jax.ops.index_mul(x, jax.ops.index[idx], y)`, and\n- returns the value of `x` that would result from the NumPy-style\n- :mod:`indexed assignment <numpy.doc.indexing>` :code:`x[idx] *= y`.\n+ ``x.at[idx].mul(y)`` is syntactic sugar for\n+ ``jax.ops.index_mul(x, jax.ops.index[idx], y)``, and\n+ returns the value of ``x`` that would result from the NumPy-style\n+ :mod:indexed assignment <numpy.doc.indexing>` ``x[idx] *= y``.\nSee :mod:`jax.ops` for details.\n\"\"\"\nreturn ops.index_mul(self.array, self.index, values)\ndef min(self, values):\n- \"\"\"Pure equivalent of :code:`x[idx] = minimum(x[idx], y)`.\n+ \"\"\"Pure equivalent of ``x[idx] = minimum(x[idx], y)``.\n- :code:`x.at[idx].min(y)` is syntactic sugar for\n- :code:`jax.ops.index_min(x, jax.ops.index[idx], y)`, and\n- returns the value of `x` that would result from the NumPy-style\n- :mod:`indexed assignment <numpy.doc.indexing>`\n- :code:`x[idx] = minimum(x[idx], y)`.\n+ ``x.at[idx].min(y)`` is syntactic sugar for\n+ ``jax.ops.index_min(x, jax.ops.index[idx], y)``, and\n+ returns the value of ``x`` that would result from the NumPy-style\n+ :mod:indexed assignment <numpy.doc.indexing>`\n+ ``x[idx] = minimum(x[idx], y)``.\nSee :mod:`jax.ops` for details.\n\"\"\"\nreturn ops.index_min(self.array, self.index, values)\ndef max(self, values):\n- \"\"\"Pure equivalent of :code:`x[idx] = maximum(x[idx], y)`.\n+ \"\"\"Pure equivalent of ``x[idx] = maximum(x[idx], y)``.\n- :code:`x.at[idx].max(y)` is syntactic sugar for\n- :code:`jax.ops.index_max(x, jax.ops.index[idx], y)`, and\n- returns the value of `x` that would result from the NumPy-style\n- :mod:`indexed assignment <numpy.doc.indexing>`\n- :code:`x[idx] = maximum(x[idx], y)`.\n+ ``x.at[idx].max(y)`` is syntactic sugar for\n+ ``jax.ops.index_max(x, jax.ops.index[idx], y)``, and\n+ returns the value of ``x`` that would result from the NumPy-style\n+ :mod:indexed assignment <numpy.doc.indexing>`\n+ ``x[idx] = maximum(x[idx], y)``.\nSee :mod:`jax.ops` for details.\n\"\"\"\nreturn ops.index_max(self.array, self.index, values)\n-setattr(DeviceArray, \"at\", property(IndexUpdateHelper))\n-setattr(ShapedArray, \"at\", core.aval_property(IndexUpdateHelper))\n+setattr(DeviceArray, \"at\", property(_IndexUpdateHelper))\n+setattr(ShapedArray, \"at\", core.aval_property(_IndexUpdateHelper))\n" } ]
Python
Apache License 2.0
google/jax
Update names and documentation.
260,287
17.04.2020 11:20:54
0
8c00b35f21e51163dcbd8b427d4c861b6d6cc13a
Add FIXMEs for AD type errors
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/ad.py", "new_path": "jax/interpreters/ad.py", "diff": "@@ -137,7 +137,8 @@ def unpair_pval(pval):\naval_1, aval_2 = aval\nreturn (aval_1, const_1), (aval_2, const_2)\n-def backward_pass(jaxpr: core.Jaxpr, consts, args, cotangents_in):\n+# NOTE: The FIXMEs below are caused by primal/tangent mixups (type errors if you will)\n+def backward_pass(jaxpr: core.Jaxpr, consts, primals_in, cotangents_in):\nif all(ct is zero for ct in cotangents_in):\nreturn [zero] * len(jaxpr.invars)\n@@ -165,7 +166,9 @@ def backward_pass(jaxpr: core.Jaxpr, consts, args, cotangents_in):\nprimal_env: Dict[Any, Any] = {}\nwrite_primal(core.unitvar, core.unit)\nmap(write_primal, jaxpr.constvars, consts)\n- map(write_primal, jaxpr.invars, args)\n+ # FIXME: invars can contain both primal and tangent values, and this line\n+ # forces primal_in to contain UndefinedPrimals for tangent values!\n+ map(write_primal, jaxpr.invars, primals_in)\ndef is_linear(var):\nif type(var) is Literal:\n@@ -190,6 +193,7 @@ def backward_pass(jaxpr: core.Jaxpr, consts, args, cotangents_in):\nif any(is_linear(v) for v in eqn.invars):\nlinear_eqns.append(eqn)\nif any(not is_linear(v) for v in eqn.invars):\n+ # FIXME: Some invars correspond to tangents\nans = _eval_subjaxpr_primals(eqn.primitive, call_jaxpr,\nmap(read_primal, eqn.invars), params)\nmap(write_primal, eqn.outvars, ans)\n@@ -197,6 +201,7 @@ def backward_pass(jaxpr: core.Jaxpr, consts, args, cotangents_in):\nct_env: Dict[Any, Any] = {}\nmap(write_cotangent, jaxpr.outvars, cotangents_in)\nfor eqn in linear_eqns[::-1]:\n+ # FIXME: Some invars correspond to tangents\ninvals = map(read_primal, eqn.invars)\nif eqn.primitive.multiple_results:\ncts_in = map(read_cotangent, eqn.outvars)\n@@ -209,6 +214,7 @@ def backward_pass(jaxpr: core.Jaxpr, consts, args, cotangents_in):\nelse:\ncts_out = get_primitive_transpose(eqn.primitive)(cts_in, *invals, **eqn.params)\ncts_out = [zero] * len(eqn.invars) if cts_out is zero else cts_out\n+ # FIXME: Some invars correspond to primals!\nmap(write_cotangent, eqn.invars, cts_out)\ncotangents_out = map(read_cotangent, jaxpr.invars)\n" } ]
Python
Apache License 2.0
google/jax
Add FIXMEs for AD type errors
260,411
19.04.2020 12:13:07
-10,800
3ca7f6e4f120853a5f0c9f98c6b8b2db30300820
Fixes in the FAQ for RST
[ { "change_type": "MODIFY", "old_path": "docs/CHANGELOG.rst", "new_path": "docs/CHANGELOG.rst", "diff": "@@ -12,6 +12,10 @@ These are the release notes for JAX.\njax 0.1.64 (unreleased)\n---------------------------\n+* `GitHub commits <https://github.com/google/jax/compare/jax-v0.1.63...master>`_.\n+* Improves error message for reverse-mode differentiation of :func:`lax.while_loop`\n+ `#2129 <https://github.com/google/jax/issues/2129>`_.\n+\njaxlib 0.1.45 (unreleased)\n------------------------------\n@@ -29,6 +33,7 @@ jaxlib 0.1.44 (April 16, 2020)\njax 0.1.63\n---------------------------\n+* `GitHub commits <https://github.com/google/jax/compare/jax-v0.1.62...jax-v0.1.63>`_.\n* Added ``jax.custom_jvp`` and ``jax.custom_vjp`` from `#2026 <https://github.com/google/jax/pull/2026>`_, see the `tutorial notebook <https://jax.readthedocs.io/en/latest/notebooks/Custom_derivative_rules_for_Python_code.html>`_. Deprecated ``jax.custom_transforms`` and removed it from the docs (though it still works).\n* Add ``scipy.sparse.linalg.cg`` `#2566 <https://github.com/google/jax/pull/2566>`_.\n* Changed how Tracers are printed to show more useful information for debugging `#2591 <https://github.com/google/jax/pull/2591>`_.\n@@ -51,6 +56,7 @@ jaxlib 0.1.43 (March 31, 2020)\njax 0.1.62 (March 21, 2020)\n---------------------------\n+* `GitHub commits <https://github.com/google/jax/compare/jax-v0.1.61...jax-v0.1.62>`_.\n* JAX has dropped support for Python 3.5. Please upgrade to Python 3.6 or newer.\n* Removed the internal function ``lax._safe_mul``, which implemented the\nconvention ``0. * nan == 0.``. This change means some programs when\n@@ -69,14 +75,14 @@ jaxlib 0.1.42 (March 19, 2020)\njax 0.1.61 (March 17, 2020)\n---------------------------\n-\n+* `GitHub commits <https://github.com/google/jax/compare/jax-v0.1.60...jax-v0.1.61>`_.\n* Fixes Python 3.5 support. This will be the last JAX or jaxlib release that\nsupports Python 3.5.\njax 0.1.60 (March 17, 2020)\n---------------------------\n-* `GitHub commits <https://github.com/google/jax/compare/jax-v0.1.59...master>`_.\n+* `GitHub commits <https://github.com/google/jax/compare/jax-v0.1.59...jax-v0.1.60>`_.\n* New features:\n* :py:func:`jax.pmap` has ``static_broadcast_argnums`` argument which allows\n" }, { "change_type": "MODIFY", "old_path": "docs/conf.py", "new_path": "docs/conf.py", "diff": "@@ -98,7 +98,9 @@ exclude_patterns = [\n'notebooks/score_matching.ipynb',\n'notebooks/maml.ipynb',\n# Fails with shape error in XL\n- 'notebooks/XLA_in_Python.ipynb'\n+ 'notebooks/XLA_in_Python.ipynb',\n+ # Sometimes sphinx reads its own outputs as inputs!\n+ 'build/html',\n]\n# The name of the Pygments (syntax highlighting) style to use.\n" }, { "change_type": "MODIFY", "old_path": "docs/faq.rst", "new_path": "docs/faq.rst", "diff": "@@ -13,8 +13,8 @@ JAX's NumPy::\nimport numpy as np\nnp.array([0] * int(1e6))\n-The reason is that in NumPy the `numpy.array` function is implemented in C, while\n-the `jax.numpy.array` is implemented in Python, and it needs to iterate over a long\n+The reason is that in NumPy the ``numpy.array`` function is implemented in C, while\n+the :func:`jax.numpy.array` is implemented in Python, and it needs to iterate over a long\nlist to convert each list element to an array element.\nAn alternative would be to create the array with original NumPy and then convert\n@@ -23,13 +23,12 @@ it to a JAX array::\nfrom jax import numpy as jnp\njnp.array(np.array([0] * int(1e6)))\n-\n`jit` changes the behavior of my function\n-----------------------------------------\n-If you have a Python function that changes behavior after using `jit`, perhaps\n+If you have a Python function that changes behavior after using :func:`jax.jit`, perhaps\nyour function uses global state, or has side-effects. In the following code, the\n-`impure_func` uses the global `y` and has a side-effect due to `print`::\n+``impure_func`` uses the global ``y`` and has a side-effect due to ``print``::\ny = 0\n@@ -41,7 +40,7 @@ your function uses global state, or has side-effects. In the following code, the\nfor y in range(3):\nprint(\"Result:\", impure_func(y))\n-Without `jit` the output is::\n+Without ``jit`` the output is::\nInside: 0\nResult: 0\n@@ -50,27 +49,28 @@ Without `jit` the output is::\nInside: 2\nResult: 4\n-and with `jit` it is:\n+and with ``jit`` it is::\nInside: 0\nResult: 0\nResult: 1\nResult: 2\n-For `jit` the function is executed once using the Python interpreter, at which time the\n-`Inside` printing happens, and the first value of `y` is observed. Then the function\n-is compiled and cached, and executed multiple times with different values of `x`, but\n-with the same first value of `y`.\n+For :func:`jax.jit`, the function is executed once using the Python interpreter, at which time the\n+``Inside`` printing happens, and the first value of ``y`` is observed. Then the function\n+is compiled and cached, and executed multiple times with different values of ``x``, but\n+with the same first value of ``y``.\nAdditional reading:\n- * [JAX - The Sharp Bits: Pure Functions](https://jax.readthedocs.io/en/latest/notebooks/Common_Gotchas_in_JAX.html#%F0%9F%94%AA-Pure-functions)\n+ * `JAX - The Sharp Bits: Pure Functions <https://jax.readthedocs.io/en/latest/notebooks/Common_Gotchas_in_JAX.html#%F0%9F%94%AA-Pure-functions>`_.\n+\nGradients contain `NaN` where using ``where``\n------------------------------------------------\nIf you define a function using ``where`` to avoid an undefined value, if you\n-are not careful you may obtain a `NaN` for reverse differentiation::\n+are not careful you may obtain a ``NaN`` for reverse differentiation::\ndef my_log(x):\nreturn np.where(x > 0., np.log(x), 0.)\n@@ -90,7 +90,7 @@ that the adjoint is always finite::\nsafe_for_grad_log(0.) ==> 0. # Ok\njax.grad(safe_for_grad_log)(0.) ==> 0. # Ok\n-The inner ``np.where`` may be needed in addition to the original one, e.g.:\n+The inner ``np.where`` may be needed in addition to the original one, e.g.::\ndef my_log_or_y(x, y):\n\"\"\"Return log(x) if x > 0 or y\"\"\"\n@@ -99,5 +99,5 @@ The inner ``np.where`` may be needed in addition to the original one, e.g.:\nAdditional reading:\n- * [Issue: gradients through np.where when one of branches is nan](https://github.com/google/jax/issues/1052#issuecomment-514083352)\n- * [How to avoid NaN gradients when using ``where``](https://github.com/tensorflow/probability/blob/master/discussion/where-nan.pdf)\n\\ No newline at end of file\n+ * `Issue: gradients through np.where when one of branches is nan <https://github.com/google/jax/issues/1052#issuecomment-514083352>`_.\n+ * `How to avoid NaN gradients when using where <https://github.com/tensorflow/probability/blob/master/discussion/where-nan.pdf>`_.\n\\ No newline at end of file\n" } ]
Python
Apache License 2.0
google/jax
Fixes in the FAQ for RST (#2761)
260,700
19.04.2020 16:02:17
14,400
61fc2bf2c1ddf4a2c5847a9dc3ca24f35a4e6fae
add adamax test
[ { "change_type": "MODIFY", "old_path": "tests/optimizers_test.py", "new_path": "tests/optimizers_test.py", "diff": "@@ -149,6 +149,13 @@ class OptimizerTests(jtu.JaxTestCase):\nx0 = (np.ones(2), np.ones((2, 2)))\nself._CheckOptimizer(optimizers.sm3, loss, x0, num_iters, step_size)\n+ def testAdaMaxVector(self):\n+ def loss(x): return np.dot(x, x)\n+ x0 = np.ones(2)\n+ num_iters = 100\n+ step_size = 0.1\n+ self._CheckOptimizer(optimizers.adamax, loss, x0, num_iters, step_size)\n+\ndef testSgdVectorExponentialDecaySchedule(self):\ndef loss(x): return np.dot(x, x)\nx0 = np.ones(2)\n" } ]
Python
Apache License 2.0
google/jax
add adamax test
260,411
20.04.2020 11:30:52
-10,800
ff24f9a2f7bca0b463c1bcd27e138b79695553a7
Added FAQ entry about relationship between VJP and JVP
[ { "change_type": "MODIFY", "old_path": "docs/faq.rst", "new_path": "docs/faq.rst", "diff": "JAX Frequently Asked Questions\n==============================\n+.. comment RST primer for Sphinx: https://thomas-cokelaer.info/tutorials/sphinx/rest_syntax.html\n+.. comment Some links referenced here. Use JAX_sharp_bits_ (underscore at the end) to reference\n+\n+\n+.. _JAX_sharp_bits: https://jax.readthedocs.io/en/latest/notebooks/Common_Gotchas_in_JAX.html\n+.. _How_JAX_primitives_work: https://jax.readthedocs.io/en/latest/notebooks/How_JAX_primitives_work.html\n+\nWe are collecting here answers to frequently asked questions.\nContributions welcome!\n@@ -63,7 +70,7 @@ with the same first value of ``y``.\nAdditional reading:\n- * `JAX - The Sharp Bits: Pure Functions <https://jax.readthedocs.io/en/latest/notebooks/Common_Gotchas_in_JAX.html#%F0%9F%94%AA-Pure-functions>`_.\n+ * JAX_sharp_bits_\nGradients contain `NaN` where using ``where``\n@@ -101,3 +108,20 @@ Additional reading:\n* `Issue: gradients through np.where when one of branches is nan <https://github.com/google/jax/issues/1052#issuecomment-514083352>`_.\n* `How to avoid NaN gradients when using where <https://github.com/tensorflow/probability/blob/master/discussion/where-nan.pdf>`_.\n+\n+Why do I get forward-mode differentiation error when I am trying to do reverse-mode differentiation?\n+-----------------------------------------------------------------------------------------------------\n+\n+JAX implements reverse-mode differentiation as a composition of two operations:\n+linearization and transposition. The linearization step (see :func:`jax.linearize`)\n+uses the JVP rules to form the forward-computation of tangents along with the intermediate\n+forward computations of intermediate values on which the tangents depend.\n+The transposition step will turn the forward-computation of tangents\n+into a reverse-mode computation.\n+\n+If the JVP rule is not implemented for a primitive, then neither the forward-mode\n+nor the reverse-mode differentiation will work, but the error given will refer\n+to the forward-mode because that is the one that fails.\n+\n+You can read more details at How_JAX_primitives_work_.\n+\n" } ]
Python
Apache License 2.0
google/jax
Added FAQ entry about relationship between VJP and JVP (#2762)
260,346
20.04.2020 20:37:05
-7,200
685c0de99bbf7224d5644feab93f850e3e0a96c7
Fix confusing documentation typo.
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -919,7 +919,7 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None,\nEach host passes in a different length-4 array, corresponding to its 4 local\ndevices, and the psum operates over all 8 values. Conceptually, the two\n- length-4 arrays can be thought of as sharded length-16 array (in this example\n+ length-4 arrays can be thought of as sharded length-8 array (in this example\nequivalent to np.arange(8)) that is mapped over, with the length-8 mapped axis\ngiven name 'i'. The pmap call on each host then returns the corresponding\nlength-4 output shard.\n" } ]
Python
Apache License 2.0
google/jax
Fix confusing documentation typo. (#2773)
260,471
20.04.2020 22:55:23
14,400
2bc3c7985e4da55d21c0915d0b455e29a92c3394
Fix distribution name in docstring
[ { "change_type": "MODIFY", "old_path": "jax/random.py", "new_path": "jax/random.py", "diff": "@@ -626,7 +626,7 @@ def beta(key: np.ndarray,\nb: Union[float, np.ndarray],\nshape: Optional[Sequence[int]] = None,\ndtype: onp.dtype = onp.float64) -> np.ndarray:\n- \"\"\"Sample Bernoulli random values with given shape and mean.\n+ \"\"\"Sample Beta random values with given shape and float dtype.\nArgs:\nkey: a PRNGKey used as the random key.\n" } ]
Python
Apache License 2.0
google/jax
Fix distribution name in docstring (#2764)
260,335
20.04.2020 23:47:49
25,200
94178d98a055eca94f836ecc2ca7a690c5186b66
also skip jax.numpy.unique test on tpu
[ { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -1095,7 +1095,7 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nfor return_index in [False, True]\nfor return_inverse in [False, True]\nfor return_counts in [False, True]))\n- @jtu.skip_on_devices(\"gpu\") # https://github.com/google/jax/issues/2779\n+ @jtu.skip_on_devices(\"gpu\", \"tpu\") # https://github.com/google/jax/issues/2779\ndef testUnique(self, shape, dtype, return_index, return_inverse, return_counts, rng):\nargs_maker = lambda: [rng(shape, dtype)]\nonp_fun = lambda x: onp.unique(x, return_index, return_inverse, return_counts)\n@@ -1104,7 +1104,8 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\ndef testIssue1233(self):\n'''\n- Following numpy test suite from `test_repeat` at https://github.com/numpy/numpy/blob/master/numpy/core/tests/test_multiarray.py\n+ Following numpy test suite from `test_repeat` at\n+ https://github.com/numpy/numpy/blob/master/numpy/core/tests/test_multiarray.py\n'''\ntol = 1e-5\n" } ]
Python
Apache License 2.0
google/jax
also skip jax.numpy.unique test on tpu
260,335
21.04.2020 14:08:26
25,200
964cf4fb6acca7d172545e9baf4e5a163ff34b26
autodiff cookbook: assume continuous second derivatives fixes
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/autodiff_cookbook.ipynb", "new_path": "docs/notebooks/autodiff_cookbook.ipynb", "diff": "\"\\n\",\n\"A Hessian-vector product function can be useful in a [truncated Newton Conjugate-Gradient algorithm](https://en.wikipedia.org/wiki/Truncated_Newton_method) for minimizing smooth convex functions, or for studying the curvature of neural network training objectives (e.g. [1](https://arxiv.org/abs/1406.2572), [2](https://arxiv.org/abs/1811.07062), [3](https://arxiv.org/abs/1706.04454), [4](https://arxiv.org/abs/1802.03451)).\\n\",\n\"\\n\",\n- \"For a scalar-valued function $f : \\\\mathbb{R}^n \\\\to \\\\mathbb{R}$, the Hessian at a point $x \\\\in \\\\mathbb{R}^n$ is written as $\\\\partial^2 f(x)$. A Hessian-vector product function is then able to evaluate\\n\",\n+ \"For a scalar-valued function $f : \\\\mathbb{R}^n \\\\to \\\\mathbb{R}$ with continuous second derivatives (so that the Hessian matrix is symmetric), the Hessian at a point $x \\\\in \\\\mathbb{R}^n$ is written as $\\\\partial^2 f(x)$. A Hessian-vector product function is then able to evaluate\\n\",\n\"\\n\",\n\"$\\\\qquad v \\\\mapsto \\\\partial^2 f(x) \\\\cdot v$\\n\",\n\"\\n\",\n\"id\": \"YG3g5C3KdW7H\"\n},\n\"source\": [\n- \"In a previous section, we implemented a Hessian-vector product function just using reverse-mode:\"\n+ \"In a previous section, we implemented a Hessian-vector product function just using reverse-mode (assuming continuous second derivatives):\"\n]\n},\n{\n" } ]
Python
Apache License 2.0
google/jax
autodiff cookbook: assume continuous second derivatives fixes #2772
260,335
21.04.2020 14:27:59
25,200
18f967420c12765467db5c66d2e54f5580d08269
attempt to fix changelog formatting bugs
[ { "change_type": "MODIFY", "old_path": "docs/CHANGELOG.rst", "new_path": "docs/CHANGELOG.rst", "diff": "@@ -26,21 +26,24 @@ jax 0.1.64 (April 21, 2020)\n* `GitHub commits <https://github.com/google/jax/compare/jax-v0.1.63...master>`_.\n* New features:\n+\n* Add syntactic sugar for functional indexed updates\n`#2684 <https://github.com/google/jax/issues/2684>`_.\n* Add :func:`jax.numpy.linalg.multi_dot` `#2726 <https://github.com/google/jax/issues/2726>`_.\n* Add :func:`jax.numpy.unique` `#2760 <https://github.com/google/jax/issues/2760>`_.\n* Add :func:`jax.numpy.rint` `#2724 <https://github.com/google/jax/issues/2724>`_.\n* Add :func:`jax.numpy.rint` `#2724 <https://github.com/google/jax/issues/2724>`_.\n- * Add more primitive rules for :`func:`jax.experimental.jet`.\n+ * Add more primitive rules for :func:`jax.experimental.jet`.\n* Bug fixes:\n+\n* Fix :func:`logaddexp` and :func:`logaddexp2` differentiation at zero `#2107\n<https://github.com/google/jax/issues/2107>`_.\n* Improve memory usage in reverse-mode autodiff without :func:`jit`\n`#2719 <https://github.com/google/jax/issues/2719>`_.\n* Better errors:\n+\n* Improves error message for reverse-mode differentiation of :func:`lax.while_loop`\n`#2129 <https://github.com/google/jax/issues/2129>`_.\n" } ]
Python
Apache License 2.0
google/jax
attempt to fix changelog formatting bugs
260,335
21.04.2020 15:01:30
25,200
783e24fd4557c35254772f6a255f30b5bba69b67
try optimize=True with einsum closes can revert if this ends up problematic for some reason!
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -2558,7 +2558,7 @@ def tensordot(a, b, axes=2, precision=None):\n@_wraps(onp.einsum, lax_description=_PRECISION_DOC)\ndef einsum(*operands, **kwargs):\n- optimize = kwargs.pop('optimize', 'auto')\n+ optimize = kwargs.pop('optimize', True)\noptimize = 'greedy' if optimize is True else optimize\nprecision = kwargs.pop('precision', None)\nif kwargs:\n" } ]
Python
Apache License 2.0
google/jax
try optimize=True with einsum closes #2583 can revert if this ends up problematic for some reason!
260,335
21.04.2020 17:47:28
25,200
1bcaef142f9744ead76cc4641821373ed9facb59
apply is_stable=True to sort translation rules fixes
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -4494,6 +4494,7 @@ def _sort_batch_rule(batched_args, batch_dims, *, dimension):\nsort_p = standard_primitive(sort_shape, _input_dtype, 'sort')\nad.defjvp(sort_p, _sort_jvp_rule)\n+xla.translations[sort_p] = partial(standard_translate, 'sort', is_stable=True)\nbatching.primitive_batchers[sort_p] = _sort_batch_rule\ndef _sort_key_val_abstract_eval(keys, values, *, dimension):\n@@ -4562,7 +4563,8 @@ sort_key_val_p = Primitive('sort_key_val')\nsort_key_val_p.multiple_results = True\nsort_key_val_p.def_impl(partial(xla.apply_primitive, sort_key_val_p))\nsort_key_val_p.def_abstract_eval(_sort_key_val_abstract_eval)\n-xla.translations[sort_key_val_p] = partial(standard_translate, 'sort_key_val')\n+xla.translations[sort_key_val_p] = partial(standard_translate, 'sort_key_val',\n+ is_stable=True)\nad.primitive_jvps[sort_key_val_p] = _sort_key_val_jvp\nad.primitive_transposes[sort_key_val_p] = _sort_key_val_transpose_rule\nbatching.primitive_batchers[sort_key_val_p] = _sort_key_val_batch_rule\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -1095,7 +1095,6 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nfor return_index in [False, True]\nfor return_inverse in [False, True]\nfor return_counts in [False, True]))\n- @jtu.skip_on_devices(\"gpu\", \"tpu\") # https://github.com/google/jax/issues/2779\ndef testUnique(self, shape, dtype, return_index, return_inverse, return_counts, rng):\nargs_maker = lambda: [rng(shape, dtype)]\nonp_fun = lambda x: onp.unique(x, return_index, return_inverse, return_counts)\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_scipy_sparse_test.py", "new_path": "tests/lax_scipy_sparse_test.py", "diff": "@@ -101,12 +101,14 @@ class LaxBackedScipyTests(jtu.JaxTestCase):\ncheck_dtypes=True,\ntol=3e-5)\n+ # TODO(shoyer,mattjj): I had to loosen the tolerance for complex64[7,7]\n+ # with preconditioner=random\nself._CheckAgainstNumpy(\npartial(scipy_cg, M=M, maxiter=3),\npartial(lax_cg, M=M, maxiter=3),\nargs_maker,\ncheck_dtypes=True,\n- tol=1e-4)\n+ tol=3e-3)\nself._CheckAgainstNumpy(\nnp.linalg.solve,\n" } ]
Python
Apache License 2.0
google/jax
apply is_stable=True to sort translation rules (#2789) fixes #2779
260,335
21.04.2020 18:12:02
25,200
b1cb3a5deae7f682541acc8591a6367a7bcfaa2f
factor out process_map / post_process_map * factor out process_map / post_process_map Also fix a bug from reusing post_process_call for pmap. Fixes * consolidate call_bind / map_bind code
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -942,8 +942,8 @@ def canonicalize_shape(shape):\n\"smaller subfunctions.\")\nraise TypeError(msg.format(shape))\n-# ------------------- Call -------------------\n+# ------------------- Call and map -------------------\ndef apply_todos(todos, outs):\ntodos_list = list(todos)\n@@ -952,7 +952,8 @@ def apply_todos(todos, outs):\nreturn outs\n@lu.transformation_with_aux\n-def process_env_traces(primitive, level, params_tuple, *args):\n+def process_env_traces(post_processor: str, primitive: Primitive,\n+ level: int, params_tuple: tuple, *args):\nouts = yield args, {}\nparams = dict(params_tuple)\ntodo = []\n@@ -964,29 +965,34 @@ def process_env_traces(primitive, level, params_tuple, *args):\nbreak\ntrace = type(ans._trace)(ans._trace.master, cur_sublevel())\nouts = map(trace.full_raise, outs)\n- outs, cur_todo = trace.post_process_call(primitive, outs, params)\n+ post_process = getattr(trace, post_processor)\n+ outs, cur_todo = post_process(primitive, outs, params)\ntodo.append(cur_todo)\nyield outs, tuple(todo) # Ensure the aux output is immutable\n-def call_bind(primitive, f: lu.WrappedFun, *args, **params):\n+def _call_bind(processor: str, post_processor: str, primitive: Primitive,\n+ f: lu.WrappedFun, *args, **params):\ntop_trace = find_top_trace(args)\nlevel = trace_state.trace_stack.next_level(True) if top_trace is None else top_trace.level\nparams_tuple = tuple(params.items())\n- f, env_trace_todo = process_env_traces(f, primitive, level, params_tuple)\n+ f, env_trace_todo = process_env_traces(f, post_processor, primitive, level, params_tuple)\nif top_trace is None:\nwith new_sublevel():\nouts = primitive.impl(f, *args, **params)\nelse:\ntracers = map(top_trace.full_raise, args)\n- outs = map(full_lower, top_trace.process_call(primitive, f, tracers, params))\n+ process = getattr(top_trace, processor)\n+ outs = map(full_lower, process(primitive, f, tracers, params))\nreturn apply_todos(env_trace_todo(), outs)\n+call_bind = partial(_call_bind, 'process_call', 'post_process_call')\n+map_bind = partial(_call_bind, 'process_map', 'post_process_map')\n+\ndef call_impl(f: lu.WrappedFun, *args, **params):\ndel params # params parameterize the call primitive, not the function\nreturn f.call_wrapped(*args)\n-\ncall_p = Primitive('call')\ncall_p.multiple_results = True\ncall_p.call_primitive = True\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/ad.py", "new_path": "jax/interpreters/ad.py", "diff": "@@ -350,6 +350,9 @@ class JVPTrace(Trace):\nreturn map(partial(JVPTracer, trace), primals, tangents)\nreturn out, todo\n+ process_map = process_call\n+ post_process_map = post_process_call\n+\ndef process_custom_jvp_call(self, _, __, f_jvp, tracers):\nprimals_in, tangents_in = unzip2((t.primal, t.tangent) for t in tracers)\nprimals_in = map(core.full_lower, primals_in)\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/batching.py", "new_path": "jax/interpreters/batching.py", "diff": "@@ -140,9 +140,6 @@ class BatchTrace(Trace):\ndef process_call(self, call_primitive, f: lu.WrappedFun, tracers, params):\nassert call_primitive.multiple_results\nparams = dict(params, name=wrap_name(params.get('name', f.__name__), 'vmap'))\n- if call_primitive in pe.map_primitives:\n- return self.process_map(call_primitive, f, tracers, params)\n- else:\nvals, dims = unzip2((t.val, t.batch_dim) for t in tracers)\nif all(bdim is not_mapped for bdim in dims):\nreturn call_primitive.bind(f, *vals, **params)\n@@ -151,6 +148,14 @@ class BatchTrace(Trace):\nvals_out = call_primitive.bind(f, *vals, **params)\nreturn [BatchTracer(self, v, d) for v, d in zip(vals_out, dims_out())]\n+ def post_process_call(self, call_primitive, out_tracers, params):\n+ vals, dims = unzip2((t.val, t.batch_dim) for t in out_tracers)\n+ master = self.master\n+ def todo(vals):\n+ trace = BatchTrace(master, core.cur_sublevel())\n+ return map(partial(BatchTracer, trace), vals, dims)\n+ return vals, todo\n+\ndef process_map(self, map_primitive, f: lu.WrappedFun, tracers, params):\nvals, dims = unzip2((t.val, t.batch_dim) for t in tracers)\nif all(dim is not_mapped for dim in dims):\n@@ -166,12 +171,13 @@ class BatchTrace(Trace):\ndims_out = tuple(d + 1 if d is not not_mapped else d for d in dims_out())\nreturn [BatchTracer(self, v, d) for v, d in zip(vals_out, dims_out)]\n- def post_process_call(self, call_primitive, out_tracers, params):\n+ def post_process_map(self, call_primitive, out_tracers, params):\nvals, dims = unzip2((t.val, t.batch_dim) for t in out_tracers)\nmaster = self.master\n- def todo(x):\n+ def todo(vals):\ntrace = BatchTrace(master, core.cur_sublevel())\n- return map(partial(BatchTracer, trace), x, dims)\n+ return [BatchTracer(trace, v, d + 1 if d is not not_mapped else d)\n+ for v, d in zip(vals, dims)]\nreturn vals, todo\ndef process_custom_jvp_call(self, prim, fun, jvp, tracers):\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/parallel.py", "new_path": "jax/interpreters/parallel.py", "diff": "@@ -111,8 +111,6 @@ class PapplyTrace(Trace):\nreturn PapplyTracer(self, name, size, val_out, axis_out)\ndef process_call(self, call_primitive, f: lu.WrappedFun, tracers, params):\n- if call_primitive in pe.map_primitives:\n- return self.process_map(call_primitive, f, tracers, params)\nnames, vals, axes = unzip3((t.name, t.val, t.axis) for t in tracers)\nif all(axis is not_sharded for axis in axes):\nreturn call_primitive.bind(f, *vals, **params)\n@@ -133,8 +131,5 @@ class PapplyTrace(Trace):\nreturn PapplyTracer(trace, name, size, x, axis)\nreturn val, todo\n- def process_map(self, map_primitive, f :lu.WrappedFun, tracers, params):\n- raise NotImplementedError # TODO(mattjj,frostig)\n-\npapply_primitive_rules: Dict[core.Primitive, Callable] = {}\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -168,10 +168,9 @@ class JaxprTrace(Trace):\nelse:\nname = wrap_name(name, 'pe')\nparams = dict(params, name=name)\n+\nif call_primitive in call_partial_eval_rules:\nreturn call_partial_eval_rules[call_primitive](self, call_primitive, f, tracers, params)\n- if call_primitive in map_primitives:\n- return self.process_map(call_primitive, f, tracers, params)\nin_pvs, in_consts = unzip2([t.pval for t in tracers])\nfun, aux = partial_eval(f, self, in_pvs)\nout_flat = call_primitive.bind(fun, *in_consts, **params)\n@@ -190,7 +189,38 @@ class JaxprTrace(Trace):\nt.recipe = eqn\nreturn out_tracers\n+ def post_process_call(self, call_primitive, out_tracers, params):\n+ jaxpr, consts, env = tracers_to_jaxpr([], out_tracers)\n+ out_pvs, out_pv_consts = unzip2(t.pval for t in out_tracers)\n+ out = out_pv_consts + consts\n+ del consts, out_pv_consts\n+ master = self.master\n+ def todo(x):\n+ n = len(jaxpr.outvars)\n+ out_pv_consts, consts = x[:n], x[n:]\n+ trace = JaxprTrace(master, core.cur_sublevel())\n+ const_tracers = map(trace.new_instantiated_const, consts)\n+ env_tracers = map(trace.full_raise, env)\n+ lifted_jaxpr = convert_constvars_jaxpr(jaxpr)\n+ out_tracers = [JaxprTracer(trace, PartialVal((out_pv, out_pv_const)), None)\n+ for out_pv, out_pv_const in zip(out_pvs, out_pv_consts)]\n+ new_params = dict(params, call_jaxpr=lifted_jaxpr)\n+ # The `jaxpr` already contains the env_vars at start of invars\n+ eqn = new_eqn_recipe(tuple(it.chain(const_tracers, env_tracers)),\n+ out_tracers, call_primitive, new_params)\n+ for t in out_tracers:\n+ t.recipe = eqn\n+ return out_tracers\n+ return out, todo\n+\ndef process_map(self, map_primitive, f: lu.WrappedFun, tracers, params):\n+ name = params.get('name', f.__name__)\n+ if self.master.trace_type is StagingJaxprTrace:\n+ tracers = map(self.instantiate_const_abstracted, tracers)\n+ else:\n+ name = wrap_name(name, 'pe')\n+\n+ params = dict(params, name=name)\nin_pvs, in_consts = unzip2([t.pval for t in tracers])\nreduced_pvs = [None if pv is None else _mapped_aval(pv) for pv in in_pvs]\nfun, aux = partial_eval(f, self, reduced_pvs)\n@@ -216,32 +246,6 @@ class JaxprTrace(Trace):\nt.recipe = eqn\nreturn out_tracers\n- def post_process_call(self, call_primitive, out_tracers, params):\n- if call_primitive in map_primitives:\n- return self.post_process_map(call_primitive, out_tracers, params)\n- jaxpr, consts, env = tracers_to_jaxpr([], out_tracers)\n- out_pvs, out_pv_consts = unzip2(t.pval for t in out_tracers)\n- out = out_pv_consts + consts\n- del consts, out_pv_consts\n- master = self.master\n- def todo(x):\n- n = len(jaxpr.outvars)\n- out_pv_consts, consts = x[:n], x[n:]\n- trace = JaxprTrace(master, core.cur_sublevel())\n- const_tracers = map(trace.new_instantiated_const, consts)\n- env_tracers = map(trace.full_raise, env)\n- lifted_jaxpr = convert_constvars_jaxpr(jaxpr)\n- out_tracers = [JaxprTracer(trace, PartialVal((out_pv, out_pv_const)), None)\n- for out_pv, out_pv_const in zip(out_pvs, out_pv_consts)]\n- new_params = dict(params, call_jaxpr=lifted_jaxpr)\n- # The `jaxpr` already contains the env_vars at start of invars\n- eqn = new_eqn_recipe(tuple(it.chain(const_tracers, env_tracers)),\n- out_tracers, call_primitive, new_params)\n- for t in out_tracers:\n- t.recipe = eqn\n- return out_tracers\n- return out, todo\n-\ndef post_process_map(self, map_primitive, out_tracers, params):\njaxpr, consts, env = tracers_to_jaxpr([], out_tracers)\nout_pvs_reduced, out_pv_consts = unzip2(t.pval for t in out_tracers)\n@@ -306,7 +310,6 @@ def _unmapped_aval(size, aval):\nelse:\nraise TypeError(aval)\n-map_primitives: Set[core.Primitive] = set()\ncustom_partial_eval_rules: Dict[core.Primitive, Callable] = {}\ncall_partial_eval_rules: Dict[core.Primitive, Callable] = {}\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -816,7 +816,7 @@ def execute_replicated(compiled, backend, in_handler, out_handler, *args):\nxla_pmap_p = core.Primitive('xla_pmap')\nxla_pmap_p.call_primitive = True\nxla_pmap_p.multiple_results = True\n-xla_pmap = partial(core.call_bind, xla_pmap_p)\n+xla_pmap = partial(core.map_bind, xla_pmap_p)\nxla_pmap_p.def_custom_bind(xla_pmap)\nxla_pmap_p.def_impl(xla_pmap_impl)\n@@ -847,7 +847,6 @@ def _pmap_translation_rule(c, axis_env,\nxla.call_translations[xla_pmap_p] = _pmap_translation_rule\nad.primitive_transposes[xla_pmap_p] = partial(ad.map_transpose, xla_pmap_p)\n-pe.map_primitives.add(xla_pmap_p)\ndef _xla_shard(c, aval, axis_env, x):\nif aval is core.abstract_unit:\n@@ -1019,9 +1018,6 @@ class SplitAxisTrace(core.Trace):\ndef process_call(self, call_primitive, f: lu.WrappedFun, tracers, params):\nassert call_primitive.multiple_results\n- if call_primitive in pe.map_primitives:\n- return self.process_map(call_primitive, f, tracers, params)\n- else:\nvals, names = unzip2((t.val, t.axis_name) for t in tracers)\nif all(name is not_mapped for name in names):\nreturn call_primitive.bind(f, *vals, **params)\n@@ -1054,5 +1050,7 @@ class SplitAxisTrace(core.Trace):\nreturn SplitAxisTracer(trace, name, x)\nreturn val, todo\n+ post_process_map = post_process_call\n+\nsplit_axis_rules: Dict[core.Primitive, Callable] = {}\n" }, { "change_type": "MODIFY", "old_path": "tests/pmap_test.py", "new_path": "tests/pmap_test.py", "diff": "@@ -1015,6 +1015,27 @@ class PmapTest(jtu.JaxTestCase):\nnp.array([sum(vals)] * ndevices),\ncheck_dtypes=True)\n+ def testPostProcessMap(self):\n+ # code from https://github.com/google/jax/issues/2787\n+ def vv(x, y):\n+ \"\"\"Vector-vector multiply\"\"\"\n+ return np.dot(x, y)\n+\n+ def distributed_matrix_vector(x, y):\n+ \"\"\"Matrix vector multiply. First batch it and then row by row\"\"\"\n+ fv = lambda z: lax.map(lambda j: vv(j, y), z)\n+ res = pmap(fv)(x.reshape((jax.device_count(), -1) + tuple(x.shape[1:])))\n+ res = res.reshape(res.shape[0] * res.shape[1], *res.shape[2:])\n+ return res\n+\n+ key = random.PRNGKey(1)\n+ x = random.normal(key, (800, 50))\n+ batched_mvm = vmap(lambda b: distributed_matrix_vector(x, b), in_axes=0)\n+ y = random.normal(key, (10, 50, 1))\n+ result = batched_mvm(y)\n+ expected = np.einsum('ij,njk->nik', x, y)\n+ self.assertAllClose(result, expected, check_dtypes=False, atol=1e-3, rtol=1e-3)\n+\nclass PmapWithDevicesTest(jtu.JaxTestCase):\n" } ]
Python
Apache License 2.0
google/jax
factor out process_map / post_process_map (#2788) * factor out process_map / post_process_map Also fix a bug from reusing post_process_call for pmap. Fixes #2787 * consolidate call_bind / map_bind code
260,335
21.04.2020 18:27:53
25,200
f753845f617a2f38eff65d41cd537b3fd25518b9
adjust test tolerance for tpu
[ { "change_type": "MODIFY", "old_path": "tests/pmap_test.py", "new_path": "tests/pmap_test.py", "diff": "@@ -1029,12 +1029,13 @@ class PmapTest(jtu.JaxTestCase):\nreturn res\nkey = random.PRNGKey(1)\n- x = random.normal(key, (800, 50))\n+ x = random.normal(key, (80, 50))\nbatched_mvm = vmap(lambda b: distributed_matrix_vector(x, b), in_axes=0)\ny = random.normal(key, (10, 50, 1))\nresult = batched_mvm(y)\nexpected = np.einsum('ij,njk->nik', x, y)\n- self.assertAllClose(result, expected, check_dtypes=False, atol=1e-3, rtol=1e-3)\n+ tol = 1e-1 if jtu.device_under_test() == \"tpu\" else 1e-3\n+ self.assertAllClose(result, expected, check_dtypes=False, atol=tol, rtol=tol)\nclass PmapWithDevicesTest(jtu.JaxTestCase):\n" } ]
Python
Apache License 2.0
google/jax
adjust test tolerance for tpu
260,335
21.04.2020 19:04:28
25,200
2e34dbc188f43150578a6f6f978ffc343f0e3213
update travis to match min jaxlib version
[ { "change_type": "MODIFY", "old_path": ".travis.yml", "new_path": ".travis.yml", "diff": "@@ -32,7 +32,7 @@ install:\n# The jaxlib version should match the minimum jaxlib version in\n# jax/lib/__init__.py. This tests JAX PRs against the oldest permitted\n# jaxlib.\n- - pip install jaxlib==0.1.43\n+ - pip install jaxlib==0.1.45\n- pip install -v .\n# The following are needed to test the Colab notebooks and the documentation building\n- if [ \"$JAX_ONLY_DOCUMENTATION\" = true ]; then\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -2296,10 +2296,6 @@ def _conv_general_dilated_transpose_lhs(\nout = _reshape_axis_into(lhs_spec[1], lhs_spec[0], out)\nreturn out\n-# TODO(phawkins): remove when the minimum jaxlib version is incremented past\n-# 0.1.43.\n-_jaxlib_has_working_batch_group_count = lib.version > (0, 1, 43)\n-\ndef _conv_general_dilated_transpose_rhs(\ng, lhs, *, window_strides, padding, lhs_dilation, rhs_dilation,\ndimension_numbers: ConvDimensionNumbers, feature_group_count: int,\n@@ -2315,12 +2311,8 @@ def _conv_general_dilated_transpose_rhs(\nfeature_group_count = batch_group_count\nbatch_group_count = 1\nelif feature_group_count > 1:\n- if _jaxlib_has_working_batch_group_count:\nbatch_group_count = feature_group_count\nfeature_group_count = 1\n- else:\n- lhs = _reshape_axis_out_of(lhs_trans[0], feature_group_count, lhs)\n- lhs = _reshape_axis_into(lhs_trans[0], lhs_trans[1], lhs)\ntrans_dimension_numbers = ConvDimensionNumbers(lhs_trans, out_trans, rhs_trans)\npadding = _conv_general_vjp_rhs_padding(\nonp.take(lhs_shape, lhs_sdims), onp.take(rhs_shape, rhs_sdims),\n" } ]
Python
Apache License 2.0
google/jax
update travis to match min jaxlib version
260,335
22.04.2020 08:59:22
25,200
7334f97da86521598a3c91ad7563e4b049848e51
attempt to fix failing travis (numerical issues)
[ { "change_type": "MODIFY", "old_path": "tests/lax_test.py", "new_path": "tests/lax_test.py", "diff": "@@ -2053,6 +2053,9 @@ class LaxAutodiffTest(jtu.JaxTestCase):\npadding, lhs_dil, rhs_dil, dimension_numbers,\nperms, feature_group_count, batch_group_count,\nrng_factory):\n+ if dtype == onp.float16:\n+ raise SkipTest(\"float16 numerical issues\") # TODO(mattjj): resolve\n+\nrng = rng_factory()\ntol = {dtypes.bfloat16: 1e-0, onp.float16: 5e-1, onp.float32: 1e-4}\n" }, { "change_type": "MODIFY", "old_path": "tests/linalg_test.py", "new_path": "tests/linalg_test.py", "diff": "@@ -714,7 +714,7 @@ class NumpyLinalgTest(jtu.JaxTestCase):\nargs_maker = lambda: [rng(shape, dtype)]\nself._CheckAgainstNumpy(onp.linalg.pinv, np.linalg.pinv, args_maker,\n- check_dtypes=True, tol=1e-3)\n+ check_dtypes=True, tol=1e-2)\nself._CompileAndCheck(np.linalg.pinv, args_maker, check_dtypes=True)\n@parameterized.named_parameters(jtu.cases_from_list(\n" } ]
Python
Apache License 2.0
google/jax
attempt to fix failing travis (numerical issues)
260,335
22.04.2020 14:39:51
25,200
fcb6ff0a15420cdfcf772f5d60cbdba59dcb4e47
loosen scipy convolve test tolerance (GPU flaky)
[ { "change_type": "MODIFY", "old_path": "tests/scipy_signal_test.py", "new_path": "tests/scipy_signal_test.py", "diff": "@@ -44,7 +44,7 @@ class LaxBackedScipySignalTests(jtu.JaxTestCase):\n\"\"\"Tests for LAX-backed scipy.stats implementations\"\"\"\n@parameterized.named_parameters(jtu.cases_from_list(\n- {\"testcase_name\": \"op={}_xshape=[{}]_yshape=[{}]_mode={}\".format(\n+ {\"testcase_name\": \"_op={}_xshape=[{}]_yshape=[{}]_mode={}\".format(\nop,\njtu.format_shape_dtype_string(xshape, dtype),\njtu.format_shape_dtype_string(yshape, dtype),\n@@ -63,7 +63,7 @@ class LaxBackedScipySignalTests(jtu.JaxTestCase):\nargs_maker = lambda: [rng(xshape, dtype), rng(yshape, dtype)]\nosp_fun = partial(osp_op, mode=mode)\njsp_fun = partial(jsp_op, mode=mode, precision=lax.Precision.HIGHEST)\n- tol = {onp.float16: 1e-2, onp.float32: 1e-2}\n+ tol = {onp.float16: 1e-2, onp.float32: 1e-2, onp.float64: 1e-8}\nself._CheckAgainstNumpy(osp_fun, jsp_fun, args_maker, check_dtypes=False, tol=tol)\nself._CompileAndCheck(jsp_fun, args_maker, check_dtypes=True)\n" } ]
Python
Apache License 2.0
google/jax
loosen scipy convolve test tolerance (GPU flaky)
260,700
22.04.2020 20:49:10
14,400
59bdb1fb3d1329bcb7bf71e9e869ae40c4e7f05b
add tanh rule change expit taylor rule add manual expit check, check stability of expit and tanh
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jet.py", "new_path": "jax/experimental/jet.py", "diff": "@@ -17,6 +17,7 @@ from functools import partial\nimport numpy as onp\n+import jax\nfrom jax import core\nfrom jax.util import unzip2\nfrom jax.tree_util import (register_pytree_node, tree_structure,\n@@ -217,6 +218,9 @@ def fact(n):\ndef _scale(k, j):\nreturn 1. / (fact(k - j) * fact(j - 1))\n+def _scale2(k, j):\n+ return 1. / (fact(k - j) * fact(j))\n+\ndef _exp_taylor(primals_in, series_in):\nx, = primals_in\nseries, = series_in\n@@ -253,6 +257,28 @@ def _pow_taylor(primals_in, series_in):\nreturn primal_out, series_out\njet_rules[lax.pow_p] = _pow_taylor\n+def _expit_taylor(primals_in, series_in):\n+ x, = primals_in\n+ series, = series_in\n+ u = [x] + series\n+ v = [jax.scipy.special.expit(x)] + [None] * len(series)\n+ e = [v[0] * (1 - v[0])] + [None] * len(series) # terms for sigmoid' = sigmoid * (1 - sigmoid)\n+ for k in range(1, len(v)):\n+ v[k] = fact(k-1) * sum([_scale(k, j) * e[k-j] * u[j] for j in range(1, k+1)])\n+ e[k] = (1 - v[0]) * v[k] - fact(k) * sum([_scale2(k, j)* v[j] * v[k-j] for j in range(1, k+1)])\n+\n+ primal_out, *series_out = v\n+ return primal_out, series_out\n+\n+def _tanh_taylor(primals_in, series_in):\n+ x, = primals_in\n+ series, = series_in\n+ u = [2*x] + [2 * series_ for series_ in series]\n+ primals_in, *series_in = u\n+ primal_out, series_out = _expit_taylor((primals_in, ), (series_in, ))\n+ series_out = [2 * series_ for series_ in series_out]\n+ return 2 * primal_out - 1, series_out\n+jet_rules[lax.tanh_p] = _tanh_taylor\ndef _log_taylor(primals_in, series_in):\nx, = primals_in\n" }, { "change_type": "MODIFY", "old_path": "tests/jet_test.py", "new_path": "tests/jet_test.py", "diff": "@@ -20,6 +20,7 @@ import numpy as onp\nfrom jax import test_util as jtu\nimport jax.numpy as np\n+import jax.scipy.special\nfrom jax import random\nfrom jax import jacfwd, jit\nfrom jax.experimental import stax\n@@ -153,6 +154,27 @@ class JetTest(jtu.JaxTestCase):\nelse:\nself.check_jet_finite(fun, primal_in, series_in, atol=1e-4, rtol=1e-4)\n+ def expit_check(self, lims=[-2, 2], order=3):\n+ dims = 2, 3\n+ rng = onp.random.RandomState(0)\n+ primal_in = transform(lims, rng.rand(*dims))\n+ terms_in = [rng.randn(*dims) for _ in range(order)]\n+\n+ primals = (primal_in, )\n+ series = (terms_in, )\n+\n+ y, terms = jax.experimental.jet._expit_taylor(primals, series)\n+ expected_y, expected_terms = jvp_taylor(jax.scipy.special.expit, primals, series)\n+\n+ atol = 1e-4\n+ rtol = 1e-4\n+ self.assertAllClose(y, expected_y, atol=atol, rtol=rtol,\n+ check_dtypes=True)\n+\n+ self.assertAllClose(terms, expected_terms, atol=atol, rtol=rtol,\n+ check_dtypes=True)\n+\n+\n@jtu.skip_on_devices(\"tpu\")\ndef test_exp(self): self.unary_check(np.exp)\n@jtu.skip_on_devices(\"tpu\")\n@@ -187,6 +209,12 @@ class JetTest(jtu.JaxTestCase):\ndef test_log1p(self): self.unary_check(np.log1p, lims=[0, 4.])\n@jtu.skip_on_devices(\"tpu\")\ndef test_expm1(self): self.unary_check(np.expm1)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_tanh(self): self.unary_check(np.tanh, lims=[-500, 500], order=5)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_expit(self): self.unary_check(jax.scipy.special.expit, lims=[-500, 500], order=5)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_expit2(self): self.expit_check(lims=[-500, 500], order=5)\n@jtu.skip_on_devices(\"tpu\")\ndef test_div(self): self.binary_check(lambda x, y: x / y, lims=[0.8, 4.0])\n" } ]
Python
Apache License 2.0
google/jax
add tanh rule (#2653) change expit taylor rule add manual expit check, check stability of expit and tanh
260,335
22.04.2020 22:09:56
25,200
6b5e36763c8c21912c22e033cf09db108924f0c3
skip pinv test on tpu because no svd
[ { "change_type": "MODIFY", "old_path": "tests/linalg_test.py", "new_path": "tests/linalg_test.py", "diff": "@@ -719,7 +719,7 @@ class NumpyLinalgTest(jtu.JaxTestCase):\n# TODO(phawkins): 1e-1 seems like a very loose tolerance.\njtu.check_grads(np.linalg.pinv, args_maker(), 2, rtol=1e-1)\n-\n+ @jtu.skip_on_devices(\"tpu\") # SVD is not implemented on the TPU backend\ndef testPinvGradIssue2792(self):\ndef f(p):\na = np.array([[0., 0.],[-p, 1.]], np.float32) * 1 / (1 + p**2)\n" } ]
Python
Apache License 2.0
google/jax
skip pinv test on tpu because no svd
260,335
22.04.2020 23:29:32
25,200
8ccb907d125c23efc4af119b520b24a8ebd76909
in custom_jvp/vjp stop_gradient on nondiff_argnums fixes
[ { "change_type": "MODIFY", "old_path": "jax/custom_derivatives.py", "new_path": "jax/custom_derivatives.py", "diff": "@@ -18,6 +18,7 @@ import inspect\nimport itertools as it\nimport operator as op\n+import jax\nfrom . import core\nfrom . import linear_util as lu\nfrom .tree_util import tree_flatten, tree_unflatten, tree_map, tree_multimap\n@@ -82,6 +83,15 @@ def sum_tangents(x, *xs):\ndef zeros_like_pytree(x):\nreturn tree_map(lambda _: zero, x)\n+def stop_gradient(x):\n+ return tree_map(_stop_gradient, x)\n+\n+def _stop_gradient(x):\n+ if isinstance(x, core.Tracer) or core.valid_jaxtype(x):\n+ return jax.lax.stop_gradient(x)\n+ else:\n+ return x\n+\n### JVPs\n@@ -199,7 +209,10 @@ class custom_jvp:\nraise AttributeError(msg.format(self.__name__))\nargs = _resolve_kwargs(self.fun, args, kwargs)\nif self.nondiff_argnums:\n- dyn_argnums = [i for i in range(len(args)) if i not in self.nondiff_argnums]\n+ is_nondiff = [False] * len(args)\n+ for i in self.nondiff_argnums: is_nondiff[i] = True\n+ args = [stop_gradient(x) if b else x for b, x in zip(is_nondiff, args)]\n+ dyn_argnums = [i for i, b in enumerate(is_nondiff) if not b]\nf_, dyn_args = argnums_partial(lu.wrap_init(self.fun), dyn_argnums, args)\nstatic_args = [args[i] for i in self.nondiff_argnums]\njvp = _add_args(lu.wrap_init(self.jvp), static_args, left=True)\n@@ -436,7 +449,10 @@ class custom_vjp:\nraise AttributeError(msg.format(self.__name__))\nargs = _resolve_kwargs(self.fun, args, kwargs)\nif self.nondiff_argnums:\n- dyn_argnums = [i for i in range(len(args)) if i not in self.nondiff_argnums]\n+ is_nondiff = [False] * len(args)\n+ for i in self.nondiff_argnums: is_nondiff[i] = True\n+ args = [stop_gradient(x) if b else x for b, x in zip(is_nondiff, args)]\n+ dyn_argnums = [i for i, b in enumerate(is_nondiff) if not b]\nf_, dyn_args = argnums_partial(lu.wrap_init(self.fun), dyn_argnums, args)\nstatic_args = [args[i] for i in self.nondiff_argnums]\nfwd, _ = argnums_partial(lu.wrap_init(self.fwd), dyn_argnums, args)\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -2853,6 +2853,28 @@ class CustomVJPTest(jtu.JaxTestCase):\njax.grad(g, argnums=(1,))(F(2.0), 0.) # doesn't crash\n+ def test_nondiff_argnums_stop_gradient(self):\n+ # https://github.com/google/jax/issues/2784\n+ @partial(api.custom_vjp, nondiff_argnums=(0, 1))\n+ def _clip_gradient(lo, hi, x):\n+ return x # identity function\n+\n+ def clip_gradient_fwd(lo, hi, x):\n+ # return x, None\n+ return x, (hi, )\n+\n+ def clip_gradient_bwd(lo, hi, _, g):\n+ return (np.clip(g, lo, hi),)\n+\n+ _clip_gradient.defvjp(clip_gradient_fwd, clip_gradient_bwd)\n+\n+ def clip_gradient(x):\n+ lo = -1\n+ hi = x + 1 # causes things to break\n+ return _clip_gradient(lo, hi, x)\n+\n+ jax.grad(clip_gradient)(1.) # doesn't crash\n+\nclass DeprecatedCustomTransformsTest(jtu.JaxTestCase):\n" } ]
Python
Apache License 2.0
google/jax
in custom_jvp/vjp stop_gradient on nondiff_argnums (#2804) fixes #2784
260,335
23.04.2020 09:28:14
25,200
903010b7b90f5af6ff4d724721f4bc1f10237dc4
disable mypy checks causing new errors
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax_control_flow.py", "new_path": "jax/lax/lax_control_flow.py", "diff": "@@ -1388,7 +1388,7 @@ def _memcpy(axis, num, src, dst, offset):\nreturn lax.dynamic_update_index_in_dim(dst, update, i + offset, axis)\nreturn fori_loop(0, num, body, dst)\n-masking.masking_rules[lax.concatenate_p] = _concat_masking_rule\n+masking.masking_rules[lax.concatenate_p] = _concat_masking_rule # type: ignore\ndef _check_tree(func_name, expected_name, actual_tree, expected_tree):\n" }, { "change_type": "MODIFY", "old_path": "mypy.ini", "new_path": "mypy.ini", "diff": "@@ -12,3 +12,5 @@ ignore_missing_imports = True\nignore_missing_imports = True\n[mypy-jax.interpreters.autospmd]\nignore_errors = True\n+[mypy-jax.lax.lax_parallel]\n+ignore_errors = True\n" } ]
Python
Apache License 2.0
google/jax
disable mypy checks causing new errors
260,335
23.04.2020 13:12:24
25,200
13a17286df985648257f4da2a06bd47f06c09a55
stop_gradient_p -> ad_util.py, re-enable some mypy
[ { "change_type": "MODIFY", "old_path": "jax/ad_util.py", "new_path": "jax/ad_util.py", "diff": "# limitations under the License.\n-from .core import lattice_join, Primitive, Unit, unit, AbstractUnit\n+from .core import (lattice_join, Primitive, Unit, unit, AbstractUnit,\n+ valid_jaxtype)\nfrom .tree_util import register_pytree_node\nfrom typing import Any, Dict\nfrom .util import safe_map\n@@ -64,3 +65,14 @@ class Zero(object):\nzero = Zero()\nregister_pytree_node(Zero, lambda z: ((), None), lambda _, xs: zero)\n+\n+\n+def _stop_gradient_impl(x):\n+ if not valid_jaxtype(x):\n+ raise TypeError(\"stop_gradient only works on valid JAX arrays, but \"\n+ f\"input argument is: {x}\")\n+ return x\n+\n+stop_gradient_p = Primitive('stop_gradient')\n+stop_gradient_p.def_impl(_stop_gradient_impl)\n+stop_gradient_p.def_abstract_eval(lambda x: x)\n" }, { "change_type": "MODIFY", "old_path": "jax/custom_derivatives.py", "new_path": "jax/custom_derivatives.py", "diff": "@@ -18,14 +18,13 @@ import inspect\nimport itertools as it\nimport operator as op\n-import jax\nfrom . import core\nfrom . import linear_util as lu\nfrom .tree_util import tree_flatten, tree_unflatten, tree_map, tree_multimap\nfrom .util import safe_zip, safe_map, unzip2, split_list, curry\nfrom .api_util import flatten_fun_nokwargs, argnums_partial, wrap_hashably\nfrom .abstract_arrays import raise_to_shaped\n-from .ad_util import zero\n+from .ad_util import zero, stop_gradient_p\nfrom .interpreters import partial_eval as pe\nfrom .interpreters import ad\nfrom .interpreters import batching\n@@ -88,7 +87,7 @@ def stop_gradient(x):\ndef _stop_gradient(x):\nif isinstance(x, core.Tracer) or core.valid_jaxtype(x):\n- return jax.lax.stop_gradient(x)\n+ return stop_gradient_p.bind(x)\nelse:\nreturn x\n" }, { "change_type": "MODIFY", "old_path": "jax/experimental/jet.py", "new_path": "jax/experimental/jet.py", "diff": "@@ -20,6 +20,7 @@ import numpy as onp\nimport jax\nfrom jax import core\nfrom jax.util import unzip2\n+from jax import ad_util\nfrom jax.tree_util import (register_pytree_node, tree_structure,\ntreedef_is_leaf, tree_flatten, tree_unflatten)\nimport jax.linear_util as lu\n@@ -177,7 +178,7 @@ defzero(lax.floor_p)\ndefzero(lax.ceil_p)\ndefzero(lax.round_p)\ndefzero(lax.sign_p)\n-defzero(lax.stop_gradient_p)\n+defzero(ad_util.stop_gradient_p)\ndef deflinear(prim):\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -1292,7 +1292,7 @@ def stop_gradient(x):\n>>> jax.grad(jax.grad(lambda x: jax.lax.stop_gradient(x)**2))(3.)\narray(0., dtype=float32)\n\"\"\"\n- return tree_map(stop_gradient_p.bind, x)\n+ return tree_map(ad_util.stop_gradient_p.bind, x)\n### convenience wrappers around traceables\n@@ -4600,14 +4600,6 @@ masking.shape_rules[tie_in_p] = lambda x, y: y.shape\nmasking.masking_rules[tie_in_p] = lambda vals, logical_shapes: vals[1]\n-### stop-gradient\n-\n-def _stop_gradient_impl(x):\n- if not core.valid_jaxtype(x):\n- raise TypeError(\"stop_gradient only works on valid JAX arrays, but \"\n- f\"input argument is: {x}\")\n- return x\n-\ndef _stop_gradient_jvp_rule(primals, tangents):\n# if we don't call stop_gradient here, we'd only peel off one autodiff tracer\nx, = primals\n@@ -4618,12 +4610,10 @@ def _stop_gradient_batch_rule(batched_args, batch_dims):\ndim, = batch_dims\nreturn stop_gradient(x), dim\n-stop_gradient_p = Primitive('stop_gradient')\n-stop_gradient_p.def_impl(_stop_gradient_impl)\n-stop_gradient_p.def_abstract_eval(_identity)\n-xla.translations[stop_gradient_p] = lambda c, x: x\n-ad.primitive_jvps[stop_gradient_p] = _stop_gradient_jvp_rule\n-batching.primitive_batchers[stop_gradient_p] = _stop_gradient_batch_rule\n+xla.translations[ad_util.stop_gradient_p] = lambda c, x: x\n+ad.primitive_jvps[ad_util.stop_gradient_p] = _stop_gradient_jvp_rule\n+batching.primitive_batchers[ad_util.stop_gradient_p] = _stop_gradient_batch_rule\n+\ndef create_token(x):\n\"\"\"Creates an XLA token value with no preconditions for sequencing effects.\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/lax_control_flow.py", "new_path": "jax/lax/lax_control_flow.py", "diff": "@@ -1388,7 +1388,7 @@ def _memcpy(axis, num, src, dst, offset):\nreturn lax.dynamic_update_index_in_dim(dst, update, i + offset, axis)\nreturn fori_loop(0, num, body, dst)\n-masking.masking_rules[lax.concatenate_p] = _concat_masking_rule # type: ignore\n+masking.masking_rules[lax.concatenate_p] = _concat_masking_rule\ndef _check_tree(func_name, expected_name, actual_tree, expected_tree):\n" }, { "change_type": "MODIFY", "old_path": "mypy.ini", "new_path": "mypy.ini", "diff": "@@ -12,5 +12,3 @@ ignore_missing_imports = True\nignore_missing_imports = True\n[mypy-jax.interpreters.autospmd]\nignore_errors = True\n-[mypy-jax.lax.lax_parallel]\n-ignore_errors = True\n" } ]
Python
Apache License 2.0
google/jax
stop_gradient_p -> ad_util.py, re-enable some mypy (#2806)
260,335
23.04.2020 13:34:01
25,200
d2653a1e8a29dda3f2df40098076c83978a0098f
rewrite axis_index implementation, use custom bind * rewrite axis_index implementation, use custom bind fixes * add test for
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -383,26 +383,27 @@ def axis_index(axis_name):\n[0 1]\n[0 1]]\n\"\"\"\n+ return axis_index_p.bind(axis_name=axis_name)\n+\n+def _axis_index_bind(*, axis_name):\ndynamic_axis_env = _thread_local_state.dynamic_axis_env\nframe = dynamic_axis_env[axis_name]\nsizes = dynamic_axis_env.sizes[:dynamic_axis_env.index(frame)+1]\nnreps = dynamic_axis_env.nreps\n- dummy_arg = frame.pmap_trace.pure(core.unit)\n- if frame.soft_trace:\n- dummy_arg = frame.soft_trace.pure(dummy_arg)\n-\n- return axis_index_p.bind(dummy_arg, nreps=nreps, sizes=sizes,\n- soft_size=frame.soft_size, axis_name=axis_name)\n+ trace = frame.pmap_trace\n-def _axis_index_partial_eval(trace, _, **params):\n- # This partial_eval rule adds the axis_index primitive into the jaxpr formed\n- # during pmap lowering. It is like the standard JaxprTrace.process_primitive\n- # rule except that we don't attempt to lower out of the trace.\nout_aval = ShapedArray((), onp.int32)\nout_tracer = pe.JaxprTracer(trace, pe.PartialVal.unknown(out_aval), None)\n- eqn = pe.new_eqn_recipe([], [out_tracer], axis_index_p, params)\n+ eqn = pe.new_eqn_recipe([], [out_tracer], axis_index_p,\n+ dict(nreps=nreps, sizes=sizes,\n+ soft_size=frame.soft_size, axis_name=axis_name))\nout_tracer.recipe = eqn\n+\n+ if not frame.soft_trace:\nreturn out_tracer\n+ else:\n+ val_out = out_tracer * frame.soft_size + onp.arange(frame.soft_size)\n+ return SplitAxisTracer(frame.soft_trace, axis_name, val_out)\ndef _axis_index_translation_rule(c, nreps, sizes, soft_size, axis_name):\ndiv = c.Constant(onp.array(nreps // prod(sizes), dtype=onp.uint32))\n@@ -411,8 +412,8 @@ def _axis_index_translation_rule(c, nreps, sizes, soft_size, axis_name):\nreturn c.ConvertElementType(unsigned_index, xb.dtype_to_etype(onp.int32))\naxis_index_p = core.Primitive('axis_index')\n+axis_index_p.def_custom_bind(_axis_index_bind)\nxla.translations[axis_index_p] = _axis_index_translation_rule\n-pe.custom_partial_eval_rules[axis_index_p] = _axis_index_partial_eval\n### lazy device-memory persistence and result handling\n" }, { "change_type": "MODIFY", "old_path": "tests/pmap_test.py", "new_path": "tests/pmap_test.py", "diff": "@@ -1037,6 +1037,17 @@ class PmapTest(jtu.JaxTestCase):\ntol = 1e-1 if jtu.device_under_test() == \"tpu\" else 1e-3\nself.assertAllClose(result, expected, check_dtypes=False, atol=tol, rtol=tol)\n+ def testAxisIndexRemat(self):\n+ # https://github.com/google/jax/issues/2716\n+ n = len(jax.devices())\n+\n+ def f(key):\n+ key = random.fold_in(key, jax.lax.axis_index('i'))\n+ return random.bernoulli(key, p=0.5)\n+\n+ keys = random.split(random.PRNGKey(0), n)\n+ jax.pmap(jax.remat(f), axis_name='i')(keys)\n+\nclass PmapWithDevicesTest(jtu.JaxTestCase):\n" } ]
Python
Apache License 2.0
google/jax
rewrite axis_index implementation, use custom bind (#2807) * rewrite axis_index implementation, use custom bind fixes #2716 Co-authored-by: Trevor Cai <tycai@google.com> * add test for #2716 Co-authored-by: Trevor Cai <tycai@google.com>
260,335
23.04.2020 18:07:51
25,200
251834367f3093c669612ddc119e8fefdcf53387
use static_argnums in xla_computation * use static_argnums in xla_computation fixes * add static_argnums to make_jaxpr * fix type error: handle int case
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -294,8 +294,9 @@ def xla_computation(fun: Callable,\nROOT tuple.18 = (f32[], f32[], f32[]) tuple(all-reduce.7, all-reduce.12, all-reduce.17)\n}\n\"\"\"\n- del static_argnums # Unused.\n_check_callable(fun)\n+ if isinstance(static_argnums, int):\n+ static_argnums = (static_argnums,)\nfun_name = getattr(fun, '__name__', 'unknown')\ndef make_axis_env(nreps):\n@@ -312,6 +313,11 @@ def xla_computation(fun: Callable,\n@wraps(fun)\ndef computation_maker(*args, **kwargs):\nwrapped = lu.wrap_init(fun)\n+ if static_argnums:\n+ dyn_argnums = [i for i in range(len(args)) if i not in static_argnums]\n+ wrapped, dyn_args = argnums_partial(wrapped, dyn_argnums, args)\n+ else:\n+ dyn_args = args\njax_args, in_tree = tree_flatten((args, kwargs))\njaxtree_fun, out_tree = flatten_fun(wrapped, in_tree)\navals = map(abstractify, jax_args)\n@@ -1403,17 +1409,20 @@ def _vjp(fun: lu.WrappedFun, *primals, **kwargs):\nreturn out_primal_py, vjp_py, tree_unflatten(aux_tree, aux)\n-def make_jaxpr(fun: Callable) -> Callable[..., core.TypedJaxpr]:\n+def make_jaxpr(fun: Callable,\n+ static_argnums: Union[int, Iterable[int]] = ()\n+ ) -> Callable[..., core.TypedJaxpr]:\n\"\"\"Creates a function that produces its jaxpr given example args.\nArgs:\nfun: The function whose ``jaxpr`` is to be computed. Its positional\narguments and return value should be arrays, scalars, or standard Python\ncontainers (tuple/list/dict) thereof.\n+ static_argnums: See the ``jax.jit`` docstring.\nReturns:\n- A wrapped version of ``fun`` that when applied to example arguments returns a\n- ``TypedJaxpr`` representation of ``fun`` on those arguments.\n+ A wrapped version of ``fun`` that when applied to example arguments returns\n+ a ``TypedJaxpr`` representation of ``fun`` on those arguments.\nA ``jaxpr`` is JAX's intermediate representation for program traces. The\n``jaxpr`` language is based on the simply-typed first-order lambda calculus\n@@ -1444,6 +1453,8 @@ def make_jaxpr(fun: Callable) -> Callable[..., core.TypedJaxpr]:\nin [g] }\n\"\"\"\n_check_callable(fun)\n+ if isinstance(static_argnums, int):\n+ static_argnums = (static_argnums,)\ndef pv_like(x):\naval = xla.abstractify(x)\n@@ -1452,6 +1463,11 @@ def make_jaxpr(fun: Callable) -> Callable[..., core.TypedJaxpr]:\n@wraps(fun)\ndef jaxpr_maker(*args, **kwargs):\nwrapped = lu.wrap_init(fun)\n+ if static_argnums:\n+ dyn_argnums = [i for i in range(len(args)) if i not in static_argnums]\n+ wrapped, dyn_args = argnums_partial(wrapped, dyn_argnums, args)\n+ else:\n+ dyn_args = args\njax_args, in_tree = tree_flatten((args, kwargs))\njaxtree_fun, out_tree = flatten_fun(wrapped, in_tree)\nin_pvals = map(pv_like, jax_args)\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -879,6 +879,13 @@ class APITest(jtu.JaxTestCase):\nout_shape, = xla_comp.GetProgramShape().result_shape().tuple_shapes()\nself.assertEqual(out_shape.dimensions(), (3, 4))\n+ def test_xla_computation_static_argnums(self):\n+ def f(x, y):\n+ return x + y\n+\n+ xla_comp = api.xla_computation(f, static_argnums=(1,))(2, 3)\n+ self.assertIn('constant(3)', xla_comp.GetHloText())\n+\ndef test_jit_device(self):\ndevice = xb.devices()[-1]\nx = api.jit(lambda x: x, device=device)(3.)\n@@ -1845,6 +1852,14 @@ class JaxprTest(jtu.JaxTestCase):\nin (d,) }\n\"\"\", str(jaxpr))\n+ def test_make_jaxpr_static_argnums(self):\n+ def f(x, y):\n+ return x + y\n+\n+ jaxpr = api.make_jaxpr(f, static_argnums=(1,))(2, 3)\n+ self.assertIn('3', str(jaxpr))\n+\n+\nclass LazyTest(jtu.JaxTestCase):\n@contextmanager\n" } ]
Python
Apache License 2.0
google/jax
use static_argnums in xla_computation (#2812) * use static_argnums in xla_computation fixes #1017 * add static_argnums to make_jaxpr * fix type error: handle int case
260,700
24.04.2020 01:07:35
14,400
fc4203c38adfe7b06477e4126ff30f307179ef61
implement jet rules by lowering to other primitives merge jet_test add jet rules use lax.square
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jet.py", "new_path": "jax/experimental/jet.py", "diff": "@@ -293,6 +293,41 @@ def _log_taylor(primals_in, series_in):\nreturn primal_out, series_out\njet_rules[lax.log_p] = _log_taylor\n+def _sqrt_taylor(primals_in, series_in):\n+ return jet(lambda x: x ** 0.5, primals_in, series_in)\n+jet_rules[lax.sqrt_p] = _sqrt_taylor\n+\n+def _rsqrt_taylor(primals_in, series_in):\n+ return jet(lambda x: x ** -0.5, primals_in, series_in)\n+jet_rules[lax.rsqrt_p] = _rsqrt_taylor\n+\n+def _asinh_taylor(primals_in, series_in):\n+ return jet(lambda x: lax.log(x + lax.sqrt(lax.square(x) + 1)), primals_in, series_in)\n+jet_rules[lax.asinh_p] = _asinh_taylor\n+\n+def _acosh_taylor(primals_in, series_in):\n+ return jet(lambda x: lax.log(x + lax.sqrt(lax.square(x) - 1)), primals_in, series_in)\n+jet_rules[lax.acosh_p] = _acosh_taylor\n+\n+def _atanh_taylor(primals_in, series_in):\n+ return jet(lambda x: 0.5 * lax.log(lax.div(1 + x, 1 - x)), primals_in, series_in)\n+jet_rules[lax.atanh_p] = _atanh_taylor\n+\n+def _atan2_taylor(primals_in, series_in):\n+ x, y = primals_in\n+ primal_out = lax.atan2(x, y)\n+\n+ x, series = jet(lax.div, primals_in, series_in)\n+ c0, cs = jet(lambda x: lax.div(1, 1 + lax.square(x)), (x, ), (series, ))\n+ c = [c0] + cs\n+ u = [x] + series\n+ v = [primal_out] + [None] * len(series)\n+ for k in range(1, len(v)):\n+ v[k] = fact(k-1) * sum(_scale(k, j) * c[k-j] * u[j] for j in range(1, k + 1))\n+ primal_out, *series_out = v\n+ return primal_out, series_out\n+jet_rules[lax.atan2_p] = _atan2_taylor\n+\ndef _log1p_taylor(primals_in, series_in):\nx, = primals_in\nseries, = series_in\n" }, { "change_type": "MODIFY", "old_path": "tests/jet_test.py", "new_path": "tests/jet_test.py", "diff": "@@ -215,6 +215,16 @@ class JetTest(jtu.JaxTestCase):\ndef test_expit(self): self.unary_check(jax.scipy.special.expit, lims=[-500, 500], order=5)\n@jtu.skip_on_devices(\"tpu\")\ndef test_expit2(self): self.expit_check(lims=[-500, 500], order=5)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_sqrt(self): self.unary_check(np.sqrt, lims=[0, 5.])\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_rsqrt(self): self.unary_check(lax.rsqrt, lims=[0, 5000.])\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_asinh(self): self.unary_check(lax.asinh, lims=[-100, 100])\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_acosh(self): self.unary_check(lax.acosh, lims=[-100, 100])\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_atanh(self): self.unary_check(lax.atanh, lims=[-1, 1])\n@jtu.skip_on_devices(\"tpu\")\ndef test_div(self): self.binary_check(lambda x, y: x / y, lims=[0.8, 4.0])\n@@ -245,6 +255,8 @@ class JetTest(jtu.JaxTestCase):\n@jtu.skip_on_devices(\"tpu\")\n@jtu.ignore_warning(message=\"overflow encountered in power\")\ndef test_pow(self): self.binary_check(lambda x, y: x ** y, lims=([0.2, 500], [-500, 500]), finite=False)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_atan2(self): self.binary_check(lax.atan2, lims=[-40, 40])\ndef test_process_call(self):\ndef f(x):\n" } ]
Python
Apache License 2.0
google/jax
implement jet rules by lowering to other primitives (#2816) merge jet_test add jet rules use lax.square
260,335
24.04.2020 01:47:20
25,200
11d7fb051c0a972d81b0a77e170f4b5dbf7719f8
add more ode tests
[ { "change_type": "MODIFY", "old_path": "tests/ode_test.py", "new_path": "tests/ode_test.py", "diff": "@@ -65,6 +65,43 @@ class JetTest(jtu.JaxTestCase):\njtu.check_grads(integrate, (y0, ts), modes=[\"rev\"], order=2,\nrtol=tol, atol=tol)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_decay(self):\n+ def decay(y, t, arg1, arg2):\n+ return -np.sqrt(t) - y + arg1 - np.mean((y + arg2)**2)\n+\n+ rng = onp.random.RandomState(0)\n+ arg1 = rng.randn(3)\n+ arg2 = rng.randn(3)\n+\n+ def integrate(y0, ts):\n+ return odeint(decay, y0, ts, arg1, arg2)\n+\n+ y0 = rng.randn(3)\n+ ts = np.linspace(0.1, 0.2, 4)\n+\n+ tol = 1e-1 if num_float_bits(onp.float64) == 32 else 1e-3\n+ jtu.check_grads(integrate, (y0, ts), modes=[\"rev\"], order=2,\n+ rtol=tol, atol=tol)\n+\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_swoop(self):\n+ def swoop(y, t, arg1, arg2):\n+ return np.array(y - np.sin(t) - np.cos(t) * arg1 + arg2)\n+\n+ ts = np.array([0.1, 0.2])\n+ tol = 1e-1 if num_float_bits(onp.float64) == 32 else 1e-3\n+\n+ y0 = np.linspace(0.1, 0.9, 10)\n+ integrate = lambda y0, ts: odeint(swoop, y0, ts, 0.1, 0.2)\n+ jtu.check_grads(integrate, (y0, ts), modes=[\"rev\"], order=2,\n+ rtol=tol, atol=tol)\n+\n+ big_y0 = np.linspace(1.1, 10.9, 10)\n+ integrate = lambda y0, ts: odeint(swoop, big_y0, ts, 0.1, 0.3)\n+ jtu.check_grads(integrate, (y0, ts), modes=[\"rev\"], order=2,\n+ rtol=tol, atol=tol)\n+\nif __name__ == '__main__':\nabsltest.main()\n" } ]
Python
Apache License 2.0
google/jax
add more ode tests (#2819)
260,265
24.04.2020 13:43:04
25,200
77901e9fa71f5b23066c70132a983ae57f655b39
Fix lax.rng_uniform.
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -4778,7 +4778,8 @@ def _rng_uniform_abstract_eval(a, b, *, shape):\nreturn ShapedArray(shape, a.dtype)\ndef _rng_uniform_translation_rule(c, a, b, *, shape):\n- return xops.RngUniform(a, b, shape)\n+ xla_shape = xc.Shape.array_shape(c.GetShape(a).xla_element_type(), shape)\n+ return xops.RngUniform(a, b, xla_shape)\nrng_uniform_p = Primitive(\"rng_uniform\")\nrng_uniform_p.def_impl(partial(xla.apply_primitive, rng_uniform_p))\n" } ]
Python
Apache License 2.0
google/jax
Fix lax.rng_uniform. (#2830)
260,335
24.04.2020 18:19:24
25,200
8f902452abd016b73d287641a5b999e96c3c8365
only maximally stage out for some call primitives fixes
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -163,7 +163,8 @@ class JaxprTrace(Trace):\ndef process_call(self, call_primitive, f: lu.WrappedFun, tracers, params):\nname = params.get('name', f.__name__)\n- if self.master.trace_type is StagingJaxprTrace:\n+ if (self.master.trace_type is StagingJaxprTrace\n+ and call_primitive in staged_out_calls):\ntracers = map(self.instantiate_const_abstracted, tracers)\nelse:\nname = wrap_name(name, 'pe')\n@@ -312,6 +313,7 @@ def _unmapped_aval(size, aval):\ncustom_partial_eval_rules: Dict[core.Primitive, Callable] = {}\ncall_partial_eval_rules: Dict[core.Primitive, Callable] = {}\n+staged_out_calls: Set[core.Primitive] = set()\ndef partial_eval(f, trace, pvs: Sequence[Optional[AbstractValue]], instantiate=False):\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/xla.py", "new_path": "jax/interpreters/xla.py", "diff": "@@ -608,6 +608,7 @@ xla_call_p.multiple_results = True\nxla_call = partial(core.call_bind, xla_call_p)\nxla_call_p.def_custom_bind(xla_call)\nxla_call_p.def_impl(_xla_call_impl)\n+pe.staged_out_calls.add(xla_call_p)\ndef _xla_call_translation_rule(c, axis_env,\nin_nodes, name_stack, backend, name,\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -1510,6 +1510,16 @@ class APITest(jtu.JaxTestCase):\njax.grad(scan_bug)(1.0) # doesn't crash\n+ def test_remat_jit_static_argnum(self):\n+ # https://github.com/google/jax/issues/2833\n+ def f(a_bool, y):\n+ if a_bool:\n+ return y + 1\n+ else:\n+ return y\n+\n+ api.jit(api.remat(f, concrete=True), static_argnums=0)(True, 1) # no crash\n+\ndef test_trivial_computations(self):\nx = np.array([1, 2, 3])\ny = api.jit(lambda x: x)(x)\n" } ]
Python
Apache License 2.0
google/jax
only maximally stage out for some call primitives (#2834) fixes #2833
260,677
25.04.2020 16:32:27
-3,600
ffaf417b1bfb3b2a2a834748b405fb3324b31c2a
Fix typo in docstring for _cofactor_solve Found a small typo in the description of _cofactor_solve
[ { "change_type": "MODIFY", "old_path": "jax/numpy/linalg.py", "new_path": "jax/numpy/linalg.py", "diff": "@@ -170,13 +170,13 @@ def _cofactor_solve(a, b):\nIf a is rank n-1, then the lower right corner of u will be zero and the\ntriangular_solve will fail.\nLet x = solve(p @ l, b) and y = det(a)*solve(a, b).\n- Then y_{nn} =\n- x_{nn} / u_{nn} * prod_{i=1...n}(u_{ii}) =\n- x_{nn} * prod_{i=1...n-1}(u_{ii})\n+ Then y_{n} =\n+ x_{n} / u_{nn} * prod_{i=1...n}(u_{ii}) =\n+ x_{n} * prod_{i=1...n-1}(u_{ii})\nSo by replacing the lower-right corner of u with prod_{i=1...n-1}(u_{ii})^-1\nwe can avoid the triangular_solve failing.\n- To correctly compute the rest of x_{ii} for i != n, we simply multiply\n- x_{ii} by det(a) for all i != n, which will be zero if rank(a) = n-1.\n+ To correctly compute the rest of y_{i} for i != n, we simply multiply\n+ x_{i} by det(a) for all i != n, which will be zero if rank(a) = n-1.\nFor the second case, a check is done on the matrix to see if `solve`\nreturns NaN or Inf, and gives a matrix of zeros as a result, as the\n" } ]
Python
Apache License 2.0
google/jax
Fix typo in docstring for _cofactor_solve (#2844) Found a small typo in the description of _cofactor_solve
260,335
25.04.2020 09:20:26
25,200
0a5cc090ee34610d6bd957d554e3c1b4026ad4f4
split testDetGradOfSingularMatrix into corank=1,2
[ { "change_type": "MODIFY", "old_path": "tests/linalg_test.py", "new_path": "tests/linalg_test.py", "diff": "@@ -133,16 +133,19 @@ class NumpyLinalgTest(jtu.JaxTestCase):\na[0] = 0\njtu.check_grads(np.linalg.det, (a,), 1, atol=1e-1, rtol=1e-1)\n- def testDetGradOfSingularMatrix(self):\n+ def testDetGradOfSingularMatrixCorank1(self):\n# Rank 2 matrix with nonzero gradient\na = np.array([[ 50, -30, 45],\n[-30, 90, -81],\n[ 45, -81, 81]], dtype=np.float32)\n+ jtu.check_grads(np.linalg.det, (a,), 1, atol=1e-1, rtol=1e-1)\n+\n+ @jtu.skip_on_devices(\"tpu\") # TODO(mattjj,pfau): nan on tpu, investigate\n+ def testDetGradOfSingularMatrixCorank2(self):\n# Rank 1 matrix with zero gradient\nb = np.array([[ 36, -42, 18],\n[-42, 49, -21],\n[ 18, -21, 9]], dtype=np.float32)\n- jtu.check_grads(np.linalg.det, (a,), 1, atol=1e-1, rtol=1e-1)\njtu.check_grads(np.linalg.det, (b,), 1, atol=1e-1, rtol=1e-1)\n@parameterized.named_parameters(jtu.cases_from_list(\n" } ]
Python
Apache License 2.0
google/jax
split testDetGradOfSingularMatrix into corank=1,2 (#2845)
260,700
27.04.2020 19:45:51
14,400
b277b55d1c85505024f6197c8cd686f7262d7d55
check step size is greater than zero loosen tols for grad test set tol only for float64
[ { "change_type": "MODIFY", "old_path": "jax/experimental/ode.py", "new_path": "jax/experimental/ode.py", "diff": "@@ -171,8 +171,8 @@ def _odeint(func, rtol, atol, mxstep, y0, ts, *args):\ndef scan_fun(carry, target_t):\ndef cond_fun(state):\n- i, _, _, t, _, _, _ = state\n- return (t < target_t) & (i < mxstep)\n+ i, _, _, t, dt, _, _ = state\n+ return (t < target_t) & (i < mxstep) & (dt > 0)\ndef body_fun(state):\ni, y, f, t, dt, last_t, interp_coeff = state\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -2863,7 +2863,10 @@ class CustomVJPTest(jtu.JaxTestCase):\nans = jax.grad(g)(2.) # don't crash\nexpected = jax.grad(f, 0)(2., 0.1) + jax.grad(f, 0)(2., 0.2)\n- self.assertAllClose(ans, expected, check_dtypes=False)\n+\n+ atol = {onp.float64: 5e-15}\n+ rtol = {onp.float64: 2e-15}\n+ self.assertAllClose(ans, expected, check_dtypes=False, atol=atol, rtol=rtol)\ndef test_lowering_out_of_traces(self):\n# https://github.com/google/jax/issues/2578\n" } ]
Python
Apache License 2.0
google/jax
check step size is greater than zero (#2857) loosen tols for grad test set tol only for float64
260,700
28.04.2020 00:53:38
14,400
cc0e9a3189731ce91919e8dbdbaa90e88959d8cb
refactor ode tests, add scipy benchmark * refactor ode tests, add scipy benchmark remove double import rename to scipy merge vmap test properly * clean up more global trace state after errors
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -623,7 +623,9 @@ def find_top_trace(xs):\n@contextmanager\ndef initial_style_staging():\nprev, trace_state.initial_style = trace_state.initial_style, True\n+ try:\nyield\n+ finally:\ntrace_state.initial_style = prev\n" }, { "change_type": "MODIFY", "old_path": "jax/experimental/ode.py", "new_path": "jax/experimental/ode.py", "diff": "@@ -25,7 +25,6 @@ Adjoint algorithm based on Appendix C of https://arxiv.org/pdf/1806.07366.pdf\nfrom functools import partial\nimport operator as op\n-import time\nimport jax\nimport jax.numpy as np\n@@ -33,11 +32,8 @@ from jax import lax\nfrom jax import ops\nfrom jax.util import safe_map, safe_zip\nfrom jax.flatten_util import ravel_pytree\n-from jax.test_util import check_grads\nfrom jax.tree_util import tree_map\nfrom jax import linear_util as lu\n-import numpy as onp\n-import scipy.integrate as osp_integrate\nmap = safe_map\nzip = safe_zip\n@@ -240,69 +236,3 @@ def _odeint_rev(func, rtol, atol, mxstep, res, g):\nreturn (y_bar, ts_bar, *args_bar)\n_odeint.defvjp(_odeint_fwd, _odeint_rev)\n-\n-\n-def pend(np, y, _, m, g):\n- theta, omega = y\n- return [omega, -m * omega - g * np.sin(theta)]\n-\n-def benchmark_odeint(fun, y0, tspace, *args):\n- \"\"\"Time performance of JAX odeint method against scipy.integrate.odeint.\"\"\"\n- n_trials = 10\n- n_repeat = 100\n- y0, tspace = onp.array(y0), onp.array(tspace)\n- onp_fun = partial(fun, onp)\n- scipy_times = []\n- for k in range(n_trials):\n- start = time.time()\n- for _ in range(n_repeat):\n- scipy_result = osp_integrate.odeint(onp_fun, y0, tspace, args)\n- end = time.time()\n- print('scipy odeint elapsed time ({} of {}): {}'.format(k+1, n_trials, end-start))\n- scipy_times.append(end - start)\n- y0, tspace = np.array(y0), np.array(tspace)\n- jax_fun = partial(fun, np)\n- jax_times = []\n- for k in range(n_trials):\n- start = time.time()\n- for _ in range(n_repeat):\n- jax_result = odeint(jax_fun, y0, tspace, *args)\n- jax_result.block_until_ready()\n- end = time.time()\n- print('JAX odeint elapsed time ({} of {}): {}'.format(k+1, n_trials, end-start))\n- jax_times.append(end - start)\n- print('(avg scipy time) / (avg jax time) = {}'.format(\n- onp.mean(scipy_times[1:]) / onp.mean(jax_times[1:])))\n- print('norm(scipy result-jax result): {}'.format(\n- np.linalg.norm(np.asarray(scipy_result) - jax_result)))\n- return scipy_result, jax_result\n-\n-def pend_benchmark_odeint():\n- _, _ = benchmark_odeint(pend, [np.pi - 0.1, 0.0], np.linspace(0., 10., 101),\n- 0.25, 9.8)\n-\n-def pend_check_grads():\n- def f(y0, ts, *args):\n- return odeint(partial(pend, np), y0, ts, *args)\n-\n- y0 = [np.pi - 0.1, 0.0]\n- ts = np.linspace(0., 1., 11)\n- args = (0.25, 9.8)\n-\n- check_grads(f, (y0, ts, *args), modes=[\"rev\"], order=2,\n- atol=1e-1, rtol=1e-1)\n-\n-def weird_time_pendulum_check_grads():\n- \"\"\"Test that gradients are correct when the dynamics depend on t.\"\"\"\n- def f(y0, ts):\n- return odeint(lambda y, t: np.array([y[1] * -t, -1 * y[1] - 9.8 * np.sin(y[0])]), y0, ts)\n-\n- y0 = [np.pi - 0.1, 0.0]\n- ts = np.linspace(0., 1., 11)\n-\n- check_grads(f, (y0, ts), modes=[\"rev\"], order=2)\n-\n-if __name__ == '__main__':\n- pend_benchmark_odeint()\n- pend_check_grads()\n- weird_time_pendulum_check_grads()\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -2834,40 +2834,6 @@ class CustomVJPTest(jtu.JaxTestCase):\nfoo = lambda x: api.vmap(np.linalg.det, (0,))(x)\napi.jit(foo)(arr) # doesn't crash\n- def test_odeint_vmap_grad(self):\n- # https://github.com/google/jax/issues/2531\n- # TODO(mattjj): factor out an ode tests file\n- try:\n- from jax.experimental.ode import odeint\n- except ImportError:\n- raise unittest.SkipTest(\"missing jax.experimental\") from None\n-\n- def dx_dt(x, *args):\n- return 0.1 * x\n-\n- def f(x, y):\n- y0 = np.array([x, y])\n- t = np.array([0., 5.])\n- y = odeint(dx_dt, y0, t)\n- return y[-1].sum()\n-\n- def g(x):\n- # Two initial values for the ODE\n- y0_arr = np.array([[x, 0.1],\n- [x, 0.2]])\n-\n- # Run ODE twice\n- t = np.array([0., 5.])\n- y = jax.vmap(lambda y0: odeint(dx_dt, y0, t))(y0_arr)\n- return y[:,-1].sum()\n-\n- ans = jax.grad(g)(2.) # don't crash\n- expected = jax.grad(f, 0)(2., 0.1) + jax.grad(f, 0)(2., 0.2)\n-\n- atol = {onp.float64: 5e-15}\n- rtol = {onp.float64: 2e-15}\n- self.assertAllClose(ans, expected, check_dtypes=False, atol=atol, rtol=rtol)\n-\ndef test_lowering_out_of_traces(self):\n# https://github.com/google/jax/issues/2578\n" }, { "change_type": "MODIFY", "old_path": "tests/ode_test.py", "new_path": "tests/ode_test.py", "diff": "# limitations under the License.\nfrom functools import partial\n+import unittest\nfrom absl.testing import absltest\nimport numpy as onp\n-from jax import jit\n+import jax\nfrom jax import dtypes\nfrom jax import test_util as jtu\nimport jax.numpy as np\nfrom jax.experimental.ode import odeint\n+import scipy.integrate as osp_integrate\n+\nfrom jax.config import config\nconfig.parse_flags_with_absl()\n@@ -31,77 +34,127 @@ def num_float_bits(dtype):\nreturn dtypes.finfo(dtypes.canonicalize_dtype(dtype)).bits\n-class JetTest(jtu.JaxTestCase):\n+class ODETest(jtu.JaxTestCase):\n+\n+ def check_against_scipy(self, fun, y0, tspace, *args, tol=1e-1):\n+ y0, tspace = onp.array(y0), onp.array(tspace)\n+ onp_fun = partial(fun, onp)\n+ scipy_result = np.asarray(osp_integrate.odeint(onp_fun, y0, tspace, args))\n+\n+ y0, tspace = np.array(y0), np.array(tspace)\n+ jax_fun = partial(fun, np)\n+ jax_result = odeint(jax_fun, y0, tspace, *args)\n+\n+ self.assertAllClose(jax_result, scipy_result, check_dtypes=False, atol=tol, rtol=tol)\n@jtu.skip_on_devices(\"tpu\")\ndef test_pend_grads(self):\n- def pend(y, _, m, g):\n+ def pend(_np, y, _, m, g):\ntheta, omega = y\n- return [omega, -m * omega - g * np.sin(theta)]\n+ return [omega, -m * omega - g * _np.sin(theta)]\n- def f(y0, ts, *args):\n- return odeint(pend, y0, ts, *args)\n+ integrate = partial(odeint, partial(pend, np))\ny0 = [np.pi - 0.1, 0.0]\nts = np.linspace(0., 1., 11)\nargs = (0.25, 9.8)\ntol = 1e-1 if num_float_bits(onp.float64) == 32 else 1e-3\n- jtu.check_grads(f, (y0, ts, *args), modes=[\"rev\"], order=2,\n+\n+ self.check_against_scipy(pend, y0, ts, *args, tol=tol)\n+\n+ jtu.check_grads(integrate, (y0, ts, *args), modes=[\"rev\"], order=2,\natol=tol, rtol=tol)\n@jtu.skip_on_devices(\"tpu\")\ndef test_weird_time_pendulum_grads(self):\n\"\"\"Test that gradients are correct when the dynamics depend on t.\"\"\"\n- def dynamics(y, t):\n- return np.array([y[1] * -t, -1 * y[1] - 9.8 * np.sin(y[0])])\n+ def dynamics(_np, y, t):\n+ return _np.array([y[1] * -t, -1 * y[1] - 9.8 * _np.sin(y[0])])\n- integrate = partial(odeint, dynamics)\n+ integrate = partial(odeint, partial(dynamics, np))\ny0 = [np.pi - 0.1, 0.0]\nts = np.linspace(0., 1., 11)\ntol = 1e-1 if num_float_bits(onp.float64) == 32 else 1e-3\n+\n+ self.check_against_scipy(dynamics, y0, ts, tol=tol)\n+\njtu.check_grads(integrate, (y0, ts), modes=[\"rev\"], order=2,\nrtol=tol, atol=tol)\n@jtu.skip_on_devices(\"tpu\")\ndef test_decay(self):\n- def decay(y, t, arg1, arg2):\n- return -np.sqrt(t) - y + arg1 - np.mean((y + arg2)**2)\n+ def decay(_np, y, t, arg1, arg2):\n+ return -_np.sqrt(t) - y + arg1 - _np.mean((y + arg2)**2)\n- rng = onp.random.RandomState(0)\n- arg1 = rng.randn(3)\n- arg2 = rng.randn(3)\n+ integrate = partial(odeint, partial(decay, np))\n- def integrate(y0, ts):\n- return odeint(decay, y0, ts, arg1, arg2)\n+ rng = onp.random.RandomState(0)\n+ args = (rng.randn(3), rng.randn(3))\ny0 = rng.randn(3)\nts = np.linspace(0.1, 0.2, 4)\ntol = 1e-1 if num_float_bits(onp.float64) == 32 else 1e-3\n- jtu.check_grads(integrate, (y0, ts), modes=[\"rev\"], order=2,\n+\n+ self.check_against_scipy(decay, y0, ts, *args, tol=tol)\n+\n+ jtu.check_grads(integrate, (y0, ts, *args), modes=[\"rev\"], order=2,\nrtol=tol, atol=tol)\n@jtu.skip_on_devices(\"tpu\")\ndef test_swoop(self):\n- def swoop(y, t, arg1, arg2):\n- return np.array(y - np.sin(t) - np.cos(t) * arg1 + arg2)\n+ def swoop(_np, y, t, arg1, arg2):\n+ return _np.array(y - _np.sin(t) - _np.cos(t) * arg1 + arg2)\n+\n+ integrate = partial(odeint, partial(swoop, np))\nts = np.array([0.1, 0.2])\ntol = 1e-1 if num_float_bits(onp.float64) == 32 else 1e-3\ny0 = np.linspace(0.1, 0.9, 10)\n- integrate = lambda y0, ts: odeint(swoop, y0, ts, 0.1, 0.2)\n- jtu.check_grads(integrate, (y0, ts), modes=[\"rev\"], order=2,\n+ args = (0.1, 0.2)\n+ self.check_against_scipy(swoop, y0, ts, *args, tol=tol)\n+ jtu.check_grads(integrate, (y0, ts, *args), modes=[\"rev\"], order=2,\nrtol=tol, atol=tol)\nbig_y0 = np.linspace(1.1, 10.9, 10)\n- integrate = lambda y0, ts: odeint(swoop, big_y0, ts, 0.1, 0.3)\n- jtu.check_grads(integrate, (y0, ts), modes=[\"rev\"], order=2,\n+ args = (0.1, 0.3)\n+ self.check_against_scipy(swoop, y0, ts, *args, tol=tol)\n+ jtu.check_grads(integrate, (big_y0, ts, *args), modes=[\"rev\"], order=2,\nrtol=tol, atol=tol)\n+ def test_odeint_vmap_grad(self):\n+ # https://github.com/google/jax/issues/2531\n+\n+ def dx_dt(x, *args):\n+ return 0.1 * x\n+\n+ def f(x, y):\n+ y0 = np.array([x, y])\n+ t = np.array([0., 5.])\n+ y = odeint(dx_dt, y0, t)\n+ return y[-1].sum()\n+\n+ def g(x):\n+ # Two initial values for the ODE\n+ y0_arr = np.array([[x, 0.1],\n+ [x, 0.2]])\n+\n+ # Run ODE twice\n+ t = np.array([0., 5.])\n+ y = jax.vmap(lambda y0: odeint(dx_dt, y0, t))(y0_arr)\n+ return y[:,-1].sum()\n+\n+ ans = jax.grad(g)(2.) # don't crash\n+ expected = jax.grad(f, 0)(2., 0.1) + jax.grad(f, 0)(2., 0.2)\n+\n+ atol = {onp.float64: 5e-15}\n+ rtol = {onp.float64: 2e-15}\n+ self.assertAllClose(ans, expected, check_dtypes=False, atol=atol, rtol=rtol)\n+\nif __name__ == '__main__':\nabsltest.main()\n" } ]
Python
Apache License 2.0
google/jax
refactor ode tests, add scipy benchmark (#2824) * refactor ode tests, add scipy benchmark remove double import rename to scipy merge vmap test properly * clean up more global trace state after errors Co-authored-by: Matthew Johnson <mattjj@google.com>
260,299
28.04.2020 06:32:52
-3,600
75617be80320184ae6a4cf12eb62b11c30240898
Add population_count primitive to lax * add population_count primitive (needs new jaxlib) fixes * Add popcount docs * Add population_count to lax_reference * Use int prng (since we're only testing uints)
[ { "change_type": "MODIFY", "old_path": "docs/jax.lax.rst", "new_path": "docs/jax.lax.rst", "diff": "@@ -38,6 +38,7 @@ Operators\nbitwise_and\nbitwise_or\nbitwise_xor\n+ population_count\nbroadcast\nbroadcasted_iota\nbroadcast_in_dim\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -278,6 +278,10 @@ def bitwise_xor(x: Array, y: Array) -> Array:\nr\"\"\"Elementwise exclusive OR: :math:`x \\oplus y`.\"\"\"\nreturn xor_p.bind(x, y)\n+def population_count(x: Array) -> Array:\n+ r\"\"\"Elementwise popcount, count the number of set bits in each element.\"\"\"\n+ return population_count_p.bind(x)\n+\ndef add(x: Array, y: Array) -> Array:\nr\"\"\"Elementwise addition: :math:`x + y`.\"\"\"\nreturn add_p.bind(x, y)\n@@ -2023,6 +2027,8 @@ ad.defjvp_zero(or_p)\nxor_p = standard_naryop([_bool_or_int, _bool_or_int], 'xor')\nad.defjvp_zero(xor_p)\n+population_count_p = standard_unop(_bool_or_int, 'population_count')\n+\ndef _add_transpose(t, x, y):\n# The following linearity assertion is morally true, but because in some cases we\n# instantiate zeros for convenience, it doesn't always hold.\n" }, { "change_type": "MODIFY", "old_path": "jax/lax_reference.py", "new_path": "jax/lax_reference.py", "diff": "@@ -111,6 +111,31 @@ shift_left = onp.left_shift\nshift_right_arithmetic = onp.right_shift\n# TODO shift_right_logical\n+def population_count(x):\n+ assert x.dtype in (onp.uint32, onp.uint64)\n+ m = [\n+ 0x5555555555555555, # binary: 0101...\n+ 0x3333333333333333, # binary: 00110011..\n+ 0x0f0f0f0f0f0f0f0f, # binary: 4 zeros, 4 ones ...\n+ 0x00ff00ff00ff00ff, # binary: 8 zeros, 8 ones ...\n+ 0x0000ffff0000ffff, # binary: 16 zeros, 16 ones ...\n+ 0x00000000ffffffff, # binary: 32 zeros, 32 ones\n+ ]\n+\n+ if x.dtype == onp.uint32:\n+ m = list(map(onp.uint32, m[:-1]))\n+ else:\n+ m = list(map(onp.uint64, m))\n+\n+ x = (x & m[0]) + ((x >> 1) & m[0]) # put count of each 2 bits into those 2 bits\n+ x = (x & m[1]) + ((x >> 2) & m[1]) # put count of each 4 bits into those 4 bits\n+ x = (x & m[2]) + ((x >> 4) & m[2]) # put count of each 8 bits into those 8 bits\n+ x = (x & m[3]) + ((x >> 8) & m[3]) # put count of each 16 bits into those 16 bits\n+ x = (x & m[4]) + ((x >> 16) & m[4]) # put count of each 32 bits into those 32 bits\n+ if x.dtype == onp.uint64:\n+ x = (x & m[5]) + ((x >> 32) & m[5]) # put count of each 64 bits into those 64 bits\n+ return x\n+\neq = onp.equal\nne = onp.not_equal\nge = onp.greater_equal\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_test.py", "new_path": "tests/lax_test.py", "diff": "@@ -152,6 +152,7 @@ LAX_OPS = [\nop_record(\"bitwise_not\", 1, bool_dtypes, jtu.rand_small),\nop_record(\"bitwise_or\", 2, bool_dtypes, jtu.rand_small),\nop_record(\"bitwise_xor\", 2, bool_dtypes, jtu.rand_small),\n+ op_record(\"population_count\", 1, uint_dtypes, partial(jtu.rand_int, 1 << 32)),\nop_record(\"add\", 2, default_dtypes + complex_dtypes, jtu.rand_small),\nop_record(\"sub\", 2, default_dtypes + complex_dtypes, jtu.rand_small),\n" } ]
Python
Apache License 2.0
google/jax
Add population_count primitive to lax (#2753) * add population_count primitive (needs new jaxlib) fixes #2263 * Add popcount docs * Add population_count to lax_reference * Use int prng (since we're only testing uints) Co-authored-by: Matthew Johnson <mattjj@google.com>
260,287
28.04.2020 15:07:08
-7,200
5fe6b069dfc1b1b91f6870519c7f3923005068fd
Correct the order of .format arguments in vjp wrapper
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -1336,7 +1336,7 @@ def _vjp_pullback_wrapper(fun, cotangent_dtypes, io_tree, py_args):\nif in_tree != in_tree_expected:\nmsg = (\"Tree structure of cotangent input {}, does not match structure of \"\n\"primal output {}\")\n- raise TypeError(msg.format(in_tree_expected, in_tree))\n+ raise TypeError(msg.format(in_tree, in_tree_expected))\nfor a, dtype in safe_zip(args, cotangent_dtypes):\nif _dtype(a) != dtype:\nmsg = (\"Type of cotangent input to vjp pullback function ({}) does not \"\n" } ]
Python
Apache License 2.0
google/jax
Correct the order of .format arguments in vjp wrapper (#2866)
260,403
19.04.2020 11:49:15
25,200
dddad2a3dc0e36f76e98288f6b50ccd65fdebced
Add top_k jvp and batching rules
[ { "change_type": "MODIFY", "old_path": "docs/jax.lax.rst", "new_path": "docs/jax.lax.rst", "diff": "@@ -123,6 +123,7 @@ Operators\nsub\ntan\ntie_in\n+ top_k\ntranspose\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -1183,7 +1183,8 @@ def sort_key_val(keys: Array, values: Array, dimension: int = -1) -> Array:\nsorted_keys, sorted_values = result\nreturn sorted_keys, sorted_values\n-def top_k(operand: Array, k: int) -> Array:\n+def top_k(operand: Array, k: int) -> Tuple[Array, Array]:\n+ \"\"\"Returns top ``k`` values and their indices along the last axis of ``operand``.\"\"\"\nk = int(k)\nif k < 0:\nraise ValueError(\"k argument to top_k must be nonnegative, got {}\".format(k))\n@@ -4618,12 +4619,53 @@ def _top_k_abstract_eval(operand, *, k):\nreturn (ShapedArray(shape, operand.dtype),\nShapedArray(shape, onp.dtype(onp.int32)))\n+def _top_k_jvp(primals, tangents, *, k):\n+ operand, = primals\n+ tangent, = tangents\n+ primals_out = top_k(operand, k)\n+ if tangent is ad_util.zero:\n+ tangents_out = (ad_util.zero, ad_util.zero)\n+ else:\n+ _, k_idxs = primals_out\n+ idx_shape = k_idxs.shape\n+ rank = len(idx_shape)\n+ gather_index_shape = idx_shape + (1,)\n+ gather_indices = []\n+ for i in range(rank-1):\n+ _iota = iota(k_idxs.dtype, idx_shape[i])\n+ _iota = tie_in(operand, _iota)\n+ _iota = broadcast_in_dim(_iota, gather_index_shape, (i,))\n+ gather_indices.append(_iota)\n+ gather_indices.append(reshape(k_idxs, gather_index_shape))\n+ gather_indices = concatenate(gather_indices, dimension=rank)\n+ slice_sizes = (1,) * rank\n+ dnums = GatherDimensionNumbers(\n+ offset_dims=(),\n+ collapsed_slice_dims=tuple(range(rank)),\n+ start_index_map=tuple(range(rank)))\n+ tangents_out = (gather(tangent, gather_indices, dnums, slice_sizes),\n+ ad_util.zero)\n+ return primals_out, tangents_out\n+\n+def _top_k_batch_rule(batched_args, batch_dims, *, k):\n+ operand, = batched_args\n+ bdim, = batch_dims\n+ if bdim == operand.ndim-1:\n+ perm = onp.arange(operand.ndim)\n+ perm[bdim-1], perm[bdim] = perm[bdim], perm[bdim-1]\n+ top_k_v, top_k_i = top_k(transpose(operand, perm), k=k)\n+ return (transpose(top_k_v, perm),\n+ transpose(top_k_i, perm)), (bdim, bdim)\n+ else:\n+ return top_k(operand, k=k), (bdim, bdim)\n+\ntop_k_p = Primitive('top_k')\ntop_k_p.multiple_results = True\ntop_k_p.def_impl(partial(xla.apply_primitive, top_k_p))\ntop_k_p.def_abstract_eval(_top_k_abstract_eval)\nxla.translations[top_k_p] = partial(standard_translate, 'top_k')\n-\n+ad.primitive_jvps[top_k_p] = _top_k_jvp\n+batching.primitive_batchers[top_k_p] = _top_k_batch_rule\ndef _tie_in_transpose_rule(t):\nreturn [ad_util.zero, t]\n" }, { "change_type": "MODIFY", "old_path": "jax/test_util.py", "new_path": "jax/test_util.py", "diff": "@@ -634,6 +634,13 @@ def rand_int(low, high=None):\nreturn randint(low, high=high, size=shape, dtype=dtype)\nreturn fn\n+def rand_unique_int():\n+ randchoice = npr.RandomState(0).choice\n+ def fn(shape, dtype):\n+ return randchoice(onp.arange(onp.prod(shape), dtype=dtype),\n+ size=shape, replace=False)\n+ return fn\n+\ndef rand_bool():\nrng = npr.RandomState(0)\ndef generator(shape, dtype):\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_test.py", "new_path": "tests/lax_test.py", "diff": "@@ -2487,6 +2487,22 @@ class LaxAutodiffTest(jtu.JaxTestCase):\nfun = lambda keys, values: lax.sort_key_val(keys, values, axis)\ncheck_grads(fun, (keys, values), 2, [\"fwd\", \"rev\"], 1e-2, 1e-2, 1e-2)\n+ @parameterized.named_parameters(jtu.cases_from_list(\n+ {\"testcase_name\": \"_shape={}_k={}\".format(\n+ jtu.format_shape_dtype_string(shape, dtype), k),\n+ \"rng_factory\": rng_factory, \"shape\": shape, \"dtype\": dtype, \"k\": k}\n+ for dtype in [onp.float32,]\n+ for shape in [(4,), (5, 5), (2, 1, 4)]\n+ for k in [1, 3]\n+ for rng_factory in [jtu.rand_default]))\n+ def testTopKGrad(self, shape, dtype, k, rng_factory):\n+ rng = rng_factory()\n+ perm_rng = onp.random.RandomState(0)\n+ flat_values = onp.arange(onp.prod(shape, dtype=int), dtype=dtype)\n+ values = perm_rng.permutation(flat_values).reshape(shape)\n+ fun = lambda vs: lax.top_k(vs, k=k)[0]\n+ check_grads(fun, (values,), 2, [\"fwd\", \"rev\"], eps=1e-2)\n+\n@parameterized.named_parameters(jtu.cases_from_list(\n{\"testcase_name\": \"_shape={}_idxs={}_axes={}\".format(\njtu.format_shape_dtype_string(shape, dtype), idxs, axes),\n@@ -3220,6 +3236,27 @@ class LaxVmapTest(jtu.JaxTestCase):\nout_shape = lax.broadcast_shapes(shape1, shape2)\nself.assertTrue(all(type(s) is int for s in out_shape))\n+ @parameterized.named_parameters(jtu.cases_from_list(\n+ {\"testcase_name\": \"_shape={}_k={}_bdims={}\".format(\n+ jtu.format_shape_dtype_string(shape, dtype), k, bdims),\n+ \"shape\": shape, \"dtype\": dtype, \"k\": k, \"bdims\": bdims, \"rng_factory\": rng_factory}\n+ for shape in [(4,), (3, 4, 5)]\n+ for k in [1, 3]\n+ for bdims in all_bdims(shape)\n+ # TODO(b/155170120): test with repeats once the XLA:CPU stable top_k bug is fixed:\n+ # The top_k indices for integer arrays with identical entries won't match between\n+ # vmap'd version and manual reference, so only test unique integer arrays for int_dtypes.\n+ for dtype, rng_factory in itertools.chain(\n+ zip(float_dtypes, itertools.repeat(jtu.rand_default)),\n+ zip(int_dtypes, itertools.repeat(jtu.rand_unique_int)))))\n+ def testTopK(self, shape, dtype, k, bdims, rng_factory):\n+ rng = rng_factory()\n+ # _CheckBatching doesn't work with tuple outputs, so test outputs separately.\n+ op1 = lambda x: lax.top_k(x, k=k)[0]\n+ self._CheckBatching(op1, 5, bdims, (shape,), (dtype,), rng)\n+ op2 = lambda x: lax.top_k(x, k=k)[1]\n+ self._CheckBatching(op2, 5, bdims, (shape,), (dtype,), rng)\n+\n# TODO Concatenate\n# TODO Reverse\n# TODO DynamicSlice\n" } ]
Python
Apache License 2.0
google/jax
Add top_k jvp and batching rules
260,403
28.04.2020 10:49:17
25,200
e599a254223f911a7d885c1a1bd4963f60a2e150
fix sort_key_val return type annotation, docstring
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -1181,7 +1181,9 @@ def sort(operand: Array, dimension: int = -1) -> Array:\n\"\"\"\nreturn sort_p.bind(operand, dimension=dimension)\n-def sort_key_val(keys: Array, values: Array, dimension: int = -1) -> Array:\n+def sort_key_val(keys: Array, values: Array,\n+ dimension: int = -1) -> Tuple[Array, Array]:\n+ \"\"\"Sorts ``keys`` along ``dimension`` and applies same permutation to ``values``.\"\"\"\n# TODO(mattjj): new sort_key_val is variadic\nresult = sort_key_val_p.bind(keys, values, dimension=dimension)\nsorted_keys, sorted_values = result\n" } ]
Python
Apache License 2.0
google/jax
fix sort_key_val return type annotation, docstring
260,458
28.04.2020 19:34:27
-3,600
2a0637a1b89142a255eb68190812bb18cc57b8d2
add spacing to numpy.gradient
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -1021,16 +1021,16 @@ def diff(a, n=1, axis=-1,):\nreturn a\n-@partial(jit, static_argnums=1)\n-def _gradient(a, axis):\n- def gradient_along_axis(a, axis):\n+@partial(jit, static_argnums=(1, 2))\n+def _gradient(a, varargs, axis):\n+ def gradient_along_axis(a, h, axis):\nsliced = partial(lax.slice_in_dim, a, axis=axis)\na_grad = concatenate((\n- sliced(1, 2) - sliced(0, 1),\n- (sliced(2, None) - sliced(0, -2)) * 0.5,\n- sliced(-1, None) - sliced(-2, -1),\n+ (sliced(1, 2) - sliced(0, 1)), # upper edge\n+ (sliced(2, None) - sliced(None, -2)) * 0.5, # inner\n+ (sliced(-1, None) - sliced(-2, -1)), # lower edge\n), axis)\n- return a_grad\n+ return a_grad / h\nif axis is None:\naxis = range(a.ndim)\n@@ -1039,14 +1039,32 @@ def _gradient(a, axis):\naxis = (axis,)\nif not isinstance(axis, tuple) and not isinstance(axis, list):\nraise ValueError(\"Give `axis` either as int or iterable\")\n+ elif len(axis) == 0:\n+ return []\naxis = [_canonicalize_axis(i, a.ndim) for i in axis]\nif min([s for i, s in enumerate(a.shape) if i in axis]) < 2:\n- raise ValueError(\n- \"Shape of array too small to calculate a numerical gradient\")\n+ raise ValueError(\"Shape of array too small to calculate \"\n+ \"a numerical gradient, \"\n+ \"at least 2 elements are required.\")\n+ len_axes = len(axis)\n+ n = len(varargs)\n+ if n == 0 or varargs is None:\n+ # no spacing\n+ dx = [1.0] * len_axes\n+ elif n == 1:\n+ # single value for all axes\n+ dx = varargs * len_axes\n+ elif n == len_axes:\n+ dx = varargs\n+ else:\n+ TypeError(\"Invalid number of spacing arguments %d\" % n)\n+\n+ if ndim(dx[0]) != 0:\n+ raise NotImplementedError(\"Non-constant spacing not implemented\")\n# TODO: use jax.lax loop tools if possible\n- a_grad = [gradient_along_axis(a, ax) for ax in axis]\n+ a_grad = [gradient_along_axis(a, h, ax) for ax, h in zip(axis, dx)]\nif len(axis) == 1:\na_grad = a_grad[0]\n@@ -1057,11 +1075,9 @@ def _gradient(a, axis):\n@_wraps(onp.gradient)\ndef gradient(a, *args, **kwargs):\naxis = kwargs.pop(\"axis\", None)\n- if not len(args) == 0:\n- raise ValueError(\"*args (sample distances) not implemented\")\nif not len(kwargs) == 0:\nraise ValueError(\"Only `axis` keyword is implemented\")\n- return _gradient(a, axis)\n+ return _gradient(a, args, axis)\n@_wraps(onp.isrealobj)\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -2992,20 +2992,23 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\n@parameterized.named_parameters(\njtu.cases_from_list(\n- {\"testcase_name\": (\"_shape={}_axis={}_dtype={}\").format(shape, axis, dtype),\n+ {\"testcase_name\": (\"_shape={}_varargs={} axis={}_dtype={}\")\n+ .format(shape, varargs, axis, dtype),\n\"shape\": shape,\n+ \"varargs\": varargs,\n\"axis\": axis,\n\"dtype\": dtype, \"rng_factory\": rng_factory}\nfor shape in [(10,), (10, 15), (10, 15, 20)]\nfor _num_axes in range(len(shape))\n+ for varargs in itertools.combinations(range(1, len(shape) + 1), _num_axes)\nfor axis in itertools.combinations(range(len(shape)), _num_axes)\nfor dtype in inexact_dtypes\nfor rng_factory in [jtu.rand_default]))\n- def testGradient(self, shape, axis, dtype, rng_factory):\n+ def testGradient(self, shape, varargs, axis, dtype, rng_factory):\nrng = rng_factory()\nargs_maker = self._GetArgsMaker(rng, [shape], [dtype])\n- jnp_fun = lambda y: jnp.gradient(y, axis=axis)\n- onp_fun = lambda y: onp.gradient(y, axis=axis)\n+ jnp_fun = lambda y: jnp.gradient(y, *varargs, axis=axis)\n+ onp_fun = lambda y: onp.gradient(y, *varargs, axis=axis)\nself._CheckAgainstNumpy(\nonp_fun, jnp_fun, args_maker, check_dtypes=False)\nself._CompileAndCheck(jnp_fun, args_maker, check_dtypes=True)\n" } ]
Python
Apache License 2.0
google/jax
add spacing to numpy.gradient (#2545)
260,568
28.04.2020 22:23:03
-10,800
56f6294e377cb4e03c8e1a6fe82dade0c965d617
Implement nanargmin-max and add tests
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/Common_Gotchas_in_JAX.ipynb", "new_path": "docs/notebooks/Common_Gotchas_in_JAX.ipynb", "diff": "}\n]\n},\n+ {\n+ \"cell_type\": \"markdown\",\n+ \"metadata\": {},\n+ \"source\": [\n+ \"Note that due to this behavior np.nanargmin and np.nanargmax return -1 for slices consisting of NaNs whereas Numpy would throw an error.\"\n+ ]\n+ },\n{\n\"cell_type\": \"markdown\",\n\"metadata\": {\n" }, { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -2830,6 +2830,21 @@ def argmax(a, axis=None):\nreturn _argminmax(max, a, axis)\n+_NANARG_DOC = \"\"\"\\\n+Warning: jax.numpy.arg{} returns -1 for all-NaN slices and does not raise\n+an error.\n+\"\"\"\n+\n+@_wraps(onp.nanargmax, lax_description=_NANARG_DOC.format(\"max\"))\n+def nanargmax(a, axis=None):\n+ if not issubdtype(_dtype(a), inexact):\n+ return argmax(a, axis=axis)\n+ nan_mask = isnan(a)\n+ a = where(nan_mask, -inf, a)\n+ res = argmax(a, axis=axis)\n+ return where(all(nan_mask, axis=axis), -1, res)\n+\n+\n@_wraps(onp.argmin)\ndef argmin(a, axis=None):\nif axis is None:\n@@ -2838,6 +2853,16 @@ def argmin(a, axis=None):\nreturn _argminmax(min, a, axis)\n+@_wraps(onp.nanargmin, lax_description=_NANARG_DOC.format(\"min\"))\n+def nanargmin(a, axis=None):\n+ if not issubdtype(_dtype(a), inexact):\n+ return argmin(a, axis=axis)\n+ nan_mask = isnan(a)\n+ a = where(nan_mask, inf, a)\n+ res = argmin(a, axis=axis)\n+ return where(all(nan_mask, axis=axis), -1, res)\n+\n+\n# TODO(mattjj): redo this lowering with a call to variadic lax.reduce\ndef _argminmax(op, a, axis):\nshape = [1] * a.ndim\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -294,6 +294,8 @@ JAX_REDUCER_NO_DTYPE_RECORDS = [\nJAX_ARGMINMAX_RECORDS = [\nop_record(\"argmin\", 1, all_dtypes, nonempty_shapes, jtu.rand_some_equal, []),\nop_record(\"argmax\", 1, all_dtypes, nonempty_shapes, jtu.rand_some_equal, []),\n+ op_record(\"nanargmin\", 1, all_dtypes, nonempty_shapes, jtu.rand_some_nan, []),\n+ op_record(\"nanargmax\", 1, all_dtypes, nonempty_shapes, jtu.rand_some_nan, []),\n]\nJAX_OPERATOR_OVERLOADS = [\n@@ -675,6 +677,8 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nrng = rng_factory()\nif dtype == onp.complex128 and jtu.device_under_test() == \"gpu\":\nraise unittest.SkipTest(\"complex128 reductions not supported on GPU\")\n+ if \"nan\" in onp_op.__name__ and dtype == jnp.bfloat16:\n+ raise unittest.SkipTest(\"NumPy doesn't correctly handle bfloat16 arrays\")\ndef onp_fun(array_to_reduce):\nreturn onp_op(array_to_reduce, axis).astype(jnp.int_)\n@@ -683,7 +687,13 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nreturn jnp_op(array_to_reduce, axis)\nargs_maker = lambda: [rng(shape, dtype)]\n+ try:\nself._CheckAgainstNumpy(onp_fun, jnp_fun, args_maker, check_dtypes=True)\n+ except ValueError as e:\n+ if str(e) == \"All-NaN slice encountered\":\n+ self.skipTest(\"JAX doesn't support checking for all-NaN slices\")\n+ else:\n+ raise\nself._CompileAndCheck(jnp_fun, args_maker, check_dtypes=True)\n@parameterized.named_parameters(jtu.cases_from_list(\n" } ]
Python
Apache License 2.0
google/jax
Implement nanargmin-max and add tests (#2398) Co-authored-by: vlad <veryfakemail@ya.ru>
260,571
29.04.2020 12:25:18
-19,080
ef963f06ae5947fa05e6d4b54fa47af127626173
Add ReLU6, Hard sigmoid, swish
[ { "change_type": "MODIFY", "old_path": "jax/nn/functions.py", "new_path": "jax/nn/functions.py", "diff": "@@ -266,3 +266,33 @@ def one_hot(x, num_classes, *, dtype=np.float64):\nx = np.asarray(x)\nreturn np.array(x[..., np.newaxis] == np.arange(num_classes, dtype=x.dtype),\ndtype=dtype)\n+\n+def relu6(x):\n+ r\"\"\"Rectified Linear Unit 6 activation function.\n+\n+ Computes the element-wise function\n+\n+ .. math::\n+ \\mathrm{relu6}(x) = \\min(\\max(x, 0), 6)\n+ \"\"\"\n+ return np.minimum(np.maximum(x, 0), 6.)\n+\n+def hard_sigmoid(x):\n+ r\"\"\"Hard Sigmoid activation function.\n+\n+ Computes the element-wise function\n+\n+ .. math::\n+ \\mathrm{hard\\_sigmoid}(x) = \\frac{\\mathrm{relu6}(x + 3)}{6}\n+ \"\"\"\n+ return relu6(x + 3.) / 6.\n+\n+def hard_swish(x):\n+ r\"\"\"Hard Swish activation function\n+\n+ Computes the element-wise function\n+\n+ .. math::\n+ \\mathrm{hard\\_swish}(x) = x \\cdot \\mathrm{hard\\_sigmoid}(x)\n+ \"\"\"\n+ return x * hard_sigmoid(x)\n" } ]
Python
Apache License 2.0
google/jax
Add ReLU6, Hard sigmoid, swish (#2709)
260,529
29.04.2020 03:16:49
14,400
52c69e88c58e3838a605eb952d9bbf3ad6195f89
Fix slices in Gated Linear Unit activation
[ { "change_type": "MODIFY", "old_path": "jax/nn/functions.py", "new_path": "jax/nn/functions.py", "diff": "@@ -189,7 +189,7 @@ def glu(x, axis=-1):\n\"\"\"Gated linear unit activation function.\"\"\"\nsize = x.shape[axis]\nassert size % 2 == 0, \"axis size must be divisible by 2\"\n- return x[..., :size] * sigmoid(x[..., size:])\n+ return x[..., :size // 2] * sigmoid(x[..., size // 2:])\n# other functions\n" } ]
Python
Apache License 2.0
google/jax
Fix slices in Gated Linear Unit activation (#2341)
260,700
29.04.2020 22:18:21
14,400
1f7ebabfc8445cba9b26f46275bbac61b8bce151
add jets for sines fns refactor remove duplicate
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jet.py", "new_path": "jax/experimental/jet.py", "diff": "@@ -354,6 +354,26 @@ def _div_taylor_rule(primals_in, series_in, **params):\nreturn primal_out, series_out\njet_rules[lax.div_p] = _div_taylor_rule\n+def _sinusoidal_rule(sign, prims, primals_in, series_in):\n+ x, = primals_in\n+ series, = series_in\n+ u = [x] + series\n+ s, c = prims\n+ s = [s(x)] + [None] * len(series)\n+ c = [c(x)] + [None] * len(series)\n+ for k in range(1, len(s)):\n+ s[k] = fact(k-1) * sum(_scale(k, j) * u[j] * c[k-j] for j in range(1, k + 1))\n+ c[k] = fact(k-1) * sum(_scale(k, j) * u[j] * s[k-j] for j in range(1, k + 1)) * sign\n+ return (s[0], s[1:]), (c[0], c[1:])\n+\n+def _get_ind(f, ind):\n+ return lambda *args: f(*args)[ind]\n+\n+jet_rules[lax.sin_p] = _get_ind(partial(_sinusoidal_rule, -1, (lax.sin, lax.cos)), 0)\n+jet_rules[lax.cos_p] = _get_ind(partial(_sinusoidal_rule, -1, (lax.sin, lax.cos)), 1)\n+jet_rules[lax.sinh_p] = _get_ind(partial(_sinusoidal_rule, 1, (lax.sinh, lax.cosh)), 0)\n+jet_rules[lax.cosh_p] = _get_ind(partial(_sinusoidal_rule, 1, (lax.sinh, lax.cosh)), 1)\n+\ndef _bilinear_taylor_rule(prim, primals_in, series_in, **params):\nx, y = primals_in\nx_terms, y_terms = series_in\n" }, { "change_type": "MODIFY", "old_path": "tests/jet_test.py", "new_path": "tests/jet_test.py", "diff": "@@ -210,6 +210,14 @@ class JetTest(jtu.JaxTestCase):\n@jtu.skip_on_devices(\"tpu\")\ndef test_expm1(self): self.unary_check(np.expm1)\n@jtu.skip_on_devices(\"tpu\")\n+ def test_sin(self): self.unary_check(np.sin)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_cos(self): self.unary_check(np.cos)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_sinh(self): self.unary_check(np.sinh)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_cosh(self): self.unary_check(np.cosh)\n+ @jtu.skip_on_devices(\"tpu\")\ndef test_tanh(self): self.unary_check(np.tanh, lims=[-500, 500], order=5)\n@jtu.skip_on_devices(\"tpu\")\ndef test_expit(self): self.unary_check(jax.scipy.special.expit, lims=[-500, 500], order=5)\n" } ]
Python
Apache License 2.0
google/jax
add jets for sines fns (#2892) refactor remove duplicate
260,411
30.04.2020 10:16:14
-10,800
b39da1f842f1363b8b36052c0837407de0be9c2d
Fix jit with device placement In setups with multiple backends, a jit happens on the default backend, unless we give a `backend` parameter. This is true even if the inputs are committed to a device on the non-default backend, or if we pass a `device` parameter to jit.
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/xla.py", "new_path": "jax/interpreters/xla.py", "diff": "@@ -482,6 +482,7 @@ def _xla_callable(fun: lu.WrappedFun, device, backend, name, *arg_specs):\nnreps = jaxpr_replicas(jaxpr)\ndevice = _xla_callable_device(nreps, backend, device, arg_devices)\n+ backend = device.platform if device else backend\nresult_handlers = tuple(map(partial(_pval_to_result_handler, device), pvals))\n# Computations that only produce constants and/or only rearrange their inputs,\n" }, { "change_type": "MODIFY", "old_path": "tests/multibackend_test.py", "new_path": "tests/multibackend_test.py", "diff": "@@ -144,6 +144,33 @@ class MultiBackendTest(jtu.JaxTestCase):\nself.assertEqual(api.jit(f, backend='cpu')().device_buffer.device(),\napi.devices('cpu')[0])\n+ @jtu.skip_on_devices(\"cpu\") # test only makes sense on non-cpu backends\n+ def test_jit_on_nondefault_backend(self):\n+ cpus = api.devices(\"cpu\")\n+ self.assertNotEmpty(cpus)\n+\n+ # Since we are not on CPU, some other backend will be the default\n+ default_dev = api.devices()[0]\n+ self.assertNotEqual(default_dev.platform, \"cpu\")\n+\n+ data_on_cpu = api.device_put(1, device=cpus[0])\n+ self.assertEqual(data_on_cpu.device_buffer.device(), cpus[0])\n+\n+ def my_sin(x): return np.sin(x)\n+ # jit without any device spec follows the data\n+ result1 = api.jit(my_sin)(2)\n+ self.assertEqual(result1.device_buffer.device(), default_dev)\n+ result2 = api.jit(my_sin)(data_on_cpu)\n+ self.assertEqual(result2.device_buffer.device(), cpus[0])\n+\n+ # jit with `device` spec places the data on the specified device\n+ result3 = api.jit(my_sin, device=cpus[0])(2)\n+ self.assertEqual(result1.device_buffer.device(), cpus[0])\n+\n+ # jit with `backend` spec places the data on the specified backend\n+ result4 = api.jit(my_sin, backend=\"cpu\")(2)\n+ self.assertEqual(result1.device_buffer.device(), cpus[0])\n+\nif __name__ == \"__main__\":\nabsltest.main()\n" } ]
Python
Apache License 2.0
google/jax
Fix jit with device placement (#2883) In setups with multiple backends, a jit happens on the default backend, unless we give a `backend` parameter. This is true even if the inputs are committed to a device on the non-default backend, or if we pass a `device` parameter to jit.
260,411
30.04.2020 19:16:05
-10,800
8d4b6857adbd52c3093f034b35c23a6dc2e20492
Fix typo in tests; caught on GPU and TPU
[ { "change_type": "MODIFY", "old_path": "tests/multibackend_test.py", "new_path": "tests/multibackend_test.py", "diff": "@@ -165,11 +165,11 @@ class MultiBackendTest(jtu.JaxTestCase):\n# jit with `device` spec places the data on the specified device\nresult3 = api.jit(my_sin, device=cpus[0])(2)\n- self.assertEqual(result1.device_buffer.device(), cpus[0])\n+ self.assertEqual(result3.device_buffer.device(), cpus[0])\n# jit with `backend` spec places the data on the specified backend\nresult4 = api.jit(my_sin, backend=\"cpu\")(2)\n- self.assertEqual(result1.device_buffer.device(), cpus[0])\n+ self.assertEqual(result4.device_buffer.device(), cpus[0])\nif __name__ == \"__main__\":\n" } ]
Python
Apache License 2.0
google/jax
Fix typo in tests; caught on GPU and TPU (#2902)
260,411
01.05.2020 09:16:31
-10,800
2e9047d3884451294874fd485209d3218ee968fb
Add flag to enable checking, and turn on checking in tests. Fix an error in check_jaxpr.
[ { "change_type": "MODIFY", "old_path": "jax/config.py", "new_path": "jax/config.py", "diff": "@@ -123,4 +123,13 @@ class NameSpace(object):\nconfig = Config()\nflags = config\n+FLAGS = flags.FLAGS\n+\nalready_configured_with_absl = False\n+\n+flags.DEFINE_bool(\n+ 'jax_enable_checks',\n+ bool_env('JAX_ENABLE_CHECKS', False),\n+ help=\n+ 'Turn on invariant checking (core.skip_checks = False)'\n+)\n\\ No newline at end of file\n" }, { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -28,6 +28,7 @@ from typing import Any, Callable, ClassVar, Dict, List, Optional, Set\nimport numpy as onp\nfrom . import dtypes\n+from .config import FLAGS\nfrom . import linear_util as lu\nfrom .util import safe_zip, safe_map, partial, curry, prod, partialmethod\n@@ -35,8 +36,19 @@ from .pprint_util import pp, vcat, hcat, pp_kv_pairs, PrettyPrint\n# TODO(dougalm): the trace cache breaks the leak detector. Consisder solving.\ncheck_leaks = False\n-# TODO(dougalm): put this behind a flag that's enabled during testing\n-skip_checks = True # not __debug__ # google doesn't use -O\n+\n+\"\"\"Disables internal invariant checks.\"\"\"\n+skip_checks = not FLAGS.jax_enable_checks # not __debug__ # google doesn't use -O\n+\n+@contextmanager\n+def skipping_checks():\n+ \"\"\"Context manager for temporarily disabling checks.\"\"\"\n+ global skip_checks\n+ old_value, skip_checks = skip_checks, True\n+ try:\n+ yield\n+ finally:\n+ skip_checks = old_value\nzip = safe_zip\nmap = safe_map\n@@ -653,7 +665,10 @@ class Bot(AbstractValue): pass\nbot = Bot()\nclass AbstractUnit(AbstractValue):\n- def join(self, other): return self\n+ def join(self, other):\n+ if not skip_checks:\n+ assert other is abstract_unit, other\n+ return self\ndef _eq(self, self_traced, other): return get_aval(other) is self\nabstract_unit = AbstractUnit()\n@@ -1048,7 +1063,7 @@ def check_jaxpr(jaxpr: Jaxpr):\nmap(write, jaxpr.constvars)\nmap(write, jaxpr.invars)\nfor eqn in jaxpr.eqns:\n- if eqn.primitive.call_primitive or eqn.map_primitive:\n+ if eqn.primitive.call_primitive or eqn.primitive.map_primitive:\nif \"call_jaxpr\" not in eqn.params:\nraise Exception(\"Call primitive {} should have a 'call_jaxpr' parameter\"\n.format(eqn.primitive))\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/partial_eval.py", "new_path": "jax/interpreters/partial_eval.py", "diff": "@@ -25,6 +25,7 @@ import numpy as onp\nfrom .. import core\nfrom .. import linear_util as lu\nfrom ..abstract_arrays import ShapedArray, ConcreteArray, raise_to_shaped\n+from ..ad_util import zero\nfrom ..util import (unzip2, safe_zip, safe_map, toposort, partial, split_list,\nwrap_name, cache)\nfrom ..core import (Trace, Tracer, new_master, Jaxpr, Literal, get_aval,\n@@ -49,7 +50,7 @@ class PartialVal(tuple):\nif not core.skip_checks:\n# type checks\nassert isinstance(pv, (AbstractValue, type(None))), xs\n- assert isinstance(const, core.Tracer) or core.valid_jaxtype(const), xs\n+ assert isinstance(const, core.Tracer) or const is zero or core.valid_jaxtype(const), xs\n# invariant checks\nif isinstance(pv, AbstractValue):\nassert const == core.unit, xs\n" }, { "change_type": "MODIFY", "old_path": "jax/test_util.py", "new_path": "jax/test_util.py", "diff": "@@ -708,6 +708,9 @@ class JaxTestCase(parameterized.TestCase):\n# def tearDown(self) -> None:\n# assert core.reset_trace_state()\n+ def setUp(self):\n+ core.skip_checks = False\n+\ndef assertArraysAllClose(self, x, y, check_dtypes, atol=None, rtol=None):\n\"\"\"Assert that x and y are close (up to numerical tolerances).\"\"\"\nself.assertEqual(x.shape, y.shape)\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -3075,7 +3075,8 @@ class DeprecatedCustomTransformsTest(jtu.JaxTestCase):\napi.grad(lambda x, y: f(x, y)[0])(1., 2.) # doesn't crash\ndef test_custom_transforms_vjp_nones(self):\n- # issue rasied by jsnoek@ and jumper@\n+ core.skip_checks = True # Fails with checks\n+ # issue raised by jsnoek@ and jumper@\n@jax.custom_transforms\ndef solve(a, b):\nreturn np.dot(np.linalg.inv(a), b)\n" }, { "change_type": "MODIFY", "old_path": "tests/dtypes_test.py", "new_path": "tests/dtypes_test.py", "diff": "@@ -24,6 +24,7 @@ from absl.testing import parameterized\nimport numpy as onp\nimport jax\n+from jax import core\nfrom jax import dtypes\nfrom jax import numpy as np\nfrom jax import test_util as jtu\n@@ -167,6 +168,8 @@ class DtypesTest(jtu.JaxTestCase):\nA = 42\nB = 101\nonp.testing.assert_equal(onp.array(42), onp.array(AnEnum.A))\n+ with core.skipping_checks():\n+ # Passing AnEnum.A to np.array fails the type check in bind\nonp.testing.assert_equal(np.array(42), np.array(AnEnum.A))\nonp.testing.assert_equal(onp.int32(101), onp.int32(AnEnum.B))\nonp.testing.assert_equal(np.int32(101), np.int32(AnEnum.B))\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_test.py", "new_path": "tests/lax_test.py", "diff": "@@ -2643,7 +2643,7 @@ class LaxAutodiffTest(jtu.JaxTestCase):\nexpected = onp.array(0.0)\nself.assertAllClose(ans, expected, check_dtypes=False)\n- with self.assertRaises(TypeError):\n+ with self.assertRaises(TypeError if core.skip_checks else AssertionError):\nlax.stop_gradient(lambda x: x)\n# TODO(mattjj): make this a more systematic test\n" }, { "change_type": "MODIFY", "old_path": "tests/nn_test.py", "new_path": "tests/nn_test.py", "diff": "@@ -22,6 +22,7 @@ from absl.testing import parameterized\nimport numpy as onp\n+from jax import core\nfrom jax import test_util as jtu\nfrom jax.test_util import check_grads\nfrom jax import nn\n@@ -92,11 +93,13 @@ class NNFunctionsTest(jtu.JaxTestCase):\n@jtu.skip_on_devices(\"gpu\", \"tpu\")\ndef testEluMemory(self):\n# see https://github.com/google/jax/pull/1640\n+ with core.skipping_checks(): # With checks we materialize the array\njax.make_jaxpr(nn.elu)(np.ones((10 ** 12,))) # don't oom\n@jtu.skip_on_devices(\"gpu\", \"tpu\")\ndef testHardTanhMemory(self):\n# see https://github.com/google/jax/pull/1640\n+ with core.skipping_checks(): # With checks we materialize the array\njax.make_jaxpr(nn.hard_tanh)(np.ones((10 ** 12,))) # don't oom\ndef testOneHot(self):\n" } ]
Python
Apache License 2.0
google/jax
Add flag to enable checking, and turn on checking in tests. (#2900) Fix an error in check_jaxpr.
260,335
01.05.2020 13:47:46
25,200
49a89016740b6fc7166d5f523707b0638bc34486
skip failing shapecheck tests cc
[ { "change_type": "MODIFY", "old_path": "tests/masking_test.py", "new_path": "tests/masking_test.py", "diff": "@@ -271,11 +271,15 @@ class ShapesTest(jtu.JaxTestCase):\nreturn x[x.shape[0] - 1:]\ndef test_iota(self):\n+ raise SkipTest(\"not yet implemented\")\n+ # https://travis-ci.org/github/google/jax/jobs/682086351\n@shapecheck(['n'], 'n')\ndef range_like(x):\nreturn lax.iota(np.int32, x.shape[0])\ndef test_arange(self):\n+ raise SkipTest(\"not yet implemented\")\n+ # https://travis-ci.org/github/google/jax/jobs/682086351\n@shapecheck(['n'], 'n')\ndef arange_like(x):\nreturn np.arange(x.shape[0], dtype=np.int32)\n" } ]
Python
Apache License 2.0
google/jax
skip failing shapecheck tests cc @JuliusKunze
260,335
01.05.2020 14:39:30
25,200
2263899d3c76f367b46ff013085e8f4f330ed168
replace accidental use of jax.numpy.min w/ builtin
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -1044,7 +1044,7 @@ def _gradient(a, varargs, axis):\nreturn []\naxis = [_canonicalize_axis(i, a.ndim) for i in axis]\n- if min([s for i, s in enumerate(a.shape) if i in axis]) < 2:\n+ if _min([s for i, s in enumerate(a.shape) if i in axis]) < 2:\nraise ValueError(\"Shape of array too small to calculate \"\n\"a numerical gradient, \"\n\"at least 2 elements are required.\")\n" } ]
Python
Apache License 2.0
google/jax
replace accidental use of jax.numpy.min w/ builtin
260,700
01.05.2020 20:10:20
14,400
a821e67d607dbcc530c5b57cd6175e20f7b07c12
instantiate zeros fix dtype remove TODO
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jet.py", "new_path": "jax/experimental/jet.py", "diff": "@@ -22,7 +22,7 @@ from jax import core\nfrom jax.util import unzip2\nfrom jax import ad_util\nfrom jax.tree_util import (register_pytree_node, tree_structure,\n- treedef_is_leaf, tree_flatten, tree_unflatten)\n+ treedef_is_leaf, tree_flatten, tree_unflatten, tree_map)\nimport jax.linear_util as lu\nfrom jax.interpreters import xla\nfrom jax.lax import lax\n@@ -59,6 +59,8 @@ def jet_fun(primals, series):\nwith core.new_master(JetTrace) as master:\nout_primals, out_terms = yield (master, primals, series), {}\ndel master\n+ out_terms = [tree_map(lambda x: onp.zeros_like(x, dtype=onp.result_type(out_primals[0])), series[0])\n+ if s is zero_series else s for s in out_terms]\nyield out_primals, out_terms\n@lu.transformation\n" }, { "change_type": "MODIFY", "old_path": "tests/jet_test.py", "new_path": "tests/jet_test.py", "diff": "@@ -61,10 +61,6 @@ class JetTest(jtu.JaxTestCase):\nself.assertAllClose(y, expected_y, atol=atol, rtol=rtol,\ncheck_dtypes=check_dtypes)\n- # TODO(duvenaud): Lower zero_series to actual zeros automatically.\n- if terms == zero_series:\n- terms = tree_map(np.zeros_like, expected_terms)\n-\nself.assertAllClose(terms, expected_terms, atol=atol, rtol=rtol,\ncheck_dtypes=check_dtypes)\n@@ -86,10 +82,6 @@ class JetTest(jtu.JaxTestCase):\nself.assertAllClose(y, expected_y, atol=atol, rtol=rtol,\ncheck_dtypes=check_dtypes)\n- # TODO(duvenaud): Lower zero_series to actual zeros automatically.\n- if terms == zero_series:\n- terms = tree_map(np.zeros_like, expected_terms)\n-\nself.assertAllClose(terms, expected_terms, atol=atol, rtol=rtol,\ncheck_dtypes=check_dtypes)\n@@ -291,6 +283,21 @@ class JetTest(jtu.JaxTestCase):\nseries_in = (terms_b, terms_x, terms_y)\nself.check_jet(np.where, primals, series_in)\n+ def test_inst_zero(self):\n+ def f(x):\n+ return 2.\n+ def g(x):\n+ return 2. + 0 * x\n+ x = np.ones(1)\n+ order = 3\n+ f_out_primals, f_out_series = jet(f, (x, ), ([np.ones_like(x) for _ in range(order)], ))\n+ assert f_out_series is not zero_series\n+\n+ g_out_primals, g_out_series = jet(g, (x, ), ([np.ones_like(x) for _ in range(order)], ))\n+\n+ assert g_out_primals == f_out_primals\n+ assert g_out_series == f_out_series\n+\nif __name__ == '__main__':\nabsltest.main()\n" } ]
Python
Apache License 2.0
google/jax
instantiate zeros (#2924) fix dtype remove TODO
260,335
02.05.2020 10:25:53
25,200
64f12a42463f90bdd35e5c969028a0e539e253e4
improve docs and error message for odeint *args cf.
[ { "change_type": "MODIFY", "old_path": "jax/experimental/ode.py", "new_path": "jax/experimental/ode.py", "diff": "@@ -28,6 +28,7 @@ import operator as op\nimport jax\nimport jax.numpy as np\n+from jax import core\nfrom jax import lax\nfrom jax import ops\nfrom jax.util import safe_map, safe_zip\n@@ -141,7 +142,9 @@ def odeint(func, y0, t, *args, rtol=1.4e-8, atol=1.4e-8, mxstep=np.inf):\ny0: array or pytree of arrays representing the initial value for the state.\nt: array of float times for evaluation, like `np.linspace(0., 10., 101)`,\nin which the values must be strictly increasing.\n- *args: tuple of additional arguments for `func`.\n+ *args: tuple of additional arguments for `func`, which must be arrays\n+ scalars, or (nested) standard Python containers (tuples, lists, dicts,\n+ namedtuples, i.e. pytrees) of those types.\nrtol: float, relative local error tolerance for solver (optional).\natol: float, absolute local error tolerance for solver (optional).\nmxstep: int, maximum number of steps to take for each timepoint (optional).\n@@ -151,6 +154,12 @@ def odeint(func, y0, t, *args, rtol=1.4e-8, atol=1.4e-8, mxstep=np.inf):\npoint in `t`, represented as an array (or pytree of arrays) with the same\nshape/structure as `y0` except with a new leading axis of length `len(t)`.\n\"\"\"\n+ def _check_arg(arg):\n+ if not isinstance(arg, core.Tracer) and not core.valid_jaxtype(arg):\n+ msg = (\"The contents of odeint *args must be arrays or scalars, but got \"\n+ \"\\n{}.\")\n+ raise TypeError(msg.format(arg))\n+ tree_map(_check_arg, args)\nreturn _odeint_wrapper(func, rtol, atol, mxstep, y0, t, *args)\n@partial(jax.jit, static_argnums=(0, 1, 2, 3))\n" } ]
Python
Apache License 2.0
google/jax
improve docs and error message for odeint *args (#2931) cf. #2920
260,380
04.05.2020 11:20:21
-3,600
525235d8c976d60ffecf4a017b6b528565a9cc7b
Fix a codeblock in the "understanding jaxpr" doc. This fixes an issue where the codeblock didn't render properly on the website.
[ { "change_type": "MODIFY", "old_path": "docs/jaxpr.rst", "new_path": "docs/jaxpr.rst", "diff": "Understanding Jaxprs\n====================\n-Updated: February 14, 2020 (for commit 9e6fe64).\n+Updated: May 3, 2020 (for commit f1a46fe).\n(Note: the code examples in this file can be seed also in\n``jax/tests/api_test::JaxprTest.testExamplesJaxprDoc``.)\n@@ -173,22 +173,17 @@ ConstVars arise when the computation ontains array constants, either\nfrom the Python program, or from constant-folding. For example, the function\n``func6`` below\n-.. testcode::\n-\n- def func5(first, second):\n- temp = first + jnp.sin(second) * 3. - jnp.ones(8)\n- return temp\n-\n- def func6(first):\n- return func5(first, jnp.ones(8))\n-\n- print(make_jaxpr(func6)(jnp.ones(8)))\n-\n+>>> def func5(first, second):\n+... temp = first + jnp.sin(second) * 3. - jnp.ones(8)\n+... return temp\n+...\n+>>> def func6(first):\n+... return func5(first, jnp.ones(8))\n+...\nJAX produces the following jaxpr\n-.. testoutput::\n-\n+>>> print(make_jaxpr(func6)(jnp.ones(8)))\n{ lambda b d ; a.\nlet c = add a b\ne = sub c d\n" } ]
Python
Apache License 2.0
google/jax
Fix a codeblock in the "understanding jaxpr" doc. (#2942) This fixes an issue where the codeblock didn't render properly on the website.
260,270
04.05.2020 19:02:13
-3,600
04102e5b9d98f2f79d3f4b5c1b9bb21216ff4c3e
Allow ConvDimensionNumbers to be passed into conv_transpose
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -520,9 +520,6 @@ def conv_general_dilated(\n(for a 2D convolution).\n\"\"\"\ndnums: ConvDimensionNumbers\n- if isinstance(dimension_numbers, ConvDimensionNumbers):\n- dnums = dimension_numbers\n- else:\ndnums = conv_dimension_numbers(lhs.shape, rhs.shape, dimension_numbers)\nif lhs_dilation is None:\nlhs_dilation = (1,) * (lhs.ndim - 2)\n@@ -5055,13 +5052,16 @@ def conv_dimension_numbers(lhs_shape, rhs_shape, dimension_numbers):\nArgs:\nlhs_shape: tuple of nonnegative integers, shape of the convolution input.\nrhs_shape: tuple of nonnegative integers, shape of the convolution kernel.\n- dimension_numbers: None or a tuple/list of strings, following the\n- convolution dimension number specification format in xla_client.py.\n+ dimension_numbers: None or a tuple/list of strings or a ConvDimensionNumbers\n+ object following the convolution dimension number specification format in\n+ xla_client.py.\nReturns:\nA `ConvDimensionNumbers` object that represents `dimension_numbers` in the\ncanonical form used by lax functions.\n\"\"\"\n+ if isinstance(dimension_numbers, ConvDimensionNumbers):\n+ return dimension_numbers\nif len(lhs_shape) != len(rhs_shape):\nmsg = \"convolution requires lhs and rhs ndim to be equal, got {} and {}.\"\nraise TypeError(msg.format(len(lhs_shape), len(rhs_shape)))\n" } ]
Python
Apache License 2.0
google/jax
Allow ConvDimensionNumbers to be passed into conv_transpose (#2915)
260,568
05.05.2020 04:29:55
-10,800
3e522373a00124eec8c996bc1ed89937095e3d86
Raise an error in np.var when array is complex and dtype is not
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -1646,6 +1646,14 @@ def var(a, axis=None, dtype=None, out=None, ddof=0, keepdims=False):\na_dtype = _dtype(a)\nif dtype:\n+ if (not issubdtype(dtype, complexfloating) and\n+ issubdtype(a_dtype, complexfloating)):\n+ msg = (\"jax.numpy.var does not yet support real dtype parameters when \"\n+ \"computing the variance of an array of complex values. The \"\n+ \"semantics of numpy.var seem unclear in this case. Please comment \"\n+ \"on https://github.com/google/jax/issues/2283 if this behavior is \"\n+ \"important to you.\")\n+ raise ValueError(msg)\na_dtype = promote_types(a_dtype, dtype)\nelse:\nif not issubdtype(a_dtype, inexact):\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -2665,6 +2665,10 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\njnp_fun = partial(jnp.var, dtype=out_dtype, axis=axis, ddof=ddof, keepdims=keepdims)\ntol = jtu.tolerance(out_dtype, {onp.float16: 1e-1, onp.float32: 1e-3,\nonp.float64: 1e-3, onp.complex128: 1e-6})\n+ if (jnp.issubdtype(dtype, jnp.complexfloating) and\n+ not jnp.issubdtype(out_dtype, jnp.complexfloating)):\n+ self.assertRaises(ValueError, lambda: jnp_fun(*args_maker()))\n+ else:\nself._CheckAgainstNumpy(onp_fun, jnp_fun, args_maker, check_dtypes=True,\ntol=tol)\nself._CompileAndCheck(jnp_fun, args_maker, check_dtypes=True, rtol=tol,\n" } ]
Python
Apache License 2.0
google/jax
Raise an error in np.var when array is complex and dtype is not (#2288) Co-authored-by: vlad <veryfakemail@ya.ru>
260,285
05.05.2020 05:12:43
-7,200
e4d8cacfc60a41ae1acbe312ae420c86d36325cb
Fix tests for random.categorical with multi-dimensional logits
[ { "change_type": "MODIFY", "old_path": "tests/random_test.py", "new_path": "tests/random_test.py", "diff": "@@ -246,8 +246,8 @@ class LaxRandomTest(jtu.JaxTestCase):\nfor (p, axis) in [\n([.25] * 4, -1),\n([.1, .2, .3, .4], -1),\n- ([[.25, .25], [.1, .9]], 1),\n- ([[.25, .1], [.25, .9]], 0),\n+ ([[.5, .5], [.1, .9]], 1),\n+ ([[.5, .1], [.5, .9]], 0),\n]\nfor sample_shape in [(10000,), (5000, 2)]\nfor dtype in [onp.float32, onp.float64]))\n@@ -255,25 +255,24 @@ class LaxRandomTest(jtu.JaxTestCase):\nkey = random.PRNGKey(0)\np = onp.array(p, dtype=dtype)\nlogits = onp.log(p) - 42 # test unnormalized\n- shape = sample_shape + tuple(onp.delete(logits.shape, axis))\n+ out_shape = tuple(onp.delete(logits.shape, axis))\n+ shape = sample_shape + out_shape\nrand = lambda key, p: random.categorical(key, logits, shape=shape, axis=axis)\ncrand = api.jit(rand)\nuncompiled_samples = rand(key, p)\ncompiled_samples = crand(key, p)\n- if p.ndim > 1:\n- self.skipTest(\"multi-dimensional categorical tests are currently broken!\")\n-\n- for samples in [uncompiled_samples, compiled_samples]:\nif axis < 0:\naxis += len(logits.shape)\n+ for samples in [uncompiled_samples, compiled_samples]:\nassert samples.shape == shape\n-\n+ samples = np.reshape(samples, (10000,) + out_shape)\nif len(p.shape[:-1]) > 0:\n- for cat_index, p_ in enumerate(p):\n- self._CheckChiSquared(samples[:, cat_index], pmf=lambda x: p_[x])\n+ ps = onp.transpose(p, (1, 0)) if axis == 0 else p\n+ for cat_samples, cat_p in zip(samples.transpose(), ps):\n+ self._CheckChiSquared(cat_samples, pmf=lambda x: cat_p[x])\nelse:\nself._CheckChiSquared(samples, pmf=lambda x: p[x])\n" } ]
Python
Apache License 2.0
google/jax
Fix tests for random.categorical with multi-dimensional logits (#2955)
260,617
06.05.2020 16:05:49
14,400
f338b0b061b69d90d58f543a8df73790faf16603
Add jnp.unravel_index
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -1154,6 +1154,25 @@ def ravel(a, order=\"C\"):\nreturn reshape(a, (size(a),), order)\n+_UNRAVEL_INDEX_DOC = \"\"\"\\\n+Unlike numpy's implementation of unravel_index, negative indices are accepted\n+and out-of-bounds indices are clipped.\n+\"\"\"\n+\n+@_wraps(onp.unravel_index, lax_description=_UNRAVEL_INDEX_DOC)\n+def unravel_index(indices, shape):\n+ indices = asarray(indices)\n+ sizes = pad(shape, (0, 1), constant_values=1)\n+ cumulative_sizes = cumprod(sizes[::-1])[::-1]\n+ total_size = cumulative_sizes[0]\n+ # Clip so raveling and unraveling an oob index will not change the behavior\n+ clipped_indices = clip(indices, -total_size, total_size - 1)\n+ # Add enough trailing dims to avoid conflict with flat_index\n+ cumulative_sizes = cumulative_sizes.reshape([-1] + [1] * indices.ndim)\n+ idx = clipped_indices % cumulative_sizes[:-1] // cumulative_sizes[1:]\n+ return tuple(idx)\n+\n+\n@_wraps(onp.squeeze)\ndef squeeze(a, axis=None):\nshape_a = shape(a)\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -1991,6 +1991,25 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nargs_maker = lambda: [rng.randn(3, 4).astype(\"float32\")]\nself._CompileAndCheck(lambda x: x.ravel(), args_maker, check_dtypes=True)\n+ @parameterized.parameters(\n+ (0, (2, 1, 3)),\n+ (5, (2, 1, 3)),\n+ (0, ()),\n+ ([0, 1, 2], (2, 2)),\n+ ([[[0, 1], [2, 3]]], (2, 2)))\n+ def testUnravelIndex(self, flat_index, shape):\n+ self._CheckAgainstNumpy(\n+ onp.unravel_index,\n+ jnp.unravel_index,\n+ lambda: (flat_index, shape),\n+ check_dtypes=True\n+ )\n+\n+ def testUnravelIndexOOB(self):\n+ self.assertEqual(jnp.unravel_index(2, (2,)), (1,))\n+ self.assertEqual(jnp.unravel_index(-2, (2, 1, 3,)), (1, 0, 1))\n+ self.assertEqual(jnp.unravel_index(-3, (2,)), (0,))\n+\ndef testAstype(self):\nrng = onp.random.RandomState(0)\nargs_maker = lambda: [rng.randn(3, 4).astype(\"float32\")]\n" } ]
Python
Apache License 2.0
google/jax
Add jnp.unravel_index (#2966)
260,411
07.05.2020 13:28:24
-10,800
804e083e664746dd4fcf1507a959e750084af14a
Fix pytype for copybara import
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -632,7 +632,7 @@ def find_top_trace(args) -> Optional[Tracer]:\nraise TypeError(f\"Argument '{arg}' of type {type(arg)} is not a valid JAX type\")\nreturn top_so_far\n- top_trace = reduce(check_arg, args, None)\n+ top_trace = reduce(check_arg, args, None) # type: ignore[wrong-arg-types]\nif top_trace is not None:\nreturn type(top_trace)(top_trace.master, cur_sublevel()) # type: ignore[call-arg]\nelse:\n" } ]
Python
Apache License 2.0
google/jax
Fix pytype for copybara import (#2995)
260,411
07.05.2020 16:16:22
-10,800
970e475e0a6b61e76006e0cc75f13d0f0c5f5250
Undo strict checking of LAX primitives This undoes
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -1590,7 +1590,6 @@ def _check_args(args):\nraise TypeError(\"Argument '{}' of type {} is not a valid JAX type\"\n.format(arg, type(arg)))\n-# TODO(necula): this duplicates code in core.valid_jaxtype\ndef _valid_jaxtype(arg):\ntry:\nxla.abstractify(arg) # faster than core.get_aval\n" }, { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -17,7 +17,7 @@ import operator\nfrom operator import attrgetter\nfrom contextlib import contextmanager\nfrom collections import namedtuple\n-from functools import total_ordering, reduce\n+from functools import total_ordering\nimport itertools as it\nfrom weakref import ref\nimport threading\n@@ -204,6 +204,8 @@ class Primitive(object):\nreturn '{}'.format(self.name)\ndef bind(self, *args, **kwargs):\n+ assert skip_checks or all(isinstance(arg, Tracer)\n+ or valid_jaxtype(arg) for arg in args), args\ntop_trace = find_top_trace(args)\nif top_trace is None:\nreturn self.impl(*args, **kwargs)\n@@ -621,22 +623,14 @@ def full_lower(val):\nelse:\nreturn val\n-def find_top_trace(args) -> Optional[Tracer]:\n- \"\"\"Find the tracer with the highest-level, or None. \"\"\"\n- def check_arg(top_so_far: Optional[Tracer], arg) -> Optional[Tracer]:\n- if isinstance(arg, Tracer):\n- return (top_so_far\n- if top_so_far and top_so_far.level >= arg._trace.level else arg._trace)\n- # Raises error here for bind on LAX primitives\n- if not valid_jaxtype(arg):\n- raise TypeError(f\"Argument '{arg}' of type {type(arg)} is not a valid JAX type\")\n- return top_so_far\n-\n- top_trace = reduce(check_arg, args, None) # type: ignore[wrong-arg-types]\n- if top_trace is not None:\n- return type(top_trace)(top_trace.master, cur_sublevel()) # type: ignore[call-arg]\n- else:\n+def find_top_trace(xs):\n+ try:\n+ top_trace = max((x._trace for x in xs if isinstance(x, Tracer)),\n+ key=attrgetter('level'))\n+ except ValueError:\nreturn None\n+ else:\n+ return type(top_trace)(top_trace.master, cur_sublevel())\n@contextmanager\ndef initial_style_staging():\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -1199,7 +1199,7 @@ def top_k(operand: Array, k: int) -> Tuple[Array, Array]:\nreturn top_k_p.bind(operand, k=k)\ndef tie_in(x: Array, y: Array) -> Array:\n- \"\"\"Returns the value of ``y`` but with a fake data dependence on ``x``.\n+ \"\"\"Gives ``y`` a fake data dependence on ``x``.\nWhen staging to XLA (e.g. running under jit or pmap), values that don't depend\non computation inputs are computed op-by-op, and folded into the XLA\n@@ -1209,10 +1209,6 @@ def tie_in(x: Array, y: Array) -> Array:\nWhen staging to XLA and ``x`` is already staged, then the result of ``tie_in``\nis ``y``, but staged to XLA. Downstream use of the result will also be staged\nto XLA.\n-\n- For example, ``lax.sin(const)`` would be constant-folded if ``const`` is\n- a constant array, but ``lax.sin(lax.tie_in(x, const))``, will be staged to\n- XLA as long as ``x`` is staged to XLA.\n\"\"\"\nreturn tie_in_p.bind(x, y)\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_test.py", "new_path": "tests/lax_test.py", "diff": "@@ -2633,7 +2633,7 @@ class LaxAutodiffTest(jtu.JaxTestCase):\nexpected = onp.array(0.0)\nself.assertAllClose(ans, expected, check_dtypes=False)\n- with self.assertRaises(TypeError):\n+ with self.assertRaises(TypeError if core.skip_checks else AssertionError):\nlax.stop_gradient(lambda x: x)\n# TODO(mattjj): make this a more systematic test\n@@ -3250,10 +3250,6 @@ class LaxVmapTest(jtu.JaxTestCase):\n# TODO Collapse\n# TODO Scatter\n- def test_tie_in_error(self):\n- with self.assertRaisesRegex(TypeError,\n- \".*tuple.* is not a valid JAX type\"):\n- api.make_jaxpr(lambda x: lax.tie_in((x, x), 1))(1.)\nif __name__ == '__main__':\nabsltest.main()\n" } ]
Python
Apache License 2.0
google/jax
Undo strict checking of LAX primitives (#2996) This undoes d08dec5d20
260,411
25.04.2020 10:19:21
-7,200
a16584d28080c7e39c751b9721d5e2610df6e2a7
Fixed scan, and grad. Added multiplexing protocol.
[ { "change_type": "MODIFY", "old_path": ".travis.yml", "new_path": ".travis.yml", "diff": "@@ -26,6 +26,7 @@ before_install:\n- conda update -q conda\ninstall:\n- conda install --yes python=$TRAVIS_PYTHON_VERSION pip absl-py opt_einsum numpy scipy pytest-xdist pytest-benchmark mypy=0.770\n+ - pip install msgpack\n- if [ \"$JAX_ONLY_CHECK_TYPES\" = true ]; then\npip install pytype ;\nfi\n" }, { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -33,12 +33,15 @@ Implementation plan:\n* Explore a simpler API that uses Python program-order, instead of\ndata dependency-order. Need to add support to JAX for stateful primitives.\n\"\"\"\n+from concurrent import futures\nfrom contextlib import contextmanager\nfrom functools import partial\nimport io\nimport itertools\nfrom jax import abstract_arrays\n+from jax import ad_util\n+from jax import api\nfrom jax import core\nfrom jax import dtypes\nfrom jax import lax\n@@ -50,15 +53,17 @@ from jaxlib import xla_client\nfrom jaxlib import xla_extension\nimport logging\n+import msgpack # type: ignore\nimport numpy as onp\nimport os\nimport threading\n-from typing import Any, Dict, List, Optional, Sequence, Tuple\n+from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple\n# TODO(necula): fix mypy errors if I define the type aliases below\nXlaOp = Any # xla_extension.XlaOp\nXlaShape = Any # xla_client.Shape\nXlaComputationBuilder = Any # xla_bridge._JaxComputationBuilder\n+XlaDevice = Any # xla_client.Device\nid_print_p = core.Primitive(\"id_print\")\nid_print_p.multiple_results = True\n@@ -152,17 +157,31 @@ def _id_print_abstract_eval(*args_a: pe.AbstractValue, **params) \\\nid_print_p.def_abstract_eval(_id_print_abstract_eval)\n-\n-# Each array sent to the outfeed is preceeded by a descriptor, which\n-# is an array s32[32], with the format:\n-# [0]: special header value\n-# [1]: an encoding of the element_type()\n-# [2]: the number of dimensions\n-# [3:...]: the size of the dimensions\n-# padded with 0s\n+# The data on the outfeed follows a protocol that allows multiplexing the\n+# outfeed among multiple consumers, and communicates in-stream shape and\n+# type of the data.\n+# Each batch of array data is preceeded by one or more header messages.\n+# A header message is of type uint32[_OUTFEED_HEADER_LENGTH // 4], with the\n+# uint32 being the big-endian encoding of the following array of bytes:\n+# [0], [1]: special header values 21 and 78\n+# [2]: a consumer id (e.g., _OUTFEED_CONSUMER_ID_PRINT)\n+# [3], [4]: big-endian encoding of metadata length (up to 2**16). The\n+# metadata is a msgpack-encoded value of type:\n+# ([ (type_code, (d0, d1, ...)), ...], # for each array, element type code\n+# # and the dimensions.\n+# { ... }) # kwargs to be passed to the consumer\n+# padded with 0s to _OUTFEED_HEADER_LENGTH\n+#\n+# If the metadata is too long to fit in one header array, several more\n+# header arrays will follow, with identical content except for the metadata\n+# bytes.\n#\n-_OUTFEED_DESCRIPTOR_PRINT_HEADER = 13579\n-_OUTFEED_DESCRIPTOR_LENGTH = 32\n+_OUTFEED_HEADER_LENGTH = 64 # In bytes\n+_OUTFEED_HEADER_START0 = 21\n+_OUTFEED_HEADER_START1 = 78\n+_OUTFEED_HEADER_METADATA_LENGTH = _OUTFEED_HEADER_LENGTH - 3 - 2\n+_OUTFEED_CONSUMER_ID_PRINT = 31\n+\n_CODE_TO_DTYPE = {\n0: onp.dtype(onp.int8),\n1: onp.dtype(onp.int16),\n@@ -180,43 +199,106 @@ _CODE_TO_DTYPE = {\n_DTYPE_STR_TO_CODE = dict([(str(d), c) for c, d in _CODE_TO_DTYPE.items()])\n+def _emit_outfeed(comp: XlaComputationBuilder, token: XlaOp,\n+ consumer_id: int, arrays: Sequence[XlaOp], kwargs: Dict) -> XlaOp:\n+ \"\"\"Emits the arrays to the outfeed for the current device.\n-def _id_print_translation_rule_outfeed(\n- comp: XlaComputationBuilder,\n- *args_op: XlaOp, **params):\n-\n- # TODO: outfeed a whole tuple at once?\n- def _outfeed_one_array(a: XlaOp, token: XlaOp) -> XlaOp:\n- a_shape = comp.GetShape(a)\n- dimensions = a_shape.dimensions()\n- descriptor = [_OUTFEED_DESCRIPTOR_PRINT_HEADER,\n- _DTYPE_STR_TO_CODE[str(onp.dtype(a_shape.element_type()))],\n- len(dimensions)] + list(dimensions)\n- if len(descriptor) > _OUTFEED_DESCRIPTOR_LENGTH:\n- raise ValueError(f\"Too many dimensions in array to print: {a_shape}\")\n- descriptor += [0] * (_OUTFEED_DESCRIPTOR_LENGTH - len(descriptor))\n-\n- data = xops.ConstantLiteral(comp, onp.array(descriptor, dtype=onp.int32))\n+ The consumer_id, arrays, and kwargs will be passed to the receiver.\n+ \"\"\"\n+ arrays_shape = [comp.GetShape(a) for a in arrays]\n+ def _array_shape_to_tuple(a_shape: XlaShape):\n+ # (element_type_code, (d0, d1, ..., dn))\n+ return (_DTYPE_STR_TO_CODE[str(onp.dtype(a_shape.element_type()))],\n+ a_shape.dimensions())\n+ metadata = msgpack.dumps((tuple(map(_array_shape_to_tuple, arrays_shape)),\n+ kwargs))\n+ metadata_len = len(metadata)\n+ if len(metadata) > 0xffff:\n+ raise ValueError(\"Outfeed metadata too long\")\n+ metadatas = [metadata[i:i + _OUTFEED_HEADER_METADATA_LENGTH]\n+ for i in range(0, metadata_len, _OUTFEED_HEADER_METADATA_LENGTH)]\n+ for meta in metadatas:\n+ header = ((_OUTFEED_HEADER_START0, _OUTFEED_HEADER_START1,\n+ consumer_id,\n+ (metadata_len >> 8) & 0xff, metadata_len & 0xff) +\n+ tuple(meta))\n+ header += (0,) * (_OUTFEED_HEADER_LENGTH - len(header))\n+ # Encode as uint32\n+ header_uint32 = [int.from_bytes(header[i:i+4], byteorder=\"big\")\n+ for i in range(0, _OUTFEED_HEADER_LENGTH, 4)]\n+ data = xops.ConstantLiteral(comp, onp.array(header_uint32, dtype=onp.uint32))\ntoken = xops.OutfeedWithToken(data, token, comp.GetShape(data))\n+\n+ # Now send the arrays\n+ for a, a_shape in zip(arrays, arrays_shape):\ntoken = xops.OutfeedWithToken(a, token, a_shape)\nreturn token\n+def _receive_outfeed(device: XlaDevice, receiver_name: str\n+ ) -> Tuple[int, List, Dict]:\n+ \"\"\"Receives a set of arrays on the outfeed for the specificied device.\n+ Args:\n+ receiver_name: a name used for debugging and logging\n+ Returns: a tuple with the consumer_id, the arrays received, and\n+ a kwargs dictionary that was passed to _emit_outfeed.\n+ \"\"\"\n+ platform = xla_client.get_local_backend(None).platform\n+ header_shape = xla_client.Shape.array_shape(onp.dtype(onp.uint32),\n+ (_OUTFEED_HEADER_LENGTH // 4,))\n+\n+ def _get_data(data_shape: XlaShape, device: XlaDevice) -> XlaShape:\n+ if platform in (\"gpu\", \"cpu\"):\n+ return xla_client.transfer_from_outfeed(data_shape, device)\n+ else:\n+ return xla_client.transfer_from_outfeed(\n+ xla_client.Shape.tuple_shape((data_shape,)), device)[0]\n+\n+ metadatas: List[bytes] = []\n+ remaining_metadata_length = 0\n+ while(True):\n+ header_uint32 = _get_data(header_shape, device)\n+ header = [b for h in header_uint32\n+ for b in int(h).to_bytes(4, byteorder=\"big\")]\n+ if header[0] != _OUTFEED_HEADER_START0 or header[1] != _OUTFEED_HEADER_START1:\n+ raise ValueError(f\"Read unexpected outfeed header {header[0:2]} [{receiver_name}]\")\n+ logging.info(f\"[{receiver_name}:{device}] Outfeed read header: {header}\")\n+ consumer_id = header[2]\n+ metadata_length = (header[3] << 8) + header[4]\n+ if not metadatas: # First header packet\n+ remaining_metadata_length = metadata_length\n+ if remaining_metadata_length <= _OUTFEED_HEADER_METADATA_LENGTH: # All here\n+ metadatas.append(bytes(header[5:5 + remaining_metadata_length]))\n+ break\n+ else:\n+ metadatas.append(bytes(header[5:5 + _OUTFEED_HEADER_METADATA_LENGTH]))\n+ remaining_metadata_length -= _OUTFEED_HEADER_METADATA_LENGTH\n+\n+ array_descriptors, kwargs = msgpack.unpackb(b\"\".join(metadatas))\n+ arrays = []\n+ for a_descr in array_descriptors:\n+ a_shape = xla_client.Shape.array_shape(_CODE_TO_DTYPE[a_descr[0]],\n+ a_descr[1])\n+ data = _get_data(a_shape, device)\n+ logging.info(f\"[{receiver_name}:{device}] Outfeed read data of shape \"\n+ f\"{data.dtype}{data.shape}\")\n+ arrays.append(data)\n+ return (consumer_id, arrays, kwargs)\n+\n+def _id_print_translation_rule_outfeed(\n+ comp: XlaComputationBuilder,\n+ *args_op: XlaOp, **params):\n+\nprev_token = xla.computation_state_carry.current_token(comp)\nnr_args_to_emit = len(args_op) - params.get(\"nr_results\", 0)\n- for i, a_op in enumerate(args_op):\n- if i < nr_args_to_emit:\n- prev_token = _outfeed_one_array(a_op, prev_token)\n- xla.computation_state_carry.set_current_token(comp, prev_token)\n+ next_token = _emit_outfeed(comp, prev_token,\n+ _OUTFEED_CONSUMER_ID_PRINT,\n+ args_op[0:nr_args_to_emit], {})\n+ xla.computation_state_carry.set_current_token(comp, next_token)\nreturn xops.Tuple(comp, args_op)\nxla.translations[id_print_p] = _id_print_translation_rule_outfeed\n-\n-# TODO: find a better way to signal the end of printing\n-_END_PRINTING = onp.int32(12345678)\n-def end_printing(res):\n- return id_print(_END_PRINTING, result=res)\n-\n+_END_PRINTING = onp.int32(987654321)\n@contextmanager\ndef print_receiver(output_stream=None,\nreceiver_name=\"\",\n@@ -229,54 +311,37 @@ def print_receiver(output_stream=None,\njax.jit(func)(args)\n\"\"\"\n- # TODO: start receivers for each device\n- platform = xla_client.get_local_backend(None).platform\n- def _get_data(data_shape: XlaShape) -> XlaShape:\n- if platform == \"gpu\":\n- return xla_client.transfer_from_outfeed(data_shape)\n- else:\n- return xla_client.transfer_from_outfeed(\n- xla_client.Shape.tuple_shape((data_shape,)))[0]\n-\n- def _consume_one_array_from_outfeed():\n- descriptor_shape = xla_client.Shape.array_shape(onp.dtype(onp.int32),\n- (_OUTFEED_DESCRIPTOR_LENGTH,))\n- descriptor = _get_data(descriptor_shape)\n-\n- logging.info(f\"[{receiver_name}] Read descriptor: {descriptor}\")\n- if descriptor[0] != _OUTFEED_DESCRIPTOR_PRINT_HEADER:\n- raise ValueError(f\"Read unexpected print descriptor {descriptor} [{receiver_name}]\")\n- data_dimensions = tuple(descriptor[3:3+descriptor[2]])\n- data_shape = xla_client.Shape.array_shape(_CODE_TO_DTYPE[descriptor[1]],\n- data_dimensions)\n- data = _get_data(data_shape)\n- logging.info(f\"[{receiver_name}] Read data of shape {data.dtype}{data.shape}\")\n- return data\n-\n- def receiver_loop():\n+ # TODO: pass the backend?\n+ devices = api.devices()\n+ executor = futures.ThreadPoolExecutor(thread_name_prefix=\"outfeed\",\n+ max_workers=len(devices))\n+\n+ def device_receiver_loop(device: XlaDevice) -> XlaDevice:\ni = 0\nwhile (True):\n- got = _consume_one_array_from_outfeed()\n- if not got.shape and got == _END_PRINTING:\n- logging.info(f\"[{receiver_name}] Received END_PRINTING\")\n- return\n- got_str = onp.array2string(got, threshold=1024)\n- logging.info(f\"[{receiver_name}] Received {i} ({got.dtype}{got.shape}): {got_str}\")\n+ consumer_id, arrays, kwargs = _receive_outfeed(device, receiver_name)\n+ if consumer_id != _OUTFEED_CONSUMER_ID_PRINT:\n+ raise NotImplementedError(f\"Encountered unexpected consumer {consumer_id}\")\n+ for a in arrays:\n+ if not a.shape and a == _END_PRINTING:\n+ logging.info(f\"[{receiver_name}:{device}] Outfeed received END_PRINTING\")\n+ return device\n+ a_str = onp.array2string(a, threshold=1024)\n+ logging.info(f\"[{receiver_name}:{device}] Outfeed received {i} \"\n+ f\"({a.dtype}{a.shape}): {a_str}\")\nif output_stream is not None:\n- output_stream.write(got_str)\n+ output_stream.write(a_str)\ni += 1\n- receiver = threading.Thread(target=receiver_loop)\n- receiver.start()\n+ receiver_futures = [executor.submit(device_receiver_loop, d) for d in devices]\ntry:\n- yield receiver\n+ yield\nfinally:\n- # TODO: proper termination\n- receiver.join(timeout=timeout_sec)\n- if receiver.is_alive():\n- logging.error(f\"[{receiver_name}] Receiver still alive\")\n- else:\n- logging.info(f\"[{receiver_name}] Receiver finished\")\n+ for d in devices:\n+ api.jit(lambda x: id_print(_END_PRINTING, result=x), device=d)(0)\n+ for f in futures.as_completed(receiver_futures, timeout=timeout_sec):\n+ finished_device = f.result()\n+ logging.info(f\"[{receiver_name}:{finished_device} Outfeed receiver finished\")\ndef _id_print_jvp_rule(primals, tangents, **params):\n@@ -291,7 +356,9 @@ ad.primitive_jvps[id_print_p] = _id_print_jvp_rule\ndef _id_print_transpose_rule(cts, *args, **params):\nassert all([ad.is_undefined_primal(x) for x in args])\nassert len(cts) == len(args)\n- ct_args = id_print_p.bind(*cts,\n+ cts_zeros = [ad.instantiate_zeros_aval(a.aval, ct)\n+ for a, ct in zip(args, cts)]\n+ ct_args = id_print_p.bind(*cts_zeros,\n**_expand_params_transform(params, \"transpose\"))\nreturn ct_args\n" }, { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -53,9 +53,9 @@ class _TestingOutputStream(object):\ndef write(self, what: str) -> None:\n# Sometimes we get floating points in the output; we round them\ndef repl(match_group):\n- # TODO: why can't we use here np.around?\nmatched = match_group.group(0)\nif matched == \".\": return matched\n+ # TODO: why can't we use here np.around?\nx = onp.around(float(matched), decimals=2)\nreturn f\"{x:.2f}\"\n@@ -80,11 +80,11 @@ testing_stream = _TestingOutputStream()\ndef fun1(a):\ny = hcb.id_print(a * 2., what=\"a * 2\", output_stream=testing_stream)\ny = hcb.id_print(y * 3., what=\"y * 3\", output_stream=testing_stream, result=y)\n- return y**4 # Some computation to make the gradient interesting\n+ return y**2 # Some computation to make the gradient interesting\ndef fun1_equiv(a): # Numerical equivalent of fun`\n- return (a * 2.)**4\n+ return (a * 2.)**2\nclass HostCallbackTest(jtu.JaxTestCase):\n@@ -107,32 +107,6 @@ class HostCallbackTest(jtu.JaxTestCase):\n# Clear any cached backends so new CPU backend will pick up the env var.\nxla_bridge.get_backend.cache_clear()\n- def helper_print_serialization(self, description, np_vals, params):\n- \"\"\"Encode a print_descriptor and print it.\n-\n- Args:\n- np_vals: a list of np.ndarray from which to extract the shapes.\n- \"\"\"\n- encoded = hcb._make_id_print_metadata(\n- [xla.aval_to_xla_shape(pe.get_aval(np_val)) for np_val in np_vals],\n- params)\n- print(f\"test_serialization: {description}\")\n- print(\", \".join([f\"{b}\" for b in encoded]))\n-\n- def test_serialization(self):\n- \"\"\"Prints encodings used in host_callback_test::TestParseDescriptor.\"\"\"\n- raise SkipTest(\"Not implemented\")\n- self.helper_print_serialization(\"no args, separator=sep, param=0\", [],\n- dict(param=0, separator=\"sep\"))\n- self.helper_print_serialization(\"1 scalar int, separator= param=1\",\n- [np.int32(0)], dict(param=1, separator=\"\"))\n- self.helper_print_serialization(\"1 array f32[2, 3], separator= param=2\",\n- [np.ones((2, 3), dtype=np.float32)],\n- dict(param=2, separator=\"\"))\n- self.helper_print_serialization(\n- \"1 array f32[2, 3] and 1 f64, separator= param=3\",\n- [np.ones((2, 3), dtype=np.float32),\n- np.float64(0)], dict(param=3, separator=\"\"))\ndef test_with_tuple_result(self):\n@@ -164,11 +138,11 @@ class HostCallbackTest(jtu.JaxTestCase):\ne f = id_print[ nr_results=1\noutput_stream=TestingOutputStream\nwhat=y * 3 ] d c\n- g = pow f 4.0\n+ g = pow f 2.0\nin (g,) }\"\"\", str(api.make_jaxpr(fun1)(5.)))\nself.assertEqual(\"\", testing_stream.output)\n- self.assertEqual((5. * 2.)**4, fun1(5.))\n+ self.assertEqual((5. * 2.)**2, fun1(5.))\nself.assertMultiLineStrippedEqual(\n\"\"\"\n(10.00,) {'what': 'a * 2'}\n@@ -176,8 +150,8 @@ class HostCallbackTest(jtu.JaxTestCase):\ntesting_stream.reset()\ndef test_jit_simple(self):\n- jit_fun1 = api.jit(lambda x: hcb.end_printing(3. * hcb.id_print(\n- 2. * x, what=\"here\")))\n+ jit_fun1 = api.jit(lambda x: 3. * hcb.id_print(\n+ 2. * x, what=\"here\"))\nself.assertMultiLineStrippedEqual(\n\"\"\"\n{ lambda ; a.\n@@ -186,8 +160,7 @@ class HostCallbackTest(jtu.JaxTestCase):\nlet b = mul a 2.0\nc = id_print[ what=here ] b\nd = mul c 3.0\n- e f = id_print[ nr_results=1 ] 12345678 d\n- in (f,) }\n+ in (d,) }\ndevice=None\nname=<lambda> ] a\nin (b,) }\"\"\", str(api.make_jaxpr(jit_fun1)(5.)))\n@@ -196,17 +169,17 @@ class HostCallbackTest(jtu.JaxTestCase):\nwith hcb.print_receiver(output_stream=testing_stream,\nreceiver_name=self._testMethodName):\nres = jit_fun1(5.)\n+\nself.assertAllClose(6. * 5., res, check_dtypes=True)\nself.assertMultiLineStrippedEqual(\n\"\"\"\n10.00\"\"\", testing_stream.output)\ntesting_stream.reset()\n- def test_simple_jit_sequencing(self):\n+ def test_jit_sequence1(self):\ndef func(x):\nx1 = hcb.id_print(x, where=\"1\")\n- x2 = hcb.id_print(x1 + 1, where=\"2\")\n- return hcb.end_printing(x2)\n+ return hcb.id_print(x1 + 1, where=\"2\")\nlogging.info(\"%s: %s\", self._testMethodName,\napi.make_jaxpr(func)(1))\n@@ -233,8 +206,6 @@ class HostCallbackTest(jtu.JaxTestCase):\nreceiver_name=self._testMethodName):\nself.assertEqual(2, api.jit(func)(1))\nself.assertEqual(11, api.jit(func)(10))\n- # Now send the end of printing\n- api.jit(lambda x: hcb.end_printing(x))(0)\nself.assertMultiLineStrippedEqual(\n\"\"\"\n@@ -248,24 +219,40 @@ class HostCallbackTest(jtu.JaxTestCase):\ndef func(x):\nx1 = hcb.id_print(x, where=\"1\")\ndef func_nested(x):\n- x2 = hcb.id_print(x, where=\"nested\")\n+ x2 = hcb.id_print(x + 1, where=\"nested\")\nreturn x2\nx3 = api.jit(func_nested)(x1)\n- x2 = hcb.id_print(x3 + 1, where=\"2\")\n- return hcb.end_printing(x2)\n+ return hcb.id_print(x3 + 1, where=\"2\")\nlogging.warning(\"%s: %s\", self._testMethodName,\napi.make_jaxpr(func)(1))\nlogging.warning(\"%s: %s\", self._testMethodName,\napi.xla_computation(func)(1).GetHloText())\n- return\nwith hcb.print_receiver(output_stream=testing_stream,\nreceiver_name=self._testMethodName):\n- pass # self.assertEqual(2, api.jit(func)(1))\n+ self.assertEqual(3, api.jit(func)(1))\nself.assertMultiLineStrippedEqual(\n\"\"\"\n1\n-2\"\"\", testing_stream.output)\n+2\n+3\"\"\", testing_stream.output)\n+ testing_stream.reset()\n+\n+ def test_jit_devices(self):\n+ \"\"\"Running on multiple devices.\"\"\"\n+ devices = api.local_devices()\n+ def func(x, device_id):\n+ x1 = hcb.id_print(x, dev=str(device_id))\n+ x2 = hcb.id_print(x1 + 1, dev=str(device_id))\n+ return x2\n+\n+ with hcb.print_receiver(output_stream=testing_stream,\n+ receiver_name=self._testMethodName):\n+ for d in devices:\n+ self.assertEqual(2, api.jit(func, device=d, static_argnums=1)(1, d.id))\n+\n+ self.assertEqual(len(devices), len(re.findall(r\"1\", testing_stream.output)))\n+ self.assertEqual(len(devices), len(re.findall(r\"2\", testing_stream.output)))\ntesting_stream.reset()\ndef test_jit_cond1(self):\n@@ -278,7 +265,7 @@ class HostCallbackTest(jtu.JaxTestCase):\nx2 + 1, lambda x: hcb.id_print(x, where=\"cond_t\"),\nx2 + 1, lambda x: hcb.id_print(-1, where=\"cond_f\", result=x))\nx5 = hcb.id_print(x4 + 1, where=\"w.2\")\n- return hcb.end_printing(x5)\n+ return x5\nlogging.warning(\"%s: %s\", self._testMethodName, api.make_jaxpr(func)(1))\nlogging.warning(\"%s: %s\", self._testMethodName,\n@@ -307,7 +294,7 @@ class HostCallbackTest(jtu.JaxTestCase):\nreturn hcb.id_print(x4 + 1, where=\"w.2\")\nx10 = lax.while_loop(lambda x: x < 10, body, x2)\nres = hcb.id_print(x10, where=\"10\")\n- return hcb.end_printing(res)\n+ return res\nlogging.warning(\"%s: %s\", self._testMethodName, api.make_jaxpr(func)(1))\nlogging.warning(\"%s: %s\", self._testMethodName,\napi.xla_computation(func)(1).GetHloText())\n@@ -334,6 +321,45 @@ class HostCallbackTest(jtu.JaxTestCase):\n10\"\"\", testing_stream.output)\ntesting_stream.reset()\n+ def test_jit_scan_cond(self):\n+ def func(x):\n+ x1 = hcb.id_print(x, where=\"1\")\n+ x2 = hcb.id_print(x1 + 1, where=\"2\")\n+\n+ def body(c, x):\n+ x3 = hcb.id_print(x, where=\"s.1\")\n+ x4 = lax.cond(x % 2 == 0,\n+ x3 + 1, lambda x: hcb.id_print(x, where=\"s.t\"),\n+ x3 + 1, lambda x: hcb.id_print(-1, where=\"s.f\", result=x))\n+ return (c, hcb.id_print(x4 + 1, where=\"w.2\"))\n+\n+ _, x10 = lax.scan(body, x2, np.arange(3))\n+ res = hcb.id_print(x10, where=\"10\")\n+ return res\n+\n+ logging.warning(\"%s: %s\", self._testMethodName, api.make_jaxpr(func)(1))\n+ logging.warning(\"%s: %s\", self._testMethodName,\n+ api.xla_computation(func)(1).GetHloText())\n+\n+ with hcb.print_receiver(output_stream=testing_stream,\n+ receiver_name=self._testMethodName):\n+ res = api.jit(func)(1)\n+ self.assertAllClose(np.array([2, 3, 4]), res, check_dtypes=True)\n+ self.assertMultiLineStrippedEqual(\n+ \"\"\"\n+1\n+2\n+0\n+1\n+2\n+1\n+-1\n+3\n+2\n+3\n+4\n+[2 3 4]\"\"\", testing_stream.output)\n+ testing_stream.reset()\n@parameterized.named_parameters(\njtu.cases_from_list(\n@@ -354,11 +380,10 @@ class HostCallbackTest(jtu.JaxTestCase):\nargs = [np.arange(np.prod(shape), dtype=dtype).reshape(shape)]\nif nr_args > 1:\nargs = args * nr_args\n- jit_fun1 = api.jit(lambda xs: hcb.end_printing(\n- hcb.id_print(\n+ jit_fun1 = api.jit(lambda xs: hcb.id_print(\n*xs,\na_new_test=\"************\",\n- testcase_name=f\"shape_{shape}_dtype_{dtype}_nr_args={nr_args}\")))\n+ testcase_name=f\"shape_{shape}_dtype_{dtype}_nr_args={nr_args}\"))\nwith hcb.print_receiver(receiver_name=self._testMethodName):\nres = jit_fun1(args)\n# self.assertAllClose(args, res, check_dtypes=True)\n@@ -367,7 +392,7 @@ class HostCallbackTest(jtu.JaxTestCase):\narg = np.arange(10000, dtype=np.int32).reshape((10, 10, 5, -1))\nwith hcb.print_receiver(output_stream=testing_stream,\nreceiver_name=self._testMethodName):\n- api.jit(lambda x: hcb.end_printing(hcb.id_print(x)))(arg)\n+ api.jit(hcb.id_print)(arg)\ndef test_jvp(self):\njvp_fun1 = lambda x, xt: api.jvp(fun1, (x,), (xt,))\n@@ -381,7 +406,7 @@ class HostCallbackTest(jtu.JaxTestCase):\nf g = id_print[ nr_results=1\noutput_stream=TestingOutputStream\nwhat=y * 3 ] e d\n- h = pow g 4.0\n+ h = pow g 2.0\ni = mul b 2.0\nj = id_print[ output_stream=TestingOutputStream\ntransforms=('jvp',)\n@@ -391,8 +416,8 @@ class HostCallbackTest(jtu.JaxTestCase):\noutput_stream=TestingOutputStream\ntransforms=('jvp',)\nwhat=y * 3 ] k j\n- n = pow g 3.0\n- o = mul 4.0 n\n+ n = pow g 1.0\n+ o = mul 2.0 n\np = mul m o\nin (h, p) }\"\"\",\nstr(api.make_jaxpr(jvp_fun1)(np.float32(5.), np.float32(0.1))))\n@@ -408,7 +433,6 @@ class HostCallbackTest(jtu.JaxTestCase):\ntesting_stream.reset()\ndef test_grad(self):\n- raise SkipTest(\"failing with new implementation\")\ngrad_fun1 = api.grad(fun1)\nself.assertMultiLineStrippedEqual(\n\"\"\"\n@@ -417,32 +441,32 @@ class HostCallbackTest(jtu.JaxTestCase):\nc = id_print[ output_stream=TestingOutputStream\nwhat=a * 2 ] b\nd = mul c 3.0\n- e = id_print[ output_stream=TestingOutputStream\n- what=y * 3 ] d\n- f = tie_in e c\n- g = pow f 3.0\n- h = mul 4.0 g\n+ e f = id_print[ nr_results=1\n+ output_stream=TestingOutputStream\n+ what=y * 3 ] d c\n+ g = pow f 1.0\n+ h = mul 2.0 g\ni = mul 1.0 h\n- j = id_print[ output_stream=TestingOutputStream\n+ j k = id_print[ nr_results=1\n+ output_stream=TestingOutputStream\ntransforms=('jvp', 'transpose')\n- what=a * 2 ] i\n- k = mul j 2.0\n- in (k,) }\"\"\", str(api.make_jaxpr(grad_fun1)(5.)))\n+ what=y * 3 ] 0.0 i\n+ l = mul j 3.0\n+ m = add_any k l\n+ n = id_print[ output_stream=TestingOutputStream\n+ transforms=('jvp', 'transpose')\n+ what=a * 2 ] m\n+ o = mul n 2.0\n+ in (o,) }\"\"\", str(api.make_jaxpr(grad_fun1)(5.)))\n- # This comes from the actual partial evaluation\n- self.assertMultiLineStrippedEqual(\n- \"\"\"\n-(Zero,) {'what': 'y * 3', 'transforms': ('jvp', 'transpose')}\n- \"\"\", testing_stream.output)\n- testing_stream.reset()\nres_grad = grad_fun1(np.float32(5.))\nself.assertMultiLineStrippedEqual(\n\"\"\"\n(DeviceArray(10.00, dtype=float32),) {'what': 'a * 2'}\n-(DeviceArray(30.00, dtype=float32),) {'what': 'y * 3'}\n-(Zero,) {'what': 'y * 3', 'transforms': ('jvp', 'transpose')}\n-(DeviceArray(4000.00, dtype=float32),) {'what': 'a * 2', 'transforms': ('jvp', 'transpose')}\n+(DeviceArray(30.00, dtype=float32), DeviceArray(10.00, dtype=float32)) {'what': 'y * 3', 'nr_results': 1}\n+(array(0.00, dtype=float32), DeviceArray(20.00, dtype=float32)) {'what': 'y * 3', 'nr_results': 1, 'transforms': ('jvp', 'transpose')}\n+(DeviceArray(20.00, dtype=float32),) {'what': 'a * 2', 'transforms': ('jvp', 'transpose')}\n\"\"\", testing_stream.output)\ntesting_stream.reset()\n@@ -461,7 +485,7 @@ class HostCallbackTest(jtu.JaxTestCase):\noutput_stream=TestingOutputStream\ntransforms=('batch',)\nwhat=y * 3 ] d c\n- g = pow f 4.0\n+ g = pow f 2.0\nin (g,) }\"\"\", str(api.make_jaxpr(vmap_fun1)(vargs)))\nres_vmap = vmap_fun1(vargs)\n@@ -473,12 +497,14 @@ class HostCallbackTest(jtu.JaxTestCase):\ntesting_stream.reset()\ndef test_pmap(self):\n- skip_if_jit_not_enabled()\n- self.helper_set_devices(4)\n- vargs = np.arange(api.local_device_count(), dtype=np.float32)\n+ vargs = 2. + np.arange(api.local_device_count(), dtype=np.float32)\npmap_fun1 = api.pmap(fun1, axis_name=\"i\")\n+ with hcb.print_receiver(output_stream=testing_stream,\n+ receiver_name=self._testMethodName):\nres = pmap_fun1(vargs)\n+ expected_res = np.stack([fun1_equiv(2. + a) for a in range(api.local_device_count())])\n+ self.assertAllClose(expected_res, res, check_dtypes=False)\nif __name__ == \"__main__\":\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_control_flow_test.py", "new_path": "tests/lax_control_flow_test.py", "diff": "@@ -514,7 +514,7 @@ class LaxControlFlowTest(jtu.JaxTestCase):\nx, lambda x: lax.cond(lax.lt(x, 5),\nx, lambda x: lax.mul(3, x),\n4, lambda y: lax.mul(y, x)))\n-\n+ print(api.xla_computation(cfun)(1).GetHloText())\nself.assertEqual(cfun(1), 2)\nself.assertEqual(cfun(3), 9)\nself.assertEqual(cfun(6), 24)\n" } ]
Python
Apache License 2.0
google/jax
Fixed scan, and grad. Added multiplexing protocol.
260,411
26.04.2020 16:31:02
-7,200
931cb3f684348584c315108b0fbfd74c74d81d96
Ensure that we carry state only for control-flow conditionals that use print
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -332,6 +332,7 @@ def xla_computation(fun: Callable,\nouts = xla.jaxpr_subcomp(\nc, jaxpr, backend, axis_env_, xla_consts,\nextend_name_stack(wrap_name(fun_name, 'xla_computation')), *xla_args)\n+ xla.state_carry.end_computation()\nreturn c.Build(xc.ops.Tuple(c, outs))\nreturn computation_maker\n" }, { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -67,6 +67,7 @@ XlaDevice = Any # xla_client.Device\nid_print_p = core.Primitive(\"id_print\")\nid_print_p.multiple_results = True\n+xla.stateful_primitives.add(id_print_p)\nxops = xla_client._xla.ops\n@@ -80,7 +81,7 @@ def id_print(*args, result=None, **kwargs):\nThe positional arguments must be JAX values. The keyword arguments are\nserialized to a string and printed along with the positional arguments.\n- There are a few special keywork arguments that are not printed:\n+ There are a few special keyword arguments that are not printed:\n* `result`: is the result of `id_print`, must be a JAX value or a\npytree of values.\n@@ -91,7 +92,7 @@ def id_print(*args, result=None, **kwargs):\n>>> y = id_print(x * 2) # prints and returns 2x\n>>> y, z = id_print(x * 2, x * 3) # prints and returns 2x and 3x\n>>> y = id_print(x * 2, result=y) # prints 2x and returns y\n- >>> y = id_print(x * 2, what='x') # prints what=x followed by 2x\n+ >>> y = id_print(x * 2, what='x') # prints \"what=x\" followed by 2x\nThe order of execution is by data dependency: after all the arguments are\ncomputed and before the result is used. At least one of the returned values\n@@ -110,7 +111,6 @@ def id_print(*args, result=None, **kwargs):\nadjoints of the results, with transforms=('vjp').\n\"\"\"\nflat_args, args_treedef = pytree.flatten(args)\n-\nparams = dict(kwargs) # copy\nif result is not None:\nflat_results, results_treedef = pytree.flatten(result)\n@@ -118,7 +118,7 @@ def id_print(*args, result=None, **kwargs):\nall_args = flat_args + flat_results\nelse:\nall_args = flat_args\n- flat_outs = id_print_p.bind(*all_args, **params) # Always returns a tuple of all args\n+ flat_outs = id_print_p.bind(*all_args, **params) # Always a tuple of all args\nif result is not None:\nreturn results_treedef.unflatten(flat_outs[-params[\"nr_results\"]:])\nelse:\n@@ -288,17 +288,20 @@ def _id_print_translation_rule_outfeed(\ncomp: XlaComputationBuilder,\n*args_op: XlaOp, **params):\n- prev_token = xla.computation_state_carry.current_token(comp)\n+ prev_token = xla.state_carry.current_token(comp)\nnr_args_to_emit = len(args_op) - params.get(\"nr_results\", 0)\nnext_token = _emit_outfeed(comp, prev_token,\n_OUTFEED_CONSUMER_ID_PRINT,\nargs_op[0:nr_args_to_emit], {})\n- xla.computation_state_carry.set_current_token(comp, next_token)\n+ xla.state_carry.set_current_token(comp, next_token)\n+ if xla.USE_ADD_DEPENDENCY:\n+ args_op = tuple([xops.AddDependency(a, next_token)\n+ for a in args_op])\nreturn xops.Tuple(comp, args_op)\nxla.translations[id_print_p] = _id_print_translation_rule_outfeed\n-_END_PRINTING = onp.int32(987654321)\n+\n@contextmanager\ndef print_receiver(output_stream=None,\nreceiver_name=\"\",\n@@ -306,6 +309,10 @@ def print_receiver(output_stream=None,\n# TODO: better timeout management\n\"\"\"Starts a receiver for the id_print outfeed.\n+ Args:\n+ output_stream: (optional) a Python stream to write the output to\n+ receiver_name: (optional) a name to use with debuging logging\n+\nUsage:\nwith print_receiver():\njax.jit(func)(args)\n@@ -315,9 +322,9 @@ def print_receiver(output_stream=None,\ndevices = api.devices()\nexecutor = futures.ThreadPoolExecutor(thread_name_prefix=\"outfeed\",\nmax_workers=len(devices))\n-\n+ _END_PRINTING = onp.int32(987654321)\ndef device_receiver_loop(device: XlaDevice) -> XlaDevice:\n- i = 0\n+ \"\"\"Polls the outfeed for a device in a loop.\"\"\"\nwhile (True):\nconsumer_id, arrays, kwargs = _receive_outfeed(device, receiver_name)\nif consumer_id != _OUTFEED_CONSUMER_ID_PRINT:\n@@ -327,20 +334,19 @@ def print_receiver(output_stream=None,\nlogging.info(f\"[{receiver_name}:{device}] Outfeed received END_PRINTING\")\nreturn device\na_str = onp.array2string(a, threshold=1024)\n- logging.info(f\"[{receiver_name}:{device}] Outfeed received {i} \"\n+ logging.info(f\"[{receiver_name}:{device}] Outfeed received \"\nf\"({a.dtype}{a.shape}): {a_str}\")\nif output_stream is not None:\noutput_stream.write(a_str)\n- i += 1\nreceiver_futures = [executor.submit(device_receiver_loop, d) for d in devices]\ntry:\nyield\nfinally:\n- for d in devices:\n+ for d in devices: # Signal the end of printing\napi.jit(lambda x: id_print(_END_PRINTING, result=x), device=d)(0)\nfor f in futures.as_completed(receiver_futures, timeout=timeout_sec):\n- finished_device = f.result()\n+ finished_device = f.result() # Throw exceptions here\nlogging.info(f\"[{receiver_name}:{finished_device} Outfeed receiver finished\")\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -697,6 +697,7 @@ def parallel_callable(fun, backend, axis_name, axis_size, global_axis_size,\nxla_args = xla._xla_callable_args(c, sharded_avals, tuple_args)\nout_nodes = xla.jaxpr_subcomp(c, jaxpr, backend, axis_env, xla_consts,\nextend_name_stack(wrap_name(name, 'pmap')), *xla_args)\n+ xla.state_carry.end_computation()\nbuilt = c.Build(xops.Tuple(c, out_nodes))\nif devices is None:\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/sharded_jit.py", "new_path": "jax/interpreters/sharded_jit.py", "diff": "@@ -156,6 +156,7 @@ def _sharded_callable(fun: lu.WrappedFun, partitions, name, *abstract_args):\nc._builder.SetSharding(_sharding_to_proto(partitions[1]))\nout_tuple = c.Tuple(*out_nodes)\nc._builder.ClearSharding()\n+ xla.state_carry.end_computation()\nbuilt = c.Build(out_tuple)\nnum_partitions = _get_num_partitions(partitions[0])\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/xla.py", "new_path": "jax/interpreters/xla.py", "diff": "from collections import defaultdict\n-from contextlib import contextmanager\nimport itertools as it\nimport operator as op\nimport threading\n-from typing import Any, Callable, Dict, List, Optional, Sequence, Type, Tuple\n+from typing import Any, Callable, Dict, List, Optional, Sequence, Set, Type, Tuple\nfrom absl import logging\nimport numpy as onp\n@@ -35,14 +34,13 @@ from ..abstract_arrays import (ConcreteArray, ShapedArray, AbstractToken,\nfrom ..core import Literal, pp_eqn_compact\nfrom ..pprint_util import pp\nfrom ..util import (partial, partialmethod, cache, prod, unzip2, memoize,\n- extend_name_stack, wrap_name, split_list)\n+ extend_name_stack, wrap_name)\nfrom ..lib import xla_bridge as xb\nfrom ..lib import xla_client as xc\nfrom . import partial_eval as pe\nfrom . import ad\nfrom . import masking\n-\nxe = xc._xla\nxops = xc._xla.ops\n@@ -170,13 +168,32 @@ pytype_aval_mappings.update((t, make_shaped_array) for t in array_types)\npytype_aval_mappings.update(\n(t, partial(_make_abstract_python_scalar, t)) for t in _scalar_types)\n+USE_ADD_DEPENDENCY = False # We can only use AddDependency once we land the HLO change\n+\n+# Keep track of which primitives are stateful (they read or read/write state).\n+# This helps avoid threading the state through control-flow primitives that\n+# do not need it. This is a worthwhile optimization because it seems that XLA\n+# may not be good at dealing with tokens (b/154992062).\n+stateful_primitives: Set[core.Primitive] = set()\n+def jaxpr_uses_state(jaxpr):\n+ \"\"\"Whether there is a stateful primitive anywhere inside a Jaxpr.\"\"\"\n+ if type(jaxpr) is core.TypedJaxpr:\n+ jaxpr = jaxpr.jaxpr\n+ for eqn in jaxpr.eqns:\n+ if eqn.primitive in stateful_primitives:\n+ return True\n+ for subjaxpr in core.subjaxprs(jaxpr):\n+ if jaxpr_uses_state(subjaxpr):\n+ return True\n+ return False\n+\nclass _ComputationStateCarry(threading.local):\n- \"\"\"Carries some state in global state.\n+ \"\"\"Carries some state globally as we build the HLO.\nFor now the state is only a token, obtained from the last OutFeed.\nThe translation rules for primitives can read-write from this class.\n- This assumes that the primitives are processed in order!\n+ This assumes that the primitives are processed in program-order!\n\"\"\"\n_current_computation: Optional[XlaComputationBuilder]\n_current_token: Optional[XlaOp]\n@@ -185,47 +202,47 @@ class _ComputationStateCarry(threading.local):\nself._current_computation = None\nself._current_token = None\n- def state_len(self):\n- return 1\n+ def end_computation(self):\n+ self._current_computation = None\n+ self._current_token = None\n+\n+ def decompose_tuple_op(self, comp: XlaComputationBuilder,\n+ tuple_op: XlaOp, nr_regular: int,\n+ uses_state: bool) -> Tuple[Sequence[XlaOp], Sequence[XlaOp]]:\n+ \"\"\"Decomposes a tuple, returns the regular elements and the state.\n+\n+ We assume that the `tuple_op` represents a tuple with `nr_regular` regular\n+ elements, followed by some elements encoding the state.\n+ Stores the new state.\n+ \"\"\"\n+ regular_ops = [xops.GetTupleElement(tuple_op, i) for i in range(nr_regular)]\n+ if uses_state:\n+ self._current_computation = comp\n+ self._current_token = xops.GetTupleElement(tuple_op, nr_regular)\n+ return regular_ops, self.current_state(comp, uses_state)\n+ else:\n+ return regular_ops, []\n- def current_state(self, comp: XlaComputationBuilder) -> List[XlaOp]:\n- if comp is not self._current_computation:\n- if self._current_computation is not None:\n- # TODO: add more error checking\n- logging.warning(\"Overwriting previous computation\")\n+ def current_state(self, comp: XlaComputationBuilder, uses_state: bool) -> List[XlaOp]:\n+ \"\"\"Create new state if new computation, returns the state. \"\"\"\n+ if not uses_state:\n+ return []\n+ if self._current_computation is None:\nself._current_computation = comp\nself._current_token = xops.CreateToken(comp)\n+ else:\n+ assert comp is self._current_computation, \"Overwriting previous computation\"\nreturn [self._current_token]\ndef current_token(self, comp: XlaComputationBuilder) -> XlaOp:\n- state = self.current_state(comp)\n+ state = self.current_state(comp, True)\nreturn state[0]\ndef set_current_token(self, comp: XlaComputationBuilder, token: XlaOp):\nassert comp == self._current_computation\nself._current_token = token\n- def set_comp_and_current_state(self, comp: XlaComputationBuilder,\n- state: Sequence[XlaOp]):\n- self._current_computation = comp\n- self._current_token = state[0]\n-\n- def extract_state_from_tuple_op(self, comp: XlaComputationBuilder,\n- tuple_op: XlaOp, nr_regular: int):\n- \"\"\"Given a tuple extract the state from its tail.\n-\n- We assume that the `tuple_op` represents a tuple with `nr_outs` regular\n- elements, followed by some elements encoding the state. Stores the new state,\n- and returns a Tuple with the regular elemennts.\n- \"\"\"\n- regular_ops = [xops.GetTupleElement(tuple_op, i) for i in range(nr_regular)]\n- self._current_computation = comp\n- current_state = [xops.GetTupleElement(tuple_op, i)\n- for i in range(nr_regular, nr_regular + self.state_len())]\n- self._current_token = current_state[0]\n- return xops.Tuple(comp, regular_ops)\n-\n-computation_state_carry = _ComputationStateCarry()\n+state_carry = _ComputationStateCarry()\n### op-by-op execution\n@@ -583,6 +600,7 @@ def _xla_callable(fun: lu.WrappedFun, device, backend, name, *arg_specs):\nout_nodes = jaxpr_subcomp(\nc, jaxpr, backend, AxisEnv(nreps, (), ()), xla_consts,\nextend_name_stack(wrap_name(name, 'jit')), *xla_args)\n+ state_carry.end_computation()\nbuilt = c.Build(xops.Tuple(c, out_nodes))\noptions = xb.get_compile_options(\n@@ -687,22 +705,26 @@ def _xla_call_translation_rule(c, axis_env,\nin_nodes, name_stack, backend, name,\ncall_jaxpr, device=None):\ndel device # Ignored.\n- prev_state = computation_state_carry.current_state(c)\n- all_in_nodes = list(in_nodes) + list(prev_state)\n-\nsubc = xb.make_computation_builder(f\"jit_{name}\")\n- all_args = [xb.parameter(subc, i, c.GetShape(n))\n- for i, n in enumerate(all_in_nodes)]\n- args, input_state = split_list(all_args, [len(in_nodes)])\n- computation_state_carry.set_comp_and_current_state(subc, input_state)\n+ uses_state = jaxpr_uses_state(call_jaxpr)\n+ prev_state = state_carry.current_state(c, uses_state)\n+ input_op = xops.Tuple(c, list(in_nodes) + list(prev_state))\n+ arg = xb.parameter(subc, 0, c.GetShape(input_op))\n+ nr_regular_args = len(in_nodes)\n+ args, input_state = state_carry.decompose_tuple_op(subc, arg, nr_regular_args, uses_state)\nout_nodes = jaxpr_subcomp(subc, call_jaxpr, backend, axis_env, (),\n- extend_name_stack(name_stack, wrap_name(name, 'jit')), *args)\n- result_state = computation_state_carry.current_state(subc)\n+ extend_name_stack(name_stack, wrap_name(name, 'jit')),\n+ *(args + input_state))\n+ result_state = state_carry.current_state(subc, uses_state)\nsubc = subc.Build(xops.Tuple(subc, list(out_nodes) + result_state))\n- call_op = xops.Call(c, subc, all_in_nodes)\n+ call_op = xops.Call(c, subc, [input_op])\n+ if not uses_state:\n+ return call_op\nnr_outs = len(out_nodes)\n- return computation_state_carry.extract_state_from_tuple_op(c, call_op, nr_outs)\n+ regular_outs, _ = state_carry.decompose_tuple_op(c, call_op, nr_outs, uses_state)\n+ return xops.Tuple(c, regular_outs)\n+\nad.primitive_transposes[xla_call_p] = partial(ad.call_transpose, xla_call_p)\n" }, { "change_type": "MODIFY", "old_path": "jax/lax/lax_control_flow.py", "new_path": "jax/lax/lax_control_flow.py", "diff": "@@ -254,7 +254,12 @@ def _while_loop_translation_rule(c, axis_env, name_stack, avals, backend, *args,\ncond_consts, body_consts, init_vals = split_list(args, [cond_nconsts, body_nconsts])\nbatched = bool(cond_jaxpr.out_avals[0].shape)\n- prev_state = xla.computation_state_carry.current_state(c)\n+ if xla.jaxpr_uses_state(cond_jaxpr):\n+ # TODO: implement the boolean as an extra carry\n+ raise NotImplementedError(\"State not supported in while_loop conditionals\")\n+\n+ uses_state = xla.jaxpr_uses_state(body_jaxpr)\n+ prev_state = xla.state_carry.current_state(c, uses_state)\n# Since jaxprs don't have tuples and have multiple return values, but we need\n# the HLO While loop to take a single tuple input and output a single boolean\n@@ -280,9 +285,8 @@ def _while_loop_translation_rule(c, axis_env, name_stack, avals, backend, *args,\nbody_c = xb.make_computation_builder(\"body_computation\")\nbody_carry = xb.parameter(body_c, 0, c.GetShape(init_carry))\n- body_carry_elts = [xops.GetTupleElement(body_carry, i) for i in range(len(args))]\n- xla.computation_state_carry.extract_state_from_tuple_op(\n- body_c, body_carry, len(args))\n+ body_carry_elts, _ = xla.state_carry.decompose_tuple_op(\n+ body_c, body_carry, len(args), uses_state)\nx, y, z = split_list(body_carry_elts, [cond_nconsts, body_nconsts])\nnew_z = xla.jaxpr_subcomp(body_c, body_jaxpr.jaxpr, backend, axis_env,\n_map(partial(xb.constant, body_c), body_jaxpr.literals),\n@@ -293,13 +297,11 @@ def _while_loop_translation_rule(c, axis_env, name_stack, avals, backend, *args,\nextend_name_stack(name_stack, 'body_pred'), *(x + z))\nnew_z = _map(partial(_pred_bcast_select, body_c, body_pred), new_z, z)\nassert _map(body_c.GetShape, new_z) == _map(body_c.GetShape, z) # no broadcast\n- result_state = xla.computation_state_carry.current_state(body_c)\n+ result_state = xla.state_carry.current_state(body_c, uses_state)\nnew_carry = xops.Tuple(body_c, list(itertools.chain(x, y, new_z, result_state)))\nans = xops.While(cond_c.Build(pred), body_c.Build(new_carry), init_carry)\n- ans_elts = [xops.GetTupleElement(ans, i) for i in range(len(args))]\n- xla.computation_state_carry.extract_state_from_tuple_op(\n- c, ans, len(args))\n+ ans_elts, _ = xla.state_carry.decompose_tuple_op(c, ans, len(args), uses_state)\n_, _, z = split_list(ans_elts, [cond_nconsts, body_nconsts])\nreturn xops.Tuple(c, z)\n@@ -563,30 +565,34 @@ def _cond_translation_rule(c, axis_env, name_stack, avals, backend,\npred, *args, true_jaxpr, false_jaxpr, linear):\ndel linear # Unused.\ntrue_ops, false_ops = split_list(args, [len(true_jaxpr.in_avals)])\n- current_state = xla.computation_state_carry.current_state(c)\n+ uses_state = xla.jaxpr_uses_state(true_jaxpr) or xla.jaxpr_uses_state(false_jaxpr)\n+ prev_state = xla.state_carry.current_state(c, uses_state)\n+\ndef make_computation(name, jaxpr, op_shape):\nc = xb.make_computation_builder(name + '_comp')\nop = xb.parameter(c, 0, op_shape)\n- ops = [xops.GetTupleElement(op, i) for i in range(len(jaxpr.in_avals))]\n- xla.computation_state_carry.extract_state_from_tuple_op(\n- c, op, len(jaxpr.in_avals))\n+ ops, _ = xla.state_carry.decompose_tuple_op(\n+ c, op, len(jaxpr.in_avals), uses_state)\nouts = xla.jaxpr_subcomp(c, jaxpr.jaxpr, backend, axis_env,\n_map(partial(xb.constant, c), jaxpr.literals),\nextend_name_stack(name_stack, name + '_fun'), *ops)\n- result_state = xla.computation_state_carry.current_state(c)\n+ result_state = xla.state_carry.current_state(c, uses_state)\nreturn c.Build(xops.Tuple(c, list(outs) + result_state))\n- true_op = xops.Tuple(c, true_ops + current_state)\n+ true_op = xops.Tuple(c, true_ops + prev_state)\ntrue_c = make_computation('true', true_jaxpr, c.GetShape(true_op))\n- false_op = xops.Tuple(c, false_ops + current_state)\n+ false_op = xops.Tuple(c, false_ops + prev_state)\nfalse_c = make_computation('false', false_jaxpr, c.GetShape(false_op))\ncond_op = xops.Conditional(pred, true_op, true_c, false_op, false_c)\n+ if not uses_state:\n+ return cond_op\n+ else:\nnr_outs = len(true_jaxpr.out_avals)\n-\n- return xla.computation_state_carry.extract_state_from_tuple_op(\n- c, cond_op, nr_outs)\n+ regular_outs, _ = xla.state_carry.decompose_tuple_op(\n+ c, cond_op, nr_outs, uses_state)\n+ return xops.Tuple(c, regular_outs)\ndef _cond_pred_bcast_select(pred, x, y):\nif core.get_aval(x) is core.get_aval(y) is core.abstract_unit:\n" }, { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -164,7 +164,7 @@ class HostCallbackTest(jtu.JaxTestCase):\ndevice=None\nname=<lambda> ] a\nin (b,) }\"\"\", str(api.make_jaxpr(jit_fun1)(5.)))\n- logging.info(\"%s: %s\",\n+ logging.warning(\"%s: %s\",\nself._testMethodName, api.xla_computation(jit_fun1)(5.).GetHloText())\nwith hcb.print_receiver(output_stream=testing_stream,\nreceiver_name=self._testMethodName):\n@@ -321,6 +321,34 @@ class HostCallbackTest(jtu.JaxTestCase):\n10\"\"\", testing_stream.output)\ntesting_stream.reset()\n+ def test_jit_while_pred_printing(self):\n+ raise SkipTest(\"Not yet implemented\")\n+ \"\"\"While with printing in the conditional.\"\"\"\n+ def func(x):\n+ x1 = hcb.id_print(x, where=\"1\")\n+\n+ def body(x):\n+ x3 = hcb.id_print(x, where=\"w.1\")\n+ return hcb.id_print(x3 + 1, where=\"w.2\")\n+\n+ x10 = lax.while_loop(lambda x: hcb.id_print(x < 10, where=\"w.p\"),\n+ body, x1)\n+ res = hcb.id_print(x10, where=\"10\")\n+ return res\n+\n+ logging.warning(\"%s: %s\", self._testMethodName, api.make_jaxpr(func)(1))\n+ logging.warning(\"%s: %s\", self._testMethodName,\n+ api.xla_computation(func)(1).GetHloText())\n+\n+ with hcb.print_receiver(output_stream=testing_stream,\n+ receiver_name=self._testMethodName):\n+ self.assertEqual(10, api.jit(func)(1))\n+ self.assertMultiLineStrippedEqual(\n+ \"\"\"\n+\"\"\", testing_stream.output)\n+ testing_stream.reset()\n+\n+\ndef test_jit_scan_cond(self):\ndef func(x):\nx1 = hcb.id_print(x, where=\"1\")\n@@ -432,6 +460,22 @@ class HostCallbackTest(jtu.JaxTestCase):\n\"\"\", testing_stream.output)\ntesting_stream.reset()\n+\n+\n+ def test_jit_nested_cond_no_print(self):\n+ \"\"\"A nested conditional, without any prints\"\"\"\n+ # raise SkipTest(\"skip this\")\n+ @api.jit\n+ def cfun(x):\n+ return lax.cond(\n+ lax.lt(x, 2),\n+ x, lambda x: x,\n+ x, lambda x: lax.cond(x < 5,\n+ 3, lambda x: x,\n+ 4, lambda y: y))\n+ print(self._testMethodName, api.xla_computation(cfun)(1).GetHloText())\n+ cfun(1)\n+\ndef test_grad(self):\ngrad_fun1 = api.grad(fun1)\nself.assertMultiLineStrippedEqual(\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_control_flow_test.py", "new_path": "tests/lax_control_flow_test.py", "diff": "@@ -514,7 +514,6 @@ class LaxControlFlowTest(jtu.JaxTestCase):\nx, lambda x: lax.cond(lax.lt(x, 5),\nx, lambda x: lax.mul(3, x),\n4, lambda y: lax.mul(y, x)))\n- print(api.xla_computation(cfun)(1).GetHloText())\nself.assertEqual(cfun(1), 2)\nself.assertEqual(cfun(3), 9)\nself.assertEqual(cfun(6), 24)\n" } ]
Python
Apache License 2.0
google/jax
Ensure that we carry state only for control-flow conditionals that use print
260,411
28.04.2020 14:43:22
-7,200
47cb5eaa8644da6d493c4fdfbc3d2263422c2992
Added masking transformation, added batch_dims to vmap
[ { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -46,7 +46,7 @@ from jax import core\nfrom jax import dtypes\nfrom jax import lax\nfrom jax.lib import pytree, xla_bridge\n-from jax.interpreters import ad, xla, batching\n+from jax.interpreters import ad, xla, batching, masking\nfrom jax.interpreters import partial_eval as pe\nfrom jax import util\nfrom jaxlib import xla_client\n@@ -126,7 +126,7 @@ def id_print(*args, result=None, **kwargs):\nreturn res if len(args) > 1 else res[0]\n-def _expand_params_transform(params: Dict, transform: str) -> Dict:\n+def _add_transform_name(params: Dict, transform: str) -> Dict:\n\"\"\"Adds the `transform` to the params[\"transforms\"].\"\"\"\nreturn dict(params, transforms=params.get(\"transforms\", ()) + (transform,))\n@@ -142,6 +142,7 @@ def _id_print_impl(*args, **params):\nprint_params = params\n# TODO: use the JITed version to do the actual printing.\n+ # TODO: print parameters sorted\nto_print = f\"{args} {print_params}\"\noutput_stream.write(to_print)\n@@ -352,7 +353,7 @@ def print_receiver(output_stream=None,\ndef _id_print_jvp_rule(primals, tangents, **params):\nprimals_out = id_print(primals, **params)\n- tangents_out = id_print(tangents, **_expand_params_transform(params, \"jvp\"))\n+ tangents_out = id_print(tangents, **_add_transform_name(params, \"jvp\"))\nreturn primals_out, tangents_out\n@@ -365,7 +366,7 @@ def _id_print_transpose_rule(cts, *args, **params):\ncts_zeros = [ad.instantiate_zeros_aval(a.aval, ct)\nfor a, ct in zip(args, cts)]\nct_args = id_print_p.bind(*cts_zeros,\n- **_expand_params_transform(params, \"transpose\"))\n+ **_add_transform_name(params, \"transpose\"))\nreturn ct_args\n@@ -373,9 +374,24 @@ ad.primitive_transposes[id_print_p] = _id_print_transpose_rule\ndef _id_print_batching_rule(batched_args, batch_dims, **params):\n- res = id_print_p.bind(*batched_args,\n- **_expand_params_transform(params, \"batch\"))\n+ new_params = _add_transform_name(params, \"batch\")\n+ new_params[\"batch_dims\"] = batch_dims\n+ res = id_print_p.bind(*batched_args, **new_params)\nreturn res, batch_dims\nbatching.primitive_batchers[id_print_p] = _id_print_batching_rule\n+\n+def _id_print_shape_rule(*operands, **params):\n+ return tuple([op.shape for op in operands])\n+\n+\n+masking.shape_rules[id_print_p] = _id_print_shape_rule\n+\n+def _id_print_masking_rule(operands, operands_logical_shapes, **params):\n+ new_params = _add_transform_name(params, \"mask\")\n+ new_params[\"logical_shapes\"] = operands_logical_shapes\n+ return id_print_p.bind(*operands, **new_params)\n+\n+\n+masking.masking_rules[id_print_p] = _id_print_masking_rule\n\\ No newline at end of file\n" }, { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -16,6 +16,7 @@ from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n+from functools import partial\nimport logging\nimport numpy as onp\nimport os\n@@ -521,11 +522,13 @@ class HostCallbackTest(jtu.JaxTestCase):\n\"\"\"\n{ lambda ; a.\nlet b = mul a 2.0\n- c = id_print[ output_stream=TestingOutputStream\n+ c = id_print[ batch_dims=(0,)\n+ output_stream=TestingOutputStream\ntransforms=('batch',)\nwhat=a * 2 ] b\nd = mul c 3.0\n- e f = id_print[ nr_results=1\n+ e f = id_print[ batch_dims=(0, 0)\n+ nr_results=1\noutput_stream=TestingOutputStream\ntransforms=('batch',)\nwhat=y * 3 ] d c\n@@ -535,11 +538,37 @@ class HostCallbackTest(jtu.JaxTestCase):\nres_vmap = vmap_fun1(vargs)\nself.assertMultiLineStrippedEqual(\n\"\"\"\n-(DeviceArray([ 8.00, 10.00], dtype=float32),) {'what': 'a * 2', 'transforms': ('batch',)}\n-(DeviceArray([24.00, 30.00], dtype=float32), DeviceArray([ 8.00, 10.00], dtype=float32)) {'what': 'y * 3', 'nr_results': 1, 'transforms': ('batch',)}\n+(DeviceArray([ 8.00, 10.00], dtype=float32),) {'what': 'a * 2', 'transforms': ('batch',), 'batch_dims': (0,)}\n+(DeviceArray([24.00, 30.00], dtype=float32), DeviceArray([ 8.00, 10.00], dtype=float32)) {'what': 'y * 3', 'nr_results': 1, 'transforms': ('batch',), 'batch_dims': (0, 0)}\n\"\"\", testing_stream.output)\ntesting_stream.reset()\n+ def test_vmap_not_batched(self):\n+ x = 3.\n+ def func(y):\n+ # x is not mapped, y is mapped\n+ _, y = hcb.id_print(x, y, output_stream=testing_stream)\n+ return x + y\n+\n+ vmap_func = api.vmap(func)\n+ vargs = np.array([np.float32(4.), np.float32(5.)])\n+ self.assertMultiLineStrippedEqual(\n+ \"\"\"\n+{ lambda ; a.\n+ let b c = id_print[ batch_dims=(None, 0)\n+ output_stream=TestingOutputStream\n+ transforms=('batch',) ] 3.0 a\n+ d = add c 3.0\n+ in (d,) }\"\"\", str(api.make_jaxpr(vmap_func)(vargs)))\n+\n+ res_vmap = vmap_func(vargs)\n+ self.assertMultiLineStrippedEqual(\n+ \"\"\"\n+(3.00, DeviceArray([4.00, 5.00], dtype=float32)) {'transforms': ('batch',), 'batch_dims': (None, 0)}\n+ \"\"\", testing_stream.output)\n+ testing_stream.reset()\n+\n+\ndef test_pmap(self):\nvargs = 2. + np.arange(api.local_device_count(), dtype=np.float32)\n@@ -550,6 +579,30 @@ class HostCallbackTest(jtu.JaxTestCase):\nexpected_res = np.stack([fun1_equiv(2. + a) for a in range(api.local_device_count())])\nself.assertAllClose(expected_res, res, check_dtypes=False)\n+ def test_mask(self):\n+\n+ @partial(api.mask, in_shapes=['n'], out_shape='')\n+ def padded_sum(x):\n+ return np.sum(hcb.id_print(x, what=\"x\", output_stream=testing_stream))\n+ args = [np.arange(4)], dict(n=2)\n+ self.assertMultiLineStrippedEqual(\n+ \"\"\"\n+{ lambda c f ; a b.\n+ let d = lt c b\n+ e = id_print[ logical_shapes=[(Traced<ShapedArray(int32[], weak_type=True):JaxprTrace(level=0/0)>,)]\n+ output_stream=TestingOutputStream\n+ transforms=('mask',)\n+ what=x ] a\n+ g = select d e f\n+ h = reduce_sum[ axes=(0,) ] g\n+ in (h,) }\"\"\", str(api.make_jaxpr(padded_sum)(*args)))\n+\n+ res = padded_sum(*args)\n+ self.assertMultiLineStrippedEqual(\n+ \"\"\"\n+(DeviceArray([0, 1, 2, 3], dtype=int32),) {'what': 'x', 'transforms': ('mask',), 'logical_shapes': [(2,)]}\n+ \"\"\", testing_stream.output)\n+ testing_stream.reset()\nif __name__ == \"__main__\":\nabsltest.main()\n" } ]
Python
Apache License 2.0
google/jax
Added masking transformation, added batch_dims to vmap
260,411
02.05.2020 15:52:59
-10,800
0444f92795586e16e65da5f0fe253edc93b01f9c
Added support for sending all arrays in a single message
[ { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -74,6 +74,8 @@ xops = xla_client._xla.ops\n# TODO: write description for descriptor\n+_OUTFEED_ALL_TOGETHER = True # If true, then all the arrays are outfed together\n+\n# The data on the outfeed follows a protocol that allows multiplexing the\n# outfeed among multiple consumers, and communicates in-stream shape and\n# type of the data.\n@@ -298,6 +300,10 @@ def _emit_outfeed(comp: XlaComputationBuilder, token: XlaOp,\ntoken = xops.OutfeedWithToken(data, token, comp.GetShape(data))\n# Now send the arrays\n+ if _OUTFEED_ALL_TOGETHER:\n+ entire_shape = xla_client.Shape.tuple_shape(arrays_shape)\n+ token = xops.OutfeedWithToken(xops.Tuple(comp, arrays), token, entire_shape)\n+ else:\nfor a, a_shape in zip(arrays, arrays_shape):\ntoken = xops.OutfeedWithToken(a, token, a_shape)\nreturn token\n@@ -332,6 +338,13 @@ def _receive_outfeed(device: XlaDevice, receiver_name: str\nfor i in range(4, 4 + (metadata_length + 3) // 4)]\nmetadata = b\"\".join(metadatas)[:metadata_length]\narray_descriptors = msgpack.unpackb(metadata)\n+ if _OUTFEED_ALL_TOGETHER:\n+ arrays_shape = [xla_client.Shape.array_shape(_CODE_TO_DTYPE[a_descr[0]],\n+ a_descr[1])\n+ for a_descr in array_descriptors]\n+ entire_shape = xla_client.Shape.tuple_shape(arrays_shape)\n+ arrays = _get_data(entire_shape, device)\n+ else:\narrays = []\nfor a_descr in array_descriptors:\na_shape = xla_client.Shape.array_shape(_CODE_TO_DTYPE[a_descr[0]],\n" }, { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -495,6 +495,11 @@ where: 10\nwith hcb.outfeed_receiver(receiver_name=self._testMethodName):\napi.jit(hcb.id_print)(arg)\n+ def test_jit_several(self):\n+ arg = np.arange(50, dtype=np.int32).reshape((10, 5))\n+ with hcb.outfeed_receiver(receiver_name=self._testMethodName):\n+ api.jit(lambda x, y: hcb.id_print(x, y, x * 2.))(arg, np.ones(100, dtype=np.int32))\n+\ndef test_jvp(self):\njvp_fun1 = lambda x, xt: api.jvp(fun1, (x,), (xt,))\nself.assertMultiLineStrippedEqual(\n" } ]
Python
Apache License 2.0
google/jax
Added support for sending all arrays in a single message
260,411
03.05.2020 11:30:27
-10,800
cee5989a1dbe618d757801e2c0e5895c76e139dc
Implemented pytree support for arg and result. Enabled outfeed for all arrays as a tuple
[ { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -41,6 +41,7 @@ from jax import lax\nfrom jax.lib import pytree, xla_bridge\nfrom jax.interpreters import ad, xla, batching, masking\nfrom jax.interpreters import partial_eval as pe\n+from jax import pprint_util as ppu\nfrom jax import util\nfrom jaxlib import xla_client\nfrom jaxlib import xla_extension\n@@ -60,42 +61,35 @@ XlaShape = Any # xla_client.Shape\nXlaComputationBuilder = Any # xla_bridge._JaxComputationBuilder\nXlaDevice = Any # xla_client.Device\n-\"\"\"The id_tap primitive acts like the identity function. It has a number of\n-positional arguments and parameters:\n- * func: the actual (Python) function to invoke with the positional arguments\n- and the parameters.\n- * nr_untapped: how many positional arguments (from the tail) should not be\n- passed to the tap function.\n- * transforms: a tuple of the transformations that have been applied.\n- * batch_dims: a tuple of the dims that have been batched, for vmap\n- * logical_shapes: a tuple of evaluated logical shapes, for mask\n-\"\"\"\n-# TODO: handle multiple vmap and mask\n-id_tap_p = core.Primitive(\"id_tap\")\n-id_tap_p.multiple_results = True\n-xla.stateful_primitives.add(id_tap_p)\n-\n# TODO: add a flag\n_LOGGING = True\n-def id_tap(func: Callable, *args, result=None, **kwargs):\n- \"\"\"Behaves like the identity function for positional arguments, but invokes\n- the ``func`` with the arrays corresponding to ``args`` and the\n- ``kwargs``.\n+def id_tap(func: Callable, arg, *,\n+ result=None,\n+ **kwargs):\n+ \"\"\"Behaves like the identity function, but invokes ``func`` on positional\n+ argument ``arg`` and keyword arguments ``kwargs``.\n+ The return value of ``func`` is ignored.\n- The return value is a tuple with the values of `args` or the value of the\n- keyword parameter `result` if present. If there is a single positional\n- argument, it returns just that argument without packing it in a tuple.\n+ The argument can be a JAX type, or a pytree thereof (tuples/list/dict).\n+ If the ``return`` keyword argument is given, then its value must be\n+ a JAX type and it is the value being returned.\n- The positional arguments must be JAX values or pytrees.\n- * `result`: is the result of `id_tap`, must be a JAX value or a\n- pytree of values.\n+ Note that only the JAX types from ``arg`` are passed through the compiled\n+ code; all the values from ``kwargs`` are stored in Python and passed to\n+ ``func``.\nUsage:\n- >>> y = id_tap(func, x * 2) # calls func(2x) and returns 2x\n- >>> y, z = id_tap(func, x * 2, x * 3) # calls func(2x, 3x) and returns (2x, 3x)\n- >>> y = id_print(func, x * 2, result=y) # calls func(2x) and returns y\n- >>> y = id_print(func, x * 2, what='x') # calls func(2x, what='x') and returns 2x\n+ >>> # calls func(2x) and returns 2x\n+ >>> y = id_tap(func, x * 2)\n+ >>> # calls func((2x, 3x)) and returns (2x, 3x)\n+ >>> y, z = id_tap(func, (x * 2, x * 3))\n+ >>> # calls func(2x) and returns y\n+ >>> y = id_tap(func, x * 2, result=y)\n+ >>> # calls func(2x, what='x') and returns 2x\n+ >>> y = id_tap(func, x * 2, what='x')\n+ >>> # calls func(dict(x=x, y=y), what='foo') and returns dict(x=x, y=y)\n+ >>> x, y = id_tap(func, dict(x=x, y=y), what='a dict')\nThe order of execution is by data dependency: after all the arguments are\ncomputed and before the result is used. At least one of the returned values\n@@ -113,64 +107,42 @@ def id_tap(func: Callable, *args, result=None, **kwargs):\nneeded in the computation of `grad` and an ``id_print`` with the\nadjoints of the results, with transforms=('vjp').\n\"\"\"\n- flat_args, args_treedef = pytree.flatten(args)\n- params = dict(kwargs) # copy; we pass all params to the primitive\n- params[\"func\"] = func # pass the function, to have it for transforms\n+ if func not in (_end_consumer, _unknown_consumer):\n+ api._check_callable(func)\n+ flat_args, arg_treedef = pytree.flatten(arg)\n+ api._check_args(flat_args)\n+ params = dict(kwargs) # we pass a copy of params to the primitive\n+ # See definition of id_tap_p for what parameters it takes\n+ params[\"func\"] = func\n+ params[\"arg_treedef\"] = arg_treedef\nif result is not None:\n- flat_results, results_treedef = pytree.flatten(result)\n+ flat_results, result_treedef = pytree.flatten(result)\n+ api._check_args(flat_results)\nparams[\"nr_untapped\"] = len(flat_results)\nall_args = flat_args + flat_results\nelse:\nall_args = flat_args\nflat_outs = id_tap_p.bind(*all_args, **params) # Always a tuple of all args\nif result is not None:\n- return results_treedef.unflatten(flat_outs[-params[\"nr_untapped\"]:]) # type: ignore[unsupported-operands]\n+ return result_treedef.unflatten(flat_outs[-params[\"nr_untapped\"]:]) # type: ignore[unsupported-operands]\nelse:\n- res = args_treedef.unflatten(flat_outs)\n- return res if len(args) > 1 else res[0]\n+ return arg_treedef.unflatten(flat_outs)\n# TODO: clean up the docstring\n-def id_print(*args, result=None, output_stream=None, threshold=1024,\n+def id_print(arg, *, result=None, output_stream=None, threshold=1024,\n**kwargs):\n- \"\"\"Behaves like the identify function for positional arguments, but prints all\n- arguments on the host, even from transformed or compiled code.\n-\n- The return value is a tuple with the value of `args` or the value of the\n- keyword parameter `result` if present. If there is a single positional\n- argument, it returns just that argument without packing it in a tuple.\n-\n- The positional arguments must be JAX values. The keyword arguments are\n- serialized to a string and printed along with the positional arguments.\n- There are a few special keyword arguments that are not printed:\n-\n- * `result`: is the result of `id_print`, must be a JAX value or a\n- pytree of values.\n- * `output_stream`: is the output stream where the values should be\n- printed. (Note: does not yet work from under JIT).\n-\n- Usage:\n- >>> y = id_print(x * 2) # prints and returns 2x\n- >>> y, z = id_print(x * 2, x * 3) # prints and returns 2x and 3x\n- >>> y = id_print(x * 2, result=y) # prints 2x and returns y\n- >>> y = id_print(x * 2, what='x') # prints \"what=x\" followed by 2x\n+ \"\"\"Like ``id_tap`` with a printing tap function.\n- The order of execution is by data dependency: after all the arguments are\n- computed and before the result is used. At least one of the returned values\n- must be used in the rest of the computation, or else this operation has\n- no effect.\n-\n- Upon JAX transformations, the transformed values are wrapped with\n- `id_print`, and a special `transforms` tuple keyword argument is added with\n- the sequence of transformations applied:\n+ On each invocation of the printing tap, the ``kwargs`` if present\n+ will be printed first (sorted by keys). Then arg will be printed,\n+ with the arrays stringified with ``numpy.array2string``.\n- - For `vmap` the arguments are batched, and transforms=('vmap')\n- - For `jvp` there will be an id_print for the primal values, and a\n- separate `id_print` for the tangents with `transforms=('jvp')`.\n- - For `grad` there will be an `id_print` for the primal values (if\n- needed in the computation of `grad` and an `id_print` with the\n- adjoints of the results, with transforms=('vjp').\n+ Additional keyword arguments:\n+ * ``output_stream`` if given then it will be used instead of the\n+ built-in ``print``. The string will be passed as ``output_stream.write(s)``.\n+ * ``threshold`` is passed to ``numpy.array2string``.\n\"\"\"\n- return id_tap(_print_consumer, *args,\n+ return id_tap(_print_consumer, arg,\nresult=result, output_stream=output_stream, **kwargs)\n@@ -179,6 +151,7 @@ class _ConsumerCallable(NamedTuple):\n\"\"\"Host-side information for a outfeed consumer.\"\"\"\nfunc: Callable\nkwargs: Tuple[Tuple[str, Any], ...]\n+ arg_treedef: Any\n_consumer_registry: Dict[_ConsumerCallable, int] = dict()\n_consumer_registry_by_id: Dict[int, _ConsumerCallable] = dict()\n@@ -194,10 +167,10 @@ def _register_consumer(cons: _ConsumerCallable) -> int:\n_consumer_registry_by_id[cons_id] = cons\nreturn cons_id\n-def _print_consumer(*arrays, output_stream=None,\n+def _print_consumer(arg, *, output_stream=None,\nthreshold=1024, **kwargs):\n\"\"\"The consumer for id_print\"\"\"\n- def emit(s: str):\n+ def emit_str(s: str):\nif output_stream is not None:\noutput_stream.write(s + \"\\n\")\nelse:\n@@ -206,13 +179,44 @@ def _print_consumer(*arrays, output_stream=None,\nfor k, v in sorted(kwargs.items())\nif k not in (\"consumer_id\", \"nr_untapped\")])\nif kv_pairs:\n- emit(kv_pairs)\n- for a in arrays:\n- if not isinstance(a, onp.ndarray):\n- a = onp.array(a)\n- emit(onp.array2string(a, threshold=threshold))\n+ emit_str(kv_pairs)\n+\n+ def pp_val(arg) -> ppu.PrettyPrint:\n+ if isinstance(arg, (tuple, list)):\n+ return (ppu.pp('[ ') >>\n+ ppu.vcat([pp_val(e) for e in arg]) >> ppu.pp(' ]'))\n+ elif isinstance(arg, dict):\n+ return (ppu.pp('{ ') >>\n+ ppu.vcat(\n+ [ppu.pp(f\"{k}=\") >> pp_val(v)\n+ for k, v in sorted(arg.items())]) >>\n+ ppu.pp(' }'))\n+ elif isinstance(arg, onp.ndarray):\n+ return ppu.pp(onp.array2string(arg, threshold=threshold))\n+ else:\n+ return ppu.pp(str(arg))\n+\n+ emit_str(str(pp_val(arg)))\n+\n+\"\"\"The id_tap primitive acts like the identity function. It has a number of\n+positional arguments and parameters:\n+ * func: the actual (Python) function to invoke with the positional arguments\n+ and the parameters.\n+ * nr_untapped: how many positional arguments (from the tail) should not be\n+ passed to the tap function.\n+ * arg_treedef: the treedef of the tapped positional arguments\n+ * transforms: a tuple of the transformations that have been applied.\n+ * batch_dims: a tuple of the dims that have been batched, for vmap\n+ * logical_shapes: a tuple of evaluated logical shapes, for mask\n+\n+ * the remaining parameters are passed to the tap function.\n+\"\"\"\n+# TODO: handle multiple vmap and mask\n+id_tap_p = core.Primitive(\"id_tap\")\n+id_tap_p.multiple_results = True\n+xla.stateful_primitives.add(id_tap_p)\ndef _add_transform_name(params: Dict, transform: str) -> Dict:\n@@ -220,13 +224,16 @@ def _add_transform_name(params: Dict, transform: str) -> Dict:\nreturn dict(params, transforms=params.get(\"transforms\", ()) + (transform,))\n-def _id_tap_impl(*arrays, func=None, nr_untapped=0, **params):\n+def _id_tap_impl(*arrays, func=None, nr_untapped=0, arg_treedef=None,\n+ **params):\nassert isinstance(func, Callable)\nfunc_params = dict(params)\n- # TODO: handle errors in the tap consumer\n- func_arrays = arrays[:-nr_untapped] if nr_untapped > 0 else arrays\n+ # TODO: consolidate logic with the outfeed receiver\ntry:\n- func(*func_arrays, **func_params)\n+ assert nr_untapped <= len(arrays)\n+ func_arrays = arrays[:-nr_untapped] if nr_untapped > 0 else arrays\n+ arg = api.tree_unflatten(arg_treedef, func_arrays)\n+ func(arg, **func_params)\nexcept Exception as e:\nraise TapFunctionException from e\n# We continue for now, we need to keep reading the outfeed\n@@ -298,7 +305,8 @@ _unknown_consumer = 1 # for testing error cases\ndef _id_print_translation_rule_outfeed(comp: XlaComputationBuilder,\n*args_op: XlaOp, func=None,\n- nr_untapped=0, **params):\n+ nr_untapped=0, arg_treedef=None,\n+ **params):\nparams = dict(params)\nif func is _end_consumer:\nparams[\"consumer_id\"] = _end_consumer\n@@ -306,7 +314,7 @@ def _id_print_translation_rule_outfeed(comp: XlaComputationBuilder,\nparams[\"consumer_id\"] = _unknown_consumer # Will trigger an error, for testing\nelse:\nparams[\"consumer_id\"] = _register_consumer(\n- _ConsumerCallable(func, tuple(params.items())))\n+ _ConsumerCallable(func, tuple(params.items()), arg_treedef))\nprev_token = xla.state_carry.current_token(comp)\nnr_args_to_emit = len(args_op) - nr_untapped\n@@ -321,9 +329,6 @@ def _id_print_translation_rule_outfeed(comp: XlaComputationBuilder,\nxla.translations[id_tap_p] = _id_print_translation_rule_outfeed\n-# If true, then all the arrays are outfed together\n-_OUTFEED_ARRAYS_TOGETHER = True\n-\n# The data on the outfeed follows a protocol that allows multiplexing the\n# outfeed among multiple consumers, and communicates in-stream shape and\n# type of the data.\n@@ -390,12 +395,8 @@ def _emit_outfeed(comp: XlaComputationBuilder, token: XlaOp,\ntoken = xops.OutfeedWithToken(data, token, comp.GetShape(data))\n# Now send the arrays\n- if _OUTFEED_ARRAYS_TOGETHER:\nentire_shape = xla_client.Shape.tuple_shape(arrays_shape)\ntoken = xops.OutfeedWithToken(xops.Tuple(comp, arrays), token, entire_shape)\n- else:\n- for a, a_shape in zip(arrays, arrays_shape):\n- token = xops.OutfeedWithToken(a, token, a_shape)\nreturn token\ndef _receive_outfeed(device: XlaDevice, receiver_name: str\n@@ -411,14 +412,7 @@ def _receive_outfeed(device: XlaDevice, receiver_name: str\n(_OUTFEED_HEADER_LENGTH,))\ndef _get_data(data_shape: XlaShape, device: XlaDevice) -> XlaShape:\n- if platform in (\"gpu\", \"cpu\"):\n- return xla_client.transfer_from_outfeed(data_shape, device)\n- else:\n- if _OUTFEED_ARRAYS_TOGETHER:\nreturn xla_client.transfer_from_outfeed(data_shape, device)\n- else:\n- return xla_client.transfer_from_outfeed(\n- xla_client.Shape.tuple_shape((data_shape,)), device)[0]\nheader = _get_data(header_shape, device)\nif header[0] != _OUTFEED_HEADER_START:\n@@ -434,19 +428,10 @@ def _receive_outfeed(device: XlaDevice, receiver_name: str\narrays_shape = [xla_client.Shape.array_shape(_CODE_TO_DTYPE[a_descr[0]],\na_descr[1])\nfor a_descr in array_descriptors]\n- if _OUTFEED_ARRAYS_TOGETHER:\nentire_shape = xla_client.Shape.tuple_shape(arrays_shape)\narrays = _get_data(entire_shape, device)\nlogging.info(f\"[{receiver_name}:{device}] Outfeed read data of shape \"\n\",\".join([f\"{data.dtype}{data.shape}\" for data in arrays]))\n- else:\n- arrays = []\n- for a_shape in arrays_shape:\n- data = _get_data(a_shape, device)\n- if _LOGGING:\n- logging.info(f\"[{receiver_name}:{device}] Outfeed read array of shape \"\n- f\"{data.dtype}{data.shape}\")\n- arrays.append(data)\nreturn (consumer_id, arrays)\n@@ -487,7 +472,7 @@ def outfeed_receiver(*,\nthread_name_prefix=f\"outfeed_receiver_{receiver_name}\",\nmax_workers=len(devices))\n- count_tap_exceptions = False\n+ count_tap_exceptions = 0\ndef device_receiver_loop(device: XlaDevice) -> XlaDevice:\n\"\"\"Polls the outfeed for a device in a loop.\"\"\"\nnonlocal count_tap_exceptions\n@@ -507,7 +492,8 @@ def outfeed_receiver(*,\ncount_tap_exceptions += 1\ncontinue # We need to read the entire outfeed\ntry:\n- consumer.func(*arrays, **dict(consumer.kwargs)) # type: ignore[attribute-error]\n+ arg = api.tree_unflatten(consumer.arg_treedef, arrays)\n+ consumer.func(arg, **dict(consumer.kwargs)) # type: ignore[attribute-error]\nexcept Exception as e:\nlogging.error(f\"Postponing exception raised in tap function: {str(e)}\\n{traceback.format_exc()}\")\ncount_tap_exceptions += 1\n@@ -516,12 +502,13 @@ def outfeed_receiver(*,\nreceiver_futures = [executor.submit(device_receiver_loop, d) for d in devices]\n# Register a callback to raise errors if any. These exception come from\n# bugs in our code, not from the tap functions.\n- [rf.add_done_callback(lambda rf: rf.result()) for rf in receiver_futures]\n+ for rf in receiver_futures:\n+ rf.add_done_callback(lambda rf: rf.result())\ntry:\nyield\nfinally:\nfor d in devices: # Signal the end of printing\n- api.jit(lambda x: id_tap(_end_consumer, result=x), device=d)(0) # type: ignore[wrong-arg-types]\n+ api.jit(lambda x: id_tap(_end_consumer, None, result=x), device=d)(0) # type: ignore[arg-type]\nfor f in futures.as_completed(receiver_futures, timeout=timeout_sec):\nfinished_device = f.result() # Throw exceptions here\nif _LOGGING:\n" }, { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -127,11 +127,13 @@ class HostCallbackTest(jtu.JaxTestCase):\n\"\"\"\n{ lambda ; a.\nlet b = mul a 2.00\n- c = id_tap[ func=_print\n+ c = id_tap[ arg_treedef=*\n+ func=_print\noutput_stream=...\nwhat=a * 2 ] b\nd = mul c 3.00\n- e f = id_tap[ func=_print\n+ e f = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=1\noutput_stream=...\nwhat=y * 3 ] d c\n@@ -150,7 +152,7 @@ what: y * 3\ndef test_with_tuple_results(self):\ndef func2(x):\n- x1, y1 = hcb.id_print(x * 2., x * 3., output_stream=testing_stream)\n+ x1, y1 = hcb.id_print((x * 2., x * 3.), output_stream=testing_stream)\nreturn x1 + y1\nself.assertMultiLineStrippedEqual(\n@@ -158,19 +160,41 @@ what: y * 3\n{ lambda ; a.\nlet b = mul a 2.00\nc = mul a 3.00\n- d e = id_tap[ func=_print\n+ d e = id_tap[ arg_treedef=PyTreeDef(tuple, [*,*])\n+ func=_print\noutput_stream=...] b c\nf = add d e\nin (f,) }\"\"\", str(api.make_jaxpr(func2)(3.)))\nself.assertEqual(3. * (2. + 3.), func2(3.))\nself.assertMultiLineStrippedEqual(\"\"\"\n-6.00\n-9.00\"\"\", testing_stream.output)\n+[ 6.00\n+ 9.00 ]\"\"\", testing_stream.output)\n+ testing_stream.reset()\n+\n+ def test_with_dict_results(self):\n+ def func2(x):\n+ res = hcb.id_print(dict(a=x * 2., b=x * 3.), output_stream=testing_stream)\n+ return res[\"a\"] + res[\"b\"]\n+\n+ self.assertMultiLineStrippedEqual(\n+ \"\"\"\n+{ lambda ; a.\n+ let b = mul a 2.00\n+ c = mul a 3.00\n+ d e = id_tap[ arg_treedef=PyTreeDef(dict[['a', 'b']], [*,*])\n+ func=_print\n+ output_stream=...] b c\n+ f = add d e\n+ in (f,) }\"\"\", str(api.make_jaxpr(func2)(3.)))\n+ self.assertEqual(3. * (2. + 3.), func2(3.))\n+ self.assertMultiLineStrippedEqual(\"\"\"\n+{ a=6.00\n+ b=9.00 }\"\"\", testing_stream.output)\ntesting_stream.reset()\ndef test_with_result(self):\ndef func2(x):\n- x1 = hcb.id_print(x * 2., x * 3., result=x * 4.,\n+ x1 = hcb.id_print((x * 2., x * 3.), result=x * 4.,\noutput_stream=testing_stream)\nreturn x1\n@@ -180,16 +204,42 @@ what: y * 3\nlet b = mul a 2.00\nc = mul a 3.00\nd = mul a 4.00\n- e f g = id_tap[ func=_print\n+ e f g = id_tap[ arg_treedef=PyTreeDef(tuple, [*,*])\n+ func=_print\nnr_untapped=1\noutput_stream=...] b c d\nin (g,) }\"\"\", str(api.make_jaxpr(func2)(3.)))\nself.assertEqual(3. * 4., func2(3.))\nself.assertMultiLineStrippedEqual(\"\"\"\n-6.00\n-9.00\"\"\", testing_stream.output)\n+[ 6.00\n+ 9.00 ]\"\"\", testing_stream.output)\ntesting_stream.reset()\n+ def test_pytree(self):\n+ def func(x, what=\"\"):\n+ \"\"\"Returns some pytrees depending on x\"\"\"\n+ if what == \"pair_1_x\":\n+ return (1, x)\n+ elif what == \"pair_x_2x\":\n+ return (x, 2 * x)\n+ elif what == \"dict\":\n+ return dict(a=2 * x, b=3 * x)\n+ else:\n+ assert False\n+ tap_count = 0\n+ def tap_func(a, what=\"\"):\n+ nonlocal tap_count\n+ tap_count += 1\n+ self.assertEqual(func(5, what), a)\n+\n+ for what in (\"pair_1_x\", \"pair_x_2x\", \"dict\"):\n+ self.assertEqual(func(10, what),\n+ (lambda x: hcb.id_tap(tap_func, func(x, what),\n+ result=func(x * 2, what),\n+ what=what))(5))\n+ self.assertEqual(3, tap_count)\n+\n+\ndef test_eval_tap_exception(self):\n# Simulate a tap error\ndef tap_err(*args, **kwargs):\n@@ -220,7 +270,8 @@ what: x1\nlet b = xla_call[ backend=None\ncall_jaxpr={ lambda ; a.\nlet b = mul a 2.00\n- c = id_tap[ func=_print\n+ c = id_tap[ arg_treedef=*\n+ func=_print\noutput_stream=...\nwhat=here ] b\nd = mul c 3.00\n@@ -325,6 +376,32 @@ where: 2\nself.assertEqual(len(devices), len(re.findall(r\"112\", testing_stream.output)))\ntesting_stream.reset()\n+ def test_jit_pytree(self):\n+ def func(x, what=\"\"):\n+ \"\"\"Returns some pytrees depending on x\"\"\"\n+ if what == \"pair_1_x\":\n+ return (1, x)\n+ elif what == \"pair_x_2x\":\n+ return (x, 2 * x)\n+ elif what == \"dict\":\n+ return dict(a=2 * x, b=3 * x)\n+ else:\n+ assert False\n+ tap_count = 0\n+ def tap_func(a, what=\"\"):\n+ nonlocal tap_count\n+ tap_count += 1\n+ self.assertEqual(func(5, what), a)\n+\n+ with hcb.outfeed_receiver(receiver_name=self._testMethodName):\n+ for what in (\"pair_1_x\", \"pair_x_2x\", \"dict\"):\n+ self.assertEqual(func(10, what),\n+ api.jit(lambda x: hcb.id_tap(tap_func, func(x, what),\n+ result=func(x * 2, what),\n+ what=what))(5))\n+ self.assertEqual(3, tap_count)\n+\n+\ndef test_jit_cond1(self):\n\"\"\"A conditional\"\"\"\ndef func(x):\n@@ -506,7 +583,7 @@ where: 10\nif nr_args > 1:\nargs = args * nr_args\njit_fun1 = api.jit(lambda xs: hcb.id_print(\n- *xs,\n+ xs,\na_new_test=\"************\",\ntestcase_name=f\"shape_{shape}_dtype_{dtype}_nr_args={nr_args}\"))\nwith hcb.outfeed_receiver(receiver_name=self._testMethodName):\n@@ -521,20 +598,20 @@ where: 10\ndef test_jit_several_together(self):\narg = np.arange(50, dtype=np.int32).reshape((10, 5))\nwith hcb.outfeed_receiver(receiver_name=self._testMethodName):\n- api.jit(lambda x, y: hcb.id_print(x, y, x * 2.))(arg, np.ones(100, dtype=np.int32))\n+ api.jit(lambda x, y: hcb.id_print((x, y, x * 2.)))(arg, np.ones(100, dtype=np.int32))\ndef test_jit_interleaving(self):\n# Several jit's without data dependencies; they may interfere\ncount = 0 # Count tap invocations\nnr_arrays = 5\n- def tap_func(*args, **kwargs):\n+ def tap_func(arg, **kwargs):\nnonlocal count\n- assert len(args) == nr_arrays\n+ assert len(arg) == nr_arrays\ncount += 1\n# This is the function that we'll run multiple times\ndef func(x, count):\nfor i in range(count):\n- x = hcb.id_tap(tap_func, *[x + i for i in range(nr_arrays)], i=i)[-1]\n+ x = hcb.id_tap(tap_func, [x + i for i in range(nr_arrays)], i=i)[-1]\nreturn x\nwith hcb.outfeed_receiver(receiver_name=self._testMethodName):\nx = np.array(1, dtype=onp.int32)\n@@ -632,24 +709,28 @@ what: x3\n\"\"\"\n{ lambda ; a b.\nlet c = mul a 2.00\n- d = id_tap[ func=_print\n+ d = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=0\noutput_stream=...\nwhat=a * 2 ] c\ne = mul d 3.00\n- f g = id_tap[ func=_print\n+ f g = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=1\noutput_stream=...\nwhat=y * 3 ] e d\nh = pow g 2.00\ni = mul b 2.00\n- j k = id_tap[ func=_print\n+ j k = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=1\noutput_stream=...\ntransforms=('jvp',)\nwhat=a * 2 ] i d\nl = mul j 3.00\n- m n o = id_tap[ func=_print\n+ m n o = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=2\noutput_stream=...\ntransforms=('jvp',)\n@@ -710,25 +791,29 @@ transforms: ('jvp', 'transpose') what: x * 3\n\"\"\"\n{ lambda ; a.\nlet b = mul 1.00 a\n- c d = id_tap[ func=_print\n+ c d = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=1\noutput_stream=...\ntransforms=('jvp', 'transpose')\nwhat=y * 3 ] b 0.00\ne = mul c 3.00\n- f g = id_tap[ func=_print\n+ f g = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=1\noutput_stream=...\ntransforms=('jvp', 'transpose')\nwhat=x * 2 ] e 0.00\nh = mul f 2.00\ni = mul a 2.00\n- j = id_tap[ func=_print\n+ j = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=0\noutput_stream=...\nwhat=x * 2 ] i\nk = mul j 3.00\n- l = id_tap[ func=_print\n+ l = id_tap[ arg_treedef=*\n+ func=_print\nnr_untapped=0\noutput_stream=...\nwhat=y * 3 ] k\n@@ -760,13 +845,15 @@ transforms: ('jvp', 'transpose') what: x * 2\n\"\"\"\n{ lambda ; a.\nlet b = mul a 2.00\n- c = id_tap[ batch_dims=(0,)\n+ c = id_tap[ arg_treedef=*\n+ batch_dims=(0,)\nfunc=_print\noutput_stream=...\ntransforms=('batch',)\nwhat=a * 2 ] b\nd = mul c 3.00\n- e f = id_tap[ batch_dims=(0, 0)\n+ e f = id_tap[ arg_treedef=*\n+ batch_dims=(0, 0)\nfunc=_print\nnr_untapped=1\noutput_stream=...\n@@ -788,7 +875,7 @@ batch_dims: (0, 0) transforms: ('batch',) what: y * 3\nx = 3.\ndef func(y):\n# x is not mapped, y is mapped\n- _, y = hcb.id_print(x, y, output_stream=testing_stream)\n+ _, y = hcb.id_print((x, y), output_stream=testing_stream)\nreturn x + y\nvmap_func = api.vmap(func)\n@@ -796,7 +883,8 @@ batch_dims: (0, 0) transforms: ('batch',) what: y * 3\nself.assertMultiLineStrippedEqual(\n\"\"\"\n{ lambda ; a.\n- let b c = id_tap[ batch_dims=(None, 0)\n+ let b c = id_tap[ arg_treedef=PyTreeDef(tuple, [*,*])\n+ batch_dims=(None, 0)\nfunc=_print\noutput_stream=...\ntransforms=('batch',) ] 3.00 a\n@@ -807,8 +895,8 @@ batch_dims: (0, 0) transforms: ('batch',) what: y * 3\nself.assertMultiLineStrippedEqual(\n\"\"\"\nbatch_dims: (None, 0) transforms: ('batch',)\n-3.00\n-[4.00 5.00]\n+[ 3.00\n+ [4.00 5.00] ]\n\"\"\", testing_stream.output)\ntesting_stream.reset()\n" } ]
Python
Apache License 2.0
google/jax
Implemented pytree support for arg and result. Enabled outfeed for all arrays as a tuple
260,411
08.05.2020 10:01:28
-10,800
e1cb0324564d49812784590ac956d3e1f141b603
Prepare version 0.1.47 for jaxlib
[ { "change_type": "MODIFY", "old_path": "WORKSPACE", "new_path": "WORKSPACE", "diff": "@@ -28,10 +28,10 @@ http_archive(\n# and update the sha256 with the result.\nhttp_archive(\nname = \"org_tensorflow\",\n- sha256 = \"77338daa0112a62989e610fa1ee5a24afac5d139bd5bde476d6f2a82922a425e\",\n- strip_prefix = \"tensorflow-b99e88eee6a854f0a0e6012f01de87ae0248f8af\",\n+ sha256 = \"642f5a1bc191dfb96b2d7ed1cfb8f2a1515b5169b8de4381c75193cef8404b92\",\n+ strip_prefix = \"tensorflow-b25fb1fe32094b60f5a53ad5f986ad65a9f05919\",\nurls = [\n- \"https://github.com/tensorflow/tensorflow/archive/b99e88eee6a854f0a0e6012f01de87ae0248f8af.tar.gz\",\n+ \"https://github.com/tensorflow/tensorflow/archive/b25fb1fe32094b60f5a53ad5f986ad65a9f05919.tar.gz\",\n],\n)\n" }, { "change_type": "MODIFY", "old_path": "docs/CHANGELOG.rst", "new_path": "docs/CHANGELOG.rst", "diff": "@@ -20,6 +20,12 @@ jax 0.1.67 (unreleased)\ntightened. This may break code that was making use of names that were\npreviously exported accidentally.\n+jaxlib 0.1.47 (May 8, 2020)\n+------------------------------\n+\n+* Fixes crash for outfeed.\n+\n+\njax 0.1.66 (May 5, 2020)\n---------------------------\n" }, { "change_type": "MODIFY", "old_path": "jaxlib/version.py", "new_path": "jaxlib/version.py", "diff": "# See the License for the specific language governing permissions and\n# limitations under the License.\n-__version__ = \"0.1.46\"\n+__version__ = \"0.1.47\"\n" } ]
Python
Apache License 2.0
google/jax
Prepare version 0.1.47 for jaxlib (#3008)
260,335
08.05.2020 17:58:02
25,200
2b622943f4dd9e8391c546967b0df0c9d60e6d5e
improve pmap static broadcasted kwarg error msg fixes
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -138,9 +138,9 @@ def jit(fun: Callable, static_argnums: Union[int, Iterable[int]] = (),\nif _jit_is_disabled():\nreturn fun(*args, **kwargs)\nif static_argnums and max(static_argnums) >= len(args):\n- msg = (\"Jitted function has static_argnums={} but was called with only {}\"\n+ msg = (\"jitted function has static_argnums={} but was called with only {}\"\n\" positional arguments.\")\n- raise TypeError(msg.format(static_argnums, len(args)))\n+ raise ValueError(msg.format(static_argnums, len(args)))\nf = lu.wrap_init(fun)\nif static_argnums:\ndyn_argnums = [i for i in range(len(args)) if i not in static_argnums]\n@@ -771,21 +771,23 @@ def vmap(fun: Callable, in_axes=0, out_axes=0) -> Callable:\nreturn batched_fun\n-def _get_axis_size(i:int, shape: Tuple[int, ...], axis: int):\n+def _get_axis_size(name: str, i:int, shape: Tuple[int, ...], axis: int):\ntry:\nreturn shape[axis]\nexcept (IndexError, TypeError) as e:\n- raise ValueError(f\"vmap got arg {i} of rank {len(shape)} but axis to be mapped {axis}\") from e\n+ raise ValueError(f\"{name} got arg {i} of rank {len(shape)} \"\n+ f\"but axis to be mapped {axis}\") from e\ndef _mapped_axis_size(tree, vals, dims, name):\n- mapped_axis_sizes = {_get_axis_size(i, onp.shape(x), d) for i, (x, d) in enumerate(zip(vals, dims))\n+ mapped_axis_sizes = {_get_axis_size(name, i, onp.shape(x), d)\n+ for i, (x, d) in enumerate(zip(vals, dims))\nif d is not None}\ntry:\nsize, = mapped_axis_sizes\nreturn size\nexcept ValueError as e:\nif not mapped_axis_sizes:\n- raise ValueError(\"{} must have at least one non-None in_axes\".format(name)) from e\n+ raise ValueError(f\"{name} must have at least one non-None value in in_axes\") from e\nmsg = \"{} got inconsistent sizes for array axes to be mapped:\\n\".format(name) + \"{}\"\n# we switch the error message based on whether args is a tuple of arrays,\n# in which case we can produce an error message based on argument indices,\n@@ -1033,7 +1035,14 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\ndef f_pmapped(*args, **kwargs):\nf = lu.wrap_init(fun)\nif static_broadcasted_argnums:\n- dyn_argnums = [i for i in range(len(args)) if i not in static_broadcasted_argnums]\n+ if max(static_broadcasted_argnums) >= len(args):\n+ msg = (\"pmapped function has static_broadcasted_argnums={} but was \"\n+ \"called with only {} positional argument{}. All static \"\n+ \"broadcasted arguments must be passed positionally.\")\n+ raise ValueError(msg.format(static_broadcasted_argnums, len(args),\n+ \"s\" if len(args) > 1 else \"\"))\n+ dyn_argnums = [i for i in range(len(args))\n+ if i not in static_broadcasted_argnums]\nf, dyn_args = argnums_partial(f, dyn_argnums, args)\nif isinstance(in_axes, tuple):\ndyn_in_axes = tuple(in_axes[i] for i in dyn_argnums)\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -1080,7 +1080,7 @@ class APITest(jtu.JaxTestCase):\napi.vmap(lambda x: x, in_axes=(jnp.array([1., 2.]),))(jnp.array([1., 2.]))\nwith self.assertRaisesRegex(\n- ValueError, \"vmap must have at least one non-None in_axes\"):\n+ ValueError, \"vmap must have at least one non-None value in in_axes\"):\n# If the output is mapped, there must be a non-None in_axes\napi.vmap(lambda x: x, in_axes=None)(jnp.array([1., 2.]))\n@@ -1098,7 +1098,6 @@ class APITest(jtu.JaxTestCase):\n# If the output is mapped, then there must be some out_axes specified\napi.vmap(lambda x: x, out_axes=None)(jnp.array([1., 2.]))\n-\ndef test_vmap_structured_in_axes(self):\nA, B, C, D = 2, 3, 4, 5\n@@ -1655,6 +1654,18 @@ class APITest(jtu.JaxTestCase):\nre.DOTALL)):\napi.jit(func1)(2.)\n+ def test_pmap_static_kwarg_error_message(self):\n+ # https://github.com/google/jax/issues/3007\n+ def f(a, b):\n+ return a + b\n+\n+ g = jax.pmap(f, static_broadcasted_argnums=(1,))\n+\n+ msg = (r\"pmapped function has static_broadcasted_argnums=\\(1,\\) but was \"\n+ r\"called with only 1 positional argument. All static broadcasted \"\n+ r\"arguments must be passed positionally.\")\n+ with self.assertRaisesRegex(ValueError, msg):\n+ g(jnp.ones((1, 1)), b=1)\nclass JaxprTest(jtu.JaxTestCase):\n" } ]
Python
Apache License 2.0
google/jax
improve pmap static broadcasted kwarg error msg (#3018) fixes #3007
260,335
08.05.2020 17:58:25
25,200
fb685ff214d15bfdb9848bddaa4666c954a9324b
sort supported dtypes in host_callback_test.py * sort supported dtypes in host_callback_test.py This fixes issues I ran into with running `pytest -n auto tests/host_callback_test.py` or similar. * remove unused import
[ { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -46,6 +46,9 @@ def skip_if_jit_not_enabled():\nif os.getenv(\"JAX_ENABLE_JIT_PRINT\", \"false\") == \"false\":\nraise SkipTest(\"print jit not enabled yet; use JAX_ENABLE_JIT_PRINT env.\")\n+def supported_dtypes():\n+ return sorted(jtu.supported_dtypes(), key=lambda x: onp.dtype(x).name)\n+\nclass _TestingOutputStream(object):\n\"\"\"Use as `output_stream` for tests.\"\"\"\n@@ -532,7 +535,7 @@ where: 10\ndtype=dtype,\nnr_args=nr_args) for nr_args in [1, 2]\nfor shape in [(), (2,), (2, 3), (2, 3, 4)]\n- for dtype in jtu.supported_dtypes()))\n+ for dtype in supported_dtypes()))\ndef test_jit_types(self, nr_args=2, dtype=np.int16, shape=(2,)):\nif dtype in (np.complex64, np.complex128, np.bool_):\nraise SkipTest(f\"id_print jit not implemented for {dtype}.\")\n" } ]
Python
Apache License 2.0
google/jax
sort supported dtypes in host_callback_test.py (#3020) * sort supported dtypes in host_callback_test.py This fixes issues I ran into with running `pytest -n auto tests/host_callback_test.py` or similar. * remove unused import
260,568
10.05.2020 05:04:42
-10,800
7d3c886bf66a7f8ac528a46c7a168e144061efcb
Implement np.nanvar and np.nanstd * Implement nanvar & nanstd Add tests for nanvar & nanstd * Clean up bfloat16 tests for np.nanvar and np.nanstd * add nanvar & nanstd to the whitelist ignore numpy ddof warnings
[ { "change_type": "MODIFY", "old_path": "jax/numpy/__init__.py", "new_path": "jax/numpy/__init__.py", "diff": "@@ -43,10 +43,10 @@ from .lax_numpy import (\nlogical_xor, logspace, mask_indices, matmul, max, maximum, mean, median,\nmeshgrid, min, minimum, mod, moveaxis, msort, multiply, nan, nan_to_num,\nnanargmax, nanargmin, nancumprod, nancumsum, nanmax, nanmean, nanmin,\n- nanprod, nansum, ndarray, ndim, negative, newaxis, nextafter, nonzero,\n- not_equal, number, numpy_version, object_, ones, ones_like, operator_name,\n- outer, packbits, pad, percentile, pi, polyval, positive, power, prod,\n- product, promote_types, ptp, quantile, rad2deg, radians, ravel, real,\n+ nanprod, nanstd, nansum, nanvar, ndarray, ndim, negative, newaxis, nextafter,\n+ nonzero, not_equal, number, numpy_version, object_, ones, ones_like,\n+ operator_name, outer, packbits, pad, percentile, pi, polyval, positive, power,\n+ prod, product, promote_types, ptp, quantile, rad2deg, radians, ravel, real,\nreciprocal, remainder, repeat, reshape, result_type, right_shift, rint,\nroll, rollaxis, rot90, round, row_stack, save, savez, searchsorted, select,\nset_printoptions, shape, sign, signbit, signedinteger, sin, sinc, single,\n" }, { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -1585,23 +1585,7 @@ def var(a, axis=None, dtype=None, out=None, ddof=0, keepdims=False):\nif out is not None:\nraise ValueError(\"var does not support the `out` argument.\")\n- a_dtype = _dtype(a)\n- if dtype:\n- if (not issubdtype(dtype, complexfloating) and\n- issubdtype(a_dtype, complexfloating)):\n- msg = (\"jax.numpy.var does not yet support real dtype parameters when \"\n- \"computing the variance of an array of complex values. The \"\n- \"semantics of numpy.var seem unclear in this case. Please comment \"\n- \"on https://github.com/google/jax/issues/2283 if this behavior is \"\n- \"important to you.\")\n- raise ValueError(msg)\n- a_dtype = promote_types(a_dtype, dtype)\n- else:\n- if not issubdtype(a_dtype, inexact):\n- dtype = a_dtype = float_\n- else:\n- dtype = _complex_elem_type(a_dtype)\n- a_dtype = promote_types(a_dtype, float32)\n+ a_dtype, dtype = _var_promote_types(_dtype(a), dtype)\na_mean = mean(a, axis, dtype=a_dtype, keepdims=True)\ncentered = a - a_mean\nif issubdtype(centered.dtype, complexfloating):\n@@ -1620,6 +1604,25 @@ def var(a, axis=None, dtype=None, out=None, ddof=0, keepdims=False):\nreturn lax.convert_element_type(out, dtype)\n+def _var_promote_types(a_dtype, dtype):\n+ if dtype:\n+ if (not issubdtype(dtype, complexfloating) and\n+ issubdtype(a_dtype, complexfloating)):\n+ msg = (\"jax.numpy.var does not yet support real dtype parameters when \"\n+ \"computing the variance of an array of complex values. The \"\n+ \"semantics of numpy.var seem unclear in this case. Please comment \"\n+ \"on https://github.com/google/jax/issues/2283 if this behavior is \"\n+ \"important to you.\")\n+ raise ValueError(msg)\n+ a_dtype = promote_types(a_dtype, dtype)\n+ else:\n+ if not issubdtype(a_dtype, inexact):\n+ dtype = a_dtype = float_\n+ else:\n+ dtype = _complex_elem_type(a_dtype)\n+ a_dtype = promote_types(a_dtype, float32)\n+ return a_dtype, dtype\n+\n@_wraps(onp.std)\ndef std(a, axis=None, dtype=None, out=None, ddof=0, keepdims=False):\n@@ -1699,6 +1702,37 @@ def nanmean(a, axis=None, dtype=None, out=None, keepdims=False):\nreturn td\n+@_wraps(onp.nanvar)\n+def nanvar(a, axis=None, dtype=None, out=None, ddof=0, keepdims=False):\n+ if out is not None:\n+ raise ValueError(\"nanvar does not support the `out` argument.\")\n+\n+ a_dtype, dtype = _var_promote_types(_dtype(a), dtype)\n+ a_mean = nanmean(a, axis, dtype=a_dtype, keepdims=True)\n+ centered = a - a_mean\n+ if issubdtype(centered.dtype, complexfloating):\n+ centered = lax.real(lax.mul(centered, lax.conj(centered)))\n+ else:\n+ centered = lax.square(centered)\n+\n+ normalizer = sum(logical_not(isnan(a)), axis=axis, keepdims=keepdims)\n+ normalizer = normalizer - ddof\n+ normalizer_mask = lax.le(normalizer, 0)\n+\n+ result = nansum(centered, axis, keepdims=keepdims)\n+ result = where(normalizer_mask, nan, result)\n+ divisor = where(normalizer_mask, 1, normalizer)\n+ out = lax.div(result, lax.convert_element_type(divisor, result.dtype))\n+ return lax.convert_element_type(out, dtype)\n+\n+\n+@_wraps(onp.nanstd)\n+def nanstd(a, axis=None, dtype=None, out=None, ddof=0, keepdims=False):\n+ if out is not None:\n+ raise ValueError(\"nanstd does not support the `out` argument.\")\n+ return sqrt(nanvar(a, axis=axis, dtype=dtype, ddof=ddof, keepdims=keepdims))\n+\n+\ndef _make_cumulative_reduction(onp_reduction, reduction, squash_nan=False):\n# We want to allow XLA to fuse the pad and reduce-window operators to\n# avoid materializing the padded output.\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -292,6 +292,10 @@ JAX_REDUCER_NO_DTYPE_RECORDS = [\ninexact=True),\nop_record(\"std\", 1, all_dtypes, nonempty_shapes, jtu.rand_default, [],\ninexact=True),\n+ op_record(\"nanvar\", 1, all_dtypes, nonempty_shapes, jtu.rand_some_nan,\n+ [], inexact=True),\n+ op_record(\"nanstd\", 1, all_dtypes, nonempty_shapes, jtu.rand_some_nan,\n+ [], inexact=True),\n]\nJAX_ARGMINMAX_RECORDS = [\n@@ -629,7 +633,14 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\ndef testReducerNoDtype(self, onp_op, jnp_op, rng_factory, shape, dtype, axis,\nkeepdims, inexact):\nrng = rng_factory(self.rng())\n- onp_fun = lambda x: onp_op(x, axis, keepdims=keepdims)\n+ is_bf16_nan_test = dtype == jnp.bfloat16 and rng_factory.__name__ == 'rand_some_nan'\n+ @jtu.ignore_warning(category=RuntimeWarning,\n+ message=\"Degrees of freedom <= 0 for slice.*\")\n+ def onp_fun(x):\n+ x_cast = x if not is_bf16_nan_test else x.astype(onp.float32)\n+ res = onp_op(x_cast, axis, keepdims=keepdims)\n+ res = res if not is_bf16_nan_test else res.astype(jnp.bfloat16)\n+ return res\nonp_fun = _promote_like_jnp(onp_fun, inexact)\nonp_fun = jtu.ignore_warning(category=onp.ComplexWarning)(onp_fun)\njnp_fun = lambda x: jnp_op(x, axis, keepdims=keepdims)\n@@ -2780,6 +2791,39 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nself._CompileAndCheck(jnp_fun, args_maker, check_dtypes=True, rtol=tol,\natol=tol)\n+ @parameterized.named_parameters(\n+ jtu.cases_from_list(\n+ {\"testcase_name\":\n+ \"_shape={}_dtype={}_out_dtype={}_axis={}_ddof={}_keepdims={}\"\n+ .format(shape, dtype, out_dtype, axis, ddof, keepdims),\n+ \"shape\": shape, \"dtype\": dtype, \"out_dtype\": out_dtype, \"axis\": axis,\n+ \"ddof\": ddof, \"keepdims\": keepdims, \"rng_factory\": rng_factory}\n+ for shape in [(5,), (10, 5)]\n+ for dtype in all_dtypes\n+ for out_dtype in inexact_dtypes\n+ for axis in [None, 0, -1]\n+ for ddof in [0, 1, 2]\n+ for keepdims in [False, True]\n+ for rng_factory in [jtu.rand_some_nan]))\n+ def testNanVar(self, shape, dtype, out_dtype, axis, ddof, keepdims, rng_factory):\n+ rng = rng_factory(self.rng())\n+ args_maker = self._GetArgsMaker(rng, [shape], [dtype])\n+ def onp_fun(x):\n+ out = onp.nanvar(x.astype(jnp.promote_types(onp.float32, dtype)),\n+ axis=axis, ddof=ddof, keepdims=keepdims)\n+ return out.astype(out_dtype)\n+ jnp_fun = partial(jnp.nanvar, dtype=out_dtype, axis=axis, ddof=ddof, keepdims=keepdims)\n+ tol = jtu.tolerance(out_dtype, {onp.float16: 1e-1, onp.float32: 1e-3,\n+ onp.float64: 1e-3, onp.complex128: 1e-6})\n+ if (jnp.issubdtype(dtype, jnp.complexfloating) and\n+ not jnp.issubdtype(out_dtype, jnp.complexfloating)):\n+ self.assertRaises(ValueError, lambda: jnp_fun(*args_maker()))\n+ else:\n+ self._CheckAgainstNumpy(onp_fun, jnp_fun, args_maker, check_dtypes=True,\n+ tol=tol)\n+ self._CompileAndCheck(jnp_fun, args_maker, check_dtypes=True, rtol=tol,\n+ atol=tol)\n+\n@parameterized.named_parameters(\njtu.cases_from_list(\n{\"testcase_name\": \"_shape={}_dtype={}_rowvar={}_ddof={}_bias={}\".format(\n" } ]
Python
Apache License 2.0
google/jax
Implement np.nanvar and np.nanstd (#2310) * Implement nanvar & nanstd Add tests for nanvar & nanstd * Clean up bfloat16 tests for np.nanvar and np.nanstd * add nanvar & nanstd to the whitelist ignore numpy ddof warnings
260,411
10.05.2020 13:16:16
-10,800
b3ae01d1798eae7d0ce16ad4feece8bf108eb10b
Use a new variable for static_broadcasted_argnums as a tuple. * Use a new variable for static_broadcasted_argnums as a tuple. This works around a bug in pytype (b/156151503).
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -1022,7 +1022,9 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\n_check_callable(fun)\naxis_name = _TempAxisName(fun) if axis_name is None else axis_name\nif isinstance(static_broadcasted_argnums, int):\n- static_broadcasted_argnums = (static_broadcasted_argnums,)\n+ static_broadcasted_tuple: Tuple[int, ...] = (static_broadcasted_argnums,)\n+ else:\n+ static_broadcasted_tuple = tuple(static_broadcasted_argnums)\n# axis_size is an optional integer representing the global axis size.\n# The aggregate size (across all hosts) size of the mapped axis must match\n@@ -1034,15 +1036,15 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\n@wraps(fun)\ndef f_pmapped(*args, **kwargs):\nf = lu.wrap_init(fun)\n- if static_broadcasted_argnums:\n- if max(static_broadcasted_argnums) >= len(args):\n+ if static_broadcasted_tuple:\n+ if max(static_broadcasted_tuple) >= len(args):\nmsg = (\"pmapped function has static_broadcasted_argnums={} but was \"\n\"called with only {} positional argument{}. All static \"\n\"broadcasted arguments must be passed positionally.\")\n- raise ValueError(msg.format(static_broadcasted_argnums, len(args),\n+ raise ValueError(msg.format(static_broadcasted_tuple, len(args),\n\"s\" if len(args) > 1 else \"\"))\ndyn_argnums = [i for i in range(len(args))\n- if i not in static_broadcasted_argnums]\n+ if i not in static_broadcasted_tuple]\nf, dyn_args = argnums_partial(f, dyn_argnums, args)\nif isinstance(in_axes, tuple):\ndyn_in_axes = tuple(in_axes[i] for i in dyn_argnums)\n" } ]
Python
Apache License 2.0
google/jax
Use a new variable for static_broadcasted_argnums as a tuple. (#3027) * Use a new variable for static_broadcasted_argnums as a tuple. This works around a bug in pytype (b/156151503).
260,411
10.05.2020 14:25:18
-10,800
bc2d2c8ac9648262051e4daca9ab38ffcef9a03c
Fix uses of deprecated onp. in pmap_test
[ { "change_type": "MODIFY", "old_path": "tests/pmap_test.py", "new_path": "tests/pmap_test.py", "diff": "@@ -429,19 +429,19 @@ class PmapTest(jtu.JaxTestCase):\nreplicas = xla_bridge.device_count()\nif replicas % 2 != 0:\nraise SkipTest\n- axis_index_groups = onp.arange(replicas).reshape(\n+ axis_index_groups = np.arange(replicas).reshape(\n2, replicas // 2).tolist()\nf = lambda x: x - lax.psum(x, 'i', axis_index_groups=axis_index_groups)\nf = pmap(f, 'i')\nshape = (replicas, 4)\n- x = onp.arange(prod(shape), dtype=onp.float32).reshape(shape)\n+ x = np.arange(prod(shape), dtype=np.float32).reshape(shape)\ndef sum_helper(a):\n- return onp.broadcast_to(a.sum(0, keepdims=True),\n+ return np.broadcast_to(a.sum(0, keepdims=True),\n(replicas // 2, x.shape[1]))\nexpected_psum_1 = sum_helper(x[:replicas // 2])\nexpected_psum_2 = sum_helper(x[replicas // 2:])\n- expected_psum = onp.concatenate([expected_psum_1, expected_psum_2], 0)\n+ expected_psum = np.concatenate([expected_psum_1, expected_psum_2], 0)\nexpected = x - expected_psum\nans = f(x)\n@@ -451,7 +451,7 @@ class PmapTest(jtu.JaxTestCase):\nreplicas = xla_bridge.device_count()\nif replicas % 4 != 0:\nraise SkipTest\n- axis_index_groups = onp.arange(replicas // 2).reshape(\n+ axis_index_groups = np.arange(replicas // 2).reshape(\n2, replicas // 4).tolist()\nf = lambda x: x - lax.psum(x, 'i', axis_index_groups=axis_index_groups)\nf1 = pmap(pmap(f, 'i'), 'j')\n@@ -459,13 +459,13 @@ class PmapTest(jtu.JaxTestCase):\nf3 = pmap(pmap(f, 'j'), 'i')\nshape = (2, replicas // 2, 4)\n- x = onp.arange(prod(shape), dtype=onp.float32).reshape(shape)\n+ x = np.arange(prod(shape), dtype=np.float32).reshape(shape)\ndef sum_helper_f1(a):\n- return onp.broadcast_to(a.sum(1, keepdims=True),\n+ return np.broadcast_to(a.sum(1, keepdims=True),\n(shape[0], shape[1] // 2, shape[2]))\nexpected_psum_1 = sum_helper_f1(x[:, :replicas // 4])\nexpected_psum_2 = sum_helper_f1(x[:, replicas // 4:])\n- expected_psum = onp.concatenate([expected_psum_1, expected_psum_2], 1)\n+ expected_psum = np.concatenate([expected_psum_1, expected_psum_2], 1)\nexpected = x - expected_psum\nans = f1(x)\nself.assertAllClose(ans, expected, check_dtypes=True)\n@@ -475,13 +475,13 @@ class PmapTest(jtu.JaxTestCase):\nself.assertAllClose(ans, expected, check_dtypes=True)\nshape = (replicas // 2, 2, 4)\n- x = onp.arange(prod(shape), dtype=onp.float32).reshape(shape)\n+ x = np.arange(prod(shape), dtype=np.float32).reshape(shape)\ndef sum_helper_f3(a):\n- return onp.broadcast_to(a.sum(0, keepdims=True),\n+ return np.broadcast_to(a.sum(0, keepdims=True),\n(shape[0] // 2, shape[1], shape[2]))\nexpected_psum_1 = sum_helper_f3(x[:replicas // 4])\nexpected_psum_2 = sum_helper_f3(x[replicas // 4:])\n- expected_psum = onp.concatenate([expected_psum_1, expected_psum_2], 0)\n+ expected_psum = np.concatenate([expected_psum_1, expected_psum_2], 0)\nexpected = x - expected_psum\nans = f3(x)\nself.assertAllClose(ans, expected, check_dtypes=True)\n" } ]
Python
Apache License 2.0
google/jax
Fix uses of deprecated onp. in pmap_test (#3028)
260,411
10.05.2020 19:54:46
-10,800
c171c33b1c3aef0427531b4805213430f3633dbd
Update numpy references to use np. Added to Changelog
[ { "change_type": "MODIFY", "old_path": "docs/CHANGELOG.rst", "new_path": "docs/CHANGELOG.rst", "diff": "@@ -18,6 +18,9 @@ jax 0.1.67 (unreleased)\n* Support for reduction over subsets of a pmapped axis using ``axis_index_groups``\n`#2382 <https://github.com/google/jax/pull/2382>`_.\n+ * Experimental support for printing and calling host-side Python function from\n+ compiled code. See `id_print and id_tap <https://jax.readthedocs.io/en/latest/jax.experimental.host_callback.html>`_\n+ (`#3006 <https://github.com/google/jax/pull/3006>`_).\n* Notable changes:\n@@ -213,7 +216,7 @@ jax 0.1.59 (February 11, 2020)\n* Added JAX_SKIP_SLOW_TESTS environment variable to skip tests known as slow.\njaxlib 0.1.39 (February 11, 2020)\n---------------------------------\n+---------------------------------\n* Updates XLA.\n" }, { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -113,21 +113,16 @@ Still to do:\n* Explore implementation with outside compilation.\n\"\"\"\n-from collections import defaultdict, namedtuple\nfrom concurrent import futures\nfrom contextlib import contextmanager\n-from functools import partial\n-import io\nimport itertools\n-from jax import abstract_arrays\n-from jax import ad_util\nfrom jax import api\nfrom jax import core\nfrom jax import dtypes\nfrom jax import lax\n-from jax.lib import pytree, xla_bridge\n-from jax.interpreters import ad, xla, batching, masking, pxla\n+from jax.lib import pytree\n+from jax.interpreters import ad, xla, batching, masking\nfrom jax.interpreters import partial_eval as pe\nfrom jax import pprint_util as ppu\nfrom jax import util\n@@ -137,8 +132,7 @@ from jaxlib import version as jaxlib_version\nimport logging\nimport msgpack # type: ignore\n-import numpy as onp\n-import sys\n+import numpy as np\nimport traceback\nfrom typing import Any, Callable, Dict, Iterable, List, Optional, NamedTuple, Sequence, Tuple\n@@ -298,8 +292,8 @@ def _print_consumer(arg, *, output_stream=None,\n[ppu.pp(f\"{k}=\") >> pp_val(v)\nfor k, v in sorted(arg.items())]) >>\nppu.pp(' }'))\n- elif isinstance(arg, onp.ndarray):\n- return ppu.pp(onp.array2string(arg, threshold=threshold))\n+ elif isinstance(arg, np.ndarray):\n+ return ppu.pp(np.array2string(arg, threshold=threshold))\nelse:\nreturn ppu.pp(str(arg))\n@@ -605,18 +599,18 @@ _OUTFEED_HEADER_START = 271828 # [0]\n_OUTFEED_HEADER_METADATA_LENGTH = 4 * (_OUTFEED_HEADER_LENGTH - 4)\n_CODE_TO_DTYPE = {\n- 0: onp.dtype(onp.int8),\n- 1: onp.dtype(onp.int16),\n- 2: onp.dtype(onp.int32),\n- 3: onp.dtype(onp.int64),\n- 4: onp.dtype(onp.uint8),\n- 5: onp.dtype(onp.uint16),\n- 6: onp.dtype(onp.uint32),\n- 7: onp.dtype(onp.uint64),\n- 8: onp.dtype(onp.float16),\n- 9: onp.dtype(onp.float32),\n- 10: onp.dtype(onp.float64),\n- 11: onp.dtype(dtypes.bfloat16),\n+ 0: np.dtype(np.int8),\n+ 1: np.dtype(np.int16),\n+ 2: np.dtype(np.int32),\n+ 3: np.dtype(np.int64),\n+ 4: np.dtype(np.uint8),\n+ 5: np.dtype(np.uint16),\n+ 6: np.dtype(np.uint32),\n+ 7: np.dtype(np.uint64),\n+ 8: np.dtype(np.float16),\n+ 9: np.dtype(np.float32),\n+ 10: np.dtype(np.float64),\n+ 11: np.dtype(dtypes.bfloat16),\n}\n_DTYPE_STR_TO_CODE = dict([(str(d), c) for c, d in _CODE_TO_DTYPE.items()])\n@@ -627,7 +621,7 @@ def _emit_outfeed(comp: XlaComputationBuilder, token: XlaOp,\narrays_shape = [comp.GetShape(a) for a in arrays]\ndef _array_shape_to_tuple(a_shape: XlaShape):\n# (element_type_code, (d0, d1, ..., dn))\n- return (_DTYPE_STR_TO_CODE[str(onp.dtype(a_shape.element_type()))],\n+ return (_DTYPE_STR_TO_CODE[str(np.dtype(a_shape.element_type()))],\na_shape.dimensions())\nmetadata = msgpack.dumps(tuple(map(_array_shape_to_tuple, arrays_shape)))\nmetadata_len = len(metadata)\n@@ -641,7 +635,7 @@ def _emit_outfeed(comp: XlaComputationBuilder, token: XlaOp,\ntuple([int.from_bytes(metadata[i:i+4], byteorder=\"big\")\nfor i in range(0, _OUTFEED_HEADER_METADATA_LENGTH, 4)]))\nheader += (0,) * (_OUTFEED_HEADER_LENGTH - len(header))\n- data = xops.ConstantLiteral(comp, onp.array(header, dtype=onp.uint32))\n+ data = xops.ConstantLiteral(comp, np.array(header, dtype=np.uint32))\ntoken = xops.OutfeedWithToken(data, token, comp.GetShape(data))\n# Now send the arrays, all at once\n@@ -657,7 +651,7 @@ def _receive_outfeed(device: XlaDevice, receiver_name: str\nReturns: a tuple with the consumer_id, the arrays received, and\na kwargs dictionary that was passed to _emit_outfeed.\n\"\"\"\n- header_shape = xla_client.Shape.array_shape(onp.dtype(onp.uint32),\n+ header_shape = xla_client.Shape.array_shape(np.dtype(np.uint32),\n(_OUTFEED_HEADER_LENGTH,))\ndef _get_data(data_shape: XlaShape, device: XlaDevice) -> XlaShape:\n" }, { "change_type": "MODIFY", "old_path": "tests/host_callback_test.py", "new_path": "tests/host_callback_test.py", "diff": "@@ -18,10 +18,9 @@ from __future__ import print_function\nfrom functools import partial\nimport logging\n-import numpy as onp\n+import numpy as np\nimport os\nimport re\n-import threading\nfrom typing import Any, Callable, List, Sequence, Tuple\nfrom unittest import SkipTest\n@@ -30,12 +29,10 @@ from absl.testing import parameterized\nfrom jax import api\nfrom jax import lax\n-from jax import numpy as np\n+from jax import numpy as jnp\nfrom jax import test_util as jtu\nfrom jax.config import config\nfrom jax.experimental import host_callback as hcb\n-from jax.interpreters import xla\n-from jax.interpreters import partial_eval as pe\nfrom jax.lib import xla_bridge\n@@ -47,7 +44,7 @@ def skip_if_jit_not_enabled():\nraise SkipTest(\"print jit not enabled yet; use JAX_ENABLE_JIT_PRINT env.\")\ndef supported_dtypes():\n- return sorted(jtu.supported_dtypes(), key=lambda x: onp.dtype(x).name)\n+ return sorted(jtu.supported_dtypes(), key=lambda x: np.dtype(x).name)\nclass _TestingOutputStream(object):\n\"\"\"Use as `output_stream` for tests.\"\"\"\n@@ -91,7 +88,7 @@ def assertMultiLineStrippedEqual(tst: jtu.JaxTestCase, expected: str, what: str)\nmatched = match_group.group(0)\nif matched == \".\": return matched\n# TODO: why can't we use here np.around?\n- x = onp.around(float(matched), decimals=2)\n+ x = np.around(float(matched), decimals=2)\nreturn f\"{x:.2f}\"\nwhat = re.sub(r\"\\-?\\d*\\.[\\-\\def]*\", repl_floats, what)\nwhat = re.sub(r\"output_stream=[^\\]\\n]*\", \"\", what)\n@@ -487,7 +484,7 @@ where: end\nx3 + 1, lambda x: hcb.id_print(-1, where=\"s_f\", result=x, output_stream=testing_stream))\nreturn (c, hcb.id_print(x4, where=\"s_2\", output_stream=testing_stream))\n- _, x10 = lax.scan(body, x2, np.arange(3))\n+ _, x10 = lax.scan(body, x2, jnp.arange(3))\nres = hcb.id_print(x10, where=\"10\", output_stream=testing_stream)\nreturn res\n@@ -499,7 +496,7 @@ where: end\nif with_jit:\nfunc = api.jit(func)\nres = func(1)\n- self.assertAllClose(np.array([1, 2, 3]), res, check_dtypes=True)\n+ self.assertAllClose(jnp.array([1, 2, 3]), res, check_dtypes=True)\nassertMultiLineStrippedEqual(self, \"\"\"\nwhere: 1\n1\n@@ -536,13 +533,13 @@ where: 10\nnr_args=nr_args) for nr_args in [1, 2]\nfor shape in [(), (2,), (2, 3), (2, 3, 4)]\nfor dtype in supported_dtypes()))\n- def test_jit_types(self, nr_args=2, dtype=np.int16, shape=(2,)):\n- if dtype in (np.complex64, np.complex128, np.bool_):\n+ def test_jit_types(self, nr_args=2, dtype=jnp.int16, shape=(2,)):\n+ if dtype in (jnp.complex64, jnp.complex128, jnp.bool_):\nraise SkipTest(f\"id_print jit not implemented for {dtype}.\")\nif jtu.device_under_test() == \"tpu\":\n- if dtype in (np.int16,):\n+ if dtype in (jnp.int16,):\nraise SkipTest(f\"transfering {dtype} not supported on TPU\")\n- args = [np.arange(np.prod(shape), dtype=dtype).reshape(shape)]\n+ args = [jnp.arange(jnp.prod(shape), dtype=dtype).reshape(shape)]\nif nr_args > 1:\nargs = args * nr_args\njit_fun1 = api.jit(lambda xs: hcb.id_print(\n@@ -554,14 +551,14 @@ where: 10\n# self.assertAllClose(args, res, check_dtypes=True)\ndef test_jit_large(self):\n- arg = np.arange(10000, dtype=np.int32).reshape((10, 10, 5, -1))\n+ arg = jnp.arange(10000, dtype=jnp.int32).reshape((10, 10, 5, -1))\nwith hcb.outfeed_receiver(receiver_name=self._testMethodName):\napi.jit(hcb.id_print)(arg)\ndef test_jit_several_together(self):\n- arg = np.arange(50, dtype=np.int32).reshape((10, 5))\n+ arg = jnp.arange(50, dtype=jnp.int32).reshape((10, 5))\nwith hcb.outfeed_receiver(receiver_name=self._testMethodName):\n- api.jit(lambda x, y: hcb.id_print((x, y, x * 2.)))(arg, np.ones(100, dtype=np.int32))\n+ api.jit(lambda x, y: hcb.id_print((x, y, x * 2.)))(arg, jnp.ones(100, dtype=jnp.int32))\ndef test_jit_interleaving(self):\n# Several jit's without data dependencies; they may interfere\n@@ -577,7 +574,7 @@ where: 10\nx = hcb.id_tap(tap_func, [x + i for i in range(nr_arrays)], i=i)[-1]\nreturn x\nwith hcb.outfeed_receiver(receiver_name=self._testMethodName):\n- x = np.array(1, dtype=onp.int32)\n+ x = jnp.array(1, dtype=np.int32)\nres = 0\nfor i in range(10):\n# No dependencies between the jit invocations\n@@ -687,7 +684,7 @@ what: x3\ndef test_while(self):\n\"\"\"Executing while, even without JIT uses compiled code\"\"\"\n- y = np.ones(5) # captured const\n+ y = jnp.ones(5) # captured const\ndef func(x):\nreturn lax.while_loop(\n@@ -705,7 +702,7 @@ what: x3\ndef test_while_error_no_receiver(self):\n\"\"\"Executing while needs the receiver\"\"\"\n- y = np.ones(5) # captured const\n+ y = jnp.ones(5) # captured const\ndef func(x):\nreturn lax.while_loop(\nlambda c: c[1] < 5,\n@@ -747,9 +744,9 @@ what: x3\nq = mul 2.00 p\nr = mul n q\nin (h, r) }\"\"\",\n- str(api.make_jaxpr(jvp_fun1)(np.float32(5.), np.float32(0.1))))\n+ str(api.make_jaxpr(jvp_fun1)(jnp.float32(5.), jnp.float32(0.1))))\nwith hcb.outfeed_receiver():\n- res_primals, res_tangents = jvp_fun1(np.float32(5.), np.float32(0.1))\n+ res_primals, res_tangents = jvp_fun1(jnp.float32(5.), jnp.float32(0.1))\nself.assertAllClose(100., res_primals, check_dtypes=False)\nself.assertAllClose(4., res_tangents, check_dtypes=False)\nassertMultiLineStrippedEqual(self, \"\"\"\n@@ -782,7 +779,7 @@ transforms: ('jvp', 'transpose') what: x * 3\ntesting_stream.reset()\nwith hcb.outfeed_receiver():\n- res_grad = grad_func(np.float32(5.))\n+ res_grad = grad_func(jnp.float32(5.))\nself.assertAllClose(6., res_grad, check_dtypes=False)\nassertMultiLineStrippedEqual(self, \"\"\"\n@@ -827,7 +824,7 @@ transforms: ('jvp', 'transpose') what: x * 3\nin (n,) }\"\"\", str(api.make_jaxpr(grad_func)(5.)))\nwith hcb.outfeed_receiver():\n- res_grad = grad_func(np.float32(5.))\n+ res_grad = grad_func(jnp.float32(5.))\nself.assertAllClose(2. * 5. * 6., res_grad, check_dtypes=False)\nassertMultiLineStrippedEqual(self, \"\"\"\nwhat: x * 2\n@@ -858,7 +855,7 @@ transforms: ('jvp', 'transpose') what: x * 2\ntransforms: ('jvp', 'transpose', 'jvp', 'transpose') what: x * 2\n2.00\"\"\", testing_stream.output)\ntesting_stream.reset()\n- res_grad = grad_func(np.float32(5.))\n+ res_grad = grad_func(jnp.float32(5.))\nself.assertAllClose(12., res_grad, check_dtypes=False)\nassertMultiLineStrippedEqual(self, \"\"\"\n@@ -875,7 +872,7 @@ transforms: ('jvp', 'transpose') what: x * 2\ndef test_vmap(self):\nvmap_fun1 = api.vmap(fun1)\n- vargs = np.array([np.float32(4.), np.float32(5.)])\n+ vargs = jnp.array([jnp.float32(4.), jnp.float32(5.)])\nassertMultiLineStrippedEqual(self, \"\"\"\n{ lambda ; a.\nlet b = mul a 2.00\n@@ -910,7 +907,7 @@ batch_dims: (0, 0) transforms: ('batch',) what: y * 3\nreturn x + y\nvmap_func = api.vmap(func)\n- vargs = np.array([np.float32(4.), np.float32(5.)])\n+ vargs = jnp.array([jnp.float32(4.), jnp.float32(5.)])\nassertMultiLineStrippedEqual(self, \"\"\"\n{ lambda ; a.\nlet b c = id_tap[ arg_treedef=PyTreeDef(tuple, [*,*])\n@@ -929,17 +926,17 @@ batch_dims: (None, 0) transforms: ('batch',)\ntesting_stream.reset()\ndef test_pmap(self):\n- vargs = 2. + np.arange(api.local_device_count(), dtype=np.float32)\n+ vargs = 2. + jnp.arange(api.local_device_count(), dtype=jnp.float32)\npmap_fun1 = api.pmap(fun1, axis_name=\"i\")\nwith hcb.outfeed_receiver(receiver_name=self._testMethodName):\nres = pmap_fun1(vargs)\n- expected_res = np.stack([fun1_equiv(2. + a) for a in range(api.local_device_count())])\n+ expected_res = jnp.stack([fun1_equiv(2. + a) for a in range(api.local_device_count())])\nself.assertAllClose(expected_res, res, check_dtypes=False)\ndef test_pmap_error_no_receiver(self):\n# Check for errors if starting jit without a consumer active\n- vargs = 2. + np.arange(api.local_device_count(), dtype=np.float32)\n+ vargs = 2. + jnp.arange(api.local_device_count(), dtype=jnp.float32)\nwith self.assertRaisesRegex(ValueError, \"outfeed_receiver is not started\"):\napi.pmap(lambda x: hcb.id_print(x))(vargs)\n@@ -948,8 +945,8 @@ batch_dims: (None, 0) transforms: ('batch',)\nraise SkipTest(\"masking has regressed\")\n@partial(api.mask, in_shapes=['n'], out_shape='')\ndef padded_sum(x):\n- return np.sum(hcb.id_print(x, what=\"x\", output_stream=testing_stream))\n- args = [np.arange(4)], dict(n=onp.int64(2))\n+ return jnp.sum(hcb.id_print(x, what=\"x\", output_stream=testing_stream))\n+ args = [jnp.arange(4)], dict(n=np.int64(2))\nassertMultiLineStrippedEqual(self, \"\"\"\n{ lambda c f ; a b.\nlet d = lt c b\n@@ -1003,9 +1000,9 @@ class OutfeedRewriterTest(jtu.JaxTestCase):\nin (c, e) }\"\"\", lambda x: hcb.id_print(x + x), [0])\ndef test_cond(self):\n- y = np.ones(5) # captured const\n+ y = jnp.ones(5) # captured const\ndef func(x, z):\n- return lax.cond(z > 0, (1, 2), lambda a: (a[0], np.zeros(5)),\n+ return lax.cond(z > 0, (1, 2), lambda a: (a[0], jnp.zeros(5)),\nz, lambda a: (hcb.id_print(a), y))\nself.assertRewrite(\"\"\"\n{ lambda d e ; a b h.\n@@ -1022,7 +1019,7 @@ class OutfeedRewriterTest(jtu.JaxTestCase):\nin (f, g, i) }\"\"\", func, [y, 5])\ndef test_while(self):\n- y = np.ones(5) # captured const\n+ y = jnp.ones(5) # captured const\ndef func(x):\nreturn lax.while_loop(lambda c: c[1] < 5,\n@@ -1047,7 +1044,7 @@ class OutfeedRewriterTest(jtu.JaxTestCase):\nin (c, 5, f) }\"\"\", func, [y])\ndef test_scan(self):\n- y = np.ones(5) # captured const\n+ y = jnp.ones(5) # captured const\ndef func(x):\nreturn lax.scan(lambda c, a: (hcb.id_print(c), y), (1, 2), x)\nself.assertRewrite(\"\"\"\n" } ]
Python
Apache License 2.0
google/jax
Update numpy references to use np. Added to Changelog (#3029)
260,411
11.05.2020 14:37:17
-10,800
7acb41d57c92e19a4d94fe85e82b3bad05cd7782
Fixed imports for readthedocs
[ { "change_type": "MODIFY", "old_path": "docs/developer.rst", "new_path": "docs/developer.rst", "diff": "@@ -188,10 +188,11 @@ I saw in the Readthedocs logs::\ncd jax\ngit checkout --force origin/test-docs\ngit clean -d -f -f\n+ workon jax-docs\npython -m pip install --upgrade --no-cache-dir pip\npython -m pip install --upgrade --no-cache-dir -I Pygments==2.3.1 setuptools==41.0.1 docutils==0.14 mock==1.0.1 pillow==5.4.1 alabaster>=0.7,<0.8,!=0.7.5 commonmark==0.8.1 recommonmark==0.5.0 'sphinx<2' 'sphinx-rtd-theme<0.5' 'readthedocs-sphinx-ext<1.1'\npython -m pip install --exists-action=w --no-cache-dir -r docs/requirements.txt\n-\n+ cd docs\npython `which sphinx-build` -T -E -b html -d _build/doctrees-readthedocs -D language=en . _build/html\n" }, { "change_type": "MODIFY", "old_path": "docs/requirements.txt", "new_path": "docs/requirements.txt", "diff": "@@ -4,6 +4,8 @@ jaxlib\nipykernel\nnbsphinx\nsphinx-autodoc-typehints\n+# For host_callback.py\n+msgpack\n# The next packages are for notebooks\nmatplotlib\nsklearn\n" } ]
Python
Apache License 2.0
google/jax
Fixed imports for readthedocs (#3033)
260,645
11.05.2020 16:52:55
-3,600
72cd1f7dfc6cd947ae3b19393f211733d2ca0667
Fix sign error in examples.
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/Custom_derivative_rules_for_Python_code.ipynb", "new_path": "docs/notebooks/Custom_derivative_rules_for_Python_code.ipynb", "diff": "\" x, y = primals\\n\",\n\" x_dot, y_dot = tangents\\n\",\n\" primal_out = f(x, y)\\n\",\n- \" tangent_out = np.cos(x) * x_dot * y - np.sin(x) * y_dot\\n\",\n+ \" tangent_out = np.cos(x) * x_dot * y + np.sin(x) * y_dot\\n\",\n\" return primal_out, tangent_out\"\n],\n\"execution_count\": 0,\n\" return np.sin(x) * y\\n\",\n\"\\n\",\n\"f.defjvps(lambda x_dot, primal_out, x, y: np.cos(x) * x_dot * y,\\n\",\n- \" lambda y_dot, primal_out, x, y: -np.sin(x) * y_dot)\"\n+ \" lambda y_dot, primal_out, x, y: np.sin(x) * y_dot)\"\n],\n\"execution_count\": 0,\n\"outputs\": []\n\"\\n\",\n\"def f_bwd(res, g):\\n\",\n\" cos_x, sin_x, y = res\\n\",\n- \" return (cos_x * g * y, -sin_x * g)\\n\",\n+ \" return (cos_x * g * y, sin_x * g)\\n\",\n\"\\n\",\n\"f.defvjp(f_fwd, f_bwd)\"\n],\n" } ]
Python
Apache License 2.0
google/jax
Fix sign error in examples. (#3031)
260,411
11.05.2020 20:17:26
-10,800
ddf079d8f332a2095af0aecb3b9c9e2eae5383d5
Minor improvements to the script to build macos wheels
[ { "change_type": "MODIFY", "old_path": "build/build_jaxlib_wheels_macos.sh", "new_path": "build/build_jaxlib_wheels_macos.sh", "diff": "#!/bin/bash\n+set -e\n+\n# Script that builds wheels for a JAX release on Mac OS X.\n# Builds wheels for multiple Python versions, using pyenv instead of Docker.\n# Usage: run from root of JAX source tree as:\n# May also need to install XCode command line tools to fix zlib build problem:\n# https://github.com/pyenv/pyenv/issues/1219\n+if ! pyenv --version 2>/dev/null ;then\n+ echo \"Error: You need to install pyenv and pyenv-virtualenv\"\n+ exit 1\n+fi\neval \"$(pyenv init -)\"\nPLATFORM_TAG=\"macosx_10_9_x86_64\"\n@@ -21,11 +27,13 @@ build_jax () {\nPY_TAG=\"$2\"\nNUMPY_VERSION=\"$3\"\nSCIPY_VERSION=\"$4\"\n- echo \"\\nBuilding JAX for Python ${PY_VERSION}, tag ${PY_TAG}\"\n+ echo -e \"\\nBuilding JAX for Python ${PY_VERSION}, tag ${PY_TAG}\"\necho \"NumPy version ${NUMPY_VERSION}, SciPy version ${SCIPY_VERSION}\"\npyenv install -s \"${PY_VERSION}\"\nVENV=\"jax-build-${PY_VERSION}\"\n+ if pyenv virtualenvs | grep \"${VENV}\" ;then\npyenv virtualenv-delete -f \"${VENV}\"\n+ fi\npyenv virtualenv \"${PY_VERSION}\" \"${VENV}\"\npyenv activate \"${VENV}\"\n# We pin the Numpy wheel to a version < 1.16.0 for Python releases prior to\n" } ]
Python
Apache License 2.0
google/jax
Minor improvements to the script to build macos wheels (#3013)
260,443
11.05.2020 10:22:49
25,200
30c94c6ebac9f5f2b4a40d3ed939bea5574e18ac
Add lax implementation of np.indices
[ { "change_type": "MODIFY", "old_path": "docs/jax.numpy.rst", "new_path": "docs/jax.numpy.rst", "diff": "@@ -138,6 +138,7 @@ Not every function in NumPy is implemented; contributions are welcome!\nhypot\nidentity\nimag\n+ indices\ninner\nisclose\niscomplex\n" }, { "change_type": "MODIFY", "old_path": "jax/numpy/__init__.py", "new_path": "jax/numpy/__init__.py", "diff": "@@ -35,27 +35,27 @@ from .lax_numpy import (\nfloat_power, floating, floor, floor_divide, fmax, fmin, fmod, frexp, full,\nfull_like, function, gcd, geomspace, gradient, greater, greater_equal,\nhamming, hanning, heaviside, hsplit, hstack, hypot, identity, iinfo, imag,\n- inexact, inf, inner, int16, int32, int64, int8, int_, integer, isclose,\n- iscomplex, iscomplexobj, isfinite, isinf, isnan, isneginf, isposinf, isreal,\n- isrealobj, isscalar, issubdtype, issubsctype, iterable, ix_, kaiser, kron,\n- lcm, ldexp, left_shift, less, less_equal, linspace, load, log, log10, log1p,\n- log2, logaddexp, logaddexp2, logical_and, logical_not, logical_or,\n- logical_xor, logspace, mask_indices, matmul, max, maximum, mean, median,\n- meshgrid, min, minimum, mod, moveaxis, msort, multiply, nan, nan_to_num,\n- nanargmax, nanargmin, nancumprod, nancumsum, nanmax, nanmean, nanmin,\n- nanprod, nanstd, nansum, nanvar, ndarray, ndim, negative, newaxis, nextafter,\n- nonzero, not_equal, number, numpy_version, object_, ones, ones_like,\n- operator_name, outer, packbits, pad, percentile, pi, polyval, positive, power,\n- prod, product, promote_types, ptp, quantile, rad2deg, radians, ravel, real,\n- reciprocal, remainder, repeat, reshape, result_type, right_shift, rint,\n- roll, rollaxis, rot90, round, row_stack, save, savez, searchsorted, select,\n- set_printoptions, shape, sign, signbit, signedinteger, sin, sinc, single,\n- sinh, size, sometrue, sort, split, sqrt, square, squeeze, stack, std,\n- subtract, sum, swapaxes, take, take_along_axis, tan, tanh, tensordot, tile,\n- trace, transpose, tri, tril, tril_indices, triu, triu_indices, true_divide,\n- trunc, uint16, uint32, uint64, uint8, unique, unpackbits, unravel_index,\n- unsignedinteger, vander, var, vdot, vsplit, vstack, where, zeros,\n- zeros_like)\n+ indices, inexact, inf, inner, int16, int32, int64, int8, int_, integer,\n+ isclose, iscomplex, iscomplexobj, isfinite, isinf, isnan, isneginf,\n+ isposinf, isreal, isrealobj, isscalar, issubdtype, issubsctype, iterable,\n+ ix_, kaiser, kron, lcm, ldexp, left_shift, less, less_equal, linspace,\n+ load, log, log10, log1p, log2, logaddexp, logaddexp2, logical_and,\n+ logical_not, logical_or, logical_xor, logspace, mask_indices, matmul, max,\n+ maximum, mean, median, meshgrid, min, minimum, mod, moveaxis, msort,\n+ multiply, nan, nan_to_num, nanargmax, nanargmin, nancumprod, nancumsum,\n+ nanmax, nanmean, nanmin, nanprod, nanstd, nansum, nanvar, ndarray, ndim,\n+ negative, newaxis, nextafter, nonzero, not_equal, number, numpy_version,\n+ object_, ones, ones_like, operator_name, outer, packbits, pad, percentile,\n+ pi, polyval, positive, power, prod, product, promote_types, ptp, quantile,\n+ rad2deg, radians, ravel, real, reciprocal, remainder, repeat, reshape,\n+ result_type, right_shift, rint, roll, rollaxis, rot90, round, row_stack,\n+ save, savez, searchsorted, select, set_printoptions, shape, sign, signbit,\n+ signedinteger, sin, sinc, single, sinh, size, sometrue, sort, split, sqrt,\n+ square, squeeze, stack, std, subtract, sum, swapaxes, take, take_along_axis,\n+ tan, tanh, tensordot, tile, trace, transpose, tri, tril, tril_indices, triu,\n+ triu_indices, true_divide, trunc, uint16, uint32, uint64, uint8, unique,\n+ unpackbits, unravel_index, unsignedinteger, vander, var, vdot, vsplit,\n+ vstack, where, zeros, zeros_like)\nfrom .polynomial import roots\nfrom .vectorize import vectorize\n" }, { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -2319,6 +2319,21 @@ def ix_(*args):\nreturn tuple(output)\n+@_wraps(onp.indices)\n+def indices(dimensions, dtype=int32, sparse=False):\n+ dimensions = tuple(dimensions)\n+ N = len(dimensions)\n+ output = []\n+ s = dimensions\n+ for i, dim in enumerate(dimensions):\n+ idx = lax.iota(dtype, dim)\n+ if sparse:\n+ s = (1,)*i + (dim,) + (1,)*(N - i - 1)\n+ output.append(lax.broadcast_in_dim(idx, s, (i,)))\n+ if sparse:\n+ return tuple(output)\n+ return stack(output, 0) if output else array([], dtype=dtype)\n+\ndef _repeat_scalar(a, repeats, axis=None):\nif not isscalar(repeats):\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -2352,6 +2352,23 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\ncheck_dtypes=True)\nself._CompileAndCheck(jnp.ix_, args_maker, check_dtypes=True)\n+ @parameterized.named_parameters(\n+ jtu.cases_from_list(\n+ {\"testcase_name\": \"_dimensions={}_dtype={}_sparse={}\".format(\n+ dimensions, dtype, sparse),\n+ \"dimensions\": dimensions, \"dtype\": dtype, \"sparse\": sparse}\n+ for dimensions in [(), (2,), (3, 0), (4, 5, 6)]\n+ for dtype in number_dtypes\n+ for sparse in [True, False]))\n+ def testIndices(self, dimensions, dtype, sparse):\n+ def args_maker(): return []\n+ onp_fun = partial(onp.indices, dimensions=dimensions,\n+ dtype=dtype, sparse=sparse)\n+ jnp_fun = partial(jnp.indices, dimensions=dimensions,\n+ dtype=dtype, sparse=sparse)\n+ self._CheckAgainstNumpy(onp_fun, jnp_fun, args_maker, check_dtypes=True)\n+ self._CompileAndCheck(jnp_fun, args_maker, check_dtypes=True)\n+\n@parameterized.named_parameters(jtu.cases_from_list(\n{\"testcase_name\":\n\"_op={}_a_shape={}_q_shape={}_axis={}_keepdims={}_interpolation={}\".format(\n" } ]
Python
Apache License 2.0
google/jax
Add lax implementation of np.indices (#2998)
260,411
12.05.2020 09:06:22
-10,800
a2d6b1aab4c44b3fc752d537a2e429bd922f4705
Fix typo in lstsq
[ { "change_type": "MODIFY", "old_path": "jax/numpy/linalg.py", "new_path": "jax/numpy/linalg.py", "diff": "@@ -512,7 +512,7 @@ def lstsq(a, b, rcond=None, *, numpy_resid=False):\nf\"{a.ndim}-dimensional array given. Array must be two-dimensional\")\nif b.ndim != 2:\nraise TypeError(\n- f\"{b_original_ndim}-dimensional array given. Array must be one or two-dimensional\")\n+ f\"{b.ndim}-dimensional array given. Array must be one or two-dimensional\")\nm, n = a.shape\ndtype = a.dtype\nif rcond is None:\n" } ]
Python
Apache License 2.0
google/jax
Fix typo in lstsq (#3052)
260,411
12.05.2020 10:06:32
-10,800
cc9de87562d46ca8831e68193ce01c6dfb48906f
Disabled lstsq test due to numerical failures
[ { "change_type": "MODIFY", "old_path": "tests/linalg_test.py", "new_path": "tests/linalg_test.py", "diff": "@@ -861,6 +861,7 @@ class NumpyLinalgTest(jtu.JaxTestCase):\nfor dtype in float_types + complex_types\nfor rng_factory in [jtu.rand_default]))\n@jtu.skip_on_devices(\"tpu\") # SVD not implemented on TPU.\n+ @jtu.skip_on_devices(\"cpu\", \"gpu\") # TODO(jakevdp) Test fails numerically\ndef testLstsq(self, lhs_shape, rhs_shape, dtype, lowrank, rcond, rng_factory):\nrng = rng_factory(self.rng())\n_skip_if_unsupported_type(dtype)\n" } ]
Python
Apache License 2.0
google/jax
Disabled lstsq test due to numerical failures (#3054)
260,411
12.05.2020 10:09:42
-10,800
28bc4b759ee113815bfacb99546e58a91891b3f2
Adjusted lax.numpy.indices test for older versions of numpy This test was failing on numpy 1.16.4
[ { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -2362,6 +2362,12 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\nfor sparse in [True, False]))\ndef testIndices(self, dimensions, dtype, sparse):\ndef args_maker(): return []\n+ if onp.__version__ < \"1.17\":\n+ if sparse:\n+ raise SkipTest(\"indices does not have sparse on numpy < 1.17\")\n+ onp_fun = partial(onp.indices, dimensions=dimensions,\n+ dtype=dtype)\n+ else:\nonp_fun = partial(onp.indices, dimensions=dimensions,\ndtype=dtype, sparse=sparse)\njnp_fun = partial(jnp.indices, dimensions=dimensions,\n" } ]
Python
Apache License 2.0
google/jax
Adjusted lax.numpy.indices test for older versions of numpy (#3053) This test was failing on numpy 1.16.4
260,335
14.05.2020 16:06:20
25,200
dceb5787bf956dafbeb36cbac4da3076cdcbdf7f
stash order on jet master trace, fixes
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jet.py", "new_path": "jax/experimental/jet.py", "diff": "@@ -51,16 +51,17 @@ def jet(fun, primals, series):\nyield tree_flatten(ans)\nf, out_tree = flatten_fun_output(lu.wrap_init(fun))\n- out_primals, out_terms = jet_fun(jet_subtrace(f)).call_wrapped(primals, series)\n+ out_primals, out_terms = jet_fun(jet_subtrace(f), order).call_wrapped(primals, series)\nreturn tree_unflatten(out_tree(), out_primals), tree_unflatten(out_tree(), out_terms)\n@lu.transformation\n-def jet_fun(primals, series):\n+def jet_fun(order, primals, series):\nwith core.new_master(JetTrace) as master:\n+ master.order = order\nout_primals, out_terms = yield (master, primals, series), {}\ndel master\n- out_terms = [tree_map(lambda x: onp.zeros_like(x, dtype=onp.result_type(out_primals[0])), series[0])\n- if s is zero_series else s for s in out_terms]\n+ out_terms = [[onp.zeros_like(p)] * order if s is zero_series else s\n+ for p, s in zip(out_primals, out_terms)]\nyield out_primals, out_terms\n@lu.transformation\n@@ -112,8 +113,8 @@ class JetTrace(core.Trace):\ndef process_primitive(self, primitive, tracers, params):\nassert not primitive.multiple_results # TODO\n+ order = self.master.order\nprimals_in, series_in = unzip2((t.primal, t.terms) for t in tracers)\n- order, = {len(terms) for terms in series_in if terms is not zero_series}\nseries_in = [[zero_term] * order if s is zero_series else s\nfor s in series_in]\n# TODO(mattjj): avoid always instantiating zeros\n" } ]
Python
Apache License 2.0
google/jax
stash order on jet master trace, fixes #3079 (#3097)
260,565
15.05.2020 20:51:53
25,200
510af1de64e3a1390f9d6c9d983a62e64045d761
Fix documentation for `nn.elu`, `nn.celu`, and `lax.expm1`.
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -145,7 +145,7 @@ def exp(x: Array) -> Array:\nreturn exp_p.bind(x)\ndef expm1(x: Array) -> Array:\n- r\"\"\"Elementwise :math:`e^{x - 1}`.\"\"\"\n+ r\"\"\"Elementwise :math:`e^{x} - 1`.\"\"\"\nreturn expm1_p.bind(x)\ndef log(x: Array) -> Array:\n" }, { "change_type": "MODIFY", "old_path": "jax/nn/functions.py", "new_path": "jax/nn/functions.py", "diff": "@@ -95,7 +95,7 @@ def elu(x, alpha=1.0):\n.. math::\n\\mathrm{elu}(x) = \\begin{cases}\nx, & x > 0\\\\\n- \\alpha \\exp(x - 1), & x \\le 0\n+ \\alpha \\left(\\exp(x) - 1\\right), & x \\le 0\n\\end{cases}\n\"\"\"\nsafe_x = jnp.where(x > 0, 0., x)\n@@ -138,7 +138,7 @@ def celu(x, alpha=1.0):\n.. math::\n\\mathrm{celu}(x) = \\begin{cases}\nx, & x > 0\\\\\n- \\alpha \\exp(\\frac{x}{\\alpha} - 1), & x \\le 0\n+ \\alpha \\left(\\exp(\\frac{x}{\\alpha}) - 1\\right), & x \\le 0\n\\end{cases}\nFor more information, see\n" } ]
Python
Apache License 2.0
google/jax
Fix documentation for `nn.elu`, `nn.celu`, and `lax.expm1`. (#3116)
260,299
16.05.2020 14:19:24
-3,600
670fab59cfbfe19404de365f961a0321619d8d13
Test code in docs and api.py docstrings Also remove jaxpr doc tests from api_test.py.
[ { "change_type": "MODIFY", "old_path": ".travis.yml", "new_path": ".travis.yml", "diff": "@@ -12,7 +12,7 @@ jobs:\n- python: \"3.6\"\nenv: JAX_ENABLE_X64=1 JAX_NUM_GENERATED_CASES=25\n- python: \"3.7\"\n- env: JAX_ENABLE_X64=1 JAX_ONLY_DOCUMENTATION=true\n+ env: JAX_ENABLE_X64=0 JAX_ONLY_DOCUMENTATION=true\nbefore_install:\n- wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh;\n@@ -44,6 +44,8 @@ install:\nscript:\n- if [ \"$JAX_ONLY_DOCUMENTATION\" = true ]; then\nsphinx-build -b html -D nbsphinx_execute=always docs docs/build/html ;\n+ pytest docs ;\n+ pytest --doctest-modules jax/api.py ;\nelif [ \"$JAX_ONLY_CHECK_TYPES\" = true ]; then\necho \"===== Checking with mypy ====\" &&\ntime mypy --config-file=mypy.ini jax ;\n" }, { "change_type": "MODIFY", "old_path": "docs/faq.rst", "new_path": "docs/faq.rst", "diff": "@@ -88,7 +88,7 @@ By default, JAX arrays are placed uncommitted on the default device\n(``jax.devices()[0]``).\n>>> from jax import numpy as jnp\n->>> print(jnp.ones(3).device_buffer.device())\n+>>> print(jnp.ones(3).device_buffer.device()) # doctest: +SKIP\ngpu:0\nComputations involving uncommitted data are performed on the default\n@@ -97,8 +97,9 @@ device and the results are uncommitted on the default device.\nData can also be placed explicitly on a device using :func:`jax.device_put`\nwith a ``device`` parameter, in which case if becomes **committed** to the device:\n+>>> import jax\n>>> from jax import device_put\n->>> print(device_put(1, jax.devices()[2]).device_buffer.device())\n+>>> print(device_put(1, jax.devices()[2]).device_buffer.device()) # doctest: +SKIP\ngpu:2\nComputations involving some committed inputs, will happen on the\n" }, { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -119,15 +119,17 @@ def jit(fun: Callable, static_argnums: Union[int, Iterable[int]] = (),\nIn the following example, ``selu`` can be compiled into a single fused kernel\nby XLA:\n+ >>> import jax\n+ >>>\n>>> @jax.jit\n- >>> def selu(x, alpha=1.67, lmbda=1.05):\n- >>> return lmbda * jax.numpy.where(x > 0, x, alpha * jax.numpy.exp(x) - alpha)\n+ ... def selu(x, alpha=1.67, lmbda=1.05):\n+ ... return lmbda * jax.numpy.where(x > 0, x, alpha * jax.numpy.exp(x) - alpha)\n>>>\n>>> key = jax.random.PRNGKey(0)\n>>> x = jax.random.normal(key, (10,))\n- >>> print(selu(x))\n- [-0.54485154 0.27744263 -0.29255125 -0.91421586 -0.62452525 -0.2474813\n- -0.8574326 -0.7823267 0.7682731 0.59566754]\n+ >>> print(selu(x)) # doctest: +SKIP\n+ [-0.54485 0.27744 -0.29255 -0.91421 -0.62452 -0.24748\n+ -0.85743 -0.78232 0.76827 0.59566 ]\n\"\"\"\n_check_callable(fun)\nif isinstance(static_argnums, int):\n@@ -172,8 +174,10 @@ def disable_jit():\nnotice those if you use a benign side-effecting operation in a jitted\nfunction, like a print:\n+ >>> import jax\n+ >>>\n>>> @jax.jit\n- >>> def f(x):\n+ ... def f(x):\n... y = x * 2\n... print(\"Value of y is\", y)\n... return y + 3\n@@ -187,9 +191,11 @@ def disable_jit():\nalso traced. If we want to see a concrete value while debugging, and avoid the\ntracer too, we can use the ``disable_jit`` context manager:\n- >>> with jax.disable_jit():\n- >>> print(f(np.array([1, 2, 3])))\n+ >>> import jax.numpy as np\n>>>\n+ >>> with jax.disable_jit():\n+ ... print(f(np.array([1, 2, 3])))\n+ ...\nValue of y is [2 4 6]\n[5 7 9]\n\"\"\"\n@@ -243,22 +249,29 @@ def xla_computation(fun: Callable,\nFor example:\n+ >>> import jax\n+ >>>\n>>> def f(x): return jax.numpy.sin(jax.numpy.cos(x))\n>>> c = jax.xla_computation(f)(3.)\n- >>> print(c.as_hlo_text())\n- HloModule jaxpr_computation__4.5\n- ENTRY jaxpr_computation__4.5 {\n- tuple.1 = () tuple()\n- parameter.2 = f32[] parameter(0)\n- cosine.3 = f32[] cosine(parameter.2)\n- ROOT sine.4 = f32[] sine(cosine.3)\n+ >>> print(c.as_hlo_text()) # doctest: +SKIP\n+ HloModule xla_computation_f.6\n+ <BLANKLINE>\n+ ENTRY xla_computation_f.6 {\n+ constant.2 = pred[] constant(false)\n+ parameter.1 = f32[] parameter(0)\n+ cosine.3 = f32[] cosine(parameter.1)\n+ sine.4 = f32[] sine(cosine.3)\n+ ROOT tuple.5 = (f32[]) tuple(sine.4)\n}\n+ <BLANKLINE>\n+ <BLANKLINE>\n+\nHere's an example that involves a parallel collective and axis name:\n>>> def f(x): return x - jax.lax.psum(x, 'i')\n>>> c = jax.xla_computation(f, axis_env=[('i', 4)])(2)\n- >>> print(c.as_hlo_text())\n+ >>> print(c.as_hlo_text()) # doctest: +SKIP\nHloModule jaxpr_computation.9\nprimitive_computation.3 {\nparameter.4 = s32[] parameter(0)\n@@ -271,10 +284,13 @@ def xla_computation(fun: Callable,\nall-reduce.7 = s32[] all-reduce(parameter.2), replica_groups={{0,1,2,3}}, to_apply=primitive_computation.3\nROOT subtract.8 = s32[] subtract(parameter.2, all-reduce.7)\n}\n+ <BLANKLINE>\n+ <BLANKLINE>\nNotice the ``replica_groups`` that were generated. Here's an example that\ngenerates more interesting ``replica_groups``:\n+ >>> from jax import lax\n>>> def g(x):\n... rowsum = lax.psum(x, 'i')\n... colsum = lax.psum(x, 'j')\n@@ -283,7 +299,7 @@ def xla_computation(fun: Callable,\n...\n>>> axis_env = [('i', 4), ('j', 2)]\n>>> c = xla_computation(g, axis_env=axis_env)(5.)\n- >>> print(c.as_hlo_text())\n+ >>> print(c.as_hlo_text()) # doctest: +SKIP\nHloModule jaxpr_computation__1.19\n[removed uninteresting text here]\nENTRY jaxpr_computation__1.19 {\n@@ -366,6 +382,8 @@ def grad(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nFor example:\n+ >>> import jax\n+ >>>\n>>> grad_tanh = jax.grad(jax.numpy.tanh)\n>>> print(grad_tanh(0.2))\n0.961043\n@@ -485,15 +503,18 @@ def jacfwd(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nA function with the same arguments as ``fun``, that evaluates the Jacobian of\n``fun`` using forward-mode automatic differentiation.\n+ >>> import jax\n+ >>> import jax.numpy as np\n+ >>>\n>>> def f(x):\n... return jax.numpy.asarray(\n... [x[0], 5*x[2], 4*x[1]**2 - 2*x[2], x[2] * jax.numpy.sin(x[0])])\n...\n>>> print(jax.jacfwd(f)(np.array([1., 2., 3.])))\n- [[ 1. , 0. , 0. ],\n- [ 0. , 0. , 5. ],\n- [ 0. , 16. , -2. ],\n- [ 1.6209068 , 0. , 0.84147096]]\n+ [[ 1. 0. 0. ]\n+ [ 0. 0. 5. ]\n+ [ 0. 16. -2. ]\n+ [ 1.6209 0. 0.84147]]\n\"\"\"\n_check_callable(fun)\n@@ -532,15 +553,18 @@ def jacrev(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nA function with the same arguments as ``fun``, that evaluates the Jacobian of\n``fun`` using reverse-mode automatic differentiation.\n+ >>> import jax\n+ >>> import jax.numpy as np\n+ >>>\n>>> def f(x):\n... return jax.numpy.asarray(\n... [x[0], 5*x[2], 4*x[1]**2 - 2*x[2], x[2] * jax.numpy.sin(x[0])])\n...\n>>> print(jax.jacrev(f)(np.array([1., 2., 3.])))\n- [[ 1. , 0. , 0. ],\n- [ 0. , 0. , 5. ],\n- [ 0. , 16. , -2. ],\n- [ 1.6209068 , 0. , 0.84147096]]\n+ [[ 1. 0. 0. ]\n+ [ 0. 0. 5. ]\n+ [ 0. 16. -2. ]\n+ [ 1.6209 0. 0.84147]]\n\"\"\"\n_check_callable(fun)\n@@ -585,10 +609,12 @@ def hessian(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nA function with the same arguments as ``fun``, that evaluates the Hessian of\n``fun``.\n- >>> g = lambda(x): x[0]**3 - 2*x[0]*x[1] - x[1]**6\n+ >>> import jax\n+ >>>\n+ >>> g = lambda x: x[0]**3 - 2*x[0]*x[1] - x[1]**6\n>>> print(jax.hessian(g)(jax.numpy.array([1., 2.])))\n- [[ 6., -2.],\n- [ -2., -480.]]\n+ [[ 6. -2.]\n+ [ -2. -480.]]\n:py:func:`hessian` is a generalization of the usual definition of the Hessian\nthat supports nested Python containers (i.e. pytrees) as inputs and outputs.\n@@ -597,6 +623,7 @@ def hessian(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nthe structure of ``x``. A tree product of two tree structures is formed by\nreplacing each leaf of the first tree with a copy of the second. For example:\n+ >>> import jax.numpy as jnp\n>>> f = lambda dct: {\"c\": jnp.power(dct[\"a\"], dct[\"b\"])}\n>>> print(jax.hessian(f)({\"a\": jnp.arange(2.) + 1., \"b\": jnp.arange(2.) + 2.}))\n{'c': {'a': {'a': DeviceArray([[[ 2., 0.], [ 0., 0.]],\n@@ -689,6 +716,8 @@ def vmap(fun: Callable, in_axes=0, out_axes=0) -> Callable:\nFor example, we can implement a matrix-matrix product using a vector dot\nproduct:\n+ >>> import jax.numpy as np\n+ >>>\n>>> vv = lambda x, y: np.vdot(x, y) # ([a], [a]) -> []\n>>> mv = vmap(vv, (0, None), 0) # ([b,a], [a]) -> [b] (b is the mapped axis)\n>>> mm = vmap(mv, (None, 1), 1) # ([b,a], [a,c]) -> [b,c] (c is the mapped axis)\n@@ -721,7 +750,7 @@ def vmap(fun: Callable, in_axes=0, out_axes=0) -> Callable:\n>>> z = np.ones((C, D, K))\n>>> tree = (x, (y, z))\n>>> vfoo = vmap(foo, in_axes=((0, (1, 2)),))\n- >>> print(vfoo(tree)).shape\n+ >>> print(vfoo(tree).shape)\n(6, 2, 5)\nThe results of a vectorized function can be mapped or unmapped.\n@@ -730,13 +759,13 @@ def vmap(fun: Callable, in_axes=0, out_axes=0) -> Callable:\nwe can specify `out_axes` to be None (to keep it unmapped).\n>>> print(vmap(lambda x, y: (x + y, y * 2.), in_axes=(0, None), out_axes=(0, None))(np.arange(2.), 4.))\n- ([4., 5.], 8.)\n+ (DeviceArray([4., 5.], dtype=float32), 8.0)\nIf the `out_axes` is specified for an unmapped result, the result is broadcast\nacross the mapped axis:\n>>> print(vmap(lambda x, y: (x + y, y * 2.), in_axes=(0, None), out_axes=0)(np.arange(2.), 4.))\n- ([4., 5.], [8., 8.])\n+ (DeviceArray([4., 5.], dtype=float32), array([8., 8.]))\nIf the `out_axes` is specified for a mapped result, the result is\ntransposed accordingly.\n@@ -889,8 +918,10 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\nFor example, assuming 8 XLA devices are available, ``pmap`` can be used as a\nmap along a leading array axis:\n- >>> out = pmap(lambda x: x ** 2)(np.arange(8))\n- >>> print(out)\n+ >>> import jax.numpy as np\n+ >>>\n+ >>> out = pmap(lambda x: x ** 2)(np.arange(8)) # doctest: +SKIP\n+ >>> print(out) # doctest: +SKIP\n[0, 1, 4, 9, 16, 25, 36, 49]\nWhen the leading dimension is smaller than the number of available devices JAX\n@@ -898,8 +929,8 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\n>>> x = np.arange(3 * 2 * 2.).reshape((3, 2, 2))\n>>> y = np.arange(3 * 2 * 2.).reshape((3, 2, 2)) ** 2\n- >>> out = pmap(np.dot)(x, y)\n- >>> print(out)\n+ >>> out = pmap(np.dot)(x, y) # doctest: +SKIP\n+ >>> print(out) # doctest: +SKIP\n[[[ 4. 9.]\n[ 12. 29.]]\n[[ 244. 345.]\n@@ -910,7 +941,7 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\nIf your leading dimension is larger than the number of available devices you\nwill get an error:\n- >>> pmap(lambda x: x ** 2)(np.arange(9))\n+ >>> pmap(lambda x: x ** 2)(np.arange(9)) # doctest: +SKIP\nValueError: ... requires 9 replicas, but only 8 XLA devices are available\nAs with ``vmap``, using ``None`` in ``in_axes`` indicates that an argument\n@@ -918,8 +949,8 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\nacross the replicas:\n>>> x, y = np.arange(2.), 4.\n- >>> out = pmap(lambda x, y: (x + y, y * 2.), in_axes=(0, None))(x, y)\n- >>> print(out)\n+ >>> out = pmap(lambda x, y: (x + y, y * 2.), in_axes=(0, None))(x, y) # doctest: +SKIP\n+ >>> print(out) # doctest: +SKIP\n([4., 5.], [8., 8.])\nNote that ``pmap`` always returns values mapped over their leading axis,\n@@ -930,10 +961,10 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\ncollective operations. For example:\n>>> f = lambda x: x / jax.lax.psum(x, axis_name='i')\n- >>> out = pmap(f, axis_name='i')(np.arange(4.))\n- >>> print(out)\n+ >>> out = pmap(f, axis_name='i')(np.arange(4.)) # doctest: +SKIP\n+ >>> print(out) # doctest: +SKIP\n[ 0. 0.16666667 0.33333334 0.5 ]\n- >>> print(out.sum())\n+ >>> print(out.sum()) # doctest: +SKIP\n1.0\nIn this example, ``axis_name`` is a string, but it can be any Python object\n@@ -945,21 +976,23 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\ncollectives can operate over distinct axes:\n>>> from functools import partial\n+ >>> import jax\n+ >>>\n>>> @partial(pmap, axis_name='rows')\n- >>> @partial(pmap, axis_name='cols')\n- >>> def normalize(x):\n- >>> row_normed = x / jax.lax.psum(x, 'rows')\n- >>> col_normed = x / jax.lax.psum(x, 'cols')\n- >>> doubly_normed = x / jax.lax.psum(x, ('rows', 'cols'))\n- >>> return row_normed, col_normed, doubly_normed\n+ ... @partial(pmap, axis_name='cols')\n+ ... def normalize(x):\n+ ... row_normed = x / jax.lax.psum(x, 'rows')\n+ ... col_normed = x / jax.lax.psum(x, 'cols')\n+ ... doubly_normed = x / jax.lax.psum(x, ('rows', 'cols'))\n+ ... return row_normed, col_normed, doubly_normed\n>>>\n>>> x = np.arange(8.).reshape((4, 2))\n- >>> row_normed, col_normed, doubly_normed = normalize(x)\n- >>> print(row_normed.sum(0))\n+ >>> row_normed, col_normed, doubly_normed = normalize(x) # doctest: +SKIP\n+ >>> print(row_normed.sum(0)) # doctest: +SKIP\n[ 1. 1.]\n- >>> print(col_normed.sum(1))\n+ >>> print(col_normed.sum(1)) # doctest: +SKIP\n[ 1. 1. 1. 1.]\n- >>> print(doubly_normed.sum((0, 1)))\n+ >>> print(doubly_normed.sum((0, 1))) # doctest: +SKIP\n1.0\nOn multi-host platforms, collective operations operate over all devices,\n@@ -968,8 +1001,8 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\n>>> f = lambda x: x + jax.lax.psum(x, axis_name='i')\n>>> data = np.arange(4) if jax.host_id() == 0 else np.arange(4,8)\n- >>> out = pmap(f, axis_name='i')(data)\n- >>> print(out)\n+ >>> out = pmap(f, axis_name='i')(data) # doctest: +SKIP\n+ >>> print(out) # doctest: +SKIP\n[28 29 30 31] # on host 0\n[32 33 34 35] # on host 1\n@@ -987,16 +1020,16 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\n>>> from functools import partial\n>>> @partial(pmap, axis_name='i', devices=jax.devices()[:6])\n- >>> def f1(x):\n- >>> return x / jax.lax.psum(x, axis_name='i')\n+ ... def f1(x):\n+ ... return x / jax.lax.psum(x, axis_name='i')\n>>>\n>>> @partial(pmap, axis_name='i', devices=jax.devices()[-2:])\n- >>> def f2(x):\n- >>> return jax.lax.psum(x ** 2, axis_name='i')\n+ ... def f2(x):\n+ ... return jax.lax.psum(x ** 2, axis_name='i')\n>>>\n- >>> print(f1(np.arange(6.)))\n+ >>> print(f1(np.arange(6.))) # doctest: +SKIP\n[0. 0.06666667 0.13333333 0.2 0.26666667 0.33333333]\n- >>> print(f2(np.array([2., 3.])))\n+ >>> print(f2(np.array([2., 3.]))) # doctest: +SKIP\n[ 13. 13.]\n\"\"\"\n_check_callable(fun)\n@@ -1259,6 +1292,8 @@ def jvp(fun: Callable, primals, tangents) -> Tuple[Any, Any]:\nFor example:\n+ >>> import jax\n+ >>>\n>>> y, v = jax.jvp(jax.numpy.sin, (0.1,), (0.2,))\n>>> print(y)\n0.09983342\n@@ -1338,10 +1373,13 @@ def linearize(fun: Callable, *primals) -> Tuple[Any, Callable]:\nHere's a more complete example of using ``linearize``:\n+ >>> import jax\n+ >>> import jax.numpy as np\n+ >>>\n>>> def f(x): return 3. * np.sin(x) + np.cos(x / 2.)\n...\n>>> jax.jvp(f, (2.,), (3.,))\n- (array(3.2681944, dtype=float32), array(-5.007528, dtype=float32))\n+ (DeviceArray(3.26819, dtype=float32), DeviceArray(-5.00753, dtype=float32))\n>>> y, f_jvp = jax.linearize(f, 2.)\n>>> print(y)\n3.2681944\n@@ -1431,6 +1469,8 @@ def vjp(fun: Callable, *primals, **kwargs\n``(primals_out, vjpfun, aux)`` tuple where ``aux`` is the auxiliary data\nreturned by ``fun``.\n+ >>> import jax\n+ >>>\n>>> def f(x, y):\n... return jax.numpy.sin(x), jax.numpy.cos(y)\n...\n@@ -1493,23 +1533,25 @@ def make_jaxpr(fun: Callable,\nWe do not describe the semantics of the ``jaxpr`` language in detail here, but\ninstead give a few examples.\n+ >>> import jax\n+ >>>\n>>> def f(x): return jax.numpy.sin(jax.numpy.cos(x))\n>>> print(f(3.0))\n- -0.83602184\n+ -0.83602\n>>> jax.make_jaxpr(f)(3.0)\n- { lambda ; ; a.\n+ { lambda ; a.\nlet b = cos a\nc = sin b\n- in [c] }\n+ in (c,) }\n>>> jax.make_jaxpr(jax.grad(f))(3.0)\n- { lambda ; ; a.\n+ { lambda ; a.\nlet b = cos a\nc = cos b\nd = mul 1.0 c\ne = neg d\nf = sin a\ng = mul e f\n- in [g] }\n+ in (g,) }\n\"\"\"\n_check_callable(fun)\nif isinstance(static_argnums, int):\n@@ -1663,6 +1705,9 @@ def eval_shape(fun: Callable, *args, **kwargs):\nFor example:\n+ >>> import jax\n+ >>> import jax.numpy as np\n+ >>>\n>>> f = lambda A, x: np.tanh(np.dot(A, x))\n>>> class MyArgArray(object):\n... def __init__(self, shape, dtype):\n@@ -1675,7 +1720,7 @@ def eval_shape(fun: Callable, *args, **kwargs):\n>>> print(out.shape)\n(2000, 1000)\n>>> print(out.dtype)\n- dtype('float32')\n+ float32\n\"\"\"\ndef abstractify(x):\nreturn ShapedArray(onp.shape(x), dtypes.result_type(x))\n" }, { "change_type": "MODIFY", "old_path": "jax/config.py", "new_path": "jax/config.py", "diff": "@@ -67,7 +67,7 @@ class Config(object):\ndef check_exists(self, name):\nif name not in self.values:\n- raise Exception(\"Unrecognized config option: {}\".format(name))\n+ raise AttributeError(\"Unrecognized config option: {}\".format(name))\ndef DEFINE_bool(self, name, default, *args, **kwargs):\nself.add_option(name, default, bool, args, kwargs)\n" }, { "change_type": "MODIFY", "old_path": "pytest.ini", "new_path": "pytest.ini", "diff": "@@ -4,3 +4,5 @@ filterwarnings =\nignore:No GPU/TPU found, falling back to CPU.:UserWarning\nignore:Explicitly requested dtype.*is not available.*:UserWarning\nignore:jax.experimental.vectorize is deprecated.*:FutureWarning\n+doctest_optionflags = NUMBER NORMALIZE_WHITESPACE\n+addopts = --doctest-glob=\"*.rst\"\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -1707,177 +1707,6 @@ class JaxprTest(jtu.JaxTestCase):\nin (e,) }\n\"\"\", str(jaxpr))\n- def testExamplesJaxprDoc(self):\n- \"\"\"Tests examples included in the Understanding jaxprs doc (docs/jaxpr.rst).\"\"\"\n- def func1(first, second):\n- temp = first + jnp.sin(second) * 3.\n- return jnp.sum(temp)\n-\n- jaxpr = jax.make_jaxpr(func1)(jnp.zeros(8), jnp.ones(8))\n- self.assertMultiLineStrippedEqual(\"\"\"\n-{ lambda ; a b.\n- let c = sin b\n- d = mul c 3.0\n- e = add a d\n- f = reduce_sum[ axes=(0,) ] e\n- in (f,) }\n- \"\"\", str(jaxpr))\n-\n- def func5(first, second):\n- temp = first + jnp.sin(second) * 3. - jnp.ones(8)\n- return temp\n-\n- def func6(first):\n- return func5(first, jnp.ones(8))\n-\n- jaxpr = api.make_jaxpr(func6)(jnp.ones(8))\n- self.assertMultiLineStrippedEqual(\"\"\"\n-{ lambda b d ; a.\n- let c = add a b\n- e = sub c d\n- in (e,) }\n- \"\"\", str(jaxpr))\n-\n- def func7(arg):\n- return lax.cond(arg >= 0.,\n- lambda xtrue: xtrue + 3.,\n- lambda xfalse: xfalse - 3.,\n- arg)\n-\n- jaxpr = api.make_jaxpr(func7)(5.)\n- self.assertMultiLineStrippedEqual(\"\"\"\n-{ lambda ; a.\n- let b = ge a 0.0\n- c = cond[ false_jaxpr={ lambda ; a.\n- let b = sub a 3.0\n- in (b,) }\n- linear=(False,)\n- true_jaxpr={ lambda ; a.\n- let b = add a 3.0\n- in (b,) } ] b a\n- in (c,) }\n- \"\"\", str(jaxpr))\n-\n- def func8(arg1, arg2): # arg2 is a pair\n- return lax.cond(arg1 >= 0.,\n- lambda xtrue: xtrue[0],\n- lambda xfalse: jnp.ones(1) + xfalse[1],\n- arg2)\n-\n- jaxpr = api.make_jaxpr(func8)(5., (jnp.zeros(1), 2.))\n- self.assertMultiLineStrippedEqual(\"\"\"\n-{ lambda e ; a b c.\n- let d = ge a 0.0\n- f = cond[ false_jaxpr={ lambda ; c a b.\n- let d = add c b\n- in (d,) }\n- linear=(False, False, False)\n- true_jaxpr={ lambda ; a_ a b.\n- let\n- in (a,) } ] d e b c\n- in (f,) }\n- \"\"\", str(jaxpr))\n-\n- def func10(arg, n):\n- ones = jnp.ones(arg.shape) # A constant\n- return lax.fori_loop(0, n,\n- lambda i, carry: carry + ones * 3. + arg,\n- arg + ones)\n-\n- jaxpr = api.make_jaxpr(func10)(np.ones(16), 5)\n- self.assertMultiLineStrippedEqual(\"\"\"\n-{ lambda c d ; a b.\n- let e = add a d\n- f g h = while[ body_jaxpr={ lambda ; e g a b c.\n- let d = add a 1\n- f = add c e\n- h = add f g\n- in (d, b, h) }\n- body_nconsts=2\n- cond_jaxpr={ lambda ; a b c.\n- let d = lt a b\n- in (d,) }\n- cond_nconsts=0 ] c a 0 b e\n- in (h,) }\n- \"\"\", str(jaxpr))\n-\n- def func11(arr, extra):\n- ones = jnp.ones(arr.shape) # A constant\n-\n- def body(carry, aelems):\n- # carry: running dot-product of the two arrays\n- # aelems: a pair with corresponding elements from the two arrays\n- ae1, ae2 = aelems\n- return (carry + ae1 * ae2 + extra, carry)\n-\n- return lax.scan(body, 0., (arr, ones))\n-\n- jaxpr = api.make_jaxpr(func11)(np.ones(16), 5.)\n- # TODO(#2640): update docs/jaxpr.rst to reflect new jaxpr\n- self.assertMultiLineStrippedEqual(\"\"\"\n-{ lambda c ; a b.\n- let d e = scan[ jaxpr={ lambda ; f a b c.\n- let d = mul b c\n- e = add a d\n- g = add e f\n- in (g, a) }\n- length=16\n- linear=(False, False, False, False)\n- num_carry=1\n- num_consts=1\n- reverse=False ] b 0.0 a c\n- in (d, e) }\n- \"\"\", str(jaxpr))\n-\n- def func12(arg):\n- @api.jit\n- def inner(x):\n- return x + arg * jnp.ones(1) # Include a constant in the inner function\n-\n- return arg + inner(arg - 2.)\n-\n- jaxpr = api.make_jaxpr(func12)(1.)\n- self.assertMultiLineStrippedEqual(\"\"\"\n-{ lambda b ; a.\n- let c = sub a 2.0\n- d = xla_call[ backend=None\n- call_jaxpr={ lambda ; c b a.\n- let d = mul b c\n- e = add a d\n- in (e,) }\n- device=None\n- name=inner ] b a c\n- e = add a d\n- in (e,) }\n- \"\"\", str(jaxpr))\n-\n- def func13(arr, extra):\n- def inner(x):\n- # use a free variable \"extra\" and a constant jnp.ones(1)\n- return (x + extra + jnp.ones(1)) / lax.psum(x, axis_name='rows')\n-\n- return api.pmap(inner, axis_name='rows')(arr)\n-\n- jaxpr = api.make_jaxpr(func13)(jnp.ones((1, 3)), 5.)\n- self.assertMultiLineStrippedEqual(\"\"\"\n-{ lambda c ; a b.\n- let d = xla_pmap[ axis_name=rows\n- axis_size=1\n- backend=None\n- call_jaxpr={ lambda ; d b a.\n- let c = add a b\n- e = add c d\n- f = psum[ axis_index_groups=None\n- axis_name=rows ] a\n- g = div e f\n- in (g,) }\n- devices=None\n- global_axis_size=None\n- mapped_invars=(True, False, True)\n- name=inner ] c b a\n- in (d,) }\n- \"\"\", str(jaxpr))\n-\ndef test_make_jaxpr_static_argnums(self):\ndef f(x, y):\nreturn x + y\n" } ]
Python
Apache License 2.0
google/jax
Test code in docs and api.py docstrings (#2994) Also remove jaxpr doc tests from api_test.py.
260,411
18.05.2020 16:02:49
-10,800
b071b12a13ed7301f8298d0d0404cb61dea994af
Fixed link in FAQ
[ { "change_type": "MODIFY", "old_path": "docs/faq.rst", "new_path": "docs/faq.rst", "diff": "@@ -123,7 +123,7 @@ its use is not recommended.)\nFor a worked-out example, we recommend reading through\n``test_computation_follows_data`` in\n-[multi_device_test.py](https://github.com/google/jax/blob/master/tests/multi_device_test.py).\n+`multi_device_test.py <https://github.com/google/jax/blob/master/tests/multi_device_test.py>`_.\n.. comment We refer to the anchor below in JAX error messages\n" } ]
Python
Apache License 2.0
google/jax
Fixed link in FAQ (#3129)
260,361
18.05.2020 19:24:45
25,200
8d0749f13ee9e0d40af49ec0eabbc1e07ec4ff22
Fix a corner case in `repeat`. * Fixes a corner case: `jnp.repeat(jnp.array(0), 1, axis=0)` throws an error, whereas `np.repeat(np.array(0), 1, axis=0) = np.array([0])`. * Add test for `np.range(np.array(0), 1, axis=0)`.
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -2395,7 +2395,7 @@ def _repeat_scalar(a, repeats, axis=None):\nif not isscalar(repeats):\nraise NotImplementedError(\n\"_repeat_scalar implementation only supports scalar repeats\")\n- if axis is None or isscalar(a):\n+ if axis is None or isscalar(a) or len(shape(a)) == 0:\na = ravel(a)\naxis = 0\na_shape = list(shape(a))\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_numpy_test.py", "new_path": "tests/lax_numpy_test.py", "diff": "@@ -1114,7 +1114,7 @@ class LaxBackedNumpyTests(jtu.JaxTestCase):\n\"rng_factory\": jtu.rand_default}\nfor repeats in [0, 1, 2]\nfor shape, dtype in _shape_and_dtypes(all_shapes, default_dtypes)\n- for axis in [None] + list(range(-len(shape), len(shape)))))\n+ for axis in [None] + list(range(-len(shape), max(1, len(shape))))))\ndef testRepeat(self, axis, shape, dtype, repeats, rng_factory):\nrng = rng_factory(self.rng())\nonp_fun = lambda arg: onp.repeat(arg, repeats=repeats, axis=axis)\n" } ]
Python
Apache License 2.0
google/jax
Fix a corner case in `repeat`. (#3117) * Fixes a corner case: `jnp.repeat(jnp.array(0), 1, axis=0)` throws an error, whereas `np.repeat(np.array(0), 1, axis=0) = np.array([0])`. * Add test for `np.range(np.array(0), 1, axis=0)`.
260,577
19.05.2020 07:06:32
-3,600
85fe5a28f1836ed1a9e91636c224d8f3cd4eb675
Add gradients to the scatter_max and scatter_min operations. This is being done to allow the creation of a differentiable segment_max. Segment_max is an important operation for GraphNets and is an open feature request at
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax.py", "new_path": "jax/lax/lax.py", "diff": "@@ -3729,20 +3729,113 @@ ad.primitive_transposes[scatter_mul_p] = _scatter_mul_transpose_rule\nbatching.primitive_batchers[scatter_mul_p] = (\npartial(_scatter_batching_rule, scatter_mul))\n-# TODO(jlebar): Add derivatives.\n+def _scatter_extremal_jvp(scatter_op, primals, tangents, update_jaxpr,\n+ update_consts, dimension_numbers):\n+ operand, scatter_indices, updates = primals\n+ g_operand, g_scatter_indices, g_updates = tangents\n+\n+ scatter_dnums = dimension_numbers\n+ updates_shape = updates.shape\n+\n+ val_out = scatter_op.bind(\n+ operand, scatter_indices, updates, update_jaxpr=update_jaxpr,\n+ update_consts=update_consts, dimension_numbers=scatter_dnums)\n+\n+ if g_operand is ad_util.zero and g_updates is ad_util.zero:\n+ tangent_out = ad_util.zero\n+ else:\n+ g_operand = ad.instantiate_zeros(operand, g_operand)\n+ g_updates = ad.instantiate_zeros(updates, g_updates)\n+\n+ # gather_dnums and slice_sizes define the gather op that is the inverse of\n+ # the scatter op specified by scatter_dnums\n+ gather_dnums = GatherDimensionNumbers(\n+ offset_dims=scatter_dnums.update_window_dims,\n+ collapsed_slice_dims=scatter_dnums.inserted_window_dims,\n+ start_index_map=scatter_dnums.scatter_dims_to_operand_dims)\n+\n+ slice_sizes = []\n+ pos = 0\n+ for i in range(len(operand.shape)):\n+ if i in scatter_dnums.inserted_window_dims:\n+ slice_sizes.append(1)\n+ else:\n+ slice_sizes.append(updates_shape[scatter_dnums.update_window_dims[pos]])\n+ pos += 1\n+\n+ # For consistency with other max operations, if there are two or more values\n+ # in updates that are contending to replace the same index location, the\n+ # resulting tangent at that location will be the average of the associated\n+ # tangents for the values in updates.\n+\n+ initial_vals = gather(\n+ operand, scatter_indices, gather_dnums, onp.array(slice_sizes))\n+\n+ target_vals = gather(\n+ val_out, scatter_indices, gather_dnums, onp.array(slice_sizes))\n+\n+ successful_updates = (updates == target_vals)\n+ retained_values = (initial_vals == target_vals)\n+\n+ num_updates = gather(\n+ scatter_add(_zeros(operand),\n+ scatter_indices,\n+ select(successful_updates, _ones(updates), _zeros(updates)),\n+ scatter_dnums),\n+ scatter_indices,\n+ gather_dnums,\n+ onp.array(slice_sizes))\n+\n+ num_refs = gather(\n+ scatter_add(_zeros(operand),\n+ scatter_indices,\n+ _ones(updates),\n+ scatter_dnums),\n+ scatter_indices,\n+ gather_dnums,\n+ onp.array(slice_sizes))\n+\n+ updates_normalizer = select(retained_values,\n+ 1.0 / (num_updates + 1),\n+ 1.0 / num_updates)\n+\n+ updates_coef = select(successful_updates,\n+ updates_normalizer,\n+ _zeros(updates))\n+\n+ operand_normalizer = select(retained_values,\n+ 1.0 / (num_updates + 1),\n+ _zeros(num_updates))\n+\n+ operand_coef = (-1.0 + operand_normalizer) / num_refs\n+\n+ # This can be simplified once scatter has transpose implemented\n+ target_tangents = gather(\n+ g_operand, scatter_indices, gather_dnums, onp.array(slice_sizes))\n+\n+ tangent_updates = (target_tangents * operand_coef +\n+ g_updates * updates_coef)\n+\n+ tangent_out = scatter_add(g_operand,\n+ scatter_indices,\n+ tangent_updates,\n+ scatter_dnums)\n+\n+ return val_out, tangent_out\n+\nscatter_min_p = standard_primitive(\n_scatter_shape_rule, _scatter_dtype_rule, 'scatter-min',\n_scatter_translation_rule)\nbatching.primitive_batchers[scatter_min_p] = (\npartial(_scatter_batching_rule, scatter_min))\n+ad.primitive_jvps[scatter_min_p] = partial(_scatter_extremal_jvp, scatter_min_p)\n-# TODO(jlebar): Add derivatives.\nscatter_max_p = standard_primitive(\n_scatter_shape_rule, _scatter_dtype_rule, 'scatter-max',\n_scatter_translation_rule)\nbatching.primitive_batchers[scatter_max_p] = (\npartial(_scatter_batching_rule, scatter_max))\n-\n+ad.primitive_jvps[scatter_max_p] = partial(_scatter_extremal_jvp, scatter_max_p)\ndef _scatter_jvp(primals, tangents, *, update_jaxpr, update_consts,\ndimension_numbers):\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_test.py", "new_path": "tests/lax_test.py", "diff": "@@ -2629,6 +2629,68 @@ class LaxAutodiffTest(jtu.JaxTestCase):\ncheck_grads(f, (rng((5, 5), onp.float32),), 2, [\"fwd\", \"rev\"], 1e-2, 1e-2,\n1.)\n+ @parameterized.named_parameters(jtu.cases_from_list(\n+ {\"testcase_name\": \"_shape={}_idxs={}_update={}_dnums={}\".format(\n+ jtu.format_shape_dtype_string(arg_shape, dtype),\n+ idxs, update_shape, dnums),\n+ \"arg_shape\": arg_shape, \"dtype\": dtype, \"idxs\": idxs,\n+ \"update_shape\": update_shape, \"dnums\": dnums,\n+ \"rng_factory\": rng_factory, \"rng_idx_factory\": rng_idx_factory}\n+ for dtype in grad_float_dtypes\n+ for arg_shape, idxs, update_shape, dnums in [\n+ ((5,), onp.array([[0], [2]]), (2,), lax.ScatterDimensionNumbers(\n+ update_window_dims=(), inserted_window_dims=(0,),\n+ scatter_dims_to_operand_dims=(0,))),\n+ ((10,), onp.array([[0], [0], [0]]), (3, 2), lax.ScatterDimensionNumbers(\n+ update_window_dims=(1,), inserted_window_dims=(),\n+ scatter_dims_to_operand_dims=(0,))),\n+ ((10, 5,), onp.array([[0], [2], [1]]), (3, 3), lax.ScatterDimensionNumbers(\n+ update_window_dims=(1,), inserted_window_dims=(0,),\n+ scatter_dims_to_operand_dims=(0,))),\n+ ]\n+ for rng_idx_factory in [partial(jtu.rand_int, high=max(arg_shape))]\n+ for rng_factory in [jtu.rand_default]))\n+ def testScatterMax(self, arg_shape, dtype, idxs, update_shape, dnums,\n+ rng_factory, rng_idx_factory):\n+ rng = rng_factory(self.rng())\n+ rng_idx = rng_idx_factory(self.rng())\n+ idxs = rng_idx(idxs.shape, idxs.dtype)\n+ scatter_max = lambda x, y: lax.scatter_max(x, idxs, y, dnums)\n+ x = rng(arg_shape, dtype)\n+ y = rng(update_shape, dtype)\n+ check_grads(scatter_max, (x, y), 2, [\"fwd\", \"rev\"], 1e-2, 1e-2)\n+\n+ @parameterized.named_parameters(jtu.cases_from_list(\n+ {\"testcase_name\": \"_shape={}_idxs={}_update={}_dnums={}\".format(\n+ jtu.format_shape_dtype_string(arg_shape, dtype),\n+ idxs, update_shape, dnums),\n+ \"arg_shape\": arg_shape, \"dtype\": dtype, \"idxs\": idxs,\n+ \"update_shape\": update_shape, \"dnums\": dnums,\n+ \"rng_factory\": rng_factory, \"rng_idx_factory\": rng_idx_factory}\n+ for dtype in grad_float_dtypes\n+ for arg_shape, idxs, update_shape, dnums in [\n+ ((5,), onp.array([[0], [2]]), (2,), lax.ScatterDimensionNumbers(\n+ update_window_dims=(), inserted_window_dims=(0,),\n+ scatter_dims_to_operand_dims=(0,))),\n+ ((10,), onp.array([[0], [0], [0]]), (3, 2), lax.ScatterDimensionNumbers(\n+ update_window_dims=(1,), inserted_window_dims=(),\n+ scatter_dims_to_operand_dims=(0,))),\n+ ((10, 5,), onp.array([[0], [2], [1]]), (3, 3), lax.ScatterDimensionNumbers(\n+ update_window_dims=(1,), inserted_window_dims=(0,),\n+ scatter_dims_to_operand_dims=(0,))),\n+ ]\n+ for rng_idx_factory in [partial(jtu.rand_int, high=max(arg_shape))]\n+ for rng_factory in [jtu.rand_default]))\n+ def testScatterMin(self, arg_shape, dtype, idxs, update_shape, dnums,\n+ rng_factory, rng_idx_factory):\n+ rng = rng_factory(self.rng())\n+ rng_idx = rng_idx_factory(self.rng())\n+ idxs = rng_idx(idxs.shape, idxs.dtype)\n+ scatter_min = lambda x, y: lax.scatter_min(x, idxs, y, dnums)\n+ x = rng(arg_shape, dtype)\n+ y = rng(update_shape, dtype)\n+ check_grads(scatter_min, (x, y), 2, [\"fwd\", \"rev\"], 1e-2, 1e-2)\n+\ndef testStopGradient(self):\ndef f(x):\nreturn lax.sin(x) * lax.cos(lax.stop_gradient(x))\n" } ]
Python
Apache License 2.0
google/jax
Add gradients to the scatter_max and scatter_min operations. (#3111) This is being done to allow the creation of a differentiable segment_max. Segment_max is an important operation for GraphNets and is an open feature request at https://github.com/google/jax/issues/2255 Co-authored-by: Alex Davies <adavies@google.com>
260,440
19.05.2020 20:40:03
-3,600
73b76e9976ba94a9e28759faca602a4f9f295578
Exported lax from jax/__init__.py This allows to use lax functions without separately importing jax.lax.
[ { "change_type": "MODIFY", "old_path": "jax/__init__.py", "new_path": "jax/__init__.py", "diff": "# See the License for the specific language governing permissions and\n# limitations under the License.\n-from jax.version import __version__\nfrom .config import config\nfrom .api import (\nad, # TODO(phawkins): update users to avoid this.\n@@ -73,16 +72,19 @@ from .api import (\nxla, # TODO(phawkins): update users to avoid this.\nxla_computation,\n)\n-from jax import nn\n-from jax import random\n-\n+from .version import __version__\n+# These submodules are separate because they are in an import cycle with\n+# jax and rely on the names imported above.\n+from . import lax\n+from . import nn\n+from . import random\ndef _init():\nimport os\nos.environ.setdefault('TF_CPP_MIN_LOG_LEVEL', '1')\n- import jax.numpy # side-effecting import sets up operator overloads\n+ from . import numpy # side-effecting import sets up operator overloads\n_init()\ndel _init\n" } ]
Python
Apache License 2.0
google/jax
Exported lax from jax/__init__.py (#3135) This allows to use lax functions without separately importing jax.lax.
260,335
19.05.2020 15:17:03
25,200
850f1afd959917aa69337f253c435c84b5e53ebc
improve errors for complex derivs, fixes
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -150,7 +150,7 @@ def jit(fun: Callable, static_argnums: Union[int, Iterable[int]] = (),\nelse:\ndyn_args = args\nargs_flat, in_tree = tree_flatten((dyn_args, kwargs))\n- _check_args(args_flat)\n+ for arg in args_flat: _check_arg(arg)\nflat_fun, out_tree = flatten_fun(f, in_tree)\nout = xla.xla_call(flat_fun, *args_flat, device=device, backend=backend,\nname=flat_fun.__name__)\n@@ -370,7 +370,7 @@ def grad(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nfirst element is considered the output of the mathematical function to be\ndifferentiated and the second element is auxiliary data. Default False.\nholomorphic: Optional, bool. Indicates whether ``fun`` is promised to be\n- holomorphic. Default False.\n+ holomorphic. If True, inputs and outputs must be complex. Default False.\nReturns:\nA function with the same arguments as ``fun``, that evaluates the gradient\n@@ -424,7 +424,7 @@ def value_and_grad(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nfirst element is considered the output of the mathematical function to be\ndifferentiated and the second element is auxiliary data. Default False.\nholomorphic: Optional, bool. Indicates whether ``fun`` is promised to be\n- holomorphic. Default False.\n+ holomorphic. If True, inputs and outputs must be complex. Default False.\nReturns:\nA function with the same arguments as ``fun`` that evaluates both ``fun``\n@@ -454,17 +454,14 @@ def value_and_grad(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nf = lu.wrap_init(fun, kwargs)\nf_partial, dyn_args = argnums_partial(f, argnums, args)\n+ tree_map(partial(_check_input_dtype_grad, holomorphic), dyn_args)\nif not has_aux:\nans, vjp_py = _vjp(f_partial, *dyn_args)\nelse:\nans, vjp_py, aux = _vjp(f_partial, *dyn_args, has_aux=True)\n_check_scalar(ans)\ndtype = dtypes.result_type(ans)\n- if not (holomorphic or dtypes.issubdtype(dtype, onp.floating)):\n- msg = (\"Gradient only defined for real-output functions (with dtype that \"\n- \"is a subdtype of np.floating), but got dtype {}. For holomorphic \"\n- \"differentiation, pass holomorphic=True.\")\n- raise TypeError(msg.format(dtype))\n+ tree_map(partial(_check_output_dtype_grad, holomorphic), ans)\ng = vjp_py(onp.ones((), dtype=dtype))\ng = g[0] if isinstance(argnums, int) else g\nif not has_aux:\n@@ -487,6 +484,38 @@ def _check_scalar(x):\nelse:\nraise TypeError(msg(\"had abstract value {}\".format(aval)))\n+def _check_input_dtype_revderiv(name, holomorphic, x):\n+ _check_arg(x)\n+ aval = core.get_aval(x)\n+ if holomorphic:\n+ if not dtypes.issubdtype(aval.dtype, onp.complexfloating):\n+ msg = (f\"{name} with holomorphic=True requires inputs with complex dtype, \"\n+ f\"but got {aval.dtype.name}.\")\n+ raise TypeError(msg)\n+ elif not (dtypes.issubdtype(aval.dtype, onp.floating) or\n+ dtypes.issubdtype(aval.dtype, onp.complexfloating)):\n+ msg = (f\"{name} requires real- or complex-valued inputs (input dtype that \"\n+ \"is a sub-dtype of np.floating or np.complexfloating), \"\n+ f\"but got {aval.dtype.name}. \")\n+ raise TypeError(msg)\n+_check_input_dtype_grad = partial(_check_input_dtype_revderiv, \"grad\")\n+\n+def _check_output_dtype_revderiv(name, holomorphic, x):\n+ aval = core.get_aval(x)\n+ if holomorphic:\n+ if not dtypes.issubdtype(aval.dtype, onp.complexfloating):\n+ msg = (f\"{name} with holomorphic=True requires outputs with complex dtype, \"\n+ f\"but got {aval.dtype.name}.\")\n+ raise TypeError(msg)\n+ elif not dtypes.issubdtype(aval.dtype, onp.floating):\n+ msg = (f\"{name} requires real-valued outputs (output dtype that is \"\n+ f\"a sub-dtype of np.floating), but got {aval.dtype.name}. \"\n+ \"For holomorphic differentiation, pass holomorphic=True. \"\n+ \"For differentiation of non-holomorphic functions involving complex \"\n+ \"outputs, use jax.vjp directly.\")\n+ raise TypeError(msg)\n+_check_output_dtype_grad = partial(_check_output_dtype_revderiv, \"grad\")\n+\ndef jacfwd(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nholomorphic: bool = False) -> Callable:\n@@ -521,21 +550,39 @@ def jacfwd(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\ndef jacfun(*args, **kwargs):\nf = lu.wrap_init(fun, kwargs)\nf_partial, dyn_args = argnums_partial(f, argnums, args)\n- holomorphic or tree_map(_check_real_input_jacfwd, dyn_args)\n+ tree_map(partial(_check_input_dtype_jacfwd, holomorphic), dyn_args)\npushfwd = partial(_jvp, f_partial, dyn_args)\ny, jac = vmap(pushfwd, out_axes=(None, batching.last))(_std_basis(dyn_args))\n+ tree_map(partial(_check_output_dtype_jacfwd, holomorphic), y)\nexample_args = dyn_args[0] if isinstance(argnums, int) else dyn_args\nreturn tree_map(partial(_unravel_array_into_pytree, example_args, -1), jac)\nreturn jacfun\n-def _check_real_input_jacfwd(x):\n+def _check_input_dtype_jacfwd(holomorphic, x):\n+ _check_arg(x)\naval = core.get_aval(x)\n- if not dtypes.issubdtype(aval.dtype, onp.floating):\n- msg = (\"jacfwd only defined for functions with input dtypes that are \"\n- \"sub-dtypes of `np.floating` (i.e. that model real values), but \"\n- \"got {}. For holomorphic differentiation, pass holomorphic=True.\")\n- raise TypeError(msg.format(aval.dtype.name))\n+ if holomorphic:\n+ if not (dtypes.issubdtype(aval.dtype, onp.complexfloating) and\n+ not dtypes.issubdtype(aval.dtype, onp.floating)):\n+ msg = (\"jacfwd with holomorphic=True requires inputs with complex dtype, \"\n+ f\"but got {aval.dtype.name}.\")\n+ raise TypeError(msg)\n+ elif not dtypes.issubdtype(aval.dtype, onp.floating):\n+ msg = (\"jacfwd requires real-valued inputs (input dtype that is \"\n+ f\"a sub-dtype of np.floating), but got {aval.dtype.name}. \"\n+ \"For holomorphic differentiation, pass holomorphic=True. \"\n+ \"For differentiation of non-holomorphic functions involving complex \"\n+ \"inputs, use jax.jvp directly.\")\n+ raise TypeError(msg)\n+\n+def _check_output_dtype_jacfwd(holomorphic, x):\n+ aval = core.get_aval(x)\n+ if holomorphic:\n+ if not dtypes.issubdtype(aval.dtype, onp.complexfloating):\n+ msg = (\"jacfwd with holomorphic=True requires outputs with complex dtype, \"\n+ f\"but got {aval.dtype.name}.\")\n+ raise TypeError(msg)\ndef jacrev(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\n@@ -571,8 +618,9 @@ def jacrev(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\ndef jacfun(*args, **kwargs):\nf = lu.wrap_init(fun, kwargs)\nf_partial, dyn_args = argnums_partial(f, argnums, args)\n+ tree_map(partial(_check_input_dtype_jacrev, holomorphic), dyn_args)\ny, pullback = _vjp(f_partial, *dyn_args)\n- holomorphic or tree_map(_check_real_output_jacrev, y)\n+ tree_map(partial(_check_output_dtype_jacrev, holomorphic), y)\njac = vmap(pullback)(_std_basis(y))\njac = jac[0] if isinstance(argnums, int) else jac\nexample_args = dyn_args[0] if isinstance(argnums, int) else dyn_args\n@@ -582,13 +630,8 @@ def jacrev(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\nreturn jacfun\njacobian = jacrev\n-def _check_real_output_jacrev(x):\n- aval = core.get_aval(x)\n- if not dtypes.issubdtype(aval.dtype, onp.floating):\n- msg = (\"jacrev only defined for functions with output dtypes that are \"\n- \"sub-dtypes of `np.floating` (i.e. that model real values), but \"\n- \"got {}. For holomorphic differentiation, pass holomorphic=True.\")\n- raise TypeError(msg.format(aval.dtype.name))\n+_check_input_dtype_jacrev = partial(_check_input_dtype_revderiv, \"jacrev\")\n+_check_output_dtype_jacrev = partial(_check_output_dtype_revderiv, \"jacrev\")\ndef hessian(fun: Callable, argnums: Union[int, Sequence[int]] = 0,\n@@ -1070,7 +1113,7 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\nassert all(axis in (0, None) for axis in in_axes_flat), \\\n\"pmap currently only supports mapping over the leading axis\"\nlocal_axis_size = _mapped_axis_size(in_tree, args, in_axes_flat, \"pmap\")\n- _check_args(args)\n+ for arg in args: _check_arg(arg)\nflat_fun, out_tree = flatten_fun(f, in_tree)\nout = pxla.xla_pmap(\nflat_fun,\n@@ -1114,7 +1157,7 @@ def soft_pmap(fun: Callable, axis_name: Optional[AxisName] = None, *,\n\"soft_pmap currently only supports mapping over the leading axis\"\nmapped_invars = tuple(axis is not None for axis in in_axes_flat)\naxis_size = _mapped_axis_size(in_tree, args_flat, in_axes_flat, \"soft_pmap\")\n- _check_args(args_flat)\n+ for arg in args_flat: _check_arg(arg)\nflat_fun, out_tree = flatten_fun(f, in_tree)\nchunk_size, leftover = divmod(axis_size, pxla.unmapped_device_count(backend))\n@@ -1489,7 +1532,7 @@ def _vjp(fun: lu.WrappedFun, *primals, **kwargs):\nhas_aux = kwargs.pop('has_aux', False)\nassert not kwargs\nprimals_flat, in_tree = tree_flatten(primals)\n- _check_args(primals_flat)\n+ for arg in primals_flat: _check_arg(arg)\ntree_map(_check_inexact_input_vjp, primals)\nif not has_aux:\nflat_fun, out_tree = flatten_fun_nokwargs(fun, in_tree)\n@@ -1618,8 +1661,7 @@ def device_get(x):\nreturn tree_map(_device_get, x)\n-def _check_args(args):\n- for arg in args:\n+def _check_arg(arg):\nif not (isinstance(arg, core.Tracer) or _valid_jaxtype(arg)):\nraise TypeError(\"Argument '{}' of type {} is not a valid JAX type\"\n.format(arg, type(arg)))\n" }, { "change_type": "MODIFY", "old_path": "jax/experimental/host_callback.py", "new_path": "jax/experimental/host_callback.py", "diff": "@@ -193,14 +193,14 @@ def id_tap(func: Callable, arg, *,\nif func not in (_end_consumer, _unknown_testing_consumer):\napi._check_callable(func)\nflat_args, arg_treedef = pytree.flatten(arg)\n- api._check_args(flat_args)\n+ for arg in flat_args: api._check_arg(arg)\nparams = dict(kwargs) # we pass a copy of params to the primitive\n# See definition of id_tap_p for what parameters it takes\nparams[\"func\"] = func\nparams[\"arg_treedef\"] = arg_treedef\nif result is not None:\nflat_results, result_treedef = pytree.flatten(result)\n- api._check_args(flat_results)\n+ for result in flat_results: api._check_arg(result)\nall_args = flat_args + flat_results\nparams[\"nr_untapped\"] = len(flat_results)\nelse:\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -801,8 +801,49 @@ class APITest(jtu.JaxTestCase):\ndfn = grad(lambda x: x ** 2)\nself.assertRaisesRegex(\nTypeError,\n- \"Primal inputs to reverse-mode differentiation must be of float or \"\n- \"complex type, got type int..\", lambda: dfn(3))\n+ (r\"grad requires real- or complex-valued inputs \\(input dtype that is a \"\n+ r\"sub-dtype of np.floating or np.complexfloating\\), but got int.*.\"),\n+ lambda: dfn(3))\n+\n+ def test_grad_complex_result_errors(self):\n+ dfn = grad(lambda x: x ** 2 + 1j)\n+ self.assertRaisesRegex(\n+ TypeError,\n+ (r\"grad requires real-valued outputs \\(output dtype that is a \"\n+ r\"sub-dtype of np.floating\\), but got complex.*\"),\n+ lambda: dfn(3.))\n+\n+ def test_holomorphic_grad_of_float_errors(self):\n+ dfn = grad(lambda x: x ** 2, holomorphic=True)\n+ self.assertRaisesRegex(\n+ TypeError,\n+ (r\"grad with holomorphic=True requires inputs with complex dtype, \"\n+ r\"but got float.*\"),\n+ lambda: dfn(3.))\n+\n+ def test_holomorphic_jacrev_of_float_errors(self):\n+ dfn = jacrev(lambda x: x ** 2, holomorphic=True)\n+ self.assertRaisesRegex(\n+ TypeError,\n+ (r\"jacrev with holomorphic=True requires inputs with complex dtype, \"\n+ r\"but got float.*\"),\n+ lambda: dfn(3.))\n+\n+ def test_holomorphic_jacfwd_of_float_errors(self):\n+ dfn = jacfwd(lambda x: x ** 2, holomorphic=True)\n+ self.assertRaisesRegex(\n+ TypeError,\n+ (r\"jacfwd with holomorphic=True requires inputs with complex dtype, \"\n+ r\"but got float.*\"),\n+ lambda: dfn(3.))\n+\n+ def test_jacfwd_of_complex_errors(self):\n+ dfn = jacfwd(lambda x: x ** 2)\n+ self.assertRaisesRegex(\n+ TypeError,\n+ (r\"jacfwd requires real-valued inputs \\(input dtype that is a \"\n+ r\"sub-dtype of np.floating\\), but got complex.*\"),\n+ lambda: dfn(3. + 1j))\ndef test_xla_computation(self):\n# these tests basically check the examples in the xla_computation docstring\n" } ]
Python
Apache License 2.0
google/jax
improve errors for complex derivs, fixes #3121 (#3149)
260,700
19.05.2020 18:22:25
14,400
83a339e161d4016096663b8b3e4e8b991c9799f8
add erf and erfc rules refactor def comp
[ { "change_type": "MODIFY", "old_path": "jax/experimental/jet.py", "new_path": "jax/experimental/jet.py", "diff": "@@ -212,7 +212,36 @@ deflinear(lax.tie_in_p)\ndeflinear(lax_fft.fft_p)\ndeflinear(xla.device_put_p)\n+def def_deriv(prim, deriv):\n+ \"\"\"\n+ Define the jet rule for a primitive in terms of its first derivative.\n+ \"\"\"\n+ jet_rules[prim] = partial(deriv_prop, prim, deriv)\n+def deriv_prop(prim, deriv, primals_in, series_in):\n+ x, = primals_in\n+ series, = series_in\n+ primal_out = prim.bind(x)\n+ c0, cs = jet(deriv, primals_in, series_in)\n+ c = [c0] + cs\n+ u = [x] + series\n+ v = [primal_out] + [None] * len(series)\n+ for k in range(1, len(v)):\n+ v[k] = fact(k-1) * sum(_scale(k, j) * c[k-j] * u[j] for j in range(1, k + 1))\n+ primal_out, *series_out = v\n+ return primal_out, series_out\n+\n+\n+def_deriv(lax.erf_p, lambda x: lax.mul(lax._const(x, 2. / onp.sqrt(onp.pi)), lax.exp(lax.neg(lax.square(x)))))\n+\n+def def_comp(prim, comp):\n+ \"\"\"\n+ Define the jet rule for a primitive in terms of a composition of simpler primitives.\n+ \"\"\"\n+ jet_rules[prim] = partial(jet, comp)\n+\n+\n+def_comp(lax.erfc_p, lambda x: 1 - lax.erf(x))\n### More complicated rules\n" }, { "change_type": "MODIFY", "old_path": "tests/jet_test.py", "new_path": "tests/jet_test.py", "diff": "@@ -225,6 +225,10 @@ class JetTest(jtu.JaxTestCase):\ndef test_acosh(self): self.unary_check(lax.acosh, lims=[-100, 100])\n@jtu.skip_on_devices(\"tpu\")\ndef test_atanh(self): self.unary_check(lax.atanh, lims=[-1, 1])\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_erf(self): self.unary_check(lax.erf)\n+ @jtu.skip_on_devices(\"tpu\")\n+ def test_erfc(self): self.unary_check(lax.erfc)\n@jtu.skip_on_devices(\"tpu\")\ndef test_div(self): self.binary_check(lambda x, y: x / y, lims=[0.8, 4.0])\n" } ]
Python
Apache License 2.0
google/jax
add erf and erfc rules (#3051) refactor def comp
260,335
19.05.2020 15:51:07
25,200
ccb203c894e833264f2e11daf6bc24ba64c00260
improve pmap unbound axis error, fixes
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -1131,7 +1131,7 @@ def pmap(fun: Callable, axis_name: Optional[AxisName] = None, *, in_axes=0,\nf_pmapped.__name__ = namestr(f_pmapped.__name__, axis_name)\nreturn f_pmapped\n-class _TempAxisName(object):\n+class _TempAxisName:\ndef __init__(self, obj):\nself.obj = obj\ndef __repr__(self):\n@@ -1139,7 +1139,7 @@ class _TempAxisName(object):\ndef __hash__(self):\nreturn hash(self.obj)\ndef __eq__(self, other):\n- return self.obj is other.obj\n+ return type(other) is _TempAxisName and self.obj is other.obj\ndef soft_pmap(fun: Callable, axis_name: Optional[AxisName] = None, *,\n" }, { "change_type": "MODIFY", "old_path": "jax/interpreters/xla.py", "new_path": "jax/interpreters/xla.py", "diff": "@@ -435,7 +435,7 @@ def check_backend_params(params, outer_backend):\nreturn {k: params[k] for k in params if k != 'backend'}\n-class AxisEnv(object):\n+class AxisEnv:\ndef __init__(self, nreps, names=(), sizes=(), devices=None):\nassert isinstance(names, tuple)\nassert isinstance(sizes, tuple)\n@@ -448,7 +448,10 @@ def extend_axis_env(env, name, size):\nreturn AxisEnv(env.nreps, env.names + (name,), env.sizes + (size,), env.devices)\ndef axis_read(axis_env, axis_name):\n+ try:\nreturn max(i for i, name in enumerate(axis_env.names) if name == axis_name)\n+ except ValueError:\n+ raise NameError(\"unbound axis name: {}\".format(axis_name))\ndef axis_groups(axis_env, name):\nif isinstance(name, (list, tuple)):\n" }, { "change_type": "MODIFY", "old_path": "tests/pmap_test.py", "new_path": "tests/pmap_test.py", "diff": "@@ -1184,6 +1184,15 @@ class PmapTest(jtu.JaxTestCase):\nself.assertAllClose(result1, result3, check_dtypes=False, atol=1e-3, rtol=1e-3)\nself.assertAllClose(result1, result4, check_dtypes=False, atol=1e-3, rtol=1e-3)\n+ def testPmapAxisNameError(self):\n+ # https://github.com/google/jax/issues/3120\n+ a = np.arange(4)[np.newaxis,:]\n+ def test(x):\n+ return jax.lax.psum(x, axis_name='batch')\n+\n+ with self.assertRaisesRegex(NameError, \"unbound axis name: batch\"):\n+ jax.pmap(test)(a)\n+\ndef testPsumOnBooleanDtype(self):\n# https://github.com/google/jax/issues/3123\nn = xla_bridge.device_count()\n" } ]
Python
Apache License 2.0
google/jax
improve pmap unbound axis error, fixes #3120 (#3152)
260,335
20.05.2020 19:09:44
25,200
a4094f72a4435e57e2c52e27085e55659797b54a
revise "Tracer with raw numpy" error message * revise "Tracer with raw numpy" error message fixes * fix f-string typo * fix typo
[ { "change_type": "MODIFY", "old_path": "jax/core.py", "new_path": "jax/core.py", "diff": "@@ -371,11 +371,18 @@ class Tracer(object):\n__slots__ = ['_trace', '__weakref__']\ndef __array__(self, *args, **kw):\n- raise Exception(\"Tracer can't be used with raw numpy functions. \"\n- \"You might have\\n\"\n- \" import numpy as np\\n\"\n- \"instead of\\n\"\n- \" import jax.numpy as jnp\")\n+ msg = (\"The numpy.ndarray conversion method __array__() was called on \"\n+ f\"the JAX Tracer object {self}.\\n\\n\"\n+ \"This error can occur when a JAX Tracer object is passed to a raw \"\n+ \"numpy function, or a method on a numpy.ndarray object. You might \"\n+ \"want to check that you are using `jnp` together with \"\n+ \"`import jax.numpy as jnp` rather than using `np` via \"\n+ \"`import numpy as np`. If this error arises on a line that involves \"\n+ \"array indexing, like `x[idx]`, it may be that the array being \"\n+ \"indexed `x` is a raw numpy.ndarray while the indices `idx` are a \"\n+ \"JAX Tracer instance; in that case, you can instead write \"\n+ \"`jax.device_put(x)[idx]`.\")\n+ raise Exception(msg)\ndef __init__(self, trace):\nself._trace = trace\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -177,10 +177,8 @@ class APITest(jtu.JaxTestCase):\ndef f(x):\nreturn np.exp(x)\n- jtu.check_raises(lambda: grad(f)(np.zeros(3)), Exception,\n- \"Tracer can't be used with raw numpy functions. \"\n- \"You might have\\n import numpy as np\\ninstead of\\n\"\n- \" import jax.numpy as jnp\")\n+ with self.assertRaisesRegex(Exception, \"The numpy.ndarray conversion .*\"):\n+ grad(f)(np.zeros(3))\ndef test_binop_mismatch(self):\ndef f(x, y):\n" } ]
Python
Apache License 2.0
google/jax
revise "Tracer with raw numpy" error message (#3160) * revise "Tracer with raw numpy" error message fixes #3133 * fix f-string typo * fix typo Co-authored-by: James Bradbury <jekbradbury@google.com> Co-authored-by: James Bradbury <jekbradbury@google.com>
260,335
20.05.2020 19:09:54
25,200
f9c978e9d608e373965512ebc498c0a1338af3ed
improve docstring of jax.numpy.broadcast_to thanks !
[ { "change_type": "MODIFY", "old_path": "jax/numpy/lax_numpy.py", "new_path": "jax/numpy/lax_numpy.py", "diff": "@@ -1354,8 +1354,10 @@ def broadcast_arrays(*args):\nreturn [broadcast_to(arg, result_shape) for arg in args]\n+@_wraps(np.broadcast_to, lax_description=\"\"\"\\\n+The JAX version does not necessarily return a view of the input.\n+\"\"\")\ndef broadcast_to(arr, shape):\n- \"\"\"Like Numpy's broadcast_to but doesn't necessarily return views.\"\"\"\narr = arr if isinstance(arr, ndarray) else array(arr)\nshape = canonicalize_shape(shape) # check that shape is concrete\narr_shape = _shape(arr)\n" } ]
Python
Apache License 2.0
google/jax
improve docstring of jax.numpy.broadcast_to (#3173) thanks @joaogui1 !
260,335
20.05.2020 20:21:41
25,200
ae9d1753462d07d75e9e002b13538c385949d9af
fix while_loop cond function batching * fix while_loop cond function batching fixes * add test for
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax_control_flow.py", "new_path": "jax/lax/lax_control_flow.py", "diff": "@@ -375,7 +375,8 @@ def _while_loop_batching_rule(args, dims, cond_nconsts, cond_jaxpr,\nbody_jaxpr_batched, carry_bat_out = batching.batch_jaxpr(\nbody_jaxpr, size, batched, instantiate=carry_bat)\ncond_jaxpr_batched, (pred_bat,) = batching.batch_jaxpr(\n- cond_jaxpr, size, cconst_bat + carry_bat, instantiate=False)\n+ cond_jaxpr, size, cconst_bat + carry_bat,\n+ instantiate=bool(cond_jaxpr.out_avals[0].shape))\ncarry_bat_out = _map(partial(operator.or_, pred_bat), carry_bat_out)\nif carry_bat_out == carry_bat:\nbreak\n@@ -389,7 +390,7 @@ def _while_loop_batching_rule(args, dims, cond_nconsts, cond_jaxpr,\nnew_consts = [batching.moveaxis(x, d, 0) if d is not batching.not_mapped and d != 0\nelse x for x, d in zip(consts, const_dims)]\nnew_init = [batching.broadcast(x, size, 0) if now_bat and not was_bat\n- else batching.moveaxis(x, d, 0) if now_bat else x\n+ else batching.moveaxis(x, d, 0) if now_bat and d != 0 else x\nfor x, d, was_bat, now_bat in zip(init, init_dims, init_bat, carry_bat)]\nouts = while_p.bind(*(new_consts + new_init),\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_control_flow_test.py", "new_path": "tests/lax_control_flow_test.py", "diff": "@@ -1982,6 +1982,12 @@ class LaxControlFlowTest(jtu.JaxTestCase):\nwith api.disable_jit():\napi.vmap(trivial_while)(jnp.array([3.0,4.0])) # doesn't crash\n+ def test_vmaps_of_while_loop(self):\n+ # https://github.com/google/jax/issues/3164\n+ def f(x, n): return lax.fori_loop(0, n, lambda _, x: x + 1, x)\n+ x, n = jnp.arange(3), jnp.arange(4)\n+ api.vmap(api.vmap(f, (None, 0)), (0, None))(x, n) # doesn't crash\n+\nif __name__ == '__main__':\nabsltest.main()\n" } ]
Python
Apache License 2.0
google/jax
fix while_loop cond function batching (#3174) * fix while_loop cond function batching fixes #3164 * add test for #3164
260,335
21.05.2020 06:47:02
25,200
eb81d7e7ffcbfaff1e2b281ae6ee902d46653d6a
add dict in_axes example to vmap docstring * add dict in_axes example to vmap docstring fixes * fix typo
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -796,6 +796,17 @@ def vmap(fun: Callable, in_axes=0, out_axes=0) -> Callable:\n>>> print(vfoo(tree).shape)\n(6, 2, 5)\n+ Here's another example using container types in ``in_axes``, this time a\n+ dictionary, to specify the elements of the container to map over:\n+\n+ >>> dct = {'a': 0., 'b': np.arange(5.)}\n+ >>> x = 1.\n+ >>> def foo(dct, x):\n+ ... return dct['a'] + dct['b'] + x\n+ >>> out = vmap(foo, in_axes=({'a': None, 'b': 0}, None))(dct, x)\n+ >>> print(out)\n+ [1. 2. 3. 4. 5.]\n+\nThe results of a vectorized function can be mapped or unmapped.\nFor example, the function below returns a pair with the first\nelement mapped and the second unmapped. Only for unmapped results\n" } ]
Python
Apache License 2.0
google/jax
add dict in_axes example to vmap docstring (#3176) * add dict in_axes example to vmap docstring fixes #3161 * fix typo
260,335
21.05.2020 08:00:18
25,200
5c1de2836c11f5068f9df75fcce4fe887b458f1c
revise vmap in_axes/out_axes leaf type error msg from discussion
[ { "change_type": "MODIFY", "old_path": "jax/api.py", "new_path": "jax/api.py", "diff": "@@ -836,11 +836,16 @@ def vmap(fun: Callable, in_axes=0, out_axes=0) -> Callable:\n# rather than raising an error. https://github.com/google/jax/issues/2367\nin_axes = tuple(in_axes)\n- if (not isinstance(in_axes, (list, tuple, type(None), int))\n- or not isinstance(out_axes, (list, tuple, type(None), int))):\n- msg = (\"vmap arguments in_axes and out_axes must each be an integer, None, \"\n- \"or a (nested) tuple of those types, got {} and {} respectively.\")\n- raise TypeError(msg.format(type(in_axes), type(out_axes)))\n+ if not all(isinstance(l, (type(None), int, batching._Last))\n+ for l in tree_leaves(in_axes)):\n+ msg = (\"vmap in_axes must be an int, None, or (nested) container with \"\n+ \"those types as leaves, but got {}.\")\n+ raise TypeError(msg.format(in_axes))\n+ if not all(isinstance(l, (type(None), int, batching._Last))\n+ for l in tree_leaves(out_axes)):\n+ msg = (\"vmap out_axes must be an int, None, or (nested) container with \"\n+ \"those types as leaves, but got {}.\")\n+ raise TypeError(msg.format(out_axes))\n@wraps(fun, docstr=docstr)\ndef batched_fun(*args):\n" }, { "change_type": "MODIFY", "old_path": "tests/api_test.py", "new_path": "tests/api_test.py", "diff": "@@ -1066,6 +1066,16 @@ class APITest(jtu.JaxTestCase):\nlambda: api.vmap(lambda x: x, in_axes=(0, 0))(jnp.ones(3))\n)\n+ def test_vmap_in_axes_leaf_types(self):\n+ with self.assertRaisesRegex(\n+ TypeError, r\"vmap in_axes must be an int, None, or .*\"):\n+ api.vmap(lambda x: x, in_axes=(jnp.array([1., 2.]),))(jnp.array([1., 2.]))\n+\n+ def test_vmap_out_axes_leaf_types(self):\n+ with self.assertRaisesRegex(\n+ TypeError, r\"vmap out_axes must be an int, None, or .*\"):\n+ api.vmap(lambda x: x, out_axes=(jnp.array([1., 2.]),))(jnp.array([1., 2.]))\n+\ndef test_vmap_unbatched_object_passthrough_issue_183(self):\n# https://github.com/google/jax/issues/183\nfun = lambda f, x: f(x)\n@@ -1114,10 +1124,6 @@ class APITest(jtu.JaxTestCase):\n# The mapped inputs cannot be scalars\napi.vmap(lambda x: x)(1.)\n- with self.assertRaisesRegex(\n- ValueError, re.escape(\"vmap got arg 0 of rank 1 but axis to be mapped [1. 2.]\")):\n- api.vmap(lambda x: x, in_axes=(jnp.array([1., 2.]),))(jnp.array([1., 2.]))\n-\nwith self.assertRaisesRegex(\nValueError, \"vmap must have at least one non-None value in in_axes\"):\n# If the output is mapped, there must be a non-None in_axes\n" } ]
Python
Apache License 2.0
google/jax
revise vmap in_axes/out_axes leaf type error msg (#3179) from #3161 discussion
260,306
21.05.2020 14:00:58
25,200
d8ede0106a6dcf8f71d97dc5e47636398dcb9124
Update jax/interpreters/pxla.py
[ { "change_type": "MODIFY", "old_path": "jax/interpreters/pxla.py", "new_path": "jax/interpreters/pxla.py", "diff": "@@ -716,7 +716,7 @@ def parallel_callable(fun, backend, axis_name, axis_size, global_axis_size,\nif devices is None:\nif num_global_shards > xb.device_count(backend):\n- msg = (\"compiling computation that requires {} replicas, but only {} XLA \"\n+ msg = (\"compiling computation that requires {} logical devices, but only {} XLA \"\n\"devices are available (num_replicas={}, num_partitions={})\")\nraise ValueError(msg.format(num_global_shards, xb.device_count(backend),\nnum_global_replicas, num_partitions))\n" } ]
Python
Apache License 2.0
google/jax
Update jax/interpreters/pxla.py Co-authored-by: James Bradbury <jekbradbury@google.com>
260,309
22.05.2020 14:12:44
25,200
190f88dedebd90e87d73e6b19d591afaf5357b7f
Update Common_Gotchas_in_JAX.ipynb typo fix
[ { "change_type": "MODIFY", "old_path": "docs/notebooks/Common_Gotchas_in_JAX.ipynb", "new_path": "docs/notebooks/Common_Gotchas_in_JAX.ipynb", "diff": "\"colab_type\": \"text\"\n},\n\"source\": [\n- \"JAX transformation and compilation are designed to work only on Python functions that are functionally pure: all the input data is passed through the function parameters, all the results are output through the function results. A pure function will always return the same result if involved with the same inputs. \\n\",\n+ \"JAX transformation and compilation are designed to work only on Python functions that are functionally pure: all the input data is passed through the function parameters, all the results are output through the function results. A pure function will always return the same result if invoked with the same inputs. \\n\",\n\"\\n\",\n\"Here are some examples of functions that are not functially pure for which JAX behaves differently than the Python interpreter. Note that these behaviors are not guaranteed by the JAX system; the proper way to use JAX is to use it only on functionally pure Python functions.\\n\"\n]\n" } ]
Python
Apache License 2.0
google/jax
Update Common_Gotchas_in_JAX.ipynb (#3189) typo fix
260,411
26.05.2020 10:22:33
-10,800
f18f7920ba6fa830d2bdfef17dc2974f863ba907
Fix error in code generation of batched while loops Fixed the case when the value is a unit, which we do not batch.
[ { "change_type": "MODIFY", "old_path": "jax/lax/lax_control_flow.py", "new_path": "jax/lax/lax_control_flow.py", "diff": "@@ -329,7 +329,7 @@ def _while_loop_translation_rule(c, axis_env, name_stack, avals, backend, *args,\nbody_pred, = xla.jaxpr_subcomp(body_c, cond_jaxpr.jaxpr, backend, axis_env,\n_map(partial(xb.constant, body_c), cond_jaxpr.literals),\nextend_name_stack(name_stack, 'body_pred'), *(x + z))\n- new_z = _map(partial(_pred_bcast_select, body_c, body_pred), new_z, z)\n+ new_z = _map(partial(_pred_bcast_select, body_c, body_pred), new_z, z, body_jaxpr.out_avals)\nassert _map(body_c.get_shape, new_z) == _map(body_c.get_shape, z) # no broadcast\nnew_carry = xops.Tuple(body_c, list(itertools.chain(x, y, new_z)))\n@@ -338,13 +338,14 @@ def _while_loop_translation_rule(c, axis_env, name_stack, avals, backend, *args,\n_, _, z = split_list(ans_elts, [cond_nconsts, body_nconsts])\nreturn xops.Tuple(c, z)\n-def _pred_bcast_select(c, pred, x, y):\n+def _pred_bcast_select(c, pred, x, y, x_y_aval: core.AbstractValue):\npred_shape = c.get_shape(pred).dimensions()\nx_shape = c.get_shape(x).dimensions()\ny_shape = c.get_shape(y).dimensions()\nassert x_shape == y_shape\n- if not c.get_shape(x).is_array() and not c.get_shape(y).is_array():\n- # Two tokens\n+ if x_y_aval is core.abstract_unit:\n+ return x\n+ elif x_y_aval is core.abstract_token:\nreturn xops.AfterAll(c, [x, y])\nelse:\nassert pred_shape == x_shape[:len(pred_shape)] == y_shape[:len(pred_shape)]\n" }, { "change_type": "MODIFY", "old_path": "tests/lax_control_flow_test.py", "new_path": "tests/lax_control_flow_test.py", "diff": "@@ -334,6 +334,38 @@ class LaxControlFlowTest(jtu.JaxTestCase):\nexpected = (np.array([4, 3]), np.array([1, 2]))\nself.assertAllClose(ans, expected, check_dtypes=False)\n+ def test_issue_3204(self):\n+ # Error during XLA code generation for vmap of nested loops\n+ def test(a, b):\n+ val = 0\n+ i = 0\n+ j = 0\n+\n+ condfun_1 = lambda inp: inp[1] < a + 1\n+ condfun_2 = lambda inp: inp[2] < b + 1\n+\n+ def bodyfun_1(inp):\n+ val, i, j = inp\n+ j = 0\n+\n+ def bodyfun_2(inp):\n+ val, i, j = inp\n+ val += i + j\n+ j += 1\n+ return (val, i, j)\n+\n+ result = lax.while_loop(condfun_2, bodyfun_2, (val, i, j))\n+ val = result[0]\n+ i += 1\n+ return (val, i, j)\n+\n+ result = lax.while_loop(condfun_1, bodyfun_1, (val, i, j))\n+ return result[0]\n+\n+ arr = np.arange(5)\n+ vmap_test = api.vmap(test, (0, 0))\n+ vmap_test(arr, arr)\n+\ndef testForiLoopErrors(self):\n\"\"\"Test typing error messages for while.\"\"\"\nwith self.assertRaisesRegex(\n" } ]
Python
Apache License 2.0
google/jax
Fix error in code generation of batched while loops (#3207) Fixed the case when the value is a unit, which we do not batch.
260,483
26.05.2020 20:21:22
-3,600
0f230029c98ff796141873b742a38beee0f0703b
Add a JAX flag to avoid most optimizations.
[ { "change_type": "MODIFY", "old_path": "jax/lib/xla_bridge.py", "new_path": "jax/lib/xla_bridge.py", "diff": "@@ -58,6 +58,10 @@ flags.DEFINE_string(\n'Platform name for XLA. The default is to attempt to use a GPU if '\n'available, but fall back to CPU otherwise. To set the platform manually, '\n'pass \"cpu\" for CPU or \"gpu\" for GPU.')\n+flags.DEFINE_bool(\n+ 'jax_disable_most_optimizations', False,\n+ 'Try not to do much optimization work. This can be useful if the cost of '\n+ 'optimization is greater than that of running a less-optimized program.')\ndef get_compile_options(num_replicas, num_partitions, device_assignment=None):\n@@ -97,6 +101,13 @@ def get_compile_options(num_replicas, num_partitions, device_assignment=None):\nassert device_assignment.replica_count() == num_replicas\nassert device_assignment.computation_count() == num_partitions\ncompile_options.device_assignment = device_assignment\n+\n+ if FLAGS.jax_disable_most_optimizations:\n+ debug_options = compile_options.executable_build_options.debug_options\n+ debug_options.xla_backend_optimization_level = 0\n+ debug_options.xla_llvm_disable_expensive_passes = True\n+ debug_options.xla_test_all_input_layouts = False\n+\nreturn compile_options\n_backends = {}\n" } ]
Python
Apache License 2.0
google/jax
Add a JAX flag to avoid most optimizations. (#3208)