mirror of
https://github.com/saymrwulf/pytorch.git
synced 2026-05-15 21:00:47 +00:00
c028bcd466
Summary: The previous refactor of these four Ops changed their input semantics, which makes backward impatible with old models. This diff fix this problem by checking the input and define follow-up behavior by case, so that the old models can be accommodated. Reviewed By: dzhulgakov Differential Revision: D6905840 fbshipit-source-id: fc37baec407fd5eae64fc9c2b61aba3c492a90f3
189 lines
6.3 KiB
Python
189 lines
6.3 KiB
Python
# Copyright (c) 2016-present, Facebook, Inc.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
##############################################################################
|
|
|
|
from __future__ import absolute_import
|
|
from __future__ import division
|
|
from __future__ import print_function
|
|
from __future__ import unicode_literals
|
|
|
|
from caffe2.python import core, workspace
|
|
from hypothesis import given
|
|
|
|
import caffe2.python.hypothesis_test_util as hu
|
|
import hypothesis.strategies as st
|
|
import numpy as np
|
|
|
|
|
|
class TestReduceFrontReductions(hu.HypothesisTestCase):
|
|
|
|
def grad_variant_input_test(self, grad_op_name, X, ref, num_reduce_dim):
|
|
workspace.ResetWorkspace()
|
|
|
|
Y = np.array(ref(X)[0]).astype(np.float32)
|
|
dY = np.array(np.random.rand(*Y.shape)).astype(np.float32)
|
|
shape = np.array(X.shape).astype(np.int64)
|
|
|
|
workspace.FeedBlob("X", X)
|
|
workspace.FeedBlob("dY", dY)
|
|
workspace.FeedBlob("shape", shape)
|
|
|
|
grad_op = core.CreateOperator(
|
|
grad_op_name,
|
|
["dY", "X"],
|
|
["dX"],
|
|
num_reduce_dim=num_reduce_dim
|
|
)
|
|
|
|
grad_op1 = core.CreateOperator(
|
|
grad_op_name,
|
|
["dY", "shape"],
|
|
["dX1"],
|
|
num_reduce_dim=num_reduce_dim
|
|
)
|
|
|
|
workspace.RunOperatorOnce(grad_op)
|
|
workspace.RunOperatorOnce(grad_op1)
|
|
|
|
dX = workspace.FetchBlob("dX")
|
|
dX1 = workspace.FetchBlob("dX1")
|
|
np.testing.assert_array_equal(dX, dX1)
|
|
|
|
def reduce_op_test(self, op_name, op_ref, in_data, num_reduce_dims, device):
|
|
op = core.CreateOperator(
|
|
op_name,
|
|
["inputs"],
|
|
["outputs"],
|
|
num_reduce_dim=num_reduce_dims
|
|
)
|
|
|
|
self.assertReferenceChecks(
|
|
device_option=device,
|
|
op=op,
|
|
inputs=[in_data],
|
|
reference=op_ref
|
|
)
|
|
|
|
self.assertGradientChecks(
|
|
device, op, [in_data], 0, [0], stepsize=1e-2, threshold=1e-2)
|
|
|
|
@given(num_reduce_dim=st.integers(0, 4), **hu.gcs)
|
|
def test_reduce_front_sum(self, num_reduce_dim, gc, dc):
|
|
X = np.random.rand(7, 4, 3, 5).astype(np.float32)
|
|
|
|
def ref_sum(X):
|
|
return [np.sum(X, axis=(tuple(range(num_reduce_dim))))]
|
|
|
|
self.reduce_op_test("ReduceFrontSum", ref_sum, X, num_reduce_dim, gc)
|
|
self.grad_variant_input_test(
|
|
"ReduceFrontSumGradient", X, ref_sum, num_reduce_dim)
|
|
|
|
@given(num_reduce_dim=st.integers(0, 4), **hu.gcs)
|
|
def test_reduce_front_mean(self, num_reduce_dim, gc, dc):
|
|
X = np.random.rand(6, 7, 8, 2).astype(np.float32)
|
|
|
|
def ref_mean(X):
|
|
return [np.mean(X, axis=(tuple(range(num_reduce_dim))))]
|
|
|
|
self.reduce_op_test("ReduceFrontMean", ref_mean, X, num_reduce_dim, gc)
|
|
self.grad_variant_input_test(
|
|
"ReduceFrontMeanGradient", X, ref_mean, num_reduce_dim)
|
|
|
|
@given(num_reduce_dim=st.integers(0, 4), **hu.gcs)
|
|
def test_reduce_front_max(self, num_reduce_dim, gc, dc):
|
|
X = np.random.rand(6, 7, 8, 2).astype(np.float32)
|
|
|
|
def ref_frontmax(X):
|
|
return [np.max(X, axis=(tuple(range(num_reduce_dim))))]
|
|
|
|
op = core.CreateOperator(
|
|
"ReduceFrontMax",
|
|
["inputs"],
|
|
["outputs"],
|
|
num_reduce_dim=num_reduce_dim
|
|
)
|
|
|
|
self.assertReferenceChecks(
|
|
device_option=gc,
|
|
op=op,
|
|
inputs=[X],
|
|
reference=ref_frontmax,
|
|
)
|
|
|
|
# Skip gradient check because it is too unreliable with max.
|
|
# Just check CPU and CUDA have same results
|
|
Y = np.array(ref_frontmax(X)[0]).astype(np.float32)
|
|
dY = np.array(np.random.rand(*Y.shape)).astype(np.float32)
|
|
grad_op = core.CreateOperator(
|
|
"ReduceFrontMaxGradient",
|
|
["dY", "X", "Y"],
|
|
["dX"],
|
|
num_reduce_dim=num_reduce_dim
|
|
)
|
|
self.assertDeviceChecks(dc, grad_op, [dY, X, Y], [0])
|
|
|
|
@given(num_reduce_dim=st.integers(0, 4), **hu.gcs)
|
|
def test_reduce_back_max(self, num_reduce_dim, gc, dc):
|
|
X = np.random.rand(6, 7, 8, 2).astype(np.float32)
|
|
|
|
def ref_backmax(X):
|
|
return [np.max(X, axis=(0, 1, 2, 3)[4 - num_reduce_dim:])]
|
|
|
|
op = core.CreateOperator(
|
|
"ReduceBackMax",
|
|
["inputs"],
|
|
["outputs"],
|
|
num_reduce_dim=num_reduce_dim
|
|
)
|
|
|
|
self.assertReferenceChecks(
|
|
device_option=gc,
|
|
op=op,
|
|
inputs=[X],
|
|
reference=ref_backmax
|
|
)
|
|
|
|
# Skip gradient check because it is too unreliable with max
|
|
# Just check CPU and CUDA have same results
|
|
Y = np.array(ref_backmax(X)[0]).astype(np.float32)
|
|
dY = np.array(np.random.rand(*Y.shape)).astype(np.float32)
|
|
grad_op = core.CreateOperator(
|
|
"ReduceBackMaxGradient",
|
|
["dY", "X", "Y"],
|
|
["dX"],
|
|
num_reduce_dim=num_reduce_dim
|
|
)
|
|
self.assertDeviceChecks(dc, grad_op, [dY, X, Y], [0])
|
|
|
|
@given(num_reduce_dim=st.integers(0, 4), **hu.gcs)
|
|
def test_reduce_back_sum(self, num_reduce_dim, dc, gc):
|
|
X = np.random.rand(6, 7, 8, 2).astype(np.float32)
|
|
|
|
def ref_sum(X):
|
|
return [np.sum(X, axis=(0, 1, 2, 3)[4 - num_reduce_dim:])]
|
|
|
|
self.reduce_op_test("ReduceBackSum", ref_sum, X, num_reduce_dim, gc)
|
|
self.grad_variant_input_test(
|
|
"ReduceBackSumGradient", X, ref_sum, num_reduce_dim)
|
|
|
|
@given(num_reduce_dim=st.integers(0, 4), **hu.gcs)
|
|
def test_reduce_back_mean(self, num_reduce_dim, dc, gc):
|
|
X = np.random.rand(6, 7, 8, 2).astype(np.float32)
|
|
|
|
def ref_mean(X):
|
|
return [np.mean(X, axis=(0, 1, 2, 3)[4 - num_reduce_dim:])]
|
|
|
|
self.reduce_op_test("ReduceBackMean", ref_mean, X, num_reduce_dim, gc)
|
|
self.grad_variant_input_test(
|
|
"ReduceBackMeanGradient", X, ref_mean, num_reduce_dim)
|