mindspore/tests/st/pynative/test_pynative_hook_grad.py

# Copyright 2020 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
""" test_pynative_hook_grad """
import numpy as np
import pytest
import mindspore.nn as nn
import mindspore.ops.operations as P
from mindspore.nn import Cell
from mindspore import context
from mindspore.common.tensor import Tensor
from mindspore.ops.composite import GradOperation
from mindspore.common import ParameterTuple

def setup_module():
    context.set_context(mode=context.PYNATIVE_MODE, device_target="Ascend")

class MetaFactory:
    def __init__(self):
        self.device_target = context.get_context('device_target')
        self.rank_size = None
        self.device_id = None
        self.global_rank_id = None

class HookBase(MetaFactory):
    def __init__(self):
        super().__init__()
        MetaFactory.__init__(self)
        self.grad_input_list = []
        self.grad_output_list = []

    def ms_record_hook(self, cell_id, grad_input, grad_output):
        for grad in grad_input:
            self.grad_input_list.append(grad)
        for grad in grad_output:
            self.grad_output_list.append(grad)

    def ms_change_grad_double_hook(self, cell_id, grad_input, grad_output):
        y = Tensor(np.array([2.0]).astype(np.float32))
        mul = P.Mul()
        grad = grad_output[0]
        output = mul(grad, y)
        return output

class FinalNet(nn.Cell, HookBase):
    def __init__(self):
        super().__init__()
        HookBase.__init__(self)
        self.conv = nn.Conv2d(1, 3, 3)
        self.relu = nn.ReLU()

    def construct(self, x, flag):
        if flag:
            x = self.conv(x)
        else:
            x = self.relu(x)
        return self.relu(x)

class _Grad(Cell):
    def __init__(self, grad, network, wrt_params=False, real_inputs_count=None):
        super().__init__()
        self.network = network
        self.grad = grad
        self.sens_param = self.grad.sens_param
        self.wrt_params = wrt_params
        self.real_inputs_count = real_inputs_count
        if self.wrt_params:
            self.params = ParameterTuple(self.network.trainable_params())

    def construct(self, *inputs):
        if self.wrt_params:
            if self.real_inputs_count is None or self.sens_param is False:
                return self.grad(self.network, self.params)(*inputs)
            real_inputs = inputs[:self.real_inputs_count]
            sense_param_inputs = inputs[self.real_inputs_count:]
            return self.grad(self.network, self.params)(*real_inputs, sense_param_inputs)
        if self.real_inputs_count is None or self.sens_param is False:
            return self.grad(self.network)(*inputs)
        real_inputs = inputs[:self.real_inputs_count]
        sense_param_inputs = inputs[self.real_inputs_count:]
        return self.grad(self.network)(*real_inputs, sense_param_inputs)

class GradOfAllInputs(_Grad):
    def __init__(self, network, sens_param=True, real_inputs_count=None):
        super().__init__(grad=GradOperation(get_all=True, sens_param=sens_param),
                         network=network, real_inputs_count=real_inputs_count)

class MsMul4(nn.Cell):
    def construct(self, input_mul):
        out = input_mul * 2
        return out

class MsMul(nn.Cell):
    def __init__(self):
        super().__init__()
        self.mul = P.Mul()

    def construct(self, x, y):
        x = self.mul(x, y)
        return x

class MsAdd4(nn.Cell):
    def construct(self, input_add):
        out = input_add + 4
        return out

class MsOneInputNet(nn.Cell, HookBase):
    def __init__(self):
        super().__init__()
        HookBase.__init__(self)
        self.add = MsAdd4()
        self.mul = MsMul4()
        self.relu = nn.ReLU()

    def construct(self, x):
        x = self.add(x)
        x = self.mul(x)
        out = self.relu(x)
        return out

class MsMultiInputNet(nn.Cell, HookBase):
    def __init__(self):
        super().__init__()
        HookBase.__init__(self)
        self.mul1 = MsMul()
        self.mul2 = MsMul4()
    def construct(self, x, y):
        a = self.mul1(x, y)
        b = self.mul2(x)
        output = self.mul1(a, b)
        return output

class MsNetWithParameter(nn.Cell, HookBase):
    def __init__(self):
        super().__init__()
        HookBase.__init__(self)
        self.conv1 = nn.Conv2d(2, 4, kernel_size=(1, 1), has_bias=True,
                               weight_init=Tensor(np.ones([4, 2, 1, 1]).astype(np.float32)),
                               bias_init=Tensor(np.ones([4]).astype(np.float32)))
        self.conv2 = nn.Conv2d(4, 8, kernel_size=(1, 1), has_bias=True,
                               weight_init=Tensor(np.ones([8, 4, 1, 1]).astype(np.float32)),
                               bias_init=Tensor(np.ones([8]).astype(np.float32)))

    def construct(self, x):
        x = self.conv1(x)
        output = self.conv2(x)
        return output

class MsNetWithCellinCell(nn.Cell, HookBase):
    def __init__(self):
        super().__init__()
        HookBase.__init__(self)
        self.net1 = MsOneInputNet()
        self.mul = MsMul4()

    def construct(self, x):
        x = self.net1(x)
        output = self.mul(x)
        return output

class MsSingleOpNetWithBprop(nn.Cell, HookBase):
    def __init__(self):
        super().__init__()
        HookBase.__init__(self)
        self.op = nn.ReLU()

    def construct(self, x):
        return self.op(x)

    def bprop(self, x, out, dout):
        y = Tensor(np.array([5.0]).astype(np.float32))
        mul = P.Mul()
        return mul(x, y)

class MsNetHasBpropInChild(nn.Cell, HookBase):
    def __init__(self):
        super().__init__()
        HookBase.__init__(self)
        self.add = MsAdd4()
        self.bprop_net = MsSingleOpNetWithBprop()

    def construct(self, x):
        x = self.add(x)
        return self.bprop_net(x)

class MsMultiOpNetWithBprop(nn.Cell, HookBase):
    def __init__(self):
        super().__init__()
        HookBase.__init__(self)
        self.mul = MsMul4()
        self.relu = nn.ReLU()

    def construct(self, x):
        x = self.mul(x)
        return self.relu(x)

    def bprop(self, x, out, dout):
        y = Tensor(np.array([5.0]).astype(np.float32))
        mul = P.Mul()
        return mul(x, y)

def _count_unequal_element(data_expected, data_me, rtol, atol):
    assert data_expected.shape == data_me.shape
    total_count = len(data_expected.flatten())
    error = np.abs(data_expected - data_me)
    greater = np.greater(error, atol + np.abs(data_me)*rtol)
    loss_count = np.count_nonzero(greater)
    assert (loss_count/total_count) < rtol,\
        "\ndata_expected_std:{0}\ndata_me_error:{1}\nloss:{2}".\
        format(data_expected[greater], data_me[greater], error[greater])

def allclose_nparray(data_expected, data_me, rtol, atol, equal_nan=True):
    if np.any(np.isnan(data_expected)):
        assert np.allclose(data_expected, data_me, rtol, atol, equal_nan=equal_nan)
    elif not np.allclose(data_expected, data_me, rtol, atol, equal_nan=equal_nan):
        _count_unequal_element(data_expected, data_me, rtol, atol)
    else:
        assert True

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_diff_hook():
    input_np = np.ones([1, 1, 224, 224]).astype(np.float32)
    ms_net = FinalNet()
    ms_net.set_grad()
    ms_net.conv.register_backward_hook(ms_net.ms_record_hook)
    ms_net.relu.register_backward_hook(ms_net.ms_change_grad_double_hook)
    input_ms = Tensor(input_np)
    out_ms = ms_net(input_ms, Tensor(1))
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    grad_net(input_ms, Tensor(1), out_ms)

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_outermost_cell_not_change_grad():
    input_np = np.ones([2, 2]).astype(np.float32)

    ms_net = MsOneInputNet()
    ms_net.set_grad()
    ms_net.register_backward_hook(ms_net.ms_record_hook)
    input_ms = Tensor(input_np)
    out_ms = ms_net(input_ms)
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    input_ms_grad = grad_net(input_ms, out_ms)

    #input grad
    input_torch_grad = np.array([[20, 20], [20, 20]])
    allclose_nparray(input_torch_grad, input_ms_grad[0].asnumpy(), 0.001, 0.001)
    #hook record grad
    torch_net_grad_output = np.array([[10, 10], [10, 10]])
    torch_net_grad_input = np.array([[20, 20], [20, 20]])
    allclose_nparray(torch_net_grad_output, ms_net.grad_input_list[0].asnumpy(), 0.001, 0.001)
    allclose_nparray(torch_net_grad_input, ms_net.grad_output_list[0].asnumpy(), 0.001, 0.001)

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_all_cell_record_grad():
    input_np = np.ones([2, 2]).astype(np.float32)

    ms_net = MsOneInputNet()
    ms_net.set_grad()
    ms_net.mul.register_backward_hook(ms_net.ms_record_hook)
    ms_net.add.register_backward_hook(ms_net.ms_record_hook)
    ms_net.relu.register_backward_hook(ms_net.ms_record_hook)
    input_ms = Tensor(input_np)
    out_ms = ms_net(input_ms)
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    grad_net(input_ms, out_ms)

    torch_net_grad_input0 = np.array([[10, 10], [10, 10]])
    torch_net_grad_output0 = np.array([[10, 10], [10, 10]])
    torch_net_grad_input1 = np.array([[20, 20], [20, 20]])
    torch_net_grad_output1 = np.array([[10, 10], [10, 10]])
    allclose_nparray(torch_net_grad_input0, ms_net.grad_output_list[0].asnumpy(), 0.001, 0.001)
    allclose_nparray(torch_net_grad_output0, ms_net.grad_input_list[0].asnumpy(), 0.001, 0.001)
    allclose_nparray(torch_net_grad_input1, ms_net.grad_output_list[1].asnumpy(), 0.001, 0.001)
    allclose_nparray(torch_net_grad_output1, ms_net.grad_input_list[1].asnumpy(), 0.001, 0.001)

    torch_net_grad_input3 = np.array([[20, 20], [20, 20]])
    torch_net_grad_output2 = np.array([[20, 20], [20, 20]])
    allclose_nparray(torch_net_grad_input3, ms_net.grad_output_list[2].asnumpy(), 0.001, 0.001)
    allclose_nparray(torch_net_grad_output2, ms_net.grad_input_list[2].asnumpy(), 0.001, 0.001)

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_mul_change_input_grad():
    input_np = np.ones([2, 2]).astype(np.float32)

    ms_net = MsOneInputNet()
    ms_net.set_grad()
    ms_net.mul.register_backward_hook(ms_net.ms_change_grad_double_hook)
    input_ms = Tensor(input_np)
    out_ms = ms_net(input_ms)
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    input_ms_grad = grad_net(input_ms, out_ms)

    #input grad
    input_torch_grad = np.array([[40, 40], [40, 40]])
    allclose_nparray(input_torch_grad, input_ms_grad[0].asnumpy(), 0.001, 0.001)

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_mul2_change_input_grad():
    input1_np = np.array([2.0, 3.0, 4.0]).astype(np.float32)
    input2_np = np.array([2.0, 3.0, 4.0]).astype(np.float32)

    ms_net = MsMultiInputNet()
    ms_net.set_grad()
    ms_net.mul2.register_backward_hook(ms_net.ms_change_grad_double_hook)
    input1_ms = Tensor(input1_np)
    input2_ms = Tensor(input2_np)
    out_ms = ms_net(input1_ms, input2_ms)
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    input_ms_grad = grad_net(input1_ms, input2_ms, out_ms)

    #input grad
    input1_torch_grad = np.array([384, 2916, 12288])
    input2_torch_grad = np.array([128, 972, 4096])
    allclose_nparray(input1_torch_grad, input_ms_grad[0].asnumpy(), 0.001, 0.001)
    allclose_nparray(input2_torch_grad, input_ms_grad[1].asnumpy(), 0.001, 0.001)

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_outermost_cell_change_grad():
    input_np = np.ones([2, 2]).astype(np.float32)

    ms_net = MsNetWithCellinCell()
    ms_net.set_grad()
    ms_net.register_backward_hook(ms_net.ms_change_grad_double_hook)
    input_ms = Tensor(input_np)
    out_ms = ms_net(input_ms)
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    input_ms_grad = grad_net(input_ms, out_ms)

    #input grad
    out_torch = np.array([[20, 20], [20, 20]])
    input_torch_grad = np.array([[160, 160], [160, 160]])
    allclose_nparray(out_torch, out_ms.asnumpy(), 0.001, 0.001)
    allclose_nparray(input_torch_grad, input_ms_grad[0].asnumpy(), 0.001, 0.001)

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_outermost_cell_record_grad():
    input_np = np.ones([2, 2]).astype(np.float32)

    ms_net = MsSingleOpNetWithBprop()
    ms_net.set_grad()
    ms_net.bprop_debug = True
    ms_net.register_backward_hook(ms_net.ms_record_hook)
    input_ms = Tensor(input_np)
    out_ms = ms_net(input_ms)
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    input_ms_grad = grad_net(input_ms, out_ms)

    if ms_net.grad_output_list or ms_net.grad_input_list:
        assert False

    #input grad
    out_torch = np.array([[1, 1], [1, 1]])
    input_torch_grad = np.array([[5, 5], [5, 5]])
    allclose_nparray(out_torch, out_ms.asnumpy(), 0.001, 0.001)
    allclose_nparray(input_torch_grad, input_ms_grad[0].asnumpy(), 0.001, 0.001)

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_bprop_outermost_cell_record_grad():
    input_np = np.ones([2, 2]).astype(np.float32)

    ms_net = MsNetHasBpropInChild()
    ms_net.set_grad()
    ms_net.bprop_net.bprop_debug = True
    ms_net.register_backward_hook(ms_net.ms_record_hook)
    input_ms = Tensor(input_np)
    out_ms = ms_net(input_ms)
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    input_ms_grad = grad_net(input_ms, out_ms)

    if len(ms_net.grad_output_list) != len(ms_net.grad_input_list) or not ms_net.grad_output_list:
        assert False

    #input grad
    out_torch = np.array([[5, 5], [5, 5]])
    input_torch_grad = np.array([[25, 25], [25, 25]])
    allclose_nparray(out_torch, out_ms.asnumpy(), 0.001, 0.001)
    allclose_nparray(input_torch_grad, input_ms_grad[0].asnumpy(), 0.001, 0.001)
    #hook record grad
    torch_net_grad_output = np.array([[5, 5], [5, 5]])
    torch_net_grad_input = np.array([[25, 25], [25, 25]])
    allclose_nparray(torch_net_grad_output, ms_net.grad_input_list[0].asnumpy(), 0.001, 0.001)
    allclose_nparray(torch_net_grad_input, ms_net.grad_output_list[0].asnumpy(), 0.001, 0.001)

@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_pynative_hook_child_cell_record_grad():
    input_np = np.ones([2, 2]).astype(np.float32)

    ms_net = MsMultiOpNetWithBprop()
    ms_net.set_grad()
    ms_net.bprop_debug = True
    ms_net.relu.register_backward_hook(ms_net.ms_record_hook)
    ms_net.mul.register_backward_hook(ms_net.ms_record_hook)
    input_ms = Tensor(input_np)
    out_ms = ms_net(input_ms)
    grad_net = GradOfAllInputs(ms_net)
    grad_net.set_train()
    grad_net(input_ms, out_ms)

    if ms_net.grad_output_list or ms_net.grad_input_list:
        assert False