MegEngine/imperative/python/test/unit/quantization/test_fake_quant.py

# -*- coding: utf-8 -*-
# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
#
# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
import numpy as np
import pytest

import megengine as mge
import megengine.functional as F
from megengine import tensor
from megengine.core.autodiff.grad import Function, Grad
from megengine.core.tensor.dtype import QuantDtypeMeta
from megengine.core.tensor.utils import make_shape_tuple
from megengine.quantization.internal_fake_quant import *
from megengine.quantization.utils import (
    QuantMode,
    create_qparams,
    fake_quant_tensor,
    lsq_forward,
    tqt_forward,
)


class TQT_numpy:
    def __init__(self, lowerbound, upperbound):
        super().__init__()
        self.lowerbound = lowerbound
        self.upperbound = upperbound

    def forward(self, inp, scale):
        t = 2 ** scale
        # t = F.maximum(t, 1e-4)
        inp_scaled = inp / t
        inp_clipped = np.maximum(
            np.minimum(inp_scaled, self.upperbound), self.lowerbound
        )
        inp_rounded = np.round(inp_clipped)
        inp_flq = inp_rounded * t
        self.saved_tensors = (inp_scaled, inp_rounded, t)
        return inp_flq

    def backward(self, grad_inp_flq):
        (inp_scaled, inp_rounded, t) = self.saved_tensors
        mask_clip = (inp_scaled < -0.5 + self.lowerbound) + (
            inp_scaled > self.upperbound + 0.5
        )  # mask for accumulating the gradients of |data_scaled|>L
        mask_quant = np.abs(
            mask_clip - 1
        )  # mask for accumulating the gradients with |data_scaled|<=L
        grad_quant = (
            grad_inp_flq * mask_quant * (inp_rounded - inp_scaled)
        )  # gradient within |data_scaled|<=L
        grad_clip = (
            grad_inp_flq * mask_clip * inp_rounded
        )  # gradient with   | data_scaled|>L
        grad_s = grad_clip.sum() + grad_quant.sum()
        # dL/ds = dL/dt * t * ln(2)
        grad_s = grad_s * t * np.log(2)
        grad_inp = grad_inp_flq * mask_quant
        return grad_inp, grad_s


def test_tqt():

    g = []

    def cb(grad):
        g.append(grad)

    x = np.random.randint(-128, 128, size=(1, 2, 3, 4)).astype("float32")
    s = np.random.rand(1) - 1
    g_y = np.ones(shape=(1, 2, 3, 4), dtype="float32")

    n = TQT_numpy(-127, 127)
    y_np = n.forward(x, s)
    g_x_np, g_s_np = n.backward(g_y)

    x = mge.tensor(x, dtype="float32")
    s = mge.tensor(s, dtype="float32")
    g_y = mge.tensor(g_y, dtype="float32")
    with Grad() as grad:
        grad.wrt(x, s, callback=cb)
        y = tqt_forward(-127, 127, x, s)
        grad(y, g_y)
    g_x, g_s = g

    np.testing.assert_allclose(y.numpy(), y_np, rtol=1e-5, atol=1e-5)
    np.testing.assert_allclose(g_x.numpy(), g_x_np, rtol=1e-5, atol=1e-5)
    np.testing.assert_allclose(g_s.numpy(), g_s_np, rtol=5e-5, atol=5e-5)


def _save_to(self, name="grad"):
    def callback(grad):
        setattr(self, name, grad)

    return callback


class Round(Function):
    def forward(self, x):
        return F.round(x)

    def backward(self, output_grads):
        return output_grads


def fake_quant_tensor_gt(inp, scale, zero_point, qmin, qmax):
    oup = Round()(inp / scale) + zero_point
    oup = F.minimum(F.maximum(oup, qmin), qmax)
    oup = (oup - zero_point) * scale
    return oup


def test_fakequant():
    qmin = -126
    qmax = 129
    test_dtype = QuantDtypeMeta("test_qint8", None, "int8", qmin, qmax)

    def run(zero_point, scale):
        qparams = create_qparams(QuantMode.ASYMMERTIC, test_dtype, scale, zero_point)
        inp_data = np.random.uniform(low=-512.0, high=512.0, size=(1, 32, 32, 32))
        inp = tensor(inp_data, dtype=np.float32)
        # test forward
        oup = fake_quant_tensor(inp, qparams).numpy()
        oup_gt = fake_quant_tensor_gt(inp, scale, zero_point, qmin, qmax).numpy()
        assert np.allclose(oup, oup_gt)
        assert oup.shape == oup_gt.shape

        # test backward
        x = tensor(inp_data, dtype=np.float32)
        with Grad() as grad:
            grad.wrt(x, callback=_save_to(x))
            y = fake_quant_tensor(x, qparams)
            grad(y, tensor(F.ones_like(x)))

        x1 = tensor(inp_data, dtype=np.float32)
        with Grad() as grad:
            grad.wrt(x1, callback=_save_to(x1))
            y1 = fake_quant_tensor_gt(x1, scale, zero_point, qmin, qmax)
            grad(y1, tensor(F.ones_like(x1)))

        assert np.allclose(x.grad.numpy(), x1.grad.numpy())
        assert make_shape_tuple(x.grad.shape) == make_shape_tuple(x1.grad.shape)

        # test nan
        x = F.full((1, 32, 3, 3), np.nan)
        y = fake_quant_tensor(x, qparams).numpy()
        assert np.isnan(y).all()

    zero_point = tensor([1.0], dtype=np.float32)
    scale = tensor([4.0], dtype=np.float32)
    run(zero_point, scale)

    zero_point = tensor(1.0 * np.ones((1, 32, 1, 1)), dtype=np.float32)
    scale = tensor(4.0 * np.ones((1, 32, 1, 1)), dtype=np.float32)
    run(zero_point, scale)


class LSQ_numpy:
    def __init__(self, lowerbound, upperbound):
        super().__init__()
        self.lowerbound = lowerbound
        self.upperbound = upperbound

    def forward(self, inp, scale, zero_point, grad_scale):
        inp_scaled = inp / scale + zero_point
        inp_clipped = np.maximum(
            np.minimum(inp_scaled, self.upperbound), self.lowerbound
        )
        inp_rounded = np.floor(inp_clipped + 0.5)
        inp_flq = (inp_rounded - zero_point) * scale
        self.saved_tensors = (inp_scaled, inp_rounded, scale, grad_scale)
        return inp_flq

    def backward(self, grad_inp_flq):
        (inp_scaled, inp_rounded, scale, grad_scale) = self.saved_tensors

        ind_small = inp_scaled < self.lowerbound
        ind_big = inp_scaled > self.upperbound
        ind_middle = np.logical_xor(ind_small, ind_big)
        ind_middle = np.abs(ind_middle - 1)

        grad_s = (
            ind_small * self.lowerbound
            + ind_big * self.upperbound
            + ind_middle * (-inp_scaled + inp_rounded)
        )
        grad_s = grad_s * grad_scale * grad_inp_flq
        grad_s = grad_s.sum()
        grad_inp = grad_inp_flq * ind_middle

        return grad_inp, grad_s


def test_lsq():
    g = []

    def cb(grad):
        g.append(grad)

    # FIXME: use random number when LSQ is fixed
    # x = np.random.randint(-128, 128, size=(1, 2, 3, 4)).astype("float32")
    # s = np.random.rand(1)
    x = np.array(
        [
            [
                [
                    [4.0, 38.0, -121.0, 38.0],
                    [15.0, -115.0, -112.0, 24.0],
                    [23.0, -65.0, 109.0, -115.0],
                ],
                [
                    [-66.0, -90.0, -45.0, -101.0],
                    [68.0, -98.0, 108.0, -79.0],
                    [54.0, 63.0, -10.0, -50.0],
                ],
            ]
        ],
        dtype="float32",
    )
    s = np.array([0.02918224], dtype="float32")
    eps = np.array([1e-5], dtype="float32")
    s = np.abs(s) if np.abs(s) > eps else eps
    zero_point = np.array([1.0], dtype="float32")
    grad_s = np.array([2.0], dtype="float32")

    g_y = np.ones(shape=(1, 2, 3, 4), dtype="float32")

    n = LSQ_numpy(-127, 127)
    y_np = n.forward(x, s, zero_point, grad_s)
    g_x_np, g_s_np = n.backward(g_y)

    x = mge.tensor(x, dtype="float32")
    s = mge.tensor(s, dtype="float32")
    zero_point = mge.tensor(zero_point, dtype="float32")
    grad_s = mge.tensor(grad_s, dtype="float32")

    g_y = mge.tensor(g_y, dtype="float32")
    with Grad() as grad:
        grad.wrt(x, s, callback=cb)
        y = lsq_forward(-127, 127, x, s, zero_point, grad_s)
        grad(y, g_y)
    g_x, g_s = g

    np.testing.assert_allclose(y.numpy(), y_np, rtol=1e-7, atol=1e-7)
    np.testing.assert_allclose(g_x.numpy(), g_x_np, rtol=1e-7, atol=1e-7)
    np.testing.assert_allclose(g_s.numpy(), g_s_np, rtol=5e-7, atol=5e-7)