MegEngine/imperative/python/megengine/utils/module_stats.py

# 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.
from collections import namedtuple
from functools import partial

import numpy as np
import tabulate

import megengine as mge
import megengine.module as m
import megengine.module.qat as qatm
import megengine.module.quantized as qm
from megengine import Tensor
from megengine import functional as F
from megengine.core.tensor.dtype import get_dtype_bit
from megengine.functional.tensor import zeros

from .module_utils import set_module_mode_safe

try:
    mge.logger.MegEngineLogFormatter.max_lines = float("inf")
except AttributeError as e:
    raise ValueError("set logger max lines failed")

logger = mge.get_logger(__name__)
logger.setLevel("INFO")


_calc_flops_dict = {}
_calc_receptive_field_dict = {}


def _receptive_field_fallback(module, inputs, outputs):
    if not _receptive_field_enabled:
        return
    assert not hasattr(module, "_rf")
    assert not hasattr(module, "_stride")
    if len(inputs) == 0:
        # TODO: support other dimension
        module._rf = (1, 1)
        module._stride = (1, 1)
        return module._rf, module._stride
    rf, stride = preprocess_receptive_field(module, inputs, outputs)
    module._rf = rf
    module._stride = stride
    return rf, stride


# key tuple, impl_dict, fallback
_iter_list = [
    ("flops_num", _calc_flops_dict, None),
    (
        ("receptive_field", "stride"),
        _calc_receptive_field_dict,
        _receptive_field_fallback,
    ),
]

_receptive_field_enabled = False


def _register_dict(*modules, dict=None):
    def callback(impl):
        for module in modules:
            dict[module] = impl
        return impl

    return callback


def register_flops(*modules):
    return _register_dict(*modules, dict=_calc_flops_dict)


def register_receptive_field(*modules):
    return _register_dict(*modules, dict=_calc_receptive_field_dict)


def enable_receptive_field():
    global _receptive_field_enabled
    _receptive_field_enabled = True


def disable_receptive_field():
    global _receptive_field_enabled
    _receptive_field_enabled = False


@register_flops(
    m.Conv1d, m.Conv2d, m.Conv3d, m.ConvTranspose2d, m.LocalConv2d, m.DeformableConv2d
)
def flops_convNd(module: m.Conv2d, inputs, outputs):
    bias = 1 if module.bias is not None else 0
    # N x Cout x H x W x  (Cin x Kw x Kh + bias)
    return np.prod(outputs[0].shape) * (
        module.in_channels // module.groups * np.prod(module.kernel_size) + bias
    )


@register_flops(
    m.batchnorm._BatchNorm, m.SyncBatchNorm, m.GroupNorm, m.LayerNorm, m.InstanceNorm,
)
def flops_norm(module: m.Linear, inputs, outputs):
    return np.prod(inputs[0].shape) * 7


@register_flops(m.AvgPool2d, m.MaxPool2d)
def flops_pool(module: m.AvgPool2d, inputs, outputs):
    return np.prod(outputs[0].shape) * (module.kernel_size ** 2)


@register_flops(m.AdaptiveAvgPool2d, m.AdaptiveMaxPool2d)
def flops_adaptivePool(module: m.AdaptiveAvgPool2d, inputs, outputs):
    stride_h = np.floor(inputs[0].shape[2] / (inputs[0].shape[2] - 1))
    kernel_h = inputs[0].shape[2] - (inputs[0].shape[2] - 1) * stride_h
    stride_w = np.floor(inputs[0].shape[3] / (inputs[0].shape[3] - 1))
    kernel_w = inputs[0].shape[3] - (inputs[0].shape[3] - 1) * stride_w
    return np.prod(outputs[0].shape) * kernel_h * kernel_w


@register_flops(m.Linear)
def flops_linear(module: m.Linear, inputs, outputs):
    bias = module.out_features if module.bias is not None else 0
    return np.prod(outputs[0].shape) * module.in_features + bias


@register_flops(m.BatchMatMulActivation)
def flops_batchmatmul(module: m.BatchMatMulActivation, inputs, outputs):
    bias = 1 if module.bias is not None else 0
    x = inputs[0]
    w = module.weight
    batch_size = x.shape[0]
    n, p = x.shape[1:]
    _, m = w.shape[1:]
    return n * (p + bias) * m * batch_size


# does not need import qat and quantized module since they inherit from float module.
hook_modules = (
    m.conv._ConvNd,
    m.Linear,
    m.BatchMatMulActivation,
    m.batchnorm._BatchNorm,
    m.LayerNorm,
    m.GroupNorm,
    m.InstanceNorm,
    m.pooling._PoolNd,
    m.adaptive_pooling._AdaptivePoolNd,
)


def _mean(inp):
    inp = mge.tensor(inp)
    return F.mean(inp).numpy()


def _std(inp):
    inp = mge.tensor(inp)
    return F.std(inp).numpy()


def dict2table(list_of_dict, header):
    table_data = [header]
    for d in list_of_dict:
        row = []
        for h in header:
            v = ""
            if h in d:
                v = d[h]
            row.append(v)
        table_data.append(row)
    return table_data


def sizeof_fmt(num, suffix="B"):
    if suffix == "B":
        scale = 1024.0
        units = ["", "Ki", "Mi", "Gi", "Ti", "Pi", "Ei", "Zi", "Yi"]
    else:
        scale = 1000.0
        units = ["", "K", "M", "G", "T", "P", "E", "Z", "Y"]
    for unit in units:
        if abs(num) < scale or unit == units[-1]:
            return "{:3.3f} {}{}".format(num, unit, suffix)
        num /= scale


def preprocess_receptive_field(module, inputs, outputs):
    # TODO: support other dimensions
    pre_rf = (
        max(getattr(i.owner, "_rf", (1, 1))[0] for i in inputs),
        max(getattr(i.owner, "_rf", (1, 1))[1] for i in inputs),
    )
    pre_stride = (
        max(getattr(i.owner, "_stride", (1, 1))[0] for i in inputs),
        max(getattr(i.owner, "_stride", (1, 1))[1] for i in inputs),
    )
    return pre_rf, pre_stride


def get_op_stats(module, inputs, outputs):
    if not isinstance(outputs, tuple) and not isinstance(outputs, list):
        outputs = (outputs,)
    rst = {
        "input_shapes": [i.shape for i in inputs],
        "output_shapes": [o.shape for o in outputs],
    }
    valid_flag = False
    for key, _dict, fallback in _iter_list:
        for _type in _dict:
            if isinstance(module, _type):
                value = _dict[_type](module, inputs, outputs)
                valid_flag = True
                break
        else:
            if fallback is not None:
                value = fallback(module, inputs, outputs)
            continue

        if isinstance(key, tuple):
            assert isinstance(value, tuple)
            for k, v in zip(key, value):
                rst[k] = v
        else:
            rst[key] = value

    if valid_flag:
        return rst
    else:
        return None
    return


def sum_op_stats(flops, bar_length_max=20):
    max_flops_num = max([i["flops_num"] for i in flops] + [0])
    total_flops_num = 0
    for d in flops:
        total_flops_num += int(d["flops_num"])
        d["flops_cum"] = sizeof_fmt(total_flops_num, suffix="OPs")

    for d in flops:
        ratio = d["ratio"] = d["flops_num"] / total_flops_num
        d["percentage"] = "{:.2f}%".format(ratio * 100)
        bar_length = int(d["flops_num"] / max_flops_num * bar_length_max)
        d["bar"] = "#" * bar_length
        d["flops"] = sizeof_fmt(d["flops_num"], suffix="OPs")

    total_flops_str = sizeof_fmt(total_flops_num, suffix="OPs")
    total_var_size = sum(
        sum(s[1] if len(s) > 1 else 0 for s in d["output_shapes"]) for d in flops
    )
    flops.append(
        dict(name="total", flops=total_flops_str, output_shapes=total_var_size)
    )

    return total_flops_num, flops


def print_op_stats(flops):
    header = [
        "name",
        "class_name",
        "input_shapes",
        "output_shapes",
        "flops",
        "flops_cum",
        "percentage",
        "bar",
    ]
    if _receptive_field_enabled:
        header.insert(4, "receptive_field")
        header.insert(5, "stride")
    logger.info("flops stats: \n" + tabulate.tabulate(dict2table(flops, header=header)))


def get_param_stats(param: Tensor):
    nbits = get_dtype_bit(np.dtype(param.dtype).name)
    shape = param.shape
    param_dim = np.prod(param.shape)
    param_size = param_dim * nbits // 8
    return {
        "dtype": np.dtype(param.dtype),
        "shape": shape,
        "mean": "{:.3g}".format(_mean(param)),
        "std": "{:.3g}".format(_std(param)),
        "param_dim": param_dim,
        "nbits": nbits,
        "size": param_size,
    }


def sum_param_stats(params, bar_length_max=20):
    max_size = max([d["size"] for d in params] + [0])
    total_param_dims, total_param_size = 0, 0
    for d in params:
        total_param_dims += int(d["param_dim"])
        total_param_size += int(d["size"])
        d["size_cum"] = sizeof_fmt(total_param_size)

    for d in params:
        ratio = d["size"] / total_param_size
        d["ratio"] = ratio
        d["percentage"] = "{:.2f}%".format(ratio * 100)
        bar_length = int(d["size"] / max_size * bar_length_max)
        d["size_bar"] = "#" * bar_length
        d["size"] = sizeof_fmt(d["size"])

    param_size = sizeof_fmt(total_param_size)
    params.append(dict(name="total", param_dim=total_param_dims, size=param_size,))

    return total_param_dims, total_param_size, params


def print_param_stats(params):
    header = [
        "name",
        "dtype",
        "shape",
        "mean",
        "std",
        "param_dim",
        "nbits",
        "size",
        "size_cum",
        "percentage",
        "size_bar",
    ]
    logger.info(
        "param stats: \n" + tabulate.tabulate(dict2table(params, header=header))
    )


def get_activation_stats(output: Tensor):
    out_shape = output.shape
    activations_dtype = np.dtype(output.dtype)
    nbits = get_dtype_bit(activations_dtype.name)
    act_dim = np.prod(out_shape)
    act_size = act_dim * nbits // 8
    return {
        "dtype": activations_dtype,
        "shape": out_shape,
        "act_dim": act_dim,
        "mean": "{:.3g}".format(_mean(output)),
        "std": "{:.3g}".format(_std(output)),
        "nbits": nbits,
        "size": act_size,
    }


def sum_activations_stats(activations, bar_length_max=20):
    max_act_size = max([i["size"] for i in activations] + [0])
    total_act_dims, total_act_size = 0, 0
    for d in activations:
        total_act_size += int(d["size"])
        total_act_dims += int(d["act_dim"])
        d["size_cum"] = sizeof_fmt(total_act_size)

    for d in activations:
        ratio = d["ratio"] = d["size"] / total_act_size
        d["percentage"] = "{:.2f}%".format(ratio * 100)
        bar_length = int(d["size"] / max_act_size * bar_length_max)
        d["size_bar"] = "#" * bar_length
        d["size"] = sizeof_fmt(d["size"])

    act_size = sizeof_fmt(total_act_size)
    activations.append(dict(name="total", act_dim=total_act_dims, size=act_size,))

    return total_act_dims, total_act_size, activations


def print_activations_stats(activations):
    header = [
        "name",
        "class_name",
        "dtype",
        "shape",
        "mean",
        "std",
        "nbits",
        "act_dim",
        "size",
        "size_cum",
        "percentage",
        "size_bar",
    ]
    logger.info(
        "activations stats: \n"
        + tabulate.tabulate(dict2table(activations, header=header))
    )


def print_summary(**kwargs):
    data = [["item", "value"]]
    data.extend(list(kwargs.items()))
    logger.info("summary\n" + tabulate.tabulate(data))


def module_stats(
    model: m.Module,
    input_shapes: list,
    bar_length_max: int = 20,
    log_params: bool = True,
    log_flops: bool = True,
    log_activations: bool = True,
):
    r"""
    Calculate and print ``model``'s statistics by adding hook and record Module's inputs outputs size.

    :param model: model that need to get stats info.
    :param input_shapes: shapes of inputs for running model and calculating stats.
    :param bar_length_max: size of bar indicating max flops or parameter size in net stats.
    :param log_params: whether print and record params size.
    :param log_flops: whether print and record op flops.
    :param log_activations: whether print and record op activations.
    """
    disable_receptive_field()

    def module_stats_hook(module, inputs, outputs, name=""):
        class_name = str(module.__class__).split(".")[-1].split("'")[0]
        flops_stats = get_op_stats(module, inputs, outputs)
        if flops_stats is not None:
            flops_stats["name"] = name
            flops_stats["class_name"] = class_name
            flops.append(flops_stats)

        if hasattr(module, "weight") and module.weight is not None:
            w = module.weight
            param_stats = get_param_stats(w)
            param_stats["name"] = name + "-w"
            params.append(param_stats)

        if hasattr(module, "bias") and module.bias is not None:
            b = module.bias
            param_stats = get_param_stats(b)
            param_stats["name"] = name + "-b"
            params.append(param_stats)

        if not isinstance(outputs, tuple) or not isinstance(outputs, list):
            output = outputs
        else:
            output = outputs[0]
        activation_stats = get_activation_stats(output)
        activation_stats["name"] = name
        activation_stats["class_name"] = class_name
        activations.append(activation_stats)

    # multiple inputs to the network
    if not isinstance(input_shapes[0], tuple):
        input_shapes = [input_shapes]

    params = []
    flops = []
    hooks = []
    activations = []
    total_stats = namedtuple("total_stats", ["param_size", "flops", "act_size"])
    stats_details = namedtuple("module_stats", ["params", "flops", "activations"])

    for (name, module) in model.named_modules():
        if isinstance(module, hook_modules):
            hooks.append(
                module.register_forward_hook(partial(module_stats_hook, name=name))
            )

    inputs = [zeros(in_size, dtype=np.float32) for in_size in input_shapes]
    with set_module_mode_safe(model, training=False) as model:
        model(*inputs)

    for h in hooks:
        h.remove()

    extra_info = {
        "#params": len(params),
    }
    (
        total_flops,
        total_param_dims,
        total_param_size,
        total_act_dims,
        total_param_size,
    ) = (0, 0, 0, 0, 0)

    total_param_dims, total_param_size, params = sum_param_stats(params, bar_length_max)
    extra_info["total_param_dims"] = sizeof_fmt(total_param_dims, suffix="")
    extra_info["total_param_size"] = sizeof_fmt(total_param_size)
    if log_params:
        print_param_stats(params)

    total_flops, flops = sum_op_stats(flops, bar_length_max)
    extra_info["total_flops"] = sizeof_fmt(total_flops, suffix="OPs")
    if log_flops:
        print_op_stats(flops)

    total_act_dims, total_act_size, activations = sum_activations_stats(
        activations, bar_length_max
    )
    extra_info["total_act_dims"] = sizeof_fmt(total_act_dims, suffix="")
    extra_info["total_act_size"] = sizeof_fmt(total_act_size)
    if log_activations:
        print_activations_stats(activations)

    extra_info["flops/param_size"] = "{:3.3f}".format(total_flops / total_param_size)

    print_summary(**extra_info)

    return (
        total_stats(
            param_size=total_param_size, flops=total_flops, act_size=total_act_size,
        ),
        stats_details(params=params, flops=flops, activations=activations),
    )