hparams_config.py

# Copyright (c) 2022 Graphcore Ltd. All rights reserved.
# Copyright 2020 Google Research. All Rights Reserved.
#
# 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.
# ==============================================================================
"""Hparams for model architecture and trainer."""
import ast
import collections
import copy
from typing import Any, Dict, Text
import six
import tensorflow as tf
import yaml


def eval_str_fn(val):
    if val in {"true", "false"}:
        return val == "true"
    try:
        return ast.literal_eval(val)
    except (ValueError, SyntaxError):
        return val


# pylint: disable=protected-access
class Config(object):
    """A config utility class."""

    def __init__(self, config_dict=None):
        self.update(config_dict)

    def __setattr__(self, k, v):
        self.__dict__[k] = Config(v) if isinstance(v, dict) else copy.deepcopy(v)

    def __getattr__(self, k):
        return self.__dict__[k]

    def __getitem__(self, k):
        return self.__dict__[k]

    def __repr__(self):
        return repr(self.as_dict())

    def __deepcopy__(self, memodict):
        return type(self)(self.as_dict())

    def __str__(self):
        try:
            return yaml.dump(self.as_dict(), indent=4)
        except TypeError:
            return str(self.as_dict())

    def _update(self, config_dict, allow_new_keys=True):
        """Recursively update internal members."""
        if not config_dict:
            return

        for k, v in six.iteritems(config_dict):
            if k not in self.__dict__:
                if allow_new_keys:
                    self.__setattr__(k, v)
                else:
                    raise KeyError("Key `{}` does not exist for overriding. ".format(k))
            else:
                if isinstance(self.__dict__[k], Config) and isinstance(v, dict):
                    self.__dict__[k]._update(v, allow_new_keys)
                elif isinstance(self.__dict__[k], Config) and isinstance(v, Config):
                    self.__dict__[k]._update(v.as_dict(), allow_new_keys)
                else:
                    self.__setattr__(k, v)

    def get(self, k, default_value=None):
        return self.__dict__.get(k, default_value)

    def update(self, config_dict):
        """Update members while allowing new keys."""
        self._update(config_dict, allow_new_keys=True)

    def keys(self):
        return self.__dict__.keys()

    def override(self, config_dict_or_str, allow_new_keys=False):
        """Update members while disallowing new keys."""
        if isinstance(config_dict_or_str, str):
            if not config_dict_or_str:
                return
            elif "=" in config_dict_or_str:
                config_dict = self.parse_from_str(config_dict_or_str)
            elif config_dict_or_str.endswith(".yaml"):
                config_dict = self.parse_from_yaml(config_dict_or_str)
            else:
                raise ValueError('Invalid string {}, must end with .yaml or contains "=".'.format(config_dict_or_str))
        elif isinstance(config_dict_or_str, dict):
            config_dict = config_dict_or_str
        else:
            raise ValueError("Unknown value type: {}".format(config_dict_or_str))

        self._update(config_dict, allow_new_keys)

    def parse_from_yaml(self, yaml_file_path: Text) -> Dict[Any, Any]:
        """Parses a yaml file and returns a dictionary."""
        with tf.io.gfile.GFile(yaml_file_path, "r") as f:
            config_dict = yaml.load(f, Loader=yaml.FullLoader)
            return config_dict

    def save_to_yaml(self, yaml_file_path):
        """Write a dictionary into a yaml file."""
        with tf.io.gfile.GFile(yaml_file_path, "w") as f:
            yaml.dump(self.as_dict(), f, default_flow_style=False)

    def parse_from_str(self, config_str: Text) -> Dict[Any, Any]:
        """Parse a string like 'x.y=1,x.z=2' to nested dict {x: {y: 1, z: 2}}."""
        if not config_str:
            return {}
        config_dict = {}
        try:
            for kv_pair in config_str.split(","):
                if not kv_pair:  # skip empty string
                    continue
                key_str, value_str = kv_pair.split("=")
                key_str = key_str.strip()

                def add_kv_recursive(k, v):
                    """Recursively parse x.y.z=tt to {x: {y: {z: tt}}}."""
                    if "." not in k:
                        if "*" in v:
                            # we reserve * to split arrays.
                            return {k: [eval_str_fn(vv) for vv in v.split("*")]}
                        return {k: eval_str_fn(v)}
                    pos = k.index(".")
                    return {k[:pos]: add_kv_recursive(k[pos + 1 :], v)}

                def merge_dict_recursive(target, src):
                    """Recursively merge two nested dictionary."""
                    for k in src.keys():
                        if k in target and isinstance(target[k], dict) and isinstance(src[k], collections.Mapping):
                            merge_dict_recursive(target[k], src[k])
                        else:
                            target[k] = src[k]

                merge_dict_recursive(config_dict, add_kv_recursive(key_str, value_str))
            return config_dict
        except ValueError:
            raise ValueError("Invalid config_str: {}".format(config_str))

    def as_dict(self):
        """Returns a dict representation."""
        config_dict = {}
        for k, v in six.iteritems(self.__dict__):
            if isinstance(v, Config):
                config_dict[k] = v.as_dict()
            else:
                config_dict[k] = copy.deepcopy(v)
        return config_dict
        # pylint: enable=protected-access


def default_detection_configs():
    """Returns a default detection configs."""
    h = Config()

    # model name.
    h.name = "efficientdet-d1"

    # activation type: see activation_fn in utils.py.
    h.act_type = "ipu_swish"

    # input preprocessing parameters
    h.image_size = 640  # An integer or a string WxH such as 640x320.
    h.target_size = None
    h.input_rand_hflip = True
    h.jitter_min = 0.1
    h.jitter_max = 2.0
    h.autoaugment_policy = None
    h.grid_mask = False
    h.sample_image = None
    h.map_freq = 5  # AP eval frequency in epochs.

    # dataset specific parameters
    # TODO(tanmingxing): update this to be 91 for COCO, and 21 for pascal.
    h.num_classes = 90  # 1+ actual classes, 0 is reserved for background.
    h.seg_num_classes = 3  # segmentation classes
    h.heads = ["object_detection"]  # 'object_detection', 'segmentation'

    h.skip_crowd_during_training = True
    h.label_map = None  # a dict or a string of 'coco', 'voc', 'waymo'.
    h.max_instances_per_image = 100  # Default to 100 for COCO.
    h.regenerate_source_id = False

    # model architecture
    h.min_level = 3
    h.max_level = 7
    h.num_scales = 3
    # ratio w/h: 2.0 means w=1.4, h=0.7. Can be computed with k-mean per dataset.
    h.aspect_ratios = [1.0, 2.0, 0.5]  # [[0.7, 1.4], [1.0, 1.0], [1.4, 0.7]]
    h.anchor_scale = 4.0
    # is batchnorm training mode
    h.is_training_bn = True
    # optimization
    h.momentum = 0.9
    h.optimizer = "sgd"  # can be 'adam' or 'sgd'.
    h.learning_rate = 0.08  # 0.008 for adam.
    h.lr_warmup_init = 0.008  # 0.0008 for adam.
    h.lr_warmup_epoch = 1.0
    h.first_lr_drop_epoch = 200.0
    h.second_lr_drop_epoch = 250.0
    h.poly_lr_power = 0.9
    h.clip_gradients_norm = 10.0
    h.num_epochs = 300
    h.data_format = "channels_last"
    # The default image normalization is identical to Cloud TPU ResNet.
    h.mean_rgb = [0.485 * 255, 0.456 * 255, 0.406 * 255]
    h.stddev_rgb = [0.229 * 255, 0.224 * 255, 0.225 * 255]
    h.scale_range = False

    # classification loss
    h.label_smoothing = 0.0  # 0.1 is a good default
    # Behold the focal loss parameters
    h.alpha = 0.25
    h.gamma = 1.5

    # localization loss
    h.delta = 0.1  # regularization parameter of huber loss.
    # total loss = box_loss * box_loss_weight + iou_loss * iou_loss_weight
    h.box_loss_weight = 50.0
    h.iou_loss_type = None
    h.iou_loss_weight = 1.0

    # regularization l2 loss.
    h.weight_decay = 4e-5
    h.strategy = None  # 'tpu', 'gpus', 'ipu', None
    h.mixed_precision = False  # If False, use float32.
    h.loss_scale = None  # set to 2**16 enables dynamic loss scale
    h.model_optimizations = {}  # 'prune':{}

    # For detection.
    h.box_class_repeats = 3
    h.fpn_cell_repeats = 3
    h.fpn_num_filters = 88
    h.separable_conv = True
    h.apply_bn_for_resampling = True
    h.conv_after_downsample = False
    h.conv_bn_act_pattern = False
    h.drop_remainder = True  # drop remainder for the final batch eval.

    # For post-processing nms, must be a dict.
    h.nms_configs = {
        "method": "gaussian",
        "iou_thresh": None,  # use the default value based on method.
        "score_thresh": 0.0,
        "sigma": None,
        "pyfunc": False,
        "max_nms_inputs": 0,
        "max_output_size": 100,
    }
    h.tflite_max_detections = 100

    # version.
    h.fpn_name = None
    h.fpn_weight_method = None
    h.fpn_config = None

    # No stochastic depth in default.
    h.survival_prob = None
    h.img_summary_steps = None

    h.lr_decay_method = "cosine"
    h.moving_average_decay = 0.9998
    h.ckpt_var_scope = None  # ckpt variable scope.
    # If true, skip loading pretrained weights if shape mismatches.
    h.skip_mismatch = True

    h.backbone_name = "efficientnet-b1"
    h.backbone_config = None
    h.var_freeze_expr = None

    # A temporary flag to switch between legacy and keras models.
    h.use_keras_model = True
    h.dataset_type = None
    h.positives_momentum = None
    h.grad_checkpoint = False

    # Parameters for the Checkpoint Callback.
    h.verbose = 1
    h.save_freq = "epoch"

    return h


efficientdet_model_param_dict = {
    "efficientdet-d0": dict(
        name="efficientdet-d0",
        backbone_name="efficientnet-b0",
        image_size=512,
        fpn_num_filters=64,
        fpn_cell_repeats=3,
        box_class_repeats=3,
    ),
    "efficientdet-d1": dict(
        name="efficientdet-d1",
        backbone_name="efficientnet-b1",
        image_size=640,
        fpn_num_filters=88,
        fpn_cell_repeats=4,
        box_class_repeats=3,
    ),
    "efficientdet-d2": dict(
        name="efficientdet-d2",
        backbone_name="efficientnet-b2",
        image_size=768,
        fpn_num_filters=112,
        fpn_cell_repeats=5,
        box_class_repeats=3,
    ),
    "efficientdet-d3": dict(
        name="efficientdet-d3",
        backbone_name="efficientnet-b3",
        image_size=896,
        fpn_num_filters=160,
        fpn_cell_repeats=6,
        box_class_repeats=4,
    ),
    "efficientdet-d4": dict(
        name="efficientdet-d4",
        backbone_name="efficientnet-b4",
        image_size=1024,
        fpn_num_filters=224,
        fpn_cell_repeats=7,
        box_class_repeats=4,
    ),
    "efficientdet-d5": dict(
        name="efficientdet-d5",
        backbone_name="efficientnet-b5",
        image_size=1280,
        fpn_num_filters=288,
        fpn_cell_repeats=7,
        box_class_repeats=4,
    ),
    "efficientdet-d6": dict(
        name="efficientdet-d6",
        backbone_name="efficientnet-b6",
        image_size=1280,
        fpn_num_filters=384,
        fpn_cell_repeats=8,
        box_class_repeats=5,
        fpn_weight_method="sum",  # Use unweighted sum for stability.
    ),
    "efficientdet-d7": dict(
        name="efficientdet-d7",
        backbone_name="efficientnet-b6",
        image_size=1536,
        fpn_num_filters=384,
        fpn_cell_repeats=8,
        box_class_repeats=5,
        anchor_scale=5.0,
        fpn_weight_method="sum",  # Use unweighted sum for stability.
    ),
    "efficientdet-d7x": dict(
        name="efficientdet-d7x",
        backbone_name="efficientnet-b7",
        image_size=1536,
        fpn_num_filters=384,
        fpn_cell_repeats=8,
        box_class_repeats=5,
        anchor_scale=4.0,
        max_level=8,
        fpn_weight_method="sum",  # Use unweighted sum for stability.
    ),
}


lite_common_param = dict(
    mean_rgb=127.0,
    stddev_rgb=128.0,
    act_type="relu6",
    fpn_weight_method="sum",
)

efficientdet_lite_param_dict = {
    # lite models are in progress and subject to changes.
    # mean_rgb and stddev_rgb are consistent with EfficientNet-Lite models in
    # https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/lite/efficientnet_lite_builder.py#L28
    "efficientdet-lite0": dict(
        name="efficientdet-lite0",
        backbone_name="efficientnet-lite0",
        image_size=320,
        fpn_num_filters=64,
        fpn_cell_repeats=3,
        box_class_repeats=3,
        anchor_scale=3.0,
        **lite_common_param,
    ),
    "efficientdet-lite1": dict(
        name="efficientdet-lite1",
        backbone_name="efficientnet-lite1",
        image_size=384,
        fpn_num_filters=88,
        fpn_cell_repeats=4,
        box_class_repeats=3,
        anchor_scale=3.0,
        **lite_common_param,
    ),
    "efficientdet-lite2": dict(
        name="efficientdet-lite2",
        backbone_name="efficientnet-lite2",
        image_size=448,
        fpn_num_filters=112,
        fpn_cell_repeats=5,
        box_class_repeats=3,
        anchor_scale=3.0,
        **lite_common_param,
    ),
    "efficientdet-lite3": dict(
        name="efficientdet-lite3",
        backbone_name="efficientnet-lite3",
        image_size=512,
        fpn_num_filters=160,
        fpn_cell_repeats=6,
        box_class_repeats=4,
        **lite_common_param,
    ),
    "efficientdet-lite3x": dict(
        name="efficientdet-lite3x",
        backbone_name="efficientnet-lite3",
        image_size=640,
        fpn_num_filters=200,
        fpn_cell_repeats=6,
        box_class_repeats=4,
        anchor_scale=3.0,
        **lite_common_param,
    ),
    "efficientdet-lite4": dict(
        name="efficientdet-lite4",
        backbone_name="efficientnet-lite4",
        image_size=640,
        fpn_num_filters=224,
        fpn_cell_repeats=7,
        box_class_repeats=4,
        **lite_common_param,
    ),
}


def get_efficientdet_config(model_name="efficientdet-d1"):
    """Get the default config for EfficientDet based on model name."""
    h = default_detection_configs()
    if model_name in efficientdet_model_param_dict:
        h.override(efficientdet_model_param_dict[model_name])
    elif model_name in efficientdet_lite_param_dict:
        h.override(efficientdet_lite_param_dict[model_name])
    else:
        raise ValueError("Unknown model name: {}".format(model_name))

    return h


def get_detection_config(model_name):
    if model_name.startswith("efficientdet"):
        return get_efficientdet_config(model_name)
    else:
        raise ValueError("model name must start with efficientdet.")