utility.py

from torch.utils.data.sampler import Sampler
from collections import defaultdict
import copy
import random
import numpy as np
import torch
import logging
import math
import torch
from typing import Dict, Any
class RandomIdentitySampler(Sampler):
    """
    Randomly sample N identities, then for each identity,
    randomly sample K instances, therefore batch size is N*K.
    Args:
    - data_source (list): list of (img_path, pid, camid).
    - num_instances (int): number of instances per identity in a batch.
    - batch_size (int): number of examples in a batch.
    """

    def __init__(self, data_source, batch_size, num_instances):
        self.data_source = data_source
        self.batch_size = batch_size
        self.num_instances = num_instances
        self.num_pids_per_batch = self.batch_size // self.num_instances
        self.index_dic = defaultdict(list) #dict with list value
        #{783: [0, 5, 116, 876, 1554, 2041],...,}
        for index, (_, pid, _) in enumerate(self.data_source):
            self.index_dic[pid].append(index)
        self.pids = list(self.index_dic.keys())

        # estimate number of examples in an epoch
        self.length = 0
        for pid in self.pids:
            idxs = self.index_dic[pid]
            num = len(idxs)
            if num < self.num_instances:
                num = self.num_instances
            self.length += num - num % self.num_instances

    def __iter__(self):
        batch_idxs_dict = defaultdict(list)

        for pid in self.pids:
            idxs = copy.deepcopy(self.index_dic[pid])
            if len(idxs) < self.num_instances:
                idxs = np.random.choice(idxs, size=self.num_instances, replace=True)
            random.shuffle(idxs)
            batch_idxs = []
            for idx in idxs:
                batch_idxs.append(idx)
                if len(batch_idxs) == self.num_instances:
                    batch_idxs_dict[pid].append(batch_idxs)
                    batch_idxs = []

        avai_pids = copy.deepcopy(self.pids)
        final_idxs = []

        while len(avai_pids) >= self.num_pids_per_batch:
            selected_pids = random.sample(avai_pids, self.num_pids_per_batch)
            for pid in selected_pids:
                batch_idxs = batch_idxs_dict[pid].pop(0)
                final_idxs.extend(batch_idxs)
                if len(batch_idxs_dict[pid]) == 0:
                    avai_pids.remove(pid)

        return iter(final_idxs)

    def __len__(self):
        return self.length

class AverageMeter(object):
    """Computes and stores the average and current value"""

    def __init__(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

class RandomErasing3(object):
    """ Randomly selects a rectangle region in an image and erases its pixels.
        'Random Erasing Data Augmentation' by Zhong et al.
        See https://arxiv.org/pdf/1708.04896.pdf
    Args:
         probability: The probability that the Random Erasing operation will be performed.
         sl: Minimum proportion of erased area against input image.
         sh: Maximum proportion of erased area against input image.
         r1: Minimum aspect ratio of erased area.
         mean: Erasing value.
    """
    def __init__(self, probability=0.5, sl=0.02, sh=0.4, r1=0.3, mean=(0.4914, 0.4822, 0.4465)):
        self.probability = probability
        self.mean = mean
        self.sl = sl
        self.sh = sh
        self.r1 = r1
    def __call__(self, img):
        if random.uniform(0, 1) >= self.probability:
            return img , 0 
        for attempt in range(100):
            area = img.size()[1] * img.size()[2]
            target_area = random.uniform(self.sl, self.sh) * area
            aspect_ratio = random.uniform(self.r1, 1 / self.r1)
            h = int(round(math.sqrt(target_area * aspect_ratio)))
            w = int(round(math.sqrt(target_area / aspect_ratio)))
            if w < img.size()[2] and h < img.size()[1]:
                x1 = random.randint(0, img.size()[1] - h)
                y1 = random.randint(0, img.size()[2] - w)
                if img.size()[0] == 3:
                    img[0, x1:x1 + h, y1:y1 + w] = self.mean[0]
                    img[1, x1:x1 + h, y1:y1 + w] = self.mean[1]
                    img[2, x1:x1 + h, y1:y1 + w] = self.mean[2]
                else:
                    img[0, x1:x1 + h, y1:y1 + w] = self.mean[0]
                return img , 1
            return img , 0         
        

def scheduler(optimizer):
    num_epochs = 120
    
    lr_min = 0.002 * 0.008
    warmup_lr_init = 0.01 * 0.008
    
    warmup_t = 5
    noise_range = None

    lr_scheduler = CosineLRScheduler(
            optimizer,
            t_initial=num_epochs,
            lr_min=lr_min,
            t_mul= 1.,
            decay_rate=0.1,
            warmup_lr_init=warmup_lr_init,
            warmup_t=warmup_t,
            cycle_limit=1,
            t_in_epochs=True,
            noise_range_t=noise_range,
            noise_pct= 0.67,
            noise_std= 1.,
            noise_seed=42,
        )

    return lr_scheduler        




def optimizer(model):
    params = []
    for key, value in model.named_parameters():
        if not value.requires_grad:
            continue
        lr = 0.008
        weight_decay = 1e-4
        if "bias" in key:
            lr = 0.008 * 2
            weight_decay = 1e-4

        params += [{"params": [value], "lr": lr, "weight_decay": weight_decay}]

    
    optimizer = getattr(torch.optim, 'SGD')(params, momentum=0.9)

    
    
    return optimizer





class Scheduler:
    """ Parameter Scheduler Base Class
    A scheduler base class that can be used to schedule any optimizer parameter groups.

    Unlike the builtin PyTorch schedulers, this is intended to be consistently called
    * At the END of each epoch, before incrementing the epoch count, to calculate next epoch's value
    * At the END of each optimizer update, after incrementing the update count, to calculate next update's value

    The schedulers built on this should try to remain as stateless as possible (for simplicity).

    This family of schedulers is attempting to avoid the confusion of the meaning of 'last_epoch'
    and -1 values for special behaviour. All epoch and update counts must be tracked in the training
    code and explicitly passed in to the schedulers on the corresponding step or step_update call.

    Based on ideas from:
     * https://github.com/pytorch/fairseq/tree/master/fairseq/optim/lr_scheduler
     * https://github.com/allenai/allennlp/tree/master/allennlp/training/learning_rate_schedulers
    """

    def __init__(self,
                 optimizer: torch.optim.Optimizer,
                 param_group_field: str,
                 noise_range_t=None,
                 noise_type='normal',
                 noise_pct=0.67,
                 noise_std=1.0,
                 noise_seed=None,
                 initialize: bool = True) -> None:
        self.optimizer = optimizer
        self.param_group_field = param_group_field
        self._initial_param_group_field = f"initial_{param_group_field}"
        if initialize:
            for i, group in enumerate(self.optimizer.param_groups):
                if param_group_field not in group:
                    raise KeyError(f"{param_group_field} missing from param_groups[{i}]")
                group.setdefault(self._initial_param_group_field, group[param_group_field])
        else:
            for i, group in enumerate(self.optimizer.param_groups):
                if self._initial_param_group_field not in group:
                    raise KeyError(f"{self._initial_param_group_field} missing from param_groups[{i}]")
        self.base_values = [group[self._initial_param_group_field] for group in self.optimizer.param_groups]
        self.metric = None  # any point to having this for all?
        self.noise_range_t = noise_range_t
        self.noise_pct = noise_pct
        self.noise_type = noise_type
        self.noise_std = noise_std
        self.noise_seed = noise_seed if noise_seed is not None else 42
        self.update_groups(self.base_values)

    def state_dict(self) -> Dict[str, Any]:
        return {key: value for key, value in self.__dict__.items() if key != 'optimizer'}

    def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
        self.__dict__.update(state_dict)

    def get_epoch_values(self, epoch: int):
        return None

    def get_update_values(self, num_updates: int):
        return None

    def step(self, epoch: int, metric: float = None) -> None:
        self.metric = metric
        values = self.get_epoch_values(epoch)
        if values is not None:
            values = self._add_noise(values, epoch)
            self.update_groups(values)

    def step_update(self, num_updates: int, metric: float = None):
        self.metric = metric
        values = self.get_update_values(num_updates)
        if values is not None:
            values = self._add_noise(values, num_updates)
            self.update_groups(values)

    def update_groups(self, values):
        if not isinstance(values, (list, tuple)):
            values = [values] * len(self.optimizer.param_groups)
        for param_group, value in zip(self.optimizer.param_groups, values):
            param_group[self.param_group_field] = value

    def _add_noise(self, lrs, t):
        if self.noise_range_t is not None:
            if isinstance(self.noise_range_t, (list, tuple)):
                apply_noise = self.noise_range_t[0] <= t < self.noise_range_t[1]
            else:
                apply_noise = t >= self.noise_range_t
            if apply_noise:
                g = torch.Generator()
                g.manual_seed(self.noise_seed + t)
                if self.noise_type == 'normal':
                    while True:
                        # resample if noise out of percent limit, brute force but shouldn't spin much
                        noise = torch.randn(1, generator=g).item()
                        if abs(noise) < self.noise_pct:
                            break
                else:
                    noise = 2 * (torch.rand(1, generator=g).item() - 0.5) * self.noise_pct
                lrs = [v + v * noise for v in lrs]
        return lrs

class CosineLRScheduler(Scheduler):
    """
    Cosine decay with restarts.
    This is described in the paper https://arxiv.org/abs/1608.03983.

    Inspiration from
    https://github.com/allenai/allennlp/blob/master/allennlp/training/learning_rate_schedulers/cosine.py
    """

    def __init__(self,
                 optimizer: torch.optim.Optimizer,
                 t_initial: int,
                 t_mul: float = 1.,
                 lr_min: float = 0.,
                 decay_rate: float = 1.,
                 warmup_t=0,
                 warmup_lr_init=0,
                 warmup_prefix=False,
                 cycle_limit=0,
                 t_in_epochs=True,
                 noise_range_t=None,
                 noise_pct=0.67,
                 noise_std=1.0,
                 noise_seed=42,
                 initialize=True) -> None:
        super().__init__(
            optimizer, param_group_field="lr",
            noise_range_t=noise_range_t, noise_pct=noise_pct, noise_std=noise_std, noise_seed=noise_seed,
            initialize=initialize)

        assert t_initial > 0
        assert lr_min >= 0
        if t_initial == 1 and t_mul == 1 and decay_rate == 1:
            _logger.warning("Cosine annealing scheduler will have no effect on the learning "
                           "rate since t_initial = t_mul = eta_mul = 1.")
        self.t_initial = t_initial
        self.t_mul = t_mul
        self.lr_min = lr_min
        self.decay_rate = decay_rate
        self.cycle_limit = cycle_limit
        self.warmup_t = warmup_t
        self.warmup_lr_init = warmup_lr_init
        self.warmup_prefix = warmup_prefix
        self.t_in_epochs = t_in_epochs
        if self.warmup_t:
            self.warmup_steps = [(v - warmup_lr_init) / self.warmup_t for v in self.base_values]
            super().update_groups(self.warmup_lr_init)
        else:
            self.warmup_steps = [1 for _ in self.base_values]

    def _get_lr(self, t):
        if t < self.warmup_t:
            lrs = [self.warmup_lr_init + t * s for s in self.warmup_steps]
        else:
            if self.warmup_prefix:
                t = t - self.warmup_t

            if self.t_mul != 1:
                i = math.floor(math.log(1 - t / self.t_initial * (1 - self.t_mul), self.t_mul))
                t_i = self.t_mul ** i * self.t_initial
                t_curr = t - (1 - self.t_mul ** i) / (1 - self.t_mul) * self.t_initial
            else:
                i = t // self.t_initial
                t_i = self.t_initial
                t_curr = t - (self.t_initial * i)

            gamma = self.decay_rate ** i
            lr_min = self.lr_min * gamma
            lr_max_values = [v * gamma for v in self.base_values]

            if self.cycle_limit == 0 or (self.cycle_limit > 0 and i < self.cycle_limit):
                lrs = [
                    lr_min + 0.5 * (lr_max - lr_min) * (1 + math.cos(math.pi * t_curr / t_i)) for lr_max in lr_max_values
                ]
            else:
                lrs = [self.lr_min for _ in self.base_values]

        return lrs

    def get_epoch_values(self, epoch: int):
        if self.t_in_epochs:
            return self._get_lr(epoch)
        else:
            return None

    def get_update_values(self, num_updates: int):
        if not self.t_in_epochs:
            return self._get_lr(num_updates)
        else:
            return None

    def get_cycle_length(self, cycles=0):
        if not cycles:
            cycles = self.cycle_limit
        cycles = max(1, cycles)
        if self.t_mul == 1.0:
            return self.t_initial * cycles
        else:
            return int(math.floor(-self.t_initial * (self.t_mul ** cycles - 1) / (1 - self.t_mul)))