run_class_finetuning.py

# --------------------------------------------------------
# EVA-02: A Visual Representation for Neon Genesis
# Github source: https://github.com/baaivision/EVA/EVA02
# Copyright (c) 2023 Beijing Academy of Artificial Intelligence (BAAI)
# Licensed under The MIT License [see LICENSE for details]
# By Yuxin Fang
#
# Based on EVA: Exploring the Limits of Masked Visual Representation Learning at Scale (https://arxiv.org/abs/2211.07636)
# https://github.com/baaivision/EVA/tree/master/EVA-01
# --------------------------------------------------------'


import argparse
import datetime
import numpy as np
import time
import torch
import torch.backends.cudnn as cudnn
import json
import os

from pathlib import Path

from timm.data.mixup import Mixup
from timm.models import create_model
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy
from timm.utils import ModelEma
from timm.models.layers import trunc_normal_
from optim_factory import create_optimizer, get_parameter_groups, LayerDecayValueAssigner

from datasets import build_dataset
from engine_for_finetuning import train_one_epoch, evaluate, evaluate_real
from utils import NativeScalerWithGradNormCount as NativeScaler
from utils import LARS
import utils
from scipy import interpolate
import modeling_finetune
import imagenet_a_r_indices


def get_args():
    parser = argparse.ArgumentParser('EVA fine-tuning and evaluation script for image classification', add_help=False)
    parser.add_argument('--batch_size', default=64, type=int)
    parser.add_argument('--epochs', default=30, type=int)
    parser.add_argument('--update_freq', default=1, type=int)
    parser.add_argument('--save_ckpt_freq', default=5, type=int)
    parser.add_argument('--use_checkpoint', action='store_true', default=False)
    parser.add_argument('--freeze_backbone', action='store_true', default=False)
    parser.add_argument('--partial_freeze', default=0, type=int)

    # robust evaluation
    parser.add_argument('--robust_test', default=None, type=str,
                        help='robust evaluation dataset')

    # imnet real evaluation
    parser.add_argument('--real_labels', default=None, type=str,
                        help='dataset path for evaluation')

    # Model parameters
    parser.add_argument('--model', default='eva02_large_patch14_xattn_fusedLN_NaiveSwiGLU_subln_RoPE', 
                        type=str, metavar='MODEL', help='Name of model to train')

    parser.add_argument('--input_size', default=224, type=int,
                        help='images input size')
    parser.add_argument('--scale', default=[0.08, 1.0], type=float, nargs='+', metavar='BETA')

    parser.add_argument('--drop', type=float, default=0.0, metavar='PCT',
                        help='Dropout rate (default: 0.)')
    parser.add_argument('--attn_drop_rate', type=float, default=0.0, metavar='PCT',
                        help='Attention dropout rate (default: 0.)')
    parser.add_argument('--drop_path', type=float, default=0.1, metavar='PCT',
                        help='Drop path rate (default: 0.1)')

    parser.add_argument('--disable_eval_during_finetuning', action='store_true', default=False)

    parser.add_argument('--model_ema', action='store_true', default=False)
    parser.add_argument('--model_ema_decay', type=float, default=0.9999, help='')
    parser.add_argument('--model_ema_force_cpu', action='store_true', default=False, help='')
    parser.add_argument('--model_ema_eval', action='store_true')

    # Optimizer parameters
    parser.add_argument('--opt', default='adamw', type=str, metavar='OPTIMIZER',
                        help='Optimizer (default: "adamw"')
    parser.add_argument('--opt_eps', default=1e-8, type=float, metavar='EPSILON',
                        help='Optimizer Epsilon (default: 1e-8)')
    parser.add_argument('--opt_betas', default=[0.9, 0.999], type=float, nargs='+', metavar='BETA',
                        help='Optimizer Betas (default: 0.9, 0.999, use opt default)')
    parser.add_argument('--clip_grad', type=float, default=None, metavar='NORM',
                        help='Clip gradient norm (default: None, no clipping)')
    parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                        help='SGD momentum (default: 0.9)')
    parser.add_argument('--weight_decay', type=float, default=0.05,
                        help='weight decay (default: 0.05)')
    parser.add_argument('--weight_decay_end', type=float, default=None, help="""Final value of the
        weight decay. We use a cosine schedule for WD and using a larger decay by
        the end of training improves performance for ViTs.""")

    parser.add_argument('--lr', type=float, default=5e-4, metavar='LR',
                        help='learning rate (default: 5e-4)')
    parser.add_argument('--layer_decay', type=float, default=0.9)

    parser.add_argument('--warmup_lr', type=float, default=1e-6, metavar='LR',
                        help='warmup learning rate (default: 1e-6)')
    parser.add_argument('--min_lr', type=float, default=1e-6, metavar='LR',
                        help='lower lr bound for cyclic schedulers that hit 0 (1e-5)')

    parser.add_argument('--warmup_epochs', type=int, default=5, metavar='N',
                        help='epochs to warmup LR, if scheduler supports')
    parser.add_argument('--warmup_steps', type=int, default=-1, metavar='N',
                        help='num of steps to warmup LR, will overload warmup_epochs if set > 0')

    # Augmentation parameters
    parser.add_argument('--color_jitter', type=float, default=0.4, metavar='PCT',
                        help='Color jitter factor (default: 0.4)')
    parser.add_argument('--aa', type=str, default='rand-m9-mstd0.5-inc1', metavar='NAME',
                        help='Use AutoAugment policy. "v0" or "original". " + "(default: rand-m9-mstd0.5-inc1)')
    parser.add_argument('--smoothing', type=float, default=0.1,
                        help='Label smoothing (default: 0.1)')
    parser.add_argument('--train_interpolation', type=str, default='bicubic',
                        help='Training interpolation (random, bilinear, bicubic default: "bicubic")')

    # Evaluation parameters
    parser.add_argument('--crop_pct', type=float, default=None)

    # * Random Erase params
    parser.add_argument('--reprob', type=float, default=0.25, metavar='PCT',
                        help='Random erase prob (default: 0.25)')
    parser.add_argument('--remode', type=str, default='pixel',
                        help='Random erase mode (default: "pixel")')
    parser.add_argument('--recount', type=int, default=1,
                        help='Random erase count (default: 1)')
    parser.add_argument('--resplit', action='store_true', default=False,
                        help='Do not random erase first (clean) augmentation split')

    # * Mixup params
    parser.add_argument('--mixup', type=float, default=0,
                        help='mixup alpha, mixup enabled if > 0.')
    parser.add_argument('--cutmix', type=float, default=0,
                        help='cutmix alpha, cutmix enabled if > 0.')
    parser.add_argument('--cutmix_minmax', type=float, nargs='+', default=None,
                        help='cutmix min/max ratio, overrides alpha and enables cutmix if set (default: None)')
    parser.add_argument('--mixup_prob', type=float, default=1.0,
                        help='Probability of performing mixup or cutmix when either/both is enabled')
    parser.add_argument('--mixup_switch_prob', type=float, default=0.5,
                        help='Probability of switching to cutmix when both mixup and cutmix enabled')
    parser.add_argument('--mixup_mode', type=str, default='batch',
                        help='How to apply mixup/cutmix params. Per "batch", "pair", or "elem"')

    # * Finetuning params
    parser.add_argument('--no_aug', action='store_true', default=False)
    parser.add_argument('--finetune', default='',
                        help='finetune from checkpoint')
    parser.add_argument('--model_key', default='model|module', type=str)
    parser.add_argument('--model_prefix', default='', type=str)
    parser.add_argument('--init_scale', default=0.001, type=float)
    parser.add_argument('--use_mean_pooling', action='store_true')
    parser.set_defaults(use_mean_pooling=True)
    parser.add_argument('--use_cls', action='store_false', dest='use_mean_pooling')
    parser.add_argument('--disable_weight_decay_on_rel_pos_bias', action='store_true', default=False)

    # Linear probe
    parser.add_argument('--linear_probe', action='store_true')
    parser.add_argument('--probe_then_ft', action='store_true')

    # Dataset parameters
    parser.add_argument('--data_path', default='/path/to/IN-1K', type=str,
                        help='dataset path')
    parser.add_argument('--eval_data_path', default=None, type=str,
                        help='dataset path for evaluation')
    parser.add_argument('--nb_classes', default=1000, type=int,
                        help='number of the classification types')
    parser.add_argument('--imagenet_default_mean_and_std', default=False, action='store_true')


    parser.add_argument('--data_set', default='IMNET', choices=['CIFAR', 'IMNET', 'image_folder'],
                        type=str, help='ImageNet dataset path')
    parser.add_argument('--output_dir', default='',
                        help='path where to save, empty for no saving')
    parser.add_argument('--log_dir', default=None,
                        help='path where to tensorboard log')
    parser.add_argument('--device', default='cuda',
                        help='device to use for training / testing')
    parser.add_argument('--seed', default=0, type=int)
    
    parser.add_argument('--resume', default='',
                        help='resume from checkpoint')
    parser.add_argument('--auto_resume', action='store_true')
    parser.add_argument('--no_auto_resume', action='store_false', dest='auto_resume')
    parser.set_defaults(auto_resume=True)

    parser.add_argument('--save_ckpt', action='store_true')
    parser.add_argument('--no_save_ckpt', action='store_false', dest='save_ckpt')
    parser.set_defaults(save_ckpt=True)

    parser.add_argument('--start_epoch', default=0, type=int, metavar='N',
                        help='start epoch')
    parser.add_argument('--eval', action='store_true',
                        help='Perform evaluation only')
    parser.add_argument('--dist_eval', action='store_true', default=False,
                        help='Enabling distributed evaluation')
    parser.add_argument('--use_ema_ckpt_eval', action='store_true',
                        help='Perform evaluation using the ema ckpt')

    parser.add_argument('--num_workers', default=10, type=int)
    parser.add_argument('--pin_mem', action='store_true',
                        help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
    parser.add_argument('--no_pin_mem', action='store_false', dest='pin_mem')
    parser.set_defaults(pin_mem=True)

    # distributed training parameters
    parser.add_argument('--world_size', default=1, type=int,
                        help='number of distributed processes')
    parser.add_argument('--local_rank', default=-1, type=int)
    parser.add_argument('--dist_url', default='env://',
                        help='url used to set up distributed training')

    parser.add_argument('--enable_deepspeed', action='store_true', default=False)
    parser.add_argument('--zero_stage', default=0, type=int,
                        help='ZeRO optimizer stage (default: 0)')
    parser.add_argument('--bf16', default=False, action='store_true')

    known_args, _ = parser.parse_known_args()

    if known_args.enable_deepspeed:
        try:
            import deepspeed
            parser = deepspeed.add_config_arguments(parser)
            ds_init = deepspeed.initialize
        except:
            print("Please install deepspeed")
            exit(0)
    else:
        ds_init = None

    if ds_init is not None and known_args.bf16 and str(deepspeed.__version__) != '0.8.1':
        print("Please install deepspeed==0.8.1 for bf16 ft")
        exit(0)

    return parser.parse_args(), ds_init


def main(args, ds_init):
    utils.init_distributed_mode(args)

    if ds_init is not None:
        utils.create_ds_config(args)

    print("{}".format(args).replace(', ', ',\n'))

    device = torch.device(args.device)

    # fix the seed for reproducibility
    seed = args.seed + utils.get_rank()
    torch.manual_seed(seed)
    np.random.seed(seed)

    cudnn.benchmark = True

    dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)

    if args.disable_eval_during_finetuning:
        dataset_val = None
    elif args.robust_test == 'imagenet_v2':
        from robustness_dataset import RobustnessDataset
        dataset_val = RobustnessDataset(args.eval_data_path)
    elif args.robust_test == 'objectnet':
        from objectnet_dataset import ObjectNetDataset
        dataset_val = ObjectNetDataset(args.eval_data_path)
    else:
        dataset_val, _ = build_dataset(is_train=False, args=args)

    if True:  # args.distributed:
        num_tasks = utils.get_world_size()
        global_rank = utils.get_rank()
        sampler_train = torch.utils.data.DistributedSampler(
            dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True
        )
        print("Sampler_train = %s" % str(sampler_train))
        if args.dist_eval:
            if len(dataset_val) % num_tasks != 0:
                print('Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
                      'This will slightly alter validation results as extra duplicate entries are added to achieve '
                      'equal num of samples per-process.')
            sampler_val = torch.utils.data.DistributedSampler(
                dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False)
        else:
            sampler_val = torch.utils.data.SequentialSampler(dataset_val)
    else:
        sampler_train = torch.utils.data.RandomSampler(dataset_train)
        sampler_val = torch.utils.data.SequentialSampler(dataset_val)

    if global_rank == 0 and args.log_dir is not None:
        os.makedirs(args.log_dir, exist_ok=True)
        log_writer = utils.TensorboardLogger(log_dir=args.log_dir)
    else:
        log_writer = None

    data_loader_train = torch.utils.data.DataLoader(
        dataset_train, sampler=sampler_train,
        batch_size=args.batch_size,
        num_workers=args.num_workers,
        pin_memory=args.pin_mem,
        drop_last=True,
    )

    if dataset_val is not None:
        data_loader_val = torch.utils.data.DataLoader(
            dataset_val, sampler=sampler_val,
            batch_size=int(args.batch_size) if args.linear_probe else int(1.5 * args.batch_size),
            num_workers=args.num_workers,
            pin_memory=args.pin_mem,
            drop_last=False
        )
    else:
        data_loader_val = None

    mixup_fn = None
    mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
    if mixup_active:
        print("Mixup is activated!")
        mixup_fn = Mixup(
            mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax,
            prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode,
            label_smoothing=args.smoothing, num_classes=args.nb_classes)

    model = create_model(
        args.model,
        pretrained=False,
        img_size=args.input_size,
        num_classes=args.nb_classes,
        drop_rate=args.drop,
        drop_path_rate=args.drop_path,
        attn_drop_rate=args.attn_drop_rate,
        drop_block_rate=None,
        use_mean_pooling=args.use_mean_pooling,
        use_checkpoint=args.use_checkpoint, 
    )

    patch_size = model.patch_embed.patch_size
    print("Patch size = %s" % str(patch_size))
    args.window_size = (args.input_size // patch_size[0], args.input_size // patch_size[1])
    args.patch_size = patch_size

    if args.linear_probe or args.probe_then_ft:
        trunc_normal_(model.head.weight, std=0.01)
        model.head = torch.nn.Sequential(torch.nn.BatchNorm1d(model.head.in_features, affine=False, eps=1e-6), model.head)
        
    if args.finetune:
        if args.finetune.startswith('https'):
            checkpoint = torch.hub.load_state_dict_from_url(
                args.finetune, map_location='cpu', check_hash=True)
        else:
            checkpoint = torch.load(args.finetune, map_location='cpu')

        print("Load ckpt from %s" % args.finetune)
        checkpoint_model = None
        for model_key in args.model_key.split('|'):
            if model_key in checkpoint:
                checkpoint_model = checkpoint[model_key]
                print("Load state_dict by model_key = %s" % model_key)
                break
        if checkpoint_model is None:
            checkpoint_model = checkpoint
        if args.use_ema_ckpt_eval: 
            checkpoint_model = checkpoint['model_ema']
            print("Load state_dict model_ema [eval only]")

        state_dict = model.state_dict()
        for k in ['head.weight', 'head.bias']:
            if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape:
                if args.robust_test == 'imagenet_r':
                    mask = torch.tensor(imagenet_a_r_indices.imagenet_r_mask)
                    checkpoint_model[k] = checkpoint_model[k][mask]
                elif args.robust_test == 'imagenet_a':
                    mask = torch.tensor(imagenet_a_r_indices.imagenet_a_mask)
                    checkpoint_model[k] = checkpoint_model[k][mask]
                else:
                    print(f"Removing key {k} from pretrained checkpoint")
                    del checkpoint_model[k]

        if model.use_rel_pos_bias and "rel_pos_bias.relative_position_bias_table" in checkpoint_model:
            print("Expand the shared relative position embedding to each transformer block. ")
            num_layers = model.get_num_layers()
            rel_pos_bias = checkpoint_model["rel_pos_bias.relative_position_bias_table"]
            for i in range(num_layers):
                checkpoint_model["blocks.%d.attn.relative_position_bias_table" % i] = rel_pos_bias.clone()

            checkpoint_model.pop("rel_pos_bias.relative_position_bias_table")

        if model.use_decoupled_rel_pos_bias and "rel_pos_bias.relative_position_bias_for_high" in checkpoint_model:
            print("Expand the shared relative position embedding to each layers. ")
            num_layers = model.get_num_layers()
            rel_pos_bias_for_high = checkpoint_model["rel_pos_bias.relative_position_bias_for_high"]
            rel_pos_bias_for_width = checkpoint_model["rel_pos_bias.relative_position_bias_for_width"]
            for i in range(num_layers):
                checkpoint_model["blocks.%d.attn.rel_pos_bias.relative_position_bias_for_high" % i] \
                    = rel_pos_bias_for_high.clone()
                checkpoint_model["blocks.%d.attn.rel_pos_bias.relative_position_bias_for_width" % i] \
                    = rel_pos_bias_for_width.clone()

            checkpoint_model.pop("rel_pos_bias.relative_position_bias_for_high")
            checkpoint_model.pop("rel_pos_bias.relative_position_bias_for_width")

        all_keys = list(checkpoint_model.keys())
        for key in all_keys:
            if "relative_position_index" in key:
                checkpoint_model.pop(key)

            if "relative_position_high_index" in key:
                checkpoint_model.pop(key)

            if "relative_position_width_index" in key:
                checkpoint_model.pop(key)

            if "relative_position_bias_table" in key:
                rel_pos_bias = checkpoint_model[key]
                src_num_pos, num_attn_heads = rel_pos_bias.size()
                dst_num_pos, _ = model.state_dict()[key].size()
                dst_patch_shape = model.patch_embed.patch_shape
                if dst_patch_shape[0] != dst_patch_shape[1]:
                    raise NotImplementedError()
                num_extra_tokens = dst_num_pos - (dst_patch_shape[0] * 2 - 1) * (dst_patch_shape[1] * 2 - 1)
                src_size = int((src_num_pos - num_extra_tokens) ** 0.5)
                dst_size = int((dst_num_pos - num_extra_tokens) ** 0.5)
                if src_size != dst_size:
                    print("Position interpolate for %s from %dx%d to %dx%d" % (
                        key, src_size, src_size, dst_size, dst_size))
                    extra_tokens = rel_pos_bias[-num_extra_tokens:, :]
                    rel_pos_bias = rel_pos_bias[:-num_extra_tokens, :]

                    def geometric_progression(a, r, n):
                        return a * (1.0 - r ** n) / (1.0 - r)

                    left, right = 1.01, 1.5
                    while right - left > 1e-6:
                        q = (left + right) / 2.0
                        gp = geometric_progression(1, q, src_size // 2)
                        if gp > dst_size // 2:
                            right = q
                        else:
                            left = q

                    # if q > 1.090307:
                    #     q = 1.090307

                    dis = []
                    cur = 1
                    for i in range(src_size // 2):
                        dis.append(cur)
                        cur += q ** (i + 1)

                    r_ids = [-_ for _ in reversed(dis)]

                    x = r_ids + [0] + dis
                    y = r_ids + [0] + dis

                    t = dst_size // 2.0
                    dx = np.arange(-t, t + 0.1, 1.0)
                    dy = np.arange(-t, t + 0.1, 1.0)

                    print("Original positions = %s" % str(x))
                    print("Target positions = %s" % str(dx))

                    all_rel_pos_bias = []

                    for i in range(num_attn_heads):
                        z = rel_pos_bias[:, i].view(src_size, src_size).float().numpy()
                        f = interpolate.interp2d(x, y, z, kind='cubic')
                        all_rel_pos_bias.append(
                            torch.Tensor(f(dx, dy)).contiguous().view(-1, 1).to(rel_pos_bias.device))

                    rel_pos_bias = torch.cat(all_rel_pos_bias, dim=-1)

                    new_rel_pos_bias = torch.cat((rel_pos_bias, extra_tokens), dim=0)
                    checkpoint_model[key] = new_rel_pos_bias

            if "relative_position_bias_for_" in key:
                rel_pos_bias = checkpoint_model[key]
                src_num_pos, num_attn_heads = rel_pos_bias.size()
                dst_num_pos, _ = model.state_dict()[key].size()
                dst_patch_shape = model.patch_embed.patch_shape
                if dst_patch_shape[0] != dst_patch_shape[1]:
                    raise NotImplementedError()
                num_extra_tokens = dst_num_pos - (dst_patch_shape[0] * 2 - 1)
                src_size = src_num_pos - num_extra_tokens
                dst_size = dst_num_pos - num_extra_tokens
                if src_size != dst_size:
                    print("Position interpolate for %s from %d to %d" % (key, src_size, dst_size))
                    extra_tokens = rel_pos_bias[-num_extra_tokens:, :]
                    rel_pos_bias = rel_pos_bias[:-num_extra_tokens, :]
                    # q = 1.13492
                    q = 1.0903078

                    dis = []
                    cur = 1
                    for i in range(src_size // 2):
                        dis.append(cur)
                        cur += q ** (i + 1)

                    r_ids = [-_ for _ in reversed(dis)]

                    x = r_ids + [0] + dis

                    t = dst_size // 2.0
                    dx = np.arange(-t, t + 0.1, 1.0)

                    print("x = %s" % str(x))
                    print("dx = %s" % str(dx))

                    all_rel_pos_bias = []

                    for i in range(num_attn_heads):
                        z = rel_pos_bias[:, i].view(src_size).float().numpy()
                        f = interpolate.interp1d(x, z, kind='cubic', fill_value="extrapolate")
                        all_rel_pos_bias.append(
                            torch.Tensor(f(dx)).contiguous().view(-1, 1).to(rel_pos_bias.device))

                    rel_pos_bias = torch.cat(all_rel_pos_bias, dim=-1)

                    new_rel_pos_bias = torch.cat((rel_pos_bias, extra_tokens), dim=0)
                    checkpoint_model[key] = new_rel_pos_bias

        # interpolate position embedding
        if 'pos_embed' in checkpoint_model:
            pos_embed_checkpoint = checkpoint_model['pos_embed']
            embedding_size = pos_embed_checkpoint.shape[-1]
            num_patches = model.patch_embed.num_patches
            num_extra_tokens = model.pos_embed.shape[-2] - num_patches
            # height (== width) for the checkpoint position embedding
            orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
            # height (== width) for the new position embedding
            new_size = int(num_patches ** 0.5)
            # class_token and dist_token are kept unchanged
            if orig_size != new_size:
                print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
                extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
                # only the position tokens are interpolated
                pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
                pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
                pos_tokens = torch.nn.functional.interpolate(
                    pos_tokens.float(), size=(new_size, new_size), mode='bicubic', align_corners=False).type_as(pos_tokens)
                pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
                new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
                checkpoint_model['pos_embed'] = new_pos_embed

        utils.load_state_dict(model, checkpoint_model, prefix=args.model_prefix)

    if args.freeze_backbone or args.linear_probe:
        for _, p in model.named_parameters():
            p.requires_grad = False
        for _, p in model.head.named_parameters():
            p.requires_grad = True
        if args.freeze_backbone:
            for _, p in model.fc_norm.named_parameters():
                p.requires_grad = True

    if args.partial_freeze > 0:
        for n, p in model.named_parameters():
            if 'block' in n:
                num_block = n.split('.')[1]
                if int(num_block) < args.partial_freeze:
                    p.requires_grad = False
                else:
                    p.requires_grad = True
            else: 
                p.requires_grad = False
        for _, p in model.head.named_parameters():
            p.requires_grad = True
        for _, p in model.fc_norm.named_parameters():
            p.requires_grad = True



    model.to(device)
    
    model_ema = None
    if args.model_ema:
        # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
        model_ema = ModelEma(
            model,
            decay=args.model_ema_decay,
            device='cpu' if args.model_ema_force_cpu else '',
            resume='')
        print("Using EMA with decay = %.8f" % args.model_ema_decay)

    model_without_ddp = model
    n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)

    print("Model = %s" % str(model_without_ddp))
    print('number of params:', n_parameters)

    total_batch_size = args.batch_size * args.update_freq * utils.get_world_size()
    num_training_steps_per_epoch = len(dataset_train) // total_batch_size
    print("LR = %.8f" % args.lr)
    print("Batch size = %d" % total_batch_size)
    print("Update frequent = %d" % args.update_freq)
    print("Number of training examples = %d" % len(dataset_train))
    print("Number of training steps per epoch = %d" % num_training_steps_per_epoch)

    num_layers = model_without_ddp.get_num_layers()
    if args.layer_decay < 1.0:
        assigner = LayerDecayValueAssigner(list(args.layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2)))
    else:
        assigner = None

    if assigner is not None:
        print("Assigned values = %s" % str(assigner.values))

    skip_weight_decay_list = model.no_weight_decay()
    if args.disable_weight_decay_on_rel_pos_bias:
        for i in range(num_layers):
            skip_weight_decay_list.add("blocks.%d.attn.relative_position_bias_table" % i)

    if args.distributed:
        torch.distributed.barrier()
    if args.enable_deepspeed:
        loss_scaler = None
        optimizer_params = get_parameter_groups(
            model, args.weight_decay, skip_weight_decay_list,
            assigner.get_layer_id if assigner is not None else None,
            assigner.get_scale if assigner is not None else None)   

        opt_lower = args.opt.lower()
        if opt_lower == 'adamw':
            model, optimizer, _, _ = ds_init(
                args=args, 
                model=model, 
                model_parameters=None if args.eval else optimizer_params,
                dist_init_required=not args.distributed,
            )
        else:   # use other non-fused opt
            optimizer = create_optimizer(
                args, model_without_ddp, skip_list=skip_weight_decay_list,
                get_num_layer=assigner.get_layer_id if assigner is not None else None, 
                get_layer_scale=assigner.get_scale if assigner is not None else None)
            print(optimizer)
            model, optimizer, _, _ = ds_init(
                args=args,
                model=model,
                optimizer=optimizer,
                model_parameters=None if args.eval else optimizer_params,
                dist_init_required=not args.distributed,
            )
            
        print("model.gradient_accumulation_steps() = %d" % model.gradient_accumulation_steps())
        assert model.gradient_accumulation_steps() == args.update_freq
    else:
        if args.distributed:
            model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=True)
            model_without_ddp = model.module

        if args.linear_probe:
            optimizer = LARS(model_without_ddp.head.parameters(), lr=args.lr, weight_decay=args.weight_decay)
        else:
            optimizer = create_optimizer(
                args, model_without_ddp, skip_list=skip_weight_decay_list,
                get_num_layer=assigner.get_layer_id if assigner is not None else None, 
                get_layer_scale=assigner.get_scale if assigner is not None else None)
        print(optimizer)
        loss_scaler = NativeScaler()

    print("Use step level LR scheduler!")
    lr_schedule_values = utils.cosine_scheduler(
        args.lr, args.min_lr, args.epochs, num_training_steps_per_epoch,
        warmup_epochs=args.warmup_epochs, warmup_steps=args.warmup_steps,
    )
    if args.weight_decay_end is None:
        args.weight_decay_end = args.weight_decay
    wd_schedule_values = utils.cosine_scheduler(
        args.weight_decay, args.weight_decay_end, args.epochs, num_training_steps_per_epoch)
    print("Max WD = %.7f, Min WD = %.7f" % (max(wd_schedule_values), min(wd_schedule_values)))

    if mixup_fn is not None:
        # smoothing is handled with mixup label transform
        criterion = SoftTargetCrossEntropy()
    elif args.smoothing > 0.:
        criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
    else:
        criterion = torch.nn.CrossEntropyLoss() # linear

    print("criterion = %s" % str(criterion))

    if args.eval:
        if args.real_labels:
            from imagenet_real import RealLabelsImagenet
            real_labels = RealLabelsImagenet(dataset_val.filenames(), real_json=args.real_labels)
            evaluate_real(data_loader_val, model, device, real_labels, ds=args.enable_deepspeed, bf16=args.bf16)
        else:
            test_stats = evaluate(data_loader_val, model, device, ds=args.enable_deepspeed, bf16=args.bf16)
            print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
            exit(0)

    utils.auto_load_model(
        args=args, model=model, model_without_ddp=model_without_ddp,
        optimizer=optimizer, loss_scaler=loss_scaler, model_ema=model_ema)

    print(f"Start training for {args.epochs} epochs")
    start_time = time.time()
    
    max_accuracy = 0.0
    if args.model_ema and args.model_ema_eval:
        max_accuracy_ema = 0.0

    for epoch in range(args.start_epoch, args.epochs):
        if args.distributed:
            data_loader_train.sampler.set_epoch(epoch)
        if log_writer is not None:
            log_writer.set_step(epoch * num_training_steps_per_epoch * args.update_freq)
        train_stats = train_one_epoch(
            model, 
            criterion, 
            data_loader_train, 
            optimizer,
            device, 
            epoch, 
            loss_scaler, 
            args.clip_grad, 
            model_ema, 
            mixup_fn,
            log_writer=log_writer, 
            start_steps=epoch * num_training_steps_per_epoch,
            lr_schedule_values=lr_schedule_values, 
            wd_schedule_values=wd_schedule_values,
            num_training_steps_per_epoch=num_training_steps_per_epoch, 
            update_freq=args.update_freq,
            args=args,
        )
        if args.output_dir and args.save_ckpt:
            if (epoch + 1) % args.save_ckpt_freq == 0 or epoch + 1 == args.epochs:
                utils.save_model(
                    args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                    loss_scaler=loss_scaler, epoch=epoch, model_ema=model_ema)
        if data_loader_val is not None:
            test_stats = evaluate(data_loader_val, model, device, ds=args.enable_deepspeed, bf16=args.bf16)
            print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
            if max_accuracy < test_stats["acc1"]:
                max_accuracy = test_stats["acc1"]
                if args.output_dir and args.save_ckpt:
                    utils.save_model(
                        args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                        loss_scaler=loss_scaler, epoch="best", model_ema=model_ema)
            print(f'Max accuracy: {max_accuracy:.2f}%')

            if args.model_ema and args.model_ema_eval:
                ema_test_stats = evaluate(data_loader_val, model_ema.ema, device)
                print(f"Accuracy of the EMA network on the {len(dataset_val)} test images: {ema_test_stats['acc1']:.1f}%")
                if max_accuracy_ema < ema_test_stats["acc1"]:
                    max_accuracy_ema = ema_test_stats["acc1"]
                    if args.output_dir and args.save_ckpt:
                        utils.save_model(
                            args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                            loss_scaler=loss_scaler, epoch="best-ema", model_ema=model_ema)
                print(f'Max EMA accuracy: {max_accuracy_ema:.2f}%')

            if log_writer is not None:
                log_writer.update(test_acc1=test_stats['acc1'], head="perf", step=epoch)
                log_writer.update(test_acc5=test_stats['acc5'], head="perf", step=epoch)
                log_writer.update(test_loss=test_stats['loss'], head="perf", step=epoch)
                if args.model_ema and args.model_ema_eval:
                    log_writer.update(ema_test_acc1=ema_test_stats['acc1'], head="perf", step=epoch)
                    log_writer.update(ema_test_acc5=ema_test_stats['acc5'], head="perf", step=epoch)

            if args.model_ema and args.model_ema_eval:
                log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                            **{f'test_{k}': v for k, v in test_stats.items()},
                            **{f'ema_test_{k}': v for k, v in ema_test_stats.items()},
                            'epoch': epoch,
                            'n_parameters': n_parameters}
            else:
                log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                            **{f'test_{k}': v for k, v in test_stats.items()},
                            'epoch': epoch,
                            'n_parameters': n_parameters}
        else:
            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                         'epoch': epoch,
                         'n_parameters': n_parameters}

        if args.output_dir and utils.is_main_process():
            if log_writer is not None:
                log_writer.flush()
            with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f:
                f.write(json.dumps(log_stats) + "\n")

    total_time = time.time() - start_time
    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
    print('Training time {}'.format(total_time_str))


if __name__ == '__main__':
    opts, ds_init = get_args()
    if opts.output_dir:
        Path(opts.output_dir).mkdir(parents=True, exist_ok=True)
    main(opts, ds_init)