What are the optimum hyperparameters for ResNet50 (v1.5) with ImageNet 2012 dataset?

Question

What are the optimum hyperparameter settings (loss function, learning rate, data augmentation, learning rate scheduler, etc.) to achieve >75% accuracy with ImageNet50 (2012) dataset using ResNet50? I have tried the suggested parameters from here, but my accuracy is nowhere near 75% (it's below 30%) after 90 epochs. What I am missing?

currently, I am using the following settings-

import torch
import torchvision

batch_size = 256
epoch_n = 90 #number of epochs

#training data transformer for data augmentation
train_transform= torchvision.transforms.Compose([
    transforms.RandomResizedCrop(224),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
]) 

#loss function, optimizer, and learning rate scheduler
criterion = torch.nn.CrossEntropyLoss(label_smoothing=0.1)
optimizer = torch.optim.SGD(net.parameters(), lr=batch_size/1000, momentum=0.875, weight_decay=1/32768)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=epoch_n)