import os
import random
import shutil
import sys
from semantic_segmentation.arguments import get_args
random.seed(1)
args = get_args('args for datasplit (camvid)', mode='data')
current_path = os.path.abspath('')
perc = args.percentage
if not 0 < perc < 100:
print('Illegal usage of -p, only between 0 and 100')
sys.exit(0)
NUM = int(367 * perc / 100)
image_dir = os.path.join(current_path, 'data/CamVid/train')
mask_dir = os.path.join(current_path, 'data/CamVid/trainannot')
new_dir = os.path.join(current_path, 'data/CamVid/trainsmall' + str(perc))
new_mask_dir = os.path.join(current_path, 'data/CamVid/trainsmallannot' + str(perc))
os.makedirs(new_dir, exist_ok=True)
os.makedirs(new_mask_dir, exist_ok=True)
ls = os.listdir(image_dir)
sample = random.sample(ls, NUM)
import comet_ml
import torch
from torch.utils.data import DataLoader
import semantic_segmentation.models as models
from semantic_segmentation.arguments import get_args
from semantic_segmentation.datasets.dataset import get_dataset
from semantic_segmentation.experiments.trainer import unfreeze, train_stagewise
from semantic_segmentation.utils.utils import *
args = get_args(desc='Stagewise training using less data of UNet based on ResNet encoder')
hyper_params = {
"dataset": args.dataset,
"model": args.model,
"seed": args.seed,
"num_classes": 12,
"batch_size": 8,
"num_epochs": args.epoch,
"learning_rate": 1e-4,
"stage": 0,
"perc": str(args.percentage)
}
torch.manual_seed(hyper_params['seed'])
if args.gpu != 'cpu':
torch.cuda.set_device(args.gpu)
torch.cuda.manual_seed(hyper_params['seed'])
train_dataset, valid_dataset, num_classes = get_dataset(args.dataset, args.percentage)
hyper_params['num_classes'] = num_classes
trainloader = DataLoader(train_dataset, batch_size=hyper_params['batch_size'], shuffle=True, drop_last=True)
from comet_ml import Experiment
import torch
from torch.utils.data import DataLoader
import semantic_segmentation.models as models
from semantic_segmentation.arguments import get_args
from semantic_segmentation.datasets.dataset import get_dataset
from semantic_segmentation.experiments.trainer import pretrain
args = get_args(desc="standalone training for small dataset")
hyper_params = {
"dataset": args.dataset,
"model": args.model,
"seed": args.seed,
"perc": args.percentage,
"num_classes": 12,
"batch_size": 8,
"num_epochs": args.epoch,
"learning_rate": 1e-4,
}
torch.manual_seed(hyper_params['seed'])
if args.gpu != 'cpu':
torch.cuda.set_device(args.gpu)
torch.cuda.manual_seed(hyper_params['seed'])
train_dataset, valid_dataset, num_classes = get_dataset(
args.dataset, args.percentage)
hyper_params['num_classes'] = num_classes
import comet_ml
import torch
from torch.utils.data import DataLoader
import semantic_segmentation.models as models
from semantic_segmentation.arguments import get_args
from semantic_segmentation.datasets.dataset import get_dataset
from semantic_segmentation.experiments.trainer import unfreeze, train_traditional
from semantic_segmentation.utils.utils import get_features_trad
args = get_args(desc='traditional kd training of UNet based on ResNet encoder')
hyper_params = {
"model": args.model,
"dataset": args.dataset,
"seed": args.seed,
"num_classes": 12,
"batch_size": 8,
"num_epochs": args.epoch,
"learning_rate": 1e-4,
"stage": 0
}
torch.manual_seed(hyper_params['seed'])
if args.gpu != 'cpu':
torch.cuda.set_device(args.gpu)
torch.cuda.manual_seed(hyper_params['seed'])
train_dataset, valid_dataset, num_classes = get_dataset(
args.dataset, args.percentage)
hyper_params['num_classes'] = num_classes
import torch
from torch.utils.data import DataLoader
from semantic_segmentation.arguments import get_args
from semantic_segmentation.datasets.dataset import get_dataset
from semantic_segmentation.experiments.trainer import evaluate
args = get_args(desc="args for evaluation", mode='eval')
if args.gpu != 'cpu':
torch.cuda.set_device(args.gpu)
_, valid_dataset, test_dataset, num_classes = get_dataset(
args.dataset, None, True)
valloader = DataLoader(valid_dataset, batch_size=1, shuffle=False)
# testloader = DataLoader(test_dataset, batch_size=1, shuffle=False)
params = {"model": None, "seed": args.seed, 'num_classes': num_classes}
evaluate(valloader, args, params, mode='pretrain') # without teacher training
evaluate(valloader, args, params, mode='classifier') # stagewise training
evaluate(valloader, args, params, mode='simultaneous') # simultanous training
evaluate(valloader, args, params,
mode='traditional-kd') # traditional-kd training
import comet_ml
import torch
from torch.utils.data import DataLoader
import semantic_segmentation.models as models
from semantic_segmentation.arguments import get_args
from semantic_segmentation.datasets.dataset import get_dataset
from semantic_segmentation.experiments.trainer import unfreeze, train_simultaneous
from semantic_segmentation.utils.utils import *
args = get_args('Simultaneous training of UNet based on ResNet encoder')
hyper_params = {
"model": args.model,
"dataset": args.dataset,
"seed": args.seed,
"num_classes": 12,
"batch_size": 8,
"num_epochs": args.epoch,
"learning_rate": 1e-4
}
torch.manual_seed(hyper_params['seed'])
if args.gpu != 'cpu':
torch.cuda.set_device(args.gpu)
torch.cuda.manual_seed(hyper_params['seed'])
train_dataset, valid_dataset, num_classes = get_dataset(
args.dataset, args.percentage)
hyper_params['num_classes'] = num_classes