from pathlib import Path
from utils_model import train_resnet
from utils import (get_classes, get_log_csv_name)
# Training the ResNet.
print("\n\n+++++ Running 3_train.py +++++")
train_resnet(
batch_size=config.args.batch_size,
checkpoints_folder=Path('checkpoints_easy10'),
classes=config.classes,
color_jitter_brightness=config.args.color_jitter_brightness,
color_jitter_contrast=config.args.color_jitter_contrast,
color_jitter_hue=config.args.color_jitter_hue,
color_jitter_saturation=config.args.color_jitter_saturation,
device=config.device,
learning_rate=config.args.learning_rate,
learning_rate_decay=config.args.learning_rate_decay,
log_csv=get_log_csv_name(log_folder=Path('logs/easy10')),
num_classes=config.num_classes,
num_layers=config.args.num_layers,
num_workers=config.args.num_workers,
path_mean=config.path_mean,
path_std=config.path_std,
pretrain=config.args.pretrain,
resume_checkpoint=True,
resume_checkpoint_path=Path('checkpoints/resnet18_e10_va0.55588.pt'),
save_interval=config.args.save_interval,
num_epochs=config.args.num_epochs,
train_folder=Path('data/voc_trainval_easy_0.1/'),
weight_decay=config.args.weight_decay)
print("+++++ Finished running 3_train.py +++++\n\n")
from utils import (get_classes, get_log_csv_name, get_log_csv_train_order)
from utils_model import train_resnet
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
path_mean, path_std = ([
0.40853017568588257, 0.4573926329612732, 0.48035722970962524
], [0.28722450137138367, 0.27334490418434143, 0.2799932360649109])
exp_num = 43
# Named with date and time.
train_folder = Path(
"/home/brenta/scratch/jason/data/imagenet/grad/first_test/")
log_folder = Path("/home/brenta/scratch/jason/logs/imagenet/grad_cl/exp_" +
str(exp_num))
log_csv = get_log_csv_name(log_folder=log_folder)
train_order_csv = get_log_csv_train_order(log_folder=log_folder)
classes = get_classes(train_folder.joinpath("train"))
num_classes = len(classes)
# Training the ResNet.
print("\n\n+++++ Running 3_train.py +++++")
train_resnet(
batch_size=256,
checkpoints_folder=Path(
"/home/brenta/scratch/jason/checkpoints/image_net/grad_cl/exp_" +
str(exp_num)),
classes=classes,
color_jitter_brightness=0,
color_jitter_contrast=0,
color_jitter_hue=0,
classes = get_classes(folder=args.all_wsi)
num_classes = len(classes)
# This is the input for model training, automatically built.
train_patches = args.train_folder.joinpath("train")
val_patches = args.train_folder.joinpath("val")
# Compute the mean and standard deviation for the given set of WSI for normalization.
path_mean, path_std = compute_stats(folderpath=args.all_wsi,
image_ext=args.image_ext)
# Only used is resume_ is True.
resume_checkpoint_path = args.checkpoints_folder.joinpath(args.checkpoint_file)
# Named with date and time.
log_csv = get_log_csv_name(log_folder=args.log_folder)
# Does nothing if auto_select is True.
eval_model = args.checkpoints_folder.joinpath(args.checkpoint_file)
# Find the best threshold for filtering noise (discard patches with a confidence less than this threshold).
threshold_search = (0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9)
# For visualization.
# This order is the same order as your sorted classes.
colors = ("red", "white", "blue", "green", "purple", "orange", "black", "pink",
"yellow")
# Print the configuration.
# Source: https://stackoverflow.com/questions/44689546/how-to-print-out-a-dictionary-nicely-in-python/44689627
# chr(10) and chr(9) are ways of going around the f-string limitation of