def __call__(self, X, y):
"""
given a dataset X,y we split it, in order to do cross validation,
according to the procedure explained below:
if n_folds is not None, then we do cross validation
based on stratified folds
if n_class_samples is not None, then we do cross validation
using only <n_class_samples> training samples per class
if n_test_samples is not None, then we do cross validation
using only <n_test_samples> cross validaition samples per class
assumes that each datapoint is in a column of X
"""
n_classes = len(set(y))
if self.n_folds is not None:
# generate the folds
self.folds = StratifiedKFold(y, n_folds=self.n_folds,
shuffle=False, random_state=None)
elif self.n_class_samples is not None:
self.folds = []
for i in range(self.n_tests):
if type(self.n_class_samples) is not list:
self.n_class_samples = (np.ones(n_classes) * self.n_class_samples).astype(int)
if self.n_test_samples is not None:
self.n_test_samples = (np.ones(n_classes) * self.n_test_samples).astype(int)
data_idx = split_dataset(self.n_class_samples, self.n_test_samples, y)
train_idx = data_idx[0]
test_idx = data_idx[1]
self.folds.append((train_idx, test_idx))
self.cross_validate(X, y)
def load_data(model_type, data_path, image_path, heatmaps_path, input_size, class_names, batch_size, num_workers, rseed, heatmaps_threshold):
# ImageNet normalization
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
train_file, valid_file, test_file = split_dataset(data_path, random_state=rseed)
seq = iaa.Sequential([iaa.Resize((input_size, input_size))])
image_transform = transforms.Compose([seq.augment_image, transforms.ToTensor(), transforms.Normalize(mean=mean, std=std)])
heatmap_static_transform = transforms.Compose([transforms.Resize([input_size, input_size]),
transforms.Grayscale(num_output_channels=1),
transforms.ToTensor()])
static_heatmap_path = heatmaps_path
train_dataset = EyegazeDataset(train_file, image_path, class_names, static_heatmap_path=static_heatmap_path,
heatmaps_threshold=heatmaps_threshold, heatmap_static_transform=heatmap_static_transform,
image_transform=image_transform)
valid_dataset = EyegazeDataset(valid_file, image_path, class_names, static_heatmap_path=static_heatmap_path,
heatmaps_threshold = heatmaps_threshold, heatmap_static_transform=heatmap_static_transform,
image_transform=image_transform)
train_dl = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, drop_last=False, num_workers=num_workers)
valid_dl = DataLoader(valid_dataset, batch_size=batch_size, shuffle=True, drop_last=False, num_workers=num_workers)
test_dataset = EyegazeDataset(test_file, image_path, class_names, static_heatmap_path=static_heatmap_path,
heatmaps_threshold = heatmaps_threshold, heatmap_static_transform=heatmap_static_transform,
image_transform=image_transform)
test_dl = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, drop_last=False, num_workers=32)
return train_dl, valid_dl, test_dl
def __call__(self, X, y):
"""
given a dataset X,y we split it, in order to do cross validation,
according to the procedure explained below:
if n_folds is not None, then we do cross validation
based on stratified folds
if n_class_samples is not None, then we do cross validation
using only <n_class_samples> training samples per class
if n_test_samples is not None, then we do cross validation
using only <n_test_samples> cross validaition samples per class
assumes that each datapoint is in a column of X
"""
n_classes = len(set(y))
if self.n_folds is not None:
# generate the folds
self.folds = StratifiedKFold(y,
n_folds=self.n_folds,
shuffle=False,
random_state=None)
elif self.n_class_samples is not None:
self.folds = []
for i in range(self.n_tests):
if type(self.n_class_samples) is not list:
self.n_class_samples = (np.ones(n_classes) *
self.n_class_samples).astype(int)
if self.n_test_samples is not None:
self.n_test_samples = (np.ones(n_classes) *
self.n_test_samples).astype(int)
data_idx = split_dataset(self.n_class_samples,
self.n_test_samples, y)
train_idx = data_idx[0]
test_idx = data_idx[1]
self.folds.append((train_idx, test_idx))
self.cross_validate(X, y)
def load_data(model_type, data_path, image_path, heatmaps_path, input_size,
class_names, batch_size, num_workers, rseed):
# ImageNet normalization
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
train_file, valid_file, test_file = split_dataset(data_path,
random_state=rseed)
seq = iaa.Sequential([iaa.Resize((input_size, input_size))])
image_transform = transforms.Compose([
seq.augment_image,
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
if model_type in ['temporal']:
heatmap_temporal_transform = transforms.Compose([
transforms.Resize([input_size, input_size]),
transforms.Grayscale(num_output_channels=1),
transforms.ToTensor(),
transforms.Lambda(lambda x: x.repeat(3, 1, 1)),
transforms.Normalize(mean=mean, std=std)
])
heatmap_static_transform = transforms.Compose([
transforms.Resize([input_size, input_size]),
transforms.ToTensor()
])
static_heatmap_path = heatmaps_path
train_dataset = EyegazeDataset(
train_file,
image_path,
class_names,
heatmaps_path=heatmaps_path,
static_heatmap_path=static_heatmap_path,
heatmap_temporal_transform=heatmap_temporal_transform,
heatmap_static_transform=heatmap_static_transform,
image_transform=image_transform)
valid_dataset = EyegazeDataset(
valid_file,
image_path,
class_names,
heatmaps_path=heatmaps_path,
static_heatmap_path=static_heatmap_path,
heatmap_temporal_transform=heatmap_temporal_transform,
heatmap_static_transform=heatmap_static_transform,
image_transform=image_transform)
test_dataset = EyegazeDataset(
test_file,
image_path,
class_names,
heatmaps_path=heatmaps_path,
static_heatmap_path=static_heatmap_path,
heatmap_temporal_transform=heatmap_temporal_transform,
heatmap_static_transform=heatmap_static_transform,
image_transform=image_transform)
# drop_last=True for batchnorm issue: https://discuss.pytorch.org/t/error-expected-more-than-1-value-per-channel-when-training/26274
# this did not resolve the issue for all cases
train_dl = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
collate_fn=collate_fn,
drop_last=True)
valid_dl = DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
collate_fn=collate_fn,
drop_last=True)
test_dl = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False,
collate_fn=collate_fn,
num_workers=32)
else:
train_dataset = EyegazeDataset(train_file,
image_path,
class_names,
image_transform=image_transform)
valid_dataset = EyegazeDataset(valid_file,
image_path,
class_names,
image_transform=image_transform)
train_dl = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True,
num_workers=num_workers)
valid_dl = DataLoader(valid_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True,
num_workers=num_workers)
test_dataset = EyegazeDataset(test_file,
image_path,
class_names,
image_transform=image_transform)
test_dl = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False,
drop_last=True,
num_workers=32)
return train_dl, valid_dl, test_dl
if opt.deviceIds[1] >= 0 else "Use CPU as target nlg torch device")
##### Vocab and Dataset Reader #####
slu_vocab, nlg_vocab = Vocab(dataset=opt.dataset,
task='slu'), Vocab(dataset=opt.dataset,
task='nlg')
lm_vocab = Vocab(dataset=opt.dataset, task='lm')
slu_evaluator, nlg_evaluator = Evaluator.get_evaluator_from_task(
task='slu',
vocab=slu_vocab), Evaluator.get_evaluator_from_task(task='nlg',
vocab=nlg_vocab)
if not opt.testing:
train_dataset, dev_dataset = read_dataset(
opt.dataset, choice='train'), read_dataset(opt.dataset, choice='valid')
labeled_dataset, unlabeled_dataset = split_dataset(train_dataset,
opt.labeled)
logger.info(
"Labeled/Unlabeled train and dev dataset size is: %s/%s and %s" %
(len(labeled_dataset), len(unlabeled_dataset), len(dev_dataset)))
unlabeled_dataset = labeled_dataset + unlabeled_dataset
test_dataset = read_dataset(opt.dataset, choice='test')
logger.info("Test dataset size is: %s" % (len(test_dataset)))
##### Model Construction and Init #####
if not opt.testing:
params = vars(opt)
json.dump(params,
open(os.path.join(exp_path, 'params.json'), 'w'),
indent=4)
else:
params = json.load(open(os.path.join(exp_path, 'params.json'), 'r'))
logger.info("Random seed is set to: %d" % (opt.seed))
logger.info(
"Use GPU with index %s" %
(opt.deviceId) if opt.deviceId >= 0 else "Use CPU as target torch device")
##### Vocab and Dataset Reader #####
vocab = Vocab(dataset=opt.dataset, task=task)
logger.info("Vocab size for input utterance is: %s" % (len(vocab.word2id)))
logger.info("Vocab size for output slot label is: %s" % (len(vocab.slot2id)))
logger.info("Vocab size for output intent is: %s" % (len(vocab.int2id)))
evaluator = Evaluator.get_evaluator_from_task(task=task, vocab=vocab)
if not opt.testing:
train_dataset, dev_dataset = read_dataset(
opt.dataset, choice='train'), read_dataset(opt.dataset, choice='valid')
train_dataset, _ = split_dataset(train_dataset, opt.labeled)
logger.info("Train and dev dataset size is: %s and %s" %
(len(train_dataset), len(dev_dataset)))
test_dataset = read_dataset(opt.dataset, choice='test')
logger.info("Test dataset size is: %s" % (len(test_dataset)))
##### Model Construction and Init #####
if not opt.testing:
opt.vocab_size, opt.slot_num, opt.intent_num = len(vocab.word2id), len(
vocab.slot2id), len(vocab.int2id)
opt.pad_token_idxs = {
"word": vocab.word2id[PAD],
"slot": vocab.slot2id[PAD]
}
opt.start_idx, opt.end_idx = vocab.slot2id[BOS], vocab.slot2id[EOS]
params = vars(opt)
def main(image_folder, model_folder, building_folder, figure_folder, info_csv,
stats_json_file, train_results, batch_size, ideal_batch_size, epochs,
device_name):
# Configure folders
image_folder = Path(image_folder)
model_folder = Path(model_folder)
building_folder = Path(building_folder)
stats_json_file = Path(stats_json_file)
figure_folder = Path(figure_folder)
# Seed program
torch.manual_seed(0)
# retrieve information on buildings
df = pd.read_csv(info_csv)
# Get information on a TIF
n_ch, h, w = get_tif_dims(list(image_folder.iterdir())[0])
# Get means and stds
if stats_json_file.exists():
mean_channels, std_channels = load_stats(stats_json_file)
else:
stats = compute_mean_std(image_folder, n_ch)
mean_channels = stats['mean']
std_channels = stats["std"]
with open(stats_json_file, 'w') as file:
json.dump(stats, file)
# New dims for image
pad_h, pad_w = 8 - h % 8, 8 - w % 8
# Prepare padding
if pad_h % 2 == 0:
top_p = bottom_p = pad_h // 2
else:
top_p, bottom_p = pad_h // 2, pad_h - (pad_h // 2)
if pad_w % 2 == 0:
left_p = right_p = pad_w // 2
else:
left_p, right_p = pad_w // 2, pad_w - (pad_w // 2)
img_pad = (left_p, top_p, right_p, bottom_p)
# Compose transforms
transforms = SemSegCompose(img_pad, mean_channels, std_channels, 360)
# Prepare loading function
# load_tif_with_mask = partial(
# load_tif,
# df=df,
# mean_vec=mean_channels,
# std_vec=std_channels,
# building_folder=building_folder,
# padding=(pad_h, pad_w))
# Make dataset
ds = MulPanSharpenDataset(image_folder,
building_folder,
df,
transforms=transforms)
# ds = datasets.DatasetFolder(root=image_folder,
# loader=load_tif_with_mask,
# extensions=('.tif',))
logger.info(f"N° of images: {len(ds)}")
logger.info(f"Type of img: {ds.label}")
train_ds, val_ds = split_dataset(ds, train_size=0.8)
logger.info(f"Train set size: {len(train_ds)}")
logger.info(f"Val set size: {len(val_ds)}")
train_dl = DataLoader(train_ds, batch_size=batch_size, shuffle=True)
val_dl = DataLoader(val_ds, batch_size=batch_size, shuffle=True)
logger.info(f"N° of iterations per batch (train): {len(train_dl)}")
logger.info(f"N° of iterations per batch (val): {len(val_dl)}")
# get model
logger.info("Getting U-Net model")
model = torch.hub.load('mateuszbuda/brain-segmentation-pytorch',
'unet',
in_channels=n_ch,
out_channels=2,
init_features=32,
pretrained=False)
device = torch.device(device_name)
logger.info(f"Mounting model on {device}")
model = model.to(device)
# Define criterion
logger.info("Defining cross-entropy loss with 11/89 ratio")
criterion = nn.CrossEntropyLoss(weight=torch.tensor([.11, .89]))
criterion = criterion.to(device)
# Define optimizer
logger.info("Defining Adam optimizer")
optimizer = optim.Adam(model.parameters(), lr=1e-3)
results = train(model, train_dl, val_dl, criterion, optimizer, epochs,
device, 2, model_folder, ideal_batch_size // batch_size)
# Saves model metrics as pickle file
with open(train_results, "wb") as f:
pickle.dump(results, f)
# Saves model
torch.save(model.state_dict(), model_folder / 'unet_model')
logger.info(f"Saved model at {model_folder / 'unet_model'}")
logging.info(f"Metrics evaluation. Check {figure_folder} for results.")
# Plot metrics
plot_loss(results, figure_folder)
logger.info("Loss curve created.")
plot_last_cm(results, figure_folder)
logger.info("Last confusion matrix created")
plot_correct_preds(results, figure_folder)
logger.info("Evolution of correct predictions created;")
plot_accuracy(results, figure_folder)
logger.info("Accuracy plot created.")
plot_iou(results, figure_folder)
logger.info("IoU evolution plot created.")
for i in range(5):
generate_masks(ds, model, figure_folder, i, 5)
logger.info("Created model results.")
#!/usr/bin/env python3
# Standard import
import os
import json
# Third-party import
import torch
# Local import
from invco import DATASET_DIR, ROOT_DIR
from utils.dataset import Recipe1M
from utils.dataset import split_dataset
if '__main__' == __name__:
# dataset = Recipe1M(DATASET_DIR)
# split_csv(F'{DATASET_DIR}/recipeid.csv')
split_dataset(F'{DATASET_DIR}/recipeid.csv', F'{ROOT_DIR}/csv',
lambda r: r[4 - 1] + '.csv')
