tol=1e-3,
average=10)
clf.fit(X_train, y_train)
save_model(clf, MODEL, DATASET)
prediction = clf.predict(scaler.transform(img.reshape(-1, N_BANDS)))
prediction = prediction.reshape(img.shape[:2])
else: # 自定义算法
# Neural network
model, optimizer, loss, hyperparams = get_model(MODEL, **hyperparams)
if CLASS_BALANCING: #
weights = compute_imf_weights(train_gt, N_CLASSES, IGNORED_LABELS)
hyperparams['weights'] = torch.from_numpy(weights)
# Split train set in train/val
# train_gt, val_gt = sample_gt(train_gt, 0.9, mode='random')
# Generate the dataset
train_dataset = HyperX(img, train_gt, **hyperparams)
train_loader = data.DataLoader(train_dataset,
batch_size=hyperparams['batch_size'],
pin_memory=hyperparams['device'],
shuffle=True)
# val_dataset = HyperX(img, val_gt, **hyperparams)
# val_loader = data.DataLoader(val_dataset,
# pin_memory=hyperparams['device'],
# batch_size=hyperparams['batch_size'])
print(hyperparams)
print("Network :")
with torch.no_grad():
for input, _ in train_loader:
break
# summary(model.to(hyperparams['device']), input.size()[1:], device=hyperparams['device'])
clf, {'n_neighbors': [1, 3, 5, 10, 20]}, verbose=5, n_jobs=4)
clf.fit(X_train, y_train)
clf.fit(X_train, y_train)
save_model(clf, MODEL, DATASET)
prediction = clf.predict(img.reshape(-1, N_BANDS))
prediction = prediction.reshape(img.shape[:2])
else:
# Neural network
model, optimizer, loss, hyperparams = get_model(MODEL, **hyperparams)
if CLASS_BALANCING:
weights = compute_imf_weights(train_gt, N_CLASSES, IGNORED_LABELS)
hyperparams['weights'] = torch.from_numpy(weights)
# Split train set in train/val
train_gt, val_gt = sample_gt(train_gt, 0.95, mode='random')
# Generate the dataset
train_dataset = HyperX(img, train_gt, **hyperparams) # 生成用于训练的数据方块和标签
train_loader = data.DataLoader(
train_dataset,
batch_size=hyperparams['batch_size'],
#pin_memory=hyperparams['device'],
shuffle=True)
val_dataset = HyperX(img, val_gt, **hyperparams) # 生成用于测试的数据方块和标签
val_loader = data.DataLoader(
val_dataset,
#pin_memory=hyperparams['device'],
batch_size=hyperparams['batch_size'])
print(hyperparams)
print("Network :")
with torch.no_grad():
for input, _ in train_loader:
def main(raw_args=None):
parser = argparse.ArgumentParser(
description="Hyperspectral image classification with FixMatch")
parser.add_argument(
'--patch_size',
type=int,
default=5,
help='Size of patch around each pixel taken for classification')
parser.add_argument(
'--center_pixel',
action='store_false',
help=
'use if you only want to consider the label of the center pixel of a patch'
)
parser.add_argument('--batch_size',
type=int,
default=10,
help='Size of each batch for training')
parser.add_argument('--epochs',
type=int,
default=10,
help='number of total epochs of training to run')
parser.add_argument('--dataset',
type=str,
default='Salinas',
help='Name of dataset to run, Salinas or PaviaU')
parser.add_argument('--cuda',
type=int,
default=-1,
help='what CUDA device to run on, -1 defaults to cpu')
parser.add_argument('--warmup',
type=float,
default=0,
help='warmup epochs')
parser.add_argument('--save',
action='store_true',
help='use to save model weights when running')
parser.add_argument(
'--test_stride',
type=int,
default=1,
help='length of stride when sliding patch window over image for testing'
)
parser.add_argument(
'--sampling_percentage',
type=float,
default=0.3,
help=
'percentage of dataset to sample for training (labeled and unlabeled included)'
)
parser.add_argument(
'--sampling_mode',
type=str,
default='nalepa',
help='how to sample data, disjoint, random, nalepa, or fixed')
parser.add_argument('--lr',
type=float,
default=0.001,
help='initial learning rate')
parser.add_argument('--alpha',
type=float,
default=1.0,
help='beta distribution range')
parser.add_argument(
'--class_balancing',
action='store_false',
help='use to balance weights according to ratio in dataset')
parser.add_argument(
'--checkpoint',
type=str,
default=None,
help='use to load model weights from a certain directory')
#Augmentation arguments
parser.add_argument('--flip_augmentation',
action='store_true',
help='use to flip augmentation data for use')
parser.add_argument('--radiation_augmentation',
action='store_true',
help='use to radiation noise data for use')
parser.add_argument('--mixture_augmentation',
action='store_true',
help='use to mixture noise data for use')
parser.add_argument('--pca_augmentation',
action='store_true',
help='use to pca augment data for use')
parser.add_argument(
'--pca_strength',
type=float,
default=1.0,
help='Strength of the PCA augmentation, defaults to 1.')
parser.add_argument('--cutout_spatial',
action='store_true',
help='use to cutout spatial for data augmentation')
parser.add_argument('--cutout_spectral',
action='store_true',
help='use to cutout spectral for data augmentation')
parser.add_argument(
'--augmentation_magnitude',
type=int,
default=1,
help=
'Magnitude of augmentation (so far only for cutout). Defualts to 1, min 1 and max 10.'
)
parser.add_argument('--spatial_combinations',
action='store_true',
help='use to spatial combine for data augmentation')
parser.add_argument('--spectral_mean',
action='store_true',
help='use to spectal mean for data augmentation')
parser.add_argument(
'--moving_average',
action='store_true',
help='use to sprectral moving average for data augmentation')
parser.add_argument('--results',
type=str,
default='results',
help='where to save results to (default results)')
parser.add_argument('--save_dir',
type=str,
default='/saves/',
help='where to save models to (default /saves/)')
parser.add_argument('--data_dir',
type=str,
default='/data/',
help='where to fetch data from (default /data/)')
parser.add_argument('--load_file',
type=str,
default=None,
help='wihch file to load weights from (default None)')
parser.add_argument(
'--fold',
type=int,
default=0,
help='Which fold to sample from if using Nalepas validation scheme')
parser.add_argument(
'--sampling_fixed',
type=str,
default='True',
help=
'Use to sample a fixed amount of samples for each class from Nalepa sampling'
)
parser.add_argument(
'--samples_per_class',
type=int,
default=10,
help=
'Amount of samples to sample for each class when sampling a fixed amount. Defaults to 10.'
)
parser.add_argument(
'--supervision',
type=str,
default='full',
help=
'check this more, use to make us of all labeled or not, full or semi')
args = parser.parse_args(raw_args)
device = utils.get_device(args.cuda)
args.device = device
#vis = visdom.Visdom()
vis = None
tensorboard_dir = str(args.results + '/' +
datetime.datetime.now().strftime("%m-%d-%X"))
os.makedirs(tensorboard_dir, exist_ok=True)
writer = SummaryWriter(tensorboard_dir)
if args.sampling_mode == 'nalepa':
train_img, train_gt, test_img, test_gt, label_values, ignored_labels, rgb_bands, palette = get_patch_data(
args.dataset,
args.patch_size,
target_folder=args.data_dir,
fold=args.fold)
args.n_bands = train_img.shape[-1]
else:
img, gt, label_values, ignored_labels, rgb_bands, palette = get_dataset(
args.dataset, target_folder=args.data_dir)
args.n_bands = img.shape[-1]
args.n_classes = len(label_values) - len(ignored_labels)
args.ignored_labels = ignored_labels
if palette is None:
# Generate color palette
palette = {0: (0, 0, 0)}
for k, color in enumerate(
sns.color_palette("hls",
len(label_values) - 1)):
palette[k + 1] = tuple(
np.asarray(255 * np.array(color), dtype='uint8'))
invert_palette = {v: k for k, v in palette.items()}
def convert_to_color(x):
return utils.convert_to_color_(x, palette=palette)
def convert_from_color(x):
return utils.convert_from_color_(x, palette=invert_palette)
if args.sampling_mode == 'nalepa':
print("{} samples selected (over {})".format(
np.count_nonzero(train_gt),
np.count_nonzero(train_gt) + np.count_nonzero(test_gt)))
writer.add_text(
'Amount of training samples',
"{} samples selected (over {})".format(np.count_nonzero(train_gt),
np.count_nonzero(test_gt)))
utils.display_predictions(convert_to_color(test_gt),
vis,
writer=writer,
caption="Test ground truth")
else:
train_gt, test_gt = utils.sample_gt(gt,
args.sampling_percentage,
mode=args.sampling_mode)
print("{} samples selected (over {})".format(
np.count_nonzero(train_gt), np.count_nonzero(gt)))
writer.add_text(
'Amount of training samples',
"{} samples selected (over {})".format(np.count_nonzero(train_gt),
np.count_nonzero(gt)))
utils.display_predictions(convert_to_color(train_gt),
vis,
writer=writer,
caption="Train ground truth")
utils.display_predictions(convert_to_color(test_gt),
vis,
writer=writer,
caption="Test ground truth")
model = HamidaEtAl(args.n_bands,
args.n_classes,
patch_size=args.patch_size)
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
nesterov=True,
weight_decay=0.0005)
#loss_labeled = nn.CrossEntropyLoss(weight=weights)
#loss_unlabeled = nn.CrossEntropyLoss(weight=weights, reduction='none')
if args.sampling_mode == 'nalepa':
#Get fixed amount of random samples for validation
idx_sup, idx_val, idx_unsup = get_pixel_idx(train_img, train_gt,
args.ignored_labels,
args.patch_size)
if args.sampling_fixed == 'True':
unique_labels = np.zeros(len(label_values))
new_idx_sup = []
index = 0
for p, x, y in idx_sup:
label = train_gt[p, x, y]
if unique_labels[label] < args.samples_per_class:
unique_labels[label] += 1
new_idx_sup.append([p, x, y])
np.delete(idx_sup, index)
index += 1
idx_unsup = np.concatenate((idx_sup, idx_unsup))
idx_sup = np.asarray(new_idx_sup)
writer.add_text(
'Amount of labeled training samples',
"{} samples selected (over {})".format(idx_sup.shape[0],
np.count_nonzero(train_gt)))
train_labeled_gt = [
train_gt[p_l, x_l, y_l] for p_l, x_l, y_l in idx_sup
]
samples_class = np.zeros(args.n_classes)
for c in np.unique(train_labeled_gt):
samples_class[c - 1] = np.count_nonzero(train_labeled_gt == c)
writer.add_text('Labeled samples per class', str(samples_class))
print('Labeled samples per class: ' + str(samples_class))
val_dataset = HyperX_patches(train_img,
train_gt,
idx_val,
labeled='Val',
**vars(args))
val_loader = data.DataLoader(val_dataset, batch_size=args.batch_size)
train_dataset = HyperX_patches(train_img,
train_gt,
idx_sup,
labeled=True,
**vars(args))
train_loader = data.DataLoader(
train_dataset,
batch_size=args.batch_size,
#pin_memory=True, num_workers=5,
shuffle=True,
drop_last=True)
amount_labeled = idx_sup.shape[0]
else:
train_labeled_gt, val_gt = utils.sample_gt(train_gt,
0.95,
mode=args.sampling_mode)
val_dataset = HyperX(img, val_gt, labeled='Val', **vars(args))
val_loader = data.DataLoader(val_dataset, batch_size=args.batch_size)
writer.add_text(
'Amount of labeled training samples',
"{} samples selected (over {})".format(
np.count_nonzero(train_labeled_gt),
np.count_nonzero(train_gt)))
samples_class = np.zeros(args.n_classes)
for c in np.unique(train_labeled_gt):
samples_class[c - 1] = np.count_nonzero(train_labeled_gt == c)
writer.add_text('Labeled samples per class', str(samples_class))
train_dataset = HyperX(img,
train_labeled_gt,
labeled=True,
**vars(args))
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
pin_memory=True,
num_workers=5,
shuffle=True,
drop_last=True)
utils.display_predictions(convert_to_color(train_labeled_gt),
vis,
writer=writer,
caption="Labeled train ground truth")
utils.display_predictions(convert_to_color(val_gt),
vis,
writer=writer,
caption="Validation ground truth")
amount_labeled = np.count_nonzero(train_labeled_gt)
args.iterations = amount_labeled // args.batch_size
args.total_steps = args.iterations * args.epochs
args.scheduler = get_cosine_schedule_with_warmup(
optimizer, args.warmup * args.iterations, args.total_steps)
if args.class_balancing:
weights_balance = utils.compute_imf_weights(train_gt,
len(label_values),
args.ignored_labels)
args.weights = torch.from_numpy(weights_balance[1:])
args.weights = args.weights.to(torch.float32)
else:
weights = torch.ones(args.n_classes)
#weights[torch.LongTensor(args.ignored_labels)] = 0
args.weights = weights
args.weights = args.weights.to(args.device)
criterion = nn.CrossEntropyLoss(weight=args.weights)
loss_val = nn.CrossEntropyLoss(weight=args.weights)
print(args)
print("Network :")
writer.add_text('Arguments', str(args))
with torch.no_grad():
for input, _ in train_loader:
break
#summary(model.to(device), input.size()[1:])
#writer.add_graph(model.to(device), input)
# We would like to use device=hyperparams['device'] altough we have
# to wait for torchsummary to be fixed first.
if args.load_file is not None:
model.load_state_dict(torch.load(args.load_file))
model.zero_grad()
try:
train(model,
optimizer,
criterion,
loss_val,
train_loader,
writer,
args,
val_loader=val_loader,
display=vis)
except KeyboardInterrupt:
# Allow the user to stop the training
pass
if args.sampling_mode == 'nalepa':
probabilities = test(model, test_img, args)
else:
probabilities = test(model, img, args)
prediction = np.argmax(probabilities, axis=-1)
run_results = utils.metrics(prediction,
test_gt,
ignored_labels=args.ignored_labels,
n_classes=args.n_classes)
mask = np.zeros(test_gt.shape, dtype='bool')
for l in args.ignored_labels:
mask[test_gt == l] = True
prediction += 1
prediction[mask] = 0
color_prediction = convert_to_color(prediction)
utils.display_predictions(color_prediction,
vis,
gt=convert_to_color(test_gt),
writer=writer,
caption="Prediction vs. test ground truth")
utils.show_results(run_results,
vis,
writer=writer,
label_values=label_values)
writer.close()
return run_results
results = []
# run the experiment several times
for run in range(N_RUNS):
np.random.seed(seeds[run])
# Sample random training spectra
train_gt, test_gt = sample_gt(gt, PERCENTAGE, seed=seeds[run])
# Split train set in train/val
train_gt, val_gt = sample_gt(train_gt, 0.5, seed=seeds[run])
print("Training samples: {}, validating samples: {}, total samples: {})".
format(np.sum(train_gt > -1), np.sum(val_gt > -1), np.sum(gt > -1)))
print("Running an experiment with the {} model".format(MODEL),
"run {}/{}".format(run + 1, N_RUNS))
# data
train_dataset = HyperX(img,
train_gt,
patch_size=PATCH_SIZE,
data_aug=DATA_AUG)
train_loader = data.DataLoader(train_dataset,
batch_size=BATCH_SIZE,
drop_last=False,
shuffle=True)
val_dataset = HyperX(img, val_gt, patch_size=PATCH_SIZE)
val_loader = data.DataLoader(val_dataset,
batch_size=BATCH_SIZE,
drop_last=False)
#load model and loss
model = get_model(MODEL, args.dataset, N_CLASSES, N_BANDS, PATCH_SIZE)
if run == 0:
print("Network :")