def eval(txt, cluster_num, round, device):
eval_dataset = Eval_Dataset('images.txt', transforms.ToTensor())
eval_dataloader = DataLoader(eval_dataset)
net = SiameseNetwork().to(device)
net.load_state_dict(
torch.load("parameters/Parm_Round_" + str(round + 1) + ".pth"))
features = []
for img in eval_dataloader:
with torch.no_grad():
img = img.to(device)
output = net.forward_once(img).data.cpu().numpy().squeeze()
features.append(output)
X, new_labellist = feature2label(features, cluster_num)
return X, new_labellist
def train(labellist, batch_size, train_number_epochs, learning_rate, round,
device):
train_dataset = SiameseDataset('images.txt', labellist,
transforms.ToTensor())
train_dataloader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
# training
net = SiameseNetwork().to(device)
criterion = ContrastiveLoss()
optimizer = optim.Adam(net.parameters(), lr=learning_rate)
counter = []
loss_history = []
iteration_number = 0
for epoch in range(train_number_epochs):
total_loss = 0
start_time = datetime.now()
for i, data in enumerate(train_dataloader):
img0, img1, label = data
img0, img1, label = img0.to(device), img1.to(device), label.to(
device)
optimizer.zero_grad()
output1, output2 = net(img0, img1)
loss_contrastive = criterion(output1, output2, label)
loss_contrastive.backward()
total_loss += loss_contrastive.item()
optimizer.step()
if i % 20 == 0:
iteration_number += 20
counter.append(iteration_number)
loss_history.append(loss_contrastive.item())
end_time = datetime.now()
print("Epoch number: {} , Current loss: {:.4f}, Epoch Time: {}".format(
epoch + 1, total_loss / (i + 1), end_time - start_time))
torch.save(net.state_dict(),
"parameters/Parm_Round_" + str(round + 1) + ".pth")
return counter, loss_history
def __init__(self, path_wieght: str, path_data: str, similarity,
path_feat: str):
self.path_weight = path_wieght
self.path_data = path_data
self.similarity = similarity
self.flickr_dataset = ImageFlickrFeatures(
path_feat) #dbs/features_contrastive.db
# self.ranking = ranking
imagenet_net = ResNet34()
sketches_net = ResNet34()
# print("Adapting output layers...")
siamese_net = SiameseNetwork(sketches_net, imagenet_net)
siamese_net.load_state_dict(
torch.load(self.path_weight)
) # r'C:\Users\aleja\Desktop\Tareas\Reconocimiento Virtual con Deep Learning\T2\best_SiameseNetwork_contrastive.pth'
self.net = siamese_net
self.ranking = Ranker(self.path_data,
image_dataset_features=self.flickr_dataset,
feature_extractor=self.net,
similarity_fn=self.similarity)
def main():
"""Interface for training and evaluating using the command line"""
global args
args = parser.parse_args()
model = SiameseNetwork(1, args.embedding_size)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# If a checkpoint provided, load it's values
if args.checkpoint:
state = torch.load(args.checkpoint, map_location=device)
model.load_state_dict(state['state_dict'])
else:
state = None
# Run the model on a GPU if available
model.to(device)
# Train the network
if args.mode == 'train':
dataset = GEDDataset(args.data, which_set='train', adj_dtype=np.float32, transform=None)
model, optimiser, epoch = train(model, dataset, batch_size=args.batch_size, embed_size=args.embedding_size, num_epochs=args.epochs,
learning_rate=args.learning_rate, save_to=args.save_dir, resume_state=args.checkpoint, device=device)
if args.save_dir:
# Save the model checkpoint
state = {
'epoch': epoch,
'state_dict': model.state_dict(),
'optimiser': optimiser.state_dict(),
}
save_checkpoint(state, args.save_dir)
# Whether to store the predictions from eval for plotting
store_res = args.make_plot
if args.mode == 'train' and args.post_training_eval:
args.which_set = 'val'
if args.mode == 'eval' or args.post_training_eval:
dataset = GEDDataset(args.data, which_set=args.which_set, adj_dtype=np.float32, transform=None)
results = eval(model, dataset, batch_size=args.batch_size, store_results=store_res, device=device)
# Finally, if plotting the results:
if args.make_plot:
# Assert that the data has been evaluated
if not (args.mode == 'eval' or args.post_training_eval):
raise AttributeError('The flags provided did not specify to evaluate the dataset, which is required for'
'plotting')
# Make a plot of the results
print('Making the plot')
plot_prediction(results[0], results[1])
def __init__(self) -> None:
self.now = datetime.now().strftime("%Y.%m.%d-%H:%M:%S")
self.device = torch.device('cuda:0')
# Initializing window dataset hiperparams
self.window_size = 40
self.num_test_person = 20
# Initializing window datasets
self.train_window_dataset = SiameseDataset(
dataset='train',
is_train=False,
window_size=self.window_size,
num_test_person=self.num_test_person).base_dataset
self.val_window_dataset = SiameseDataset(
dataset="val",
is_train=False,
window_size=self.window_size,
num_test_person=self.num_test_person).base_dataset
self.test_window_dataset = SiameseDataset(
dataset='test', is_train=False,
window_size=self.window_size).base_dataset
# Initializing the trained model
#self.model_checkpoint = 'model_params/siamese_network/version_5_bs_256_baselr_1e-3_multisteplr_milestones_5-15-25_gamma_0.33_lstmoutsize_170-latentsize_170-outfcsize_32/version-5-siamese-epoch=28-val_acc_epoch=0.9355.ckpt'
self.model_checkpoint = 'model_params/siamese_network/version_6_bs_512_baselr_3e-3_multisteplr_milestones_5-10-15-25_gamma_0.33_lstmoutsize_160-latentsize_32-outfcsize_16-abs+cos+sed/version-6-siamese-epoch=33-val_acc_epoch=0.9358.ckpt'
self.model = SiameseNetwork().load_from_checkpoint(
checkpoint_path=self.model_checkpoint).to(
device=self.device).eval()
# Creating core database
self.averaging_types = ['mean', 'median']
self.averaging_type = self.averaging_types[0]
#self.database = self._create_database()
self.metric_method = 'micro'
self.similarity_thres = 0.95
self.file_log_path = f'test_logs/{self.now}--{self.model_checkpoint.rsplit("/")[-2]}--{self.model_checkpoint.rsplit("/")[-1]}/{self.now}-simthres_{self.similarity_thres}-avgtype_{self.averaging_type}/'
self.metrics_file_path = self.file_log_path + 'metrics.txt'
n_val = int(len(dataset) * args.val_size)
n_train = len(dataset) - n_val
train, val = random_split(dataset, [n_train, n_val])
train_loader = DataLoader(train, batch_size=args.batch_size)
val_loader = DataLoader(val, batch_size=args.batch_size // 4)
# load backbones
print("[*] Initializing weights...")
imagenet_net = ResNet34()
sketches_net = ResNet34()
# sketches_net.load_state_dict(torch.load(args.sketches_backbone_weights))
print("[+] Weights loaded")
print("[*] Initializing model, loss and optimizer")
contrastive_net = SiameseNetwork(sketches_net, imagenet_net)
contrastive_net.to(args.device)
if args.optimizer == 'sgd':
optimizer = torch.optim.SGD(contrastive_net.parameters(),
lr=args.lr,
momentum=args.momentum)
else:
optimizer = torch.optim.Adam(contrastive_net.parameters(), lr=args.lr)
scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
optimizer, T_0=args.t_0)
contrastive_loss = contrastive_loss()
cross_entropy_loss = torch.nn.CrossEntropyLoss()
print("[+] Model, loss and optimizer were initialized successfully")
if not args.debug:
wandb.init(project='homework1-cc7221', entity='p137')
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
import tqdm
from torch.utils.data import DataLoader
from config import Config
from models import AutoEncoder, SiameseNetwork
config = Config()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
autoencoder = AutoEncoder(config)
siamese_network = SiameseNetwork(config)
autoencoder_file = '/autoencoder_epoch175_loss1.1991.pth'
siamese_file = '/siamese_network_epoch175_loss1.1991.pth'
if config.load_model:
autoencoder.load_state_dict(torch.load(config.saved_models_folder + autoencoder_file))
siamese_network.load_state_dict(torch.load(config.saved_models_folder + siamese_file))
autoencoder.to(device)
autoencoder.train()
siamese_network.to(device)
siamese_network.train()
params = list(autoencoder.parameters()) + list(siamese_network.parameters())
description="Train model utility",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--siamese-weights',
default='weights/best_SiameseNetwork_triplet.pth',
type=str,
help='Path to Siamese network weights')
parser.add_argument('--flickr-15k',
default='data/Flickr15K',
type=str,
help='Path to flickr dataset folder')
parser.add_argument('--output',
default='features.db',
type=str,
help='Output path of feature db')
args = parser.parse_args()
imagenet_net = ResNet34()
sketches_net = ResNet34()
print("[*] Adapting output layers and loading weights...")
siamese_net = SiameseNetwork(sketches_net, imagenet_net)
siamese_net.load_state_dict(torch.load(args.siamese_weights))
print("[+] Done!")
dataset = ImageFlickr15K(path=args.flickr_15k)
saver = ImageFlickrSaver(args.output,
model=siamese_net,
image_dataset=dataset)
saver.export()
train_dataset = MadoriDataset(
) if Config.network == 'siamese' else TriMadoriDataset()
val_dataset = MadoriDataset(
train=False) if Config.network == 'siamese' else TriMadoriDataset(
train=False)
# data loaders
train_dataloader = DataLoader(train_dataset,
shuffle=True,
batch_size=Config.batch_size)
val_dataloader = DataLoader(val_dataset,
shuffle=False,
batch_size=Config.batch_size)
# models
net = SiameseNetwork() if Config.network == 'siamese' else TripletNetwork()
net = net.to(device)
criterion = ContrastiveLoss() if Config.network == 'siamese' else TripletLoss()
optimizer = optim.Adam(net.parameters(), lr=0.0005)
def train_siamese():
train_loss_history, val_loss_history = [], []
lowest_epoch_train_loss = lowest_epoch_val_loss = float('inf')
for epoch in tqdm(range(Config.train_number_epochs)):
# training
net.train()
epoch_train_loss = 0
for batch_no, data in enumerate(train_dataloader):
img0, img1, label = data
labels = list(labels.detach().cpu().numpy())
# group similarities
group_similarities = list(zip(list(paths), labels, distances))
similarities.extend(group_similarities)
similarities = sorted(similarities, key=lambda x: x[2], reverse=True)
return query_class, similarities
if __name__ == '__main__':
from models import ResNet34
from similarities import CosineSimilarity
from feats import ImageFlickrFeatures
imagenet_net = ResNet34()
sketches_net = ResNet34()
siamese_net = SiameseNetwork(sketches_net, imagenet_net)
siamese_net.load_state_dict(torch.load("weights/best_SiameseNetwork_contrastive.pth"))
s = torch.nn.CosineSimilarity()
flickr_dataset = ImageFlickrFeatures("dbs/features.db")
r = Ranker('/home/rudy/Documents/cc7221/tarea2/data/Flickr15K',
image_dataset_features=flickr_dataset,
feature_extractor=siamese_net,
similarity_fn=s)
rank = r.get_rank('data/Flickr15K/queries/1001.png')
import torchvision.transforms as transforms
import tqdm
from torch.utils.data import DataLoader
from config import Config
from models import Encoder, SiameseNetwork
batch_size = 8
threshold = 0.9
config = Config()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
encoder = Encoder(config)
siamese_network = SiameseNetwork(config)
encoder.load_state_dict(
torch.load(config.saved_models_folder +
'/encoder_epoch500_loss0.0009.pth'))
encoder.to(device)
encoder.eval()
siamese_network.load_state_dict(
torch.load(config.saved_models_folder +
'/siamese_network_epoch500_loss0.0009.pth'))
siamese_network.to(device)
siamese_network.eval()
transform = transforms.Compose(
[transforms.Grayscale(num_output_channels=1),
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
import tqdm
from torch.utils.data import DataLoader
from config import Config
from models import Encoder, SiameseNetwork
config = Config()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
encoder = Encoder(config)
siamese_network = SiameseNetwork(config)
if config.load_model:
encoder.load_state_dict(
torch.load(config.saved_models_folder + '/encoder_epoch25.pth'))
siamese_network.load_state_dict(
torch.load(config.saved_models_folder +
'/siamese_network_epoch25.pth'))
encoder.to(device)
encoder.train()
siamese_network.to(device)
siamese_network.train()
params = list(encoder.parameters()) + list(siamese_network.parameters())
train, val = random_split(dataset, [n_train, n_val])
train_loader = DataLoader(train, batch_size=args.batch_size)
val_loader = DataLoader(val, batch_size=args.batch_size)
print("[+] Dataset initialized successfully")
# load backbones
print("[*] Initializing weights...")
imagenet_net = ResNet34()
sketches_net = ResNet34()
# sketches_net.load_state_dict(torch.load(args.sketches_backbone_weights))
print("[+] Weights loaded")
print("[*] Adapting output layers...")
print("[*] Initializing model, loss and optimizer")
siamese_net = SiameseNetwork(sketches_net, imagenet_net)
siamese_net.to(args.device)
optimizer = torch.optim.Adam(siamese_net.parameters(), lr=args.lr)
triplet_loss = triplet_loss()
cross_entropy_loss = torch.nn.CrossEntropyLoss()
print("[+] Model, loss and optimizer were initialized successfully")
if not args.debug:
wandb.init(project='homework1-cc7221', entity='p137')
config = wandb.config
config.model = siamese_net.__class__.__name__ + "_triplet"
config.device = device
config.batch_size = args.batch_size
config.epochs = args.epochs
config.learning_rate = args.lr
# 5) eval
if args.eval_mode:
# +) eval dataset
print('========== eval dataset ==========')
eval_dataset = data_utils.Dataset(track=args.track,
data='eval',
size=args.input_size,
feature=args.feature,
tag=args.data_tag)
eval_loader = DataLoader(eval_dataset,
batch_size=args.eval_batch_size,
shuffle=False,
num_workers=8)
# +) load model
print('========== eval process ==========')
model = SiameseNetwork(args.embedding_size).to(device)
eval_checkpoint_path = '{}/epoch_{}.pth'.format(
model_save_path, str(args.eval_num_checkpoint))
model.load_state_dict(torch.load(eval_checkpoint_path))
print('model loaded from ', eval_checkpoint_path)
# +) eval
eval_output_path = '{}/{}.result'.format(model_save_path,
str(args.eval_num_checkpoint))
evaluate(eval_dataset, eval_loader, model, device, eval_output_path)
print('eval output saved to ', eval_output_path)
# 6) train & dev
else:
# +) dev dataset
print('========== dev dataset ==========')
dev_dataset = data_utils.Dataset(track=args.track,
import torchvision.transforms as transforms
import tqdm
from torch.utils.data import DataLoader
from config import Config
from models import AutoEncoder, SiameseNetwork
batch_size = 100
threshold = 0.7
config = Config()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
autoencoder = AutoEncoder(config)
siamese_network = SiameseNetwork(config)
autoencoder_file = '/autoencoder_epoch225_loss0.7295.pth'
siamese_file = '/siamese_network_epoch225_loss0.7295.pth'
autoencoder.load_state_dict(
torch.load(config.saved_models_folder + autoencoder_file))
autoencoder.to(device)
autoencoder.train()
siamese_network.load_state_dict(
torch.load(config.saved_models_folder + siamese_file))
siamese_network.to(device)
siamese_network.train()
transform = transforms.Compose(
n_epochs = args.n_epochs
batches = args.batchs
batch_size = args.batch_size
net = args.net
dataset = args.dataset
data_root = args.data_root
save_dir = 'outputs/{}-way-{}-shot_{}_{}/segment'.format(class_per_set,
sample_per_class,
net,
dataset)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
m = SiameseNetwork(network_type=args.net).cuda()
m.load_state_dict(torch.load(args.resume))
CNN_net = m._modules['net']
dice_all1 = np.zeros((22,2))
dice_all2 = np.zeros((22,2))
for i in range(0, 22):
print("CLASS -- ", i)
vols = np.load(os.path.join(data_root, 'testtom_vol_{}.npy'.format(str(i).zfill(2))))
segs = np.load(os.path.join(data_root, 'testtom_seg_{}.npy'.format(str(i).zfill(2))))
vols = preprocess_image(vols)
dice_c1 = np.zeros((segs.shape[0], 1))
dice_c2 = np.zeros((segs.shape[0], 1))
for j in range(segs.shape[0]):
vol = vols[:,:,j,:,:,:]
num_output_classes=num_output_classes,
hidden_size=args.lstm_hidden_dim,
attention_kernel_size=args.encoder_output_size,
use_bias=True,
num_att_layers=2,
embedding_matrix=embedding_matrix)
model_2 = StackedAttentionNetwork(desc_input_shape=(args.batch_size, 102),
img_input_shape=(args.batch_size, 512, 14, 14),
num_output_classes=2,
hidden_size=args.lstm_hidden_dim,
attention_kernel_size=args.encoder_output_size,
use_bias=True,
num_att_layers=2,
embedding_matrix=embedding_matrix)
siamese_model = SiameseNetwork(item_1_model=model_1, item_2_model=model_2, encoder_output_size=args.encoder_output_size,
fc1_size=args.fc1_size, fc2_size=args.fc2_size, use_bias=True)
experiment = ExperimentBuilder(network_model=siamese_model,
experiment_name=args.experiment_name,
num_epochs=args.num_epochs,
learning_rate=args.lr,
weight_decay_coefficient=args.weight_decay_coefficient,
continue_from_epoch=args.continue_from_epoch,
device=device,
train_data=training_data_loader,
val_data=valid_data_loader,
test_data=test_data_loader) # build an experiment object
experiment_metrics, test_metrics = experiment.run_experiment() # run experiment and return experiment metrics
dataset = args.dataset
data_root = args.data_root
if not os.path.exists('outputs/{}-way-{}-shot_{}_{}'.format(class_per_set,
sample_per_class,
net,
dataset)):
os.makedirs('outputs/{}-way-{}-shot_{}_{}'.format(class_per_set,
sample_per_class,
net,
dataset))
data = SubtomogramNShotDataset(data_root=data_root,
batch_size=batch_size, class_per_set=class_per_set, sample_per_class=sample_per_class)
model = SiameseNetwork(network_type=args.net).cuda()
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr, betas=(args.beta1, args.beta2), weight_decay=args.weight_decay)
# training phase
continue_tarining = False
if continue_tarining:
pre_model = ''
print("Using Model ", pre_model, " to continue training!")
model.load_state_dict(torch.load(pre_model))
print("====================Start Training:====================")
print("===================={}-Way-{}-Shot-{}-{} Learning====================".format(class_per_set,
sample_per_class,
net,
dataset))
