def loader(args):
if args.grouping:
if args.random_sampling:
train_loader = torch.utils.data.DataLoader(
data.DATA2(args, mode='train'),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=True)
gallery_loader = torch.utils.data.DataLoader(
data.DATA2(args, mode='gallery'),
batch_size=40,
num_workers=args.workers,
shuffle=False)
query_loader = torch.utils.data.DataLoader(
data.DATA2(args, mode='query'),
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=False)
else:
train_loader = torch.utils.data.DataLoader(
data.DATA(args, mode='train'),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=True)
gallery_loader = torch.utils.data.DataLoader(
data.DATA(args, mode='gallery'),
batch_size=40,
num_workers=args.workers,
shuffle=False)
query_loader = torch.utils.data.DataLoader(
data.DATA(args, mode='query'),
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=False)
else:
train_loader = torch.utils.data.DataLoader(data_ng.DATA(args,
mode='train'),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=True)
gallery_loader = torch.utils.data.DataLoader(data_ng.DATA(
args, mode='gallery'),
batch_size=40,
num_workers=args.workers,
shuffle=False)
query_loader = torch.utils.data.DataLoader(data_ng.DATA(args,
mode='query'),
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=False)
return train_loader, gallery_loader, query_loader
def main(args):
src_train_loader = torch.utils.data.DataLoader(data.DATA(
args, mode='train', dataset=args.source_dataset),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=True)
src_test_loader = torch.utils.data.DataLoader(data.DATA(
args, mode='test', dataset=args.source_dataset),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=False)
feature_extractor, label_predictor, adda_discriminator = models.ADDA(args)
feature_extractor.cuda(), label_predictor.cuda(), adda_discriminator.cuda()
optim = torch.optim.Adam(
list(feature_extractor.parameters()) +
list(label_predictor.parameters()))
lr_schedule = torch.optim.lr_scheduler.ReduceLROnPlateau(optim,
patience=1,
verbose=True)
criterion = torch.nn.CrossEntropyLoss()
best_accuracy = 0
for epoch in range(1, args.epochs + 1):
feature_extractor.train(), label_predictor.train()
train_loss, train_accuracy = do_epoch(feature_extractor,
label_predictor,
src_train_loader,
criterion,
optim=optim)
feature_extractor.eval(), label_predictor.eval()
with torch.no_grad():
val_loss, val_accuracy = do_epoch(feature_extractor,
label_predictor,
src_test_loader,
criterion,
optim=None)
print(
f'EPOCH {epoch:03d}: train_loss={train_loss:.4f}, train_accuracy={train_accuracy:.4f} '
f'val_loss={val_loss:.4f}, val_accuracy={val_accuracy:.4f}')
if val_accuracy > best_accuracy:
print('Saving model...')
best_accuracy = val_accuracy
torch.save(
feature_extractor.state_dict(),
os.path.join(args.save_dir, 'source_feature_extractor.pt'))
torch.save(label_predictor.state_dict(),
os.path.join(args.save_dir, 'label_predictor.pt'))
lr_schedule.step(val_loss)
def train():
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
z_h, images_h = Holder()
D_logits_real, D_logits_fake, D_logits_fake_for_G = \
arch.inference(images_h, z_h)
sampler = eval('arch.' + FLAGS.G_type)(z_h, reuse=True, bn_train=False)
if FLAGS.loss == 'lsgan':
G_loss, D_loss = arch.loss_l2(D_logits_real, D_logits_fake,
D_logits_fake_for_G)
else:
G_loss, D_loss = arch.loss_sigmoid(D_logits_real, D_logits_fake,
D_logits_fake_for_G)
G_vars, D_vars = GetVars()
G_train_op, D_train_op = arch.train(G_loss, D_loss, G_vars, D_vars,
global_step)
data_set = data.DATA(FLAGS.batch_size).load()
sess = sess_init()
tf.train.start_queue_runners(sess=sess)
saver = tf.train.Saver()
for step in xrange(0, FLAGS.max_steps):
z_v, images_v = GenValsForHolder(data_set, sess)
_, errD = sess.run([D_train_op, D_loss],
feed_dict={
z_h: z_v,
images_h: images_v
})
_, errG = sess.run([G_train_op, G_loss], feed_dict={z_h: z_v})
if step % 100 == 0:
print "step = %d, errD = %f, errG = %f" % (step, errD, errG)
if step % 1000 == 0:
samples = sess.run(sampler, feed_dict={z_h: z_v})
save_images(samples, [8, 8],
'./samples/train_{:d}.png'.format(step))
if step <= 10000:
save_images(images_v, [8, 8],
'./samples_real/train_{:d}.png'.format(step))
if step % 10000 == 0:
saver.save(sess,
'{0}/{1}.model'.format(FLAGS.checkpoint_dir,
step), global_step)
def train():
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
z_h, images_h = Holder()
D_logits_real, D_logits_fake, D_logits_fake_for_G = \
arch.inference(images_h, z_h)
sampler = eval('arch.' + FLAGS.G_type)(z_h, reuse=True, bn_train=False)
G_loss, D_loss = arch.loss_l2(D_logits_real, D_logits_fake,
D_logits_fake_for_G)
G_vars, D_vars = GetVars()
G_train_op, D_train_op = arch.train(G_loss, D_loss, G_vars, D_vars,
global_step)
data_set = data.DATA(FLAGS.batch_size).load()
sess = sess_init()
tf.train.start_queue_runners(sess=sess)
saver = tf.train.Saver()
checkpoint_dir = FLAGS.checkpoint_dir
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
ckpt_name = os.path.basename(ckpt.model_checkpoint_path)
saver.restore(sess, os.path.join(checkpoint_dir, ckpt_name))
for step in xrange(100):
z_v, images_v = GenValsForHolder(data_set, sess)
samples = sess.run(sampler, feed_dict={z_h: z_v})
save_images(samples, [8, 8],
'./samples/train_{:d}.png'.format(step))
# gts.append(gt)
# gts, preds = np.concatenate(gts), np.concatenate(preds)
# return mean_iou_score(preds, gts)
if __name__ == '__main__':
args = parser.arg_parse()
''' setup GPU '''
torch.cuda.set_device(args.gpu)
''' prepare data_loader '''
print('===> prepare data loader ...')
test_loader = torch.utils.data.DataLoader(data.DATA(args, mode='test'), # for TA checks my code
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=True)
# test_loader = torch.utils.data.DataLoader(data.DATA(args, mode='val'),
# batch_size=args.train_batch,
# num_workers=args.workers,
# shuffle=True)
''' prepare mode '''
model = models.Baseline_model(args).cuda()
# model = models.deeplabv3p(input_channel=3, num_class=9, output_stride=16).cuda()
''' resume save model '''
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint)
import data
import numpy as np
# import torch.nn as nn
# import torch.nn.functional as F
# import torch.optim as optim
from PIL import Image
if __name__ == '__main__':
args = parser.arg_parse()
''' setup GPU '''
torch.cuda.set_device(args.gpu)
''' prepare data_loader '''
print('===> prepare data loader ...')
save_loader = torch.utils.data.DataLoader(data.DATA(args, mode='save'),
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=False)
''' prepare mode '''
if args.model == 'Net':
model = models.Net(args)
model.cuda() # load model to gpu
if args.model == 'Net_improved':
model = models.Net_improved(args)
model.cuda() # load model to gpu
''' resume save model '''
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint)
model.eval()
if bn:
output = Normalize('Discriminator.BN4', [0, 2, 3], output)
output = nonlinearity(output)
output = tf.reshape(output, [-1, 4 * 4 * 8 * dim])
output = lib.ops.linear.Linear('Discriminator.Output', 4 * 4 * 8 * dim, 1,
output)
lib.ops.conv2d.unset_weights_stdev()
lib.ops.deconv2d.unset_weights_stdev()
lib.ops.linear.unset_weights_stdev()
return tf.reshape(output, [-1])
data_set = data.DATA(64).load()
Generator, Discriminator = GeneratorAndDiscriminator()
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
with tf.Session(config=config) as session:
all_real_data_conv = tf.placeholder(tf.int32,
shape=[BATCH_SIZE, 3, 64, 64])
if tf.__version__.startswith('1.'):
split_real_data_conv = tf.split(all_real_data_conv, len(DEVICES))
else:
split_real_data_conv = tf.split(0, len(DEVICES), all_real_data_conv)
gen_costs, disc_costs = [], []
"""
import os
os.environ["CUDA_VISIBLE_DEVICES"]="1"
import data
import model
import config
import datetime
if __name__ == "__main__":
with open(os.path.join(config.LOG_DIR, str(datetime.datetime.now().strftime("%Y%m%d")) + "_" + str(config.BATCH_SIZE) + "_" + str(config.NUM_EPOCHS) + ".txt"), "w") as log:
log.write(str(datetime.datetime.now()) + "\n")
log.write("Use Pretrained Weights: " + str(config.USE_PRETRAINED) + "\n")
log.write("Pretrained Model: " + config.PRETRAINED + "\n")
# READ DATA
train_data = data.DATA(config.TRAIN_DIR)
print("Train Data Loaded")
# BUILD MODEL
model = model.MODEL()
print("Model Initialized")
model.build()
print("Model Built")
# TRAIN MODEL
model.train(train_data, log)
print("Model Trained")
# TEST MODEL
test_data = data.DATA(config.TEST_DIR)
print("Test Data Loaded")
model.test(test_data, log)
print("Image Reconstruction Done")
def save_model(model, save_path):
torch.save(model.state_dict(),save_path)
save_dir = 'model_dir'
random_seed = 0
gpu_id = 0
epochs = 100
if __name__=='__main__':
torch.cuda.set_device(gpu_id)
np.random.seed(random_seed)
torch.manual_seed(random_seed)
torch.cuda.manual_seed(random_seed)
print('=====>Prepare dataloader ...')
train_loader = torch.utils.data.DataLoader(data.DATA(mode='train'),
batch_size=1,
shuffle=True)
print('=====>Prepare model ...')
model = network_new.AutoEncoder()
model.cuda()
#optimizer and log writter
optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
writer = SummaryWriter(os.path.join(save_dir, 'train_info'))
print('Start training ...')
print('=====>Start training ...')
iters = 0
for epoch in range(1, epochs+1):
model.train()
for idx, motion in enumerate(train_loader):
train_info = 'Epoch: [{0}][{1}/{2}]'.format(epoch, idx+1, len(train_loader))
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Fri Jan 26 17:56:05 2018
@author: rahul
"""
import data
import model
import config
import utilities
if __name__ == "__main__":
# READ DATA
train_data = data.DATA()
train_data.build_dataset(config.TRAIN_FILENAME)
# BUILD MODEL
net = model.MODEL()
net.build()
# TRAIN MODEL
net.train(train_data)
# PLOT EMBEDDINGS
utilities.plot_with_labels(net.embeddings, train_data)
# -*- coding: utf-8 -*-
"""
Created on Wed Jun 26 14:52:40 2019
@author: ashima.garg
"""
#Classifying Names with a Character Level RNN
import data
import model
import config
if __name__ == "__main__":
data_obj = data.DATA()
data_obj.read()
print("Train Data Loaded")
# BUILD MODEL
rnn = model.RNN(data_obj.n_letters, config.N_HIDDEN, data_obj.n_categories)
modeloperator = model.Operator(rnn, config.LEARNING_RATE)
# TRAIN MODEL
modeloperator.train(data_obj)
print("Model Trained")
# TEST MODEL
print("all categories: ", len(data_obj.categories))
modeloperator.predict(data_obj, 'Dovesky')
modeloperator.predict(data_obj, 'Satoshi')
modeloperator.predict(data_obj, 'Jackson')
args = parser.arg_parse()
device = torch.device("cuda:" + str(args.gpu) if (
torch.cuda.is_available()) else "cpu")
'''create directory to save trained model and other info'''
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
''' setup GPU '''
torch.cuda.set_device(args.gpu)
''' setup random seed '''
np.random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
''' load dataset and prepare data loader '''
print('===> prepare dataloader ...')
train_loader = torch.utils.data.DataLoader(data.DATA(args),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=True)
''' load model '''
print('===> prepare model ...')
netG, netD = models.GAN(args)
netG, netD = netG.cuda(), netD.cuda()
''' define loss '''
adversarial_loss = nn.BCELoss()
# Create batch of latent vectors that we will use to visualize
# the progression of the generator
fixed_noise = torch.randn(args.samples_num, args.nz, 1, 1, device=device)
# Establish convention for real and fake labels during training
def main(args):
src_feature_extractor, label_predictor, adda_discriminator = models.ADDA(
args)
src_feature_extractor.cuda(), label_predictor.cuda(
), adda_discriminator.cuda()
src_feature_extractor.load_state_dict(
torch.load(os.path.join(args.src_model,
'source_feature_extractor.pt')))
label_predictor.load_state_dict(
torch.load(os.path.join(args.src_model, 'label_predictor.pt')))
src_feature_extractor.eval(), label_predictor.eval()
set_requires_grad(src_feature_extractor, requires_grad=False)
set_requires_grad(label_predictor, requires_grad=False)
tar_feature_extractor, _, _ = models.ADDA(args)
tar_feature_extractor.cuda()
tar_feature_extractor.load_state_dict(
torch.load(os.path.join(args.src_model,
'source_feature_extractor.pt')))
src_train_loader = torch.utils.data.DataLoader(data.DATA(
args, mode='train', dataset=args.source_dataset),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=True)
tar_train_loader = torch.utils.data.DataLoader(data.DATA(
args, mode='train', dataset=args.target_dataset),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=True)
tar_test_loader = torch.utils.data.DataLoader(data.DATA(
args, mode='test', dataset=args.target_dataset),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=False)
discriminator_optim = torch.optim.Adam(adda_discriminator.parameters(),
lr=args.lr_adda_d)
target_optim = torch.optim.Adam(tar_feature_extractor.parameters(),
lr=args.lr_tar)
criterion = nn.BCEWithLogitsLoss()
best_acc = 0.0
for epoch in range(1, args.epochs + 1):
batch_iterator = zip(loop_iterable(src_train_loader),
loop_iterable(tar_train_loader))
tar_feature_extractor.train(), adda_discriminator.train()
total_loss = 0
total_accuracy = 0
for _ in range(args.iterations):
# Train discriminator
set_requires_grad(tar_feature_extractor, requires_grad=False)
set_requires_grad(adda_discriminator, requires_grad=True)
for _ in range(args.k_disc):
(source_x, _), (target_x, _) = next(batch_iterator)
source_x, target_x = source_x.cuda(), target_x.cuda()
source_features = src_feature_extractor(source_x).view(
source_x.shape[0], -1)
target_features = tar_feature_extractor(target_x).view(
target_x.shape[0], -1)
discriminator_x = torch.cat([source_features, target_features])
discriminator_y = torch.cat([
torch.ones(source_x.shape[0]).cuda(),
torch.zeros(target_x.shape[0]).cuda()
])
preds = adda_discriminator(discriminator_x).squeeze()
loss = criterion(preds, discriminator_y)
discriminator_optim.zero_grad()
loss.backward()
discriminator_optim.step()
total_loss += loss.item()
total_accuracy += ((
preds >
0).long() == discriminator_y.long()).float().mean().item()
# Train classifier
set_requires_grad(tar_feature_extractor, requires_grad=True)
set_requires_grad(adda_discriminator, requires_grad=False)
for _ in range(args.k_clf):
_, (target_x, _) = next(batch_iterator)
target_x = target_x.cuda()
target_features = tar_feature_extractor(target_x).view(
target_x.shape[0], -1)
# flipped labels
discriminator_y = torch.ones(target_x.shape[0]).cuda()
preds = adda_discriminator(target_features).squeeze()
loss = criterion(preds, discriminator_y)
target_optim.zero_grad()
loss.backward()
target_optim.step()
mean_loss = total_loss / (args.iterations * args.k_disc)
mean_accuracy = total_accuracy / (args.iterations * args.k_disc)
print(f'EPOCH {epoch:03d}: discriminator_loss={mean_loss:.4f}, '
f'discriminator_accuracy={mean_accuracy:.4f}')
# Create the full target model and save it
acc = eval(tar_feature_extractor, label_predictor, tar_test_loader)
print("accuracy: ", acc)
if acc > best_acc:
torch.save(tar_feature_extractor.state_dict(),
os.path.join(args.save_dir, 'tar_feature_extractor.pt'))
best_acc = acc
elif mode == "test":
if not os.path.exists(args.predictions):
os.makedirs(args.predictions)
for idx in range(len(outputs_list)):
img = Image.fromarray(outputs_list[idx])
img_path = os.path.join(args.predictions,
img_name_list[idx] + '.jpg')
img.save(img_path)
return 'Images saved'
if __name__ == "__main__":
args = argparser.arg_parse()
# Load test images
test_dataset = data.DATA(mode="test", train_status="TA")
dataloader_test = torch.utils.data.DataLoader(
dataset=test_dataset,
batch_size=args.batch_size_test,
shuffle=False,
num_workers=argparser.n_cpu)
# Set device
device = torch.device(
"cuda") if torch.cuda.is_available() else torch.device("cpu")
model = model.Net()
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint)
if torch.cuda.is_available():
transform = transforms.Compose([
transforms.ToTensor(),
])
train_loader = DataLoader(
MNIST('.', train=True, download=False, transform=transform),
batch_size=train_batch_size,
shuffle=True)
val_loader = DataLoader(
MNIST('.', train=False, download=False, transform=transform),
batch_size=train_batch_size,
shuffle=True)
input_n=784
elif opt.motion:
### Human Motion Data
data = data.DATA("h3.6m_3d", "h3.6m/dataset/")
timepoints = opt.timepoints
out_of_distribution = data.get_dct_and_sequences(input_n=timepoints, output_n=0, sample_rate=opt.sample_rate, dct_n=10, out_of_distribution_action=None)
train_loader, val_loader, OoD_val_loader, test_loader = data.get_dataloaders(train_batch=train_batch_size, test_batch=test_batch_size)
#input_n=data.node_n*timepoints
#input_n=96*timepoints
input_n=[96, timepoints]
else:
### Human PoseData
data = data.DATA("h3.6m_3d", "h3.6m/dataset/")
out_of_distribution = data.get_poses(input_n=1, output_n=1, sample_rate=opt.sample_rate, dct_n=2, out_of_distribution_action=None)
train_loader, val_loader, OoD_val_loader, test_loader = data.get_dataloaders(train_batch=train_batch_size, test_batch=test_batch_size)
input_n=data.node_n
print(">>> data loaded !")
# ===============================================================
# Instantiate model, and methods used fro training and valdation
import models
from matplotlib import pyplot as plt
import data
import parser
import torch
#torch.multiprocessing.set_sharing_strategy('file_system')
import numpy as np
from sklearn.manifold import TSNE
args = parser.arg_parse()
tar_test_loader = torch.utils.data.DataLoader(data.DATA(args, mode='test', dataset=args.target_dataset),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=False)
src_test_loader = torch.utils.data.DataLoader(data.DATA(args, mode='test', dataset=args.source_dataset),
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=False)
feature_extractor, label_predictor, domain_classifier = models.DANN(args)
feature_extractor.cuda(), label_predictor.cuda()
feature_extractor.load_state_dict(torch.load(args.resume_folder + 'feature_extractor.pth.tar'))
label_predictor.load_state_dict(torch.load(args.resume_folder + 'label_predictor.pth.tar'))
feature_extractor.eval(), label_predictor.eval()
tar_features, tar_labels, src_features, src_labels = [], [], [], []
with torch.no_grad(): # do not need to calculate information for gradient during eval
for idx, (imgs, classes) in enumerate(tar_test_loader):
imgs = imgs.cuda()
import model
import utils
import os
import tensorflow as tf
if __name__ == "__main__":
# LOAD EMBEDDING
word_to_index, index_to_word, word_to_vec, emb_matrix = utils.read_glove_vecs(
os.path.join(config.EMBEDDING_DIR, config.EMBEDDING_PATH))
print("Pretrained Embedding Loaded")
#LOAD CONFIG
train_config = config.TrainConfig()
test_config = config.TestConfig()
# LOAD DATA
train_data = data.DATA(train_config)
train_data.read_file(config.TRAIN_PATH, word_to_index)
print("Train data Loaded")
test_data = data.DATA(test_config)
test_data.read_file(config.TEST_PATH, word_to_index)
print("Test data Loaded")
# BUILD MODEL
#initializer = tf.random_uniform_initializer(train_config.init_scale, train_config.init_scale)
with tf.name_scope("Train"):
with tf.variable_scope("Model", reuse=None):
train_model = model.MODEL(train_config,
len(word_to_index),
training=True)
train_model.build()
return padded_sequence, label, n_frames
args = parser.arg_parse()
'''create directory to save trained model and other info'''
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
''' setup GPU '''
torch.cuda.set_device(args.gpu)
''' setup random seed '''
np.random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
''' load dataset and prepare data loader '''
print('===> prepare data ...')
data_loader = data.DATA(args, split='split0')
splits = ['split0', 'split1', 'split2', 'split3']
splits.remove(splits[splits.index(args.val_split)])
train_data = torch.utils.data.ConcatDataset([
data.DATA(args, split=splits[0]),
data.DATA(args, split=splits[1]),
data.DATA(args, split=splits[2])
])
val_data = data.DATA(args, split=args.val_split)
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=args.train_batch,
num_workers=4,
shuffle=True)
import resource
rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)
resource.setrlimit(resource.RLIMIT_NOFILE, (4096, rlimit[1]))
if __name__ == '__main__':
args = parser.arg_parse()
'''create directory to save trained model and other info'''
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
''' setup GPU '''
torch.cuda.set_device(args.gpu)
''' load dataset and prepare data loader '''
print('===> prepare dataloader ...')
train_loader = torch.utils.data.DataLoader(data.DATA(args, mode='train'),
batch_size=args.data_batch,
num_workers=args.workers,
shuffle=False)
val_loader = torch.utils.data.DataLoader(data.DATA(args, mode='test'),
batch_size=args.data_batch,
num_workers=args.workers,
shuffle=False)
''' load model '''
print('===> prepare model ...')
feature_extractor, FC = models.P1()
''' extract feature '''
#print('===> extract features for every videos ...')
#utils.extract_feature_p1(feature_extractor, train_loader, val_loader, args)
''' define loss '''
os.makedirs(args.save_dir)
''' setup GPU '''
torch.cuda.set_device(args.gpu)
''' setup random seed '''
np.random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
''' load dataset and prepare data loader '''
print('===> prepare data ...')
data_loader = data.DATA(args, split='split0')
splits = ['split0', 'split1', 'split2', 'split3']
splits.remove(splits[splits.index(args.val_split)])
train_data = torch.utils.data.ConcatDataset([data.DATA(args, split=splits[0]),
data.DATA(args, split=splits[1]),
data.DATA(args, split=splits[2])])
val_data = data.DATA(args, split=args.val_split)
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=args.train_batch,
num_workers=4,
shuffle=True)
val_loader = torch.utils.data.DataLoader(val_data,
# -*- coding: utf-8 -*-
"""
Created on Thu Feb 8 20:00:55 2018
@author: rahul.ghosh
"""
import data
import model
import config
if __name__ == "__main__":
# READ DATA
data = data.DATA()
data.read(config.TRAIN_FILENAME)
# BUILD MODEL
model = model.MODEL()
model.build()
# TRAIN MODEL
model.train(data)
# TEST MODEL
data.read(config.TEST_FILENAME)
model.test(data)
# -*- coding: utf-8 -*-
"""
Created on Thu Feb 12 17:55:53 2019
@author: ashima.garg
"""
import os
import data
import model
import tensorflow as tf
import utils
if __name__ == "__main__":
# READ DATA
data_ = data.DATA()
data_.read()
print("Train Data Loaded")
# BUILD MODEL
model = model.MODEL()
print("Model Initialized")
with tf.variable_scope("siamese") as scope:
model.output_1 = model.build(model.inputs_1)
scope.reuse_variables()
model.output_2 = model.build(model.inputs_2)
print("Model Built")
# TRAIN MODEL
model.train(data_)
print("Model Trained")
# TEST MODEL
if __name__ == '__main__':
args = parser.arg_parse()
# get input and output directory
input_dir = args.input_dir
''' setup GPU '''
torch.cuda.set_device(args.gpu)
''' prepare data_loader '''
print('===> prepare data loader ...')
if input_dir == "val_test":
test_loader = torch.utils.data.DataLoader(data.DATA(args, mode='val'),
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=False)
else:
test_loader = torch.utils.data.DataLoader(data_test.DATA_TEST(args, mode='test'),
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=False)
''' prepare mode '''
if args.model == "simple_baseline":
model = simple_baseline_model.SimpleBaselineModel(args).cuda()
else:
model = baseline_model.BaselineModel(args).cuda()
''' resume save model '''
def main():
''' setup '''
torch.manual_seed(args.random_seed)
''' load dataset and prepare data loader '''
print('===> prepare dataloader ...')
train_set = data.DATA(args, mode='train')
val_set = data.DATA(args, mode='val')
test_set = data.DATA(args, mode='test')
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=args.train_batch,
num_workers=args.workers,
shuffle=True)
val_loader = torch.utils.data.DataLoader(val_set,
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=False)
test_loader = torch.utils.data.DataLoader(test_set,
batch_size=args.test_batch,
num_workers=args.workers,
shuffle=False)
''' load models '''
print('===> prepare models ...')
A = Net('A')
B = Net('B')
C = Net('C')
models = {'A': A, 'B': B, 'C': C}
''' define loss '''
criterion = nn.CrossEntropyLoss()
best_acc = {'A': 0, 'B': 0, 'C': 0}
best_epoch = {'A': 0, 'B': 0, 'C': 0}
train_losses = {'A': [], 'B': [], 'C': []}
val_losses = {'A': [], 'B': [], 'C': []}
val_accs = {'A': [], 'B': [], 'C': []}
test_losses = {'A': 0, 'B': 0, 'C': 0}
test_accs = {'A': 0, 'B': 0, 'C': 0}
''' training and validation iterations'''
for mode in ['A', 'B', 'C']:
model = models[mode]
''' setup optimizer '''
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr)
if torch.cuda.is_available():
model.cuda()
for epoch in range(args.epoch):
''' train model and get averaged epoch loss '''
train_loss = train(mode, epoch, model, train_loader, criterion,
optimizer)
''' evaluate the model '''
val_loss, val_acc = test(epoch, model, val_loader, criterion)
train_losses[mode].append(train_loss)
val_losses[mode].append(val_loss)
val_accs[mode].append(val_acc)
print(
'\nMode: {} Epoch: [{}] TRAIN_LOSS: {} VAL_LOSS: {} VAL_ACC:{}'
.format(mode, epoch, train_loss, val_loss, val_acc))
''' save best model '''
if val_acc > best_acc[mode]:
save_model(
model,
os.path.join(args.save_dir,
'model_best_{}.pth.tar'.format(mode)))
best_acc[mode] = val_acc
best_epoch[mode] = epoch
print("Mode: {} Best acc: {}, epoch {}".format(mode, best_acc[mode],
best_epoch[mode]))
''' testing (best model) '''
for mode in ['A', 'B', 'C']:
model = models[mode]
if torch.cuda.is_available():
model.load_state_dict(
torch.load(
os.path.join(args.save_dir,
'model_best_{}.pth.tar'.format(mode))))
model.cuda()
else:
model.load_state_dict(
torch.load(os.path.join(args.save_dir,
'model_best_{}.pth.tar'.format(mode)),
map_location=torch.device('cpu')))
test_loss, test_acc = test(0, model, test_loader, criterion)
test_losses[mode] = test_loss
test_accs[mode] = test_acc
print('Mode: {} TEST_LOSS:{} TEST_ACC:{}'.format(
mode, test_loss, test_acc))
''' save train/val/test information as pickle files'''
with open(os.path.join(args.save_dir, 'train_losses.pkl'), 'wb') as f:
pickle.dump(train_losses, f)
with open(os.path.join(args.save_dir, 'val_losses.pkl'), 'wb') as f:
pickle.dump(val_losses, f)
with open(os.path.join(args.save_dir, 'val_accs.pkl'), 'wb') as f:
pickle.dump(val_accs, f)
with open(os.path.join(args.save_dir, 'test_losses.pkl'), 'wb') as f:
pickle.dump(test_losses, f)
with open(os.path.join(args.save_dir, 'test_accs.pkl'), 'wb') as f:
pickle.dump(test_accs, f)