def demo():
"""
demo for MNIST handwritten digits recognition
"""
model = Net()
train_dataset = MnistDataset(True)
test_dataset = MnistDataset(False)
train_dataloader = MNN.data.DataLoader(train_dataset,
batch_size=64,
shuffle=True)
test_dataloader = MNN.data.DataLoader(test_dataset,
batch_size=100,
shuffle=False)
opt = MNN.optim.SGD(model, 0.01, 0.9, 0.0005)
F.set_thread_number(4)
for epoch in range(0, 10):
opt.learning_rate = learning_rate_scheduler(opt.learning_rate, epoch)
train_func(model, train_dataloader, opt)
# save model
file_name = '%d.mnist.mnn' % epoch
model.train(False)
predict = model.forward(F.placeholder([1, 1, 28, 28], F.NC4HW4))
print("Save to " + file_name)
F.save([predict], file_name)
test_func(model, test_dataloader)
def main():
# Load the train data.
train_csv = './mnist_data/train.csv'
train_x, train_y, val_x, val_y = load_train_csv_dataset(
train_csv, validation_percent=0.1)
# Create pytorch dataloaders for train and validation sets.
train_dataset = MnistDataset(train_x, train_y)
train_dataloader = DataLoader(train_dataset,
batch_size=200,
shuffle=True,
num_workers=2)
val_dataset = MnistDataset(val_x, val_y)
val_dataloader = DataLoader(val_dataset,
batch_size=200,
shuffle=False,
num_workers=2)
# Define model, optimizer and loss function.
model = MnistCNN()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
loss_func = nn.CrossEntropyLoss()
# Train our model.
train_model(model,
train_dataloader,
loss_func,
optimizer,
epochs=NUM_EPOCHS)
val_accuracy = eval_model(model, val_dataloader)
print('Validation set accuracy: {}'.format(val_accuracy))
# Save model weights for inference.
torch.save(model.state_dict(), 'trained_model.pt')
def dataset(self):
if self.suffix == 'test':
return MnistDataset(file='data/test.csv')
return MnistDataset(file='data/train.csv',
fold_csv='data/fold.csv',
is_test=True
)
def test_file_exists(self):
# Above I tested that if the file doesn't exist it will download it
# Here I test that if the file was already downloaded, it will not download it again
# run it once to make sure the file is downloaded
MnistDataset().get_dataset()
urlretrieve = MagicMock()
MnistDataset().get_dataset(urlretrieve)
urlretrieve.method.assert_not_called()
def get_datasets(self, stage: str, **kwargs):
datasets = OrderedDict()
train = MnistDataset('data/train.csv', fold_csv='data/fold.csv')
valid = MnistDataset('data/train.csv',
fold_csv='data/fold.csv',
is_test=True)
datasets['train'] = train
datasets['valid'] = valid
return datasets
def model_train(fold: int) -> None:
# Prepare Data
df = pd.read_csv(os.path.join(config.save_dir, 'split_kfold.csv'))
df_train = df[df['kfold'] != fold].reset_index(drop=True)
df_val = df[df['kfold'] == fold].reset_index(drop=True)
df_train.drop(['kfold'], axis=1).to_csv(os.path.join(
config.save_dir, f'train-kfold-{fold}.csv'), index=False)
df_val.drop(['kfold'], axis=1).to_csv(os.path.join(
config.save_dir, f'val-kfold-{fold}.csv'), index=False)
train_dataset = MnistDataset(os.path.join(config.data_dir, 'train'), os.path.join(
config.save_dir, f'train-kfold-{fold}.csv'), transforms_train)
val_dataset = MnistDataset(
os.path.join(config.data_dir, 'train'), os.path.join(config.save_dir, f'val-kfold-{fold}.csv'), transforms_test)
model = MnistModel(EfficientNet())
checkpoint_callback = ModelCheckpoint(
monitor='val_loss',
dirpath=os.path.join(save_dir, f'{fold}'),
filename='{epoch:02d}-{val_loss:.2f}.pth',
save_top_k=5,
mode='min',
)
early_stopping = EarlyStopping(
monitor='val_loss',
mode='min',
)
if config.device == 'tpu':
train_loader = DataLoader(train_dataset, batch_size=16, num_workers=10, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=2, num_workers=10, shuffle=False)
trainer = Trainer(
tpu_cores=8,
num_sanity_val_steps=-1,
deterministic=True,
max_epochs=config.epochs,
callbacks=[checkpoint_callback, early_stopping]
)
else:
train_loader = DataLoader(train_dataset, batch_size=16, num_workers=10, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=8, num_workers=10, shuffle=False)
trainer = Trainer(
gpus=1,
num_sanity_val_steps=-1,
deterministic=True,
max_epochs=config.epochs,
callbacks=[checkpoint_callback, early_stopping]
)
trainer.fit(model, train_loader, val_loader)
def __init__(self,
model: Interface):
self.model = model
self.dataset = MnistDataset(file='data/train.csv',
fold_csv='data/fold.csv',
is_test=True
)
def get_datasets(self, stage: str, **kwargs):
datasets = OrderedDict()
train = MnistDataset('data/train.csv',
fold_csv='data/fold.csv',
transforms=Experiment.get_test_transforms())
valid = MnistDataset('data/train.csv',
fold_csv='data/fold.csv',
is_test=True,
transforms=Experiment.get_test_transforms())
datasets['train'] = train
datasets['valid'] = valid
return datasets
def test_file_doesnt_exist(self):
filename = 'mnist.npz'
if os.path.exists(filename):
os.remove(filename)
(x_train, y_train), (x_test, y_test) = MnistDataset().get_dataset()
assert x_train.shape == (60000, 28, 28)
assert y_train.shape == (60000, )
assert x_test.shape == (10000, 28, 28)
assert y_test.shape == (10000, )
def main(_):
# Import data
mnist = MnistDataset(FLAGS.data_dir)
model = Model()
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
model.train(sess, mnist.train)
# Test trained model
print(model.test(sess, mnist.test))
def main(_):
start_time = time.time()
# Import data
mnist = MnistDataset(FLAGS.data_dir)
model = Model()
sess = dm.get_session()
model.train(sess, mnist.train)
# Test trained model
result = model.test(sess, mnist.test)
end_time = time.time()
print("%0.2f percent accuracy in %d ms." %
(result * 100, end_time - start_time))
def create_dataset(path, test=False):
"""
Parameters
----------
path: path to data directory
test: specify whether the data is for training or testing
Returns
-------
Loaded dataset
"""
return MnistDataset(path, test=test)
def __init__(self, **kwargs):
cache_names = ['y']
super().__init__(cache_names=cache_names,
layout='img_classify',
**kwargs)
self.x_source = MnistDataset(
file='data/train.csv',
fold_csv='data/fold.csv',
is_test=True,
max_count=self.max_count,
transforms=Experiment.get_test_transforms())
self.builder = None
self.x = None
self.scores = []
def main(args):
analysis_dataset = MnistDataset(args.test_data_path,
num_classes=args.class_nums,
shape=(args.image_width,
args.image_height))
analysis_dataloader = torch.utils.data.DataLoader(
analysis_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
drop_last=False)
face_index, face_label = load_nn_model(
os.path.join(args.model_dir, args.faiss_model_file),
os.path.join(args.model_dir, args.faiss_label_file))
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = torch.jit.load(os.path.join(args.model_dir, args.model_file),
map_location=device)
results = {
"predicts": [],
"labels": analysis_dataset.labels,
"distances": []
}
for i, data in enumerate(analysis_dataloader):
images, _ = data
with torch.no_grad():
embeddings = model(images.to(device))
dists, inds = face_index.search(embeddings.detach().cpu().numpy(), 3)
predicts = face_label[inds[:, 0]]
results["predicts"] += predicts.tolist()
results["distances"] += dists.tolist()
del images
del embeddings
print("[+] Analysis using {} image data".format(str(
len(analysis_dataset))))
save_cm(results, args.class_nums)
print("finnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnn")
def __init__(self):
# Data and training :
self.batch_size = 64
self.dataset = MnistDataset(self.batch_size)
self.n_epochs = 10
# Model hyperparameters :
self.z_size = 25
# TODO : make this more modular, adapt to other datasets
self.images = tf.placeholder(tf.float32, [None, 28, 28, 1])
# estimators of mean and sdt in latent space
z_mu, z_sigma = self.encoder(self.images)
eps = tf.random.normal([self.batch_size, self.z_size],
0,
1,
dtype=tf.float32)
self.z = z_mu + (z_sigma * eps)
# generated images
self.gen_images = self.decoder(self.z)
tiny = 1e-7 # to avoid nan from logs
# latent loss : KL divergence between our estimators and a normal distribution
self.lat_loss = 0.5 * tf.reduce_sum(
tf.square(z_mu) + tf.square(z_sigma) - tf.log(tf.square(z_sigma)) -
1, 1)
# generative loss : distance between original image and reconstructed one
flat = tf.reshape(self.images, shape=[self.batch_size, 28 * 28])
flat = tf.clip_by_value(flat, tiny, 1 - tiny)
gen_flat = tf.reshape(self.gen_images,
shape=[self.batch_size, 28 * 28])
gen_flat = tf.clip_by_value(gen_flat, tiny, 1 - tiny)
# cross entropy between original and reconstructed image
self.gen_loss = -tf.reduce_sum(
flat * tf.log(gen_flat) + (1 - flat) * tf.log(1 - gen_flat), 1)
# total loss is the sum of both losses :
self.loss = tf.reduce_mean(self.gen_loss + self.lat_loss)
self.opt = tf.train.AdamOptimizer(0.001).minimize(self.loss)
# other stuff
self.save_path = os.getcwd() + "/saves/model"
import tensorflow as tf
from models.convolutional_model import ConvModel
from dataset import MnistDataset
epochs = 40
batch_size = 20
# Load Data from target directory
train_set = MnistDataset('flowers/train')
n_batches = len(train_set) // batch_size
# hold_prob = tf.placeholder(tf.float32)
# Construct CNN model
# model = ConvModel(resolution=[64, 64], channels=3, hold_prob=hold_prob)
model = ConvModel(resolution=[64, 64], channels=3)
# We use this to save the model. Instantiate it after all Variables have been created
saver = tf.train.Saver()
# Define tensors
label_placeholder = tf.placeholder(tf.float32, shape=[batch_size, 5])
# Define loss functions and Optimizer
loss = tf.losses.softmax_cross_entropy(label_placeholder, model.predictions)
update = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(loss)
top_predictions = tf.argmax(model.predictions,
axis=1) # probabilities -> top prediction
top_labels = tf.argmax(label_placeholder, axis=1) # one_hot -> number
correct = tf.equal(top_predictions, top_labels) # bool Tensor
accuracy = tf.reduce_mean(tf.cast(correct,
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=64, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=14, metavar='N',
help='number of epochs to train (default: 14)')
parser.add_argument('--lr', type=float, default=1.0, metavar='LR',
help='learning rate (default: 1.0)')
parser.add_argument('--gamma', type=float, default=0.7, metavar='M',
help='Learning rate step gamma (default: 0.7)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=10, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model', action='store_true', default=False,
help='For Saving the current Model')
args = parser.parse_args()
use_cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
device = torch.device("cuda" if use_cuda else "cpu")
if True:
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.test_batch_size, shuffle=True, **kwargs)
else:
kwargs = {'num_workers': 16, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(
MnistDataset(train=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
MnistDataset(train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.batch_size, shuffle=True, **kwargs)
model = Net().to(device)
optimizer = optim.Adadelta(model.parameters(), lr=args.lr)
scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
start = dt.datetime.now()
for epoch in range(1, args.epochs + 1):
train(args, model, device, train_loader, optimizer, epoch)
test(model, device, test_loader)
scheduler.step()
if args.save_model:
torch.save(model.state_dict(), "mnist_cnn.pt")
end = dt.datetime.now()
print(f"elapsed time: {(end - start).total_seconds()} secs")
if not os.path.isdir("./generate_image"):
os.mkdir("./generate_image")
checkpoint_dir = './training_checkpoints'
print("load models!!")
generator = Generator()
discriminator = Discriminator()
print("load ckpt!!")
checkpoint = tf.train.Checkpoint(generator=generator, discriminator=discriminator)
checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir))
dataset_name = 'fashion_mnist'
assert dataset_name in ['mnist', 'fashion_mnist']
if dataset_name == 'mnist':
mnist_dataset = MnistDataset()
test_images, test_labels = mnist_dataset.get_test_data()
if dataset_name == 'fashion_mnist':
fashin_mnist_dataset = FashinMnistDataset()
test_images, test_labels = fashin_mnist_dataset.get_test_data()
random_noise_test_images, _ = fashin_mnist_dataset.get_random_noise_test_data()
random_noise_test_images = random_noise_test_images.reshape(-1, 28, 28, 1).astype('float32')
test_images = test_images.reshape(-1, 28, 28, 1).astype('float32')
print("Compute anomaly scores!!")
for idx, (test_image, test_label) in enumerate(zip(test_images, test_labels)):
if dataset_name == 'fashion_mnist' and idx % 2 == 1:
test_image = random_noise_test_images[idx]
test_image = (test_image / 127.5) - 1
import tensorflow as tf
import numpy as np
from models.convolutional_model import ConvModel
from dataset import MnistDataset
epochs = 50
batch_size = 150
# Load Data from target directory
train_set = MnistDataset('kaggle_inria/train')
n_batches = len(train_set) // batch_size
# hold_prob = tf.placeholder(tf.float32)
with tf.device('/gpu:1'):
# Construct CNN model
# model = ConvModel(resolution=[64, 64], channels=3, hold_prob=hold_prob)
model = ConvModel(resolution=[128, 128], channels=3)
# We use this to save the model. Instantiate it after all Variables have been created
saver = tf.train.Saver()
# Define tensors
label_placeholder = tf.placeholder(tf.float32, shape=[batch_size, 2])
# Define loss functions and Optimizer
loss = tf.losses.softmax_cross_entropy(label_placeholder,
model.predictions)
update = tf.train.AdamOptimizer(learning_rate=0.0001).minimize(loss)
top_predictions = tf.argmax(model.predictions,
def __init__(self, network: tf.Tensor):
self.network = network
self.mnist = MnistDataset()
self.loss = None
def main():
use_cuda = False
device = torch.device('cuda' if use_cuda else 'cpu')
epochs = 50
torch.manual_seed(123)
transforms = T.Compose([
T.ToTensor(),
T.Normalize((0.5, ), (0.5, )),
])
trainSet = MnistDataset('data', 'train', transforms)
testSet = MnistDataset('data', 'test', transforms)
trainSet, validSet = torch.utils.data.random_split(trainSet, [24000, 6000])
traindataLoader = DataLoader(trainSet,
batch_size=64,
shuffle=True,
num_workers=4)
validDataLoader = DataLoader(validSet,
batch_size=64,
shuffle=True,
num_workers=4)
testDataLoader = DataLoader(testSet, batch_size=64, num_workers=4)
#for grid search, in comment since it was used just for finding best hyperparameter
#
#
#besthypepar = gridSearch(traindataLoader,validDataLoader,validSet,epochs,device)
#print(besthypepar)
besthypeloss = float('inf')
#besthyperparameters: 2 hidden layer, 512 layer size, relu function, 0.001 learning rate
lastModel = Model(512, 2, 2)
lastModel.to(device)
optimizer = torch.optim.Adam(lastModel.parameters(), lr=0.001)
lastModel, train_loss, valid_loss, best_val_loss, best_correct = train(
lastModel, optimizer, traindataLoader, validDataLoader, epochs, device,
besthypeloss)
#plot train and valid loss
fig = plt.figure()
plt.plot(train_loss, label='Training Loss')
plt.plot(valid_loss, label='Validation Loss')
plt.axvline(valid_loss.index(min(valid_loss)),
color="r",
label='Early Stop')
plt.xlabel('epoch')
plt.ylabel('loss')
plt.legend()
plt.tight_layout()
fig.savefig('best_loss_plot.png')
#test
lastModel.load_state_dict(torch.load('model_state_dict'))
lastModel.eval()
index = 0
test_results = []
for test_images, test_labels in testDataLoader:
pred = lastModel(test_images)
_, output = torch.max(pred, 1)
for i in range(output.size(0)):
head, tail = os.path.split(testSet.data[index][0])
test_results.append(f'{tail} ' + f'{output[i].item()}\n')
index = index + 1
with open("test_labels.txt", "w") as f:
f.writelines(test_results)
import tensorflow as tf
from dataset import MnistDataset
with tf.Session() as sess:
saver = tf.train.import_meta_graph(
'saved_models/model.ckpt.meta'
) # Don't have to recreate the entire graph
saver.restore(sess,
'saved_models/model.ckpt') # Restore all graph variables
model = tf.get_collection('model')[0]
inputs = tf.get_collection('model_inputs')[0]
test_inputs = [
'data/test/img_1.jpg', 'data/test/img_2.jpg', 'data/test/img_3.jpg'
]
test_inputs = [MnistDataset.read_image(input) for input in test_inputs]
predictions = sess.run(model, feed_dict={inputs: test_inputs})
print(predictions)
def main(args):
train_dataset = MnistDataset(args.train_data_path,
num_classes=args.class_nums)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
drop_last=False)
print("[+] " + str(len(train_dataset)) + " train dataset")
eval_dataset = MnistDataset(args.test_data_path,
num_classes=args.class_nums,
transforms=False)
eval_dataloader = torch.utils.data.DataLoader(eval_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
drop_last=False)
print("[+] " + str(len(eval_dataset)) + " evaluate dataset")
model = Embedding(input_shape=(args.image_width, args.image_height),
input_channel=args.image_channel,
d_embedding=args.d_embedding)
print(model)
metric = ArcMarginProduct(args.d_embedding,
args.class_nums,
s=args.scale_size)
print(metric)
if args.resume:
print("[+] Resume the model({}) and metric({})".format(
args.model_resume, args.metric_resume))
model.load_state_dict(
torch.load(args.model_resume)["model_state_dict"])
metric.load_state_dict(
torch.load(args.metric_resume)["metric_state_dict"])
criterion = torch.nn.CrossEntropyLoss()
if args.optimizer == "adam":
print("[+] Using Adam Optimizer")
optimizer = torch.optim.Adam([{
"params": model.parameters()
}, {
"params": metric.parameters()
}],
lr=args.lr,
weight_decay=args.weight_decay)
else:
print("[+] Using SGD Optimizer")
optimizer = torch.optim.SGD([{
"params": model.parameters()
}, {
"params": metric.parameters()
}],
lr=args.lr,
weight_decay=args.weight_decay)
# lr_sched = lr_scheduler.MultiStepLR(optimizer, gamma=0.1)
if args.n_gpus > 1:
net = DataParallel(net)
metric = DataParallel(metric)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)
metric.to(device)
criterion.to(device)
it = 0
for epoch in range(1, args.epoch + 1):
# logging.info('{} epoch started'.format(str(epoch).zfill(3)))
print('[+] {} epoch started'.format(str(epoch).zfill(3)))
for data in train_dataloader:
images, labels = data[0].to(device), data[1].to(device)
optimizer.zero_grad()
embeddings = model(images)
output = metric(embeddings, labels)
total_loss = criterion(output, labels)
total_loss.backward()
optimizer.step()
if it % args.log_iter == 0:
_, predict = torch.max(output.data, 1)
correct = (np.array(predict.cpu()) == np.array(
labels.data.cpu())).sum()
now_accuracy = correct / labels.size(0)
# logging.info('{} iterations Accuracy :{}'.format(str(it).zfill(5), str(now_accuracy)))
print('[+] {} iterations Accuracy :{}'.format(
str(it).zfill(5), str(now_accuracy)))
if it % args.save_iter == 0:
if args.n_gpus > 1:
model_state_dict = model.module.state_dict()
metric_state_dict = metric.module.state_dict()
else:
model_state_dict = model.state_dict()
metric_state_dict = metric.state_dict()
if not os.path.exists(args.ckpt_dir):
os.makedirs(args.ckpt_dir)
torch.save(
{
"epoch": epoch,
"iters": it,
"model_state_dict": model_state_dict
},
os.path.join(
args.ckpt_dir,
"iter_{}_".format(str(it).zfill(5)) + args.model_ckpt))
torch.save(
{
"epoch": epoch,
"iters": it,
"metric_state_dict": metric_state_dict
},
os.path.join(
args.ckpt_dir, "iter_{}_".format(str(it).zfill(5)) +
args.metric_ckpt))
print('[+] {} iterations model saved'.format(str(it).zfill(5)))
if it % args.eval_iter == 0:
model.eval()
metric.eval()
correct = 0
with torch.no_grad():
for data in eval_dataloader:
images, labels = data[0].to(device), data[1].to(device)
embeddings = model(images)
output = metric(embeddings, labels)
_, predict = torch.max(output.data, 1)
correct = correct + (np.array(predict.cpu())
== np.array(
labels.data.cpu())).sum()
acc = correct / len(eval_dataloader.dataset)
# logging.info('{} iterations Eval Accuracy :{}'.format(str(it).zfill(5), str(acc)))
print('[+] {} iterations Eval Accuracy :{}'.format(
str(it).zfill(5), str(acc)))
model.train()
metric.train()
it = it + 1
with torch.no_grad():
model.eval()
metric.eval()
if args.n_gpus > 1:
model_ = torch.jit.script(model.module)
metric_ = torch.jit.script(metric.module)
else:
model_ = torch.jit.script(model)
metric_ = torch.jit.script(metric)
if not os.path.exists(args.model_dir):
os.makedirs(args.model_dir)
torch.jit.save(model_, os.path.join(args.model_dir, args.model_file))
torch.jit.save(metric_, os.path.join(args.model_dir, args.metric_file))
print("[+] Saved final Models")
del train_dataset
del eval_dataset
print("fit")
self.model.fit(
x=self.train_dataset,
steps_per_epoch=self.train_size//batch_size,
validation_data=self.validation_dataset,
validation_steps=self.validation_size//batch_size,
epochs=epochs
)
# callbacks=callback
# )
if __name__ =="__main__":
from dataset import MnistDataset
from dataset import DatasetUtil
from model import SimpleCNN
from model import SimpleSoftmaxClassificationModel
train = MnistDataset.get_train_set().map(DatasetUtil.image_classification_util(10))
validation = MnistDataset.get_validation_set().map(DatasetUtil.image_classification_util(10))
train_size, validation_size, _ = MnistDataset.get_data_size()
base = SimpleCNN.get_base_model(28,28,1)
softmax_model = SimpleSoftmaxClassificationModel.get_classification_model(base,10)
trainer:Trainer = Trainer(
softmax_model,
train=train,
train_size=train_size,
validation=validation,
validation_size=validation_size
)
trainer.train()
from dataset import MnistDataset from keras_gan import GAN mnist = MnistDataset(64) gan = GAN(mnist)
x = F.relu(x)
x = self.fc2(x)
if self.act == 0:
x = torch.sigmoid(x)
elif self.act == 1:
x = torch.tanh(x)
elif self.act == 2:
x = F.relu(x)
x = self.fc3(x)
x = torch.log_softmax(x, dim=1)
return x
if __name__ == '__main__':
transforms = T.Compose([
T.ToTensor(),
T.Normalize((0.5, ), (0.5, )),
])
dataset = MnistDataset('data', 'train', transforms)
dataLoader = DataLoader(dataset,
batch_size=64,
shuffle=True,
num_workers=4)
model = Model(512, 2, 2)
for images, labels in dataLoader:
pred = model(images)
print(pred)
exit()
from __future__ import print_function
from tensorflow.python import keras
from tensorflow.python.keras.layers import (Conv2D, Dense, Dropout, Flatten,
MaxPooling2D)
from tensorflow.python.keras.models import Sequential
from dataset import MnistDataset
# Create the dataset
dataset = MnistDataset(augmentation='pad')
# Params
batch_size = 128
epochs = 12
# Define the model
model = Sequential()
model.add(
Conv2D(32,
kernel_size=(3, 3),
activation='relu',
input_shape=dataset.shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(dataset.num_classes, activation='softmax'))
model3.load_state_dict(
torch.load('save/dk/b5_fold-3_best.pth', map_location=device))
model4 = MnistModel().to(device)
model4.load_state_dict(
torch.load('save/dk/b5_fold-4_best.pth', map_location=device))
model0.eval()
model1.eval()
model2.eval()
model3.eval()
model4.eval()
# Prepare Data
submit = pd.read_csv('data/sample_submission.csv')
testset = MnistDataset('data/test', 'data/sample_submission.csv',
transforms_test)
test_loader = DataLoader(testset, batch_size=8, num_workers=4)
# Test time augmentation
conf = '{"augs":["NO",\
"ROT90",\
"ROT180",\
"ROT270"],\
"mean":"ARITH"}'
model0 = TTA(model0, device, conf)
model1 = TTA(model1, device, conf)
model2 = TTA(model2, device, conf)
model3 = TTA(model3, device, conf)
model4 = TTA(model4, device, conf)
def work(self):
dataset = MnistDataset(file='data/test.csv')
torch_interface = Torch('models/net.pth', 128, name='net')
prob = torch_interface({'dataset': dataset})['prob']
np.save(f'data/net_test', prob)
