def test_data(model, data_dir, device):
correct = 0
total = 0
step = 0
device = torch.device(
"cuda" if torch.cuda.is_available() and device == 'cuda' else "cpu")
model.to(device)
train_datasets, trainloader, validloader, testloader = process_data(
data_dir)
with torch.no_grad(
): #turn off gradient step to reduce computation time and use up resources
model.eval()
for images, labels in testloader:
step += 1
images, labels = images.to(device), labels.to(device)
outputs = model.forward(images)
ps = torch.exp(
outputs
) #convert to softmax probability from 0 to 1 for each image in each batch
top_p, top_class = ps.topk(1, dim=1)
equals = top_class == labels.view(*top_class.shape)
accuracy = torch.mean(equals.type(torch.FloatTensor))
print('Accuracy for batch', step,
':{:.3f}%'.format(accuracy * 100))
correct += sum(equals).item()
total += labels.size(0)
print('Number of correct classified images:', correct)
print('Number of images in test set:', total)
print('Accuracy of test set:{:.3f}%'.format(100 * correct / total))
def main():
try:
learning_rate = float(sys.argv[1])
except:
learning_rate = config.learning_rate
# number of hidden layers
hidden = int(sys.argv[2])
L = hidden + 2 # L is the total no of layers
data_file = sys.argv[3] #dataset filename
# regularization parameter
try:
lambd = float(sys.argv[4])
except:
lambd = config.lambd
# layers
try:
k = int(sys.argv[5]) # no of output nodes, ie, nodes in the Lth layers
except:
k = config.k
try:
normal_file = sys.argv[6] # normalized datset name
except:
normal_file = config.normal_file
# the number of nodes per layers excluding the biasing unit
# this will be used to build the theta array
nodes_per = []
for _ in range(2, hidden + 2):
print("Nodes in layer", _, end=" : ")
nodes_per.append(int(input()))
m, dataset, statistics = helper.process_data(data_file, normal_file, k)
print(dataset)
input()
nodes_per.append(k)
nodes_per.insert(0, len(dataset[0][0])) # the number of inputs
L, nodes_per, inital_thetas = design_thetas(nodes_per)
# final_thetas, theta_history, total_runs, final_rate, regular_param \
# = learn_thetas(inital_thetas, dataset, learning_rate, lambd)
theta_history = learn_thetas(inital_thetas, dataset, learning_rate, lambd)
final_thetas = theta_history[-1]
cont = True
while cont:
query_y(final_thetas)
print("Calculate another (Y/n) : ", end="")
ans, cont = str(input()), False
if ans == "" or ans[0] == "" or ans[0] == "y":
cont = True
def _text2vector(self, texts):
'''Given a list of Strings will convert to a numpy 3D array where each
token in the text is reprsented as a vector from the self.word2vec_model.
see semeval.helper.process_data for more details.
list of strings -> 3D numpy array (len(texts), max_number_tokens,
self.word2vec_model.vector_size)
'''
if self._max_length == 0:
raise Exception('Your model requires training first')
return helper.process_data(texts, self._word2vec_model, self._max_length)
def run(options):
# Checking if all required options are specified
if options.ensembl is None:
print '\nError: no Ensembl release specified. Use option -h to get help!\n'
quit()
try:
options.ensembl = int(options.ensembl)
except:
print '\nError: Ensembl release specified is not an integer. Use option -h to get help!\n'
quit()
if options.output is None:
print '\nError: no output file name specified. Use option -h to get help!\n'
quit()
# Must use Ensembl release >= 70 or v65
if not (options.ensembl >= 70 or options.ensembl == 65):
print '\nError: This version works with Ensembl v65 or >= v70.\n'
quit()
# Genome build
genome_build = 'GRCh37' if options.ensembl <= 75 else 'GRCh38'
# Print info
print 'Ensembl version: ' + str(options.ensembl)
print 'Reference genome: ' + genome_build
# Creating compressed output file
Nretrieved = helper.process_data(options, genome_build)
print '\nA total of ' + str(Nretrieved) + ' transcripts have been included\n'
# Indexing output file with Tabix
helper.indexFile(options)
# Removing uncompressed output file
os.remove(options.output)
# Printing out summary information
print ''
print '---------------------'
print 'Output files created:'
print '---------------------'
print options.output + '.gz (transcript database)'
print options.output + '.gz.tbi (index file)'
print options.output + '.txt (list of transcripts)'
def main():
filename = rd[1]
# no of features to represent each "thing"
# more is better, but more expensive
n = int(rd[2])
learning_rate = float(rd[3])
# regularized
try:
regular = float(rd[4])
except:
# regular = False
regular = config.regular
dataset = np.matrix(helper.process_data(filename))
# this contains the i,j and value of the filled elements
global fill_array
fill_array = getFilled(dataset)
n_m, n_u = dataset.shape
# learn the defining features as well as preferences for users
final_xi_s, final_thetas = grad_descent(n_m, n_u, n, learning_rate, regular)
print("FINAL THETAS", final_thetas, "FINAL XIS", final_xi_s, sep = "\n")
# fill in the approx blanks now that params have been learnt
new_dataset = fill_data(dataset, final_xi_s, final_thetas)
def recall(labels, pred):
recall = recall_score(labels, pred)
return recall
def extract_weighted_columns(data):
data = data[:, [1, 2, 4, 6, 8, 9, 10, 15, 17, 18, 21, 25, 27, 28, 29]]
return data
if __name__ == "__main__":
training_data = genfromtxt('training.csv', dtype=str, delimiter=',')
testing_data = genfromtxt('testing.csv', dtype=str, delimiter=',')
train_data, train_labels, train_weights = helper.process_data(
training_data)
test_data, test_labels, test_weights = helper.process_data(testing_data)
train_data_weighted = helper.normalize_data(
helper.replace_missing_values(extract_weighted_columns(train_data)))
test_data_weighted = helper.normalize_data(
helper.replace_missing_values(extract_weighted_columns(test_data)))
models.run_lr(train_data_weighted, train_labels, test_data_weighted,
test_labels, test_weights)
models.run_gnb(train_data_weighted, train_labels, test_data_weighted,
test_labels, test_weights)
models.run_gradient_boosting(train_data_weighted, train_labels,
test_data_weighted, test_labels, test_weights)
models.run_decision_tree(train_data_weighted, train_labels,
test_data_weighted, test_labels, test_weights)
import nltk
from helper import process_data
train = process_data('train')
x = train['question'][0]
parser.add_argument("-e"
"--epochs",
dest='epochs',
action="store",
type=int,
default=5,
help="number of epochs for training, default 5")
parser.add_argument("-d"
"--device",
dest='device',
action="store",
type=str,
default="cuda",
help="device for training,default cuda")
args = parser.parse_args()
train_datasets, trainloader, validloader, testloader = process_data(
args.data_dir)
model = build_model(args.arch, args.hidden_units, args.output_units)
running_losses, running_valid_losses, trained_model = train_model(
args.data_dir, model, args.learning_rate, args.epochs, args.device)
test_data(trained_model, args.data_dir, args.device)
trained_model.class_to_idx = train_datasets.class_to_idx
#device = torch.to("cuda" if torch.cuda.is_available() and args.device == 'cuda' else "cpu")
#trained_model.to(device)
torch.save(
{
'arch': args.arch,
'hidden_units': args.hidden_units,
'output_units': args.output_units,
'state_dict': trained_model.state_dict(),
'class_to_idx': trained_model.class_to_idx
def train_model(data_dir, model, learning_rate, epochs, device):
device = torch.device(
"cuda" if torch.cuda.is_available() and device == 'cuda' else "cpu")
model.to(device)
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.classifier.parameters(), lr=learning_rate)
epochs = epochs
train_datasets, trainloader, validloader, testloader = process_data(
data_dir)
running_losses, running_valid_losses = [], []
for e in range(epochs):
running_loss = 0
corrects = 0
for images, labels in trainloader:
images, labels = images.to(device), labels.to(
device) #Make sure that the training step is running on gpu
#clean up accumulated gradients before training the new batch
optimizer.zero_grad()
#Forward and backward pass
log_ps = model.forward(images)
loss = criterion(log_ps, labels)
loss.backward()
optimizer.step()
running_loss += loss.item(
) #add loss of the batch to the running loss
#use the validation datset to compare train and validation loss
else:
running_valid_loss = 0
running_accuracy = 0
with torch.no_grad():
model.eval() #set model to evaluation mode to stop dropout
for images, labels in validloader:
images, labels = images.to(device), labels.to(device)
log_ps = model.forward(images)
valid_loss = criterion(log_ps, labels)
running_valid_loss += valid_loss.item()
ps = torch.exp(log_ps)
top_p, top_class = ps.topk(1, dim=1)
equals = top_class == labels.view(*top_class.shape)
accuracy = torch.mean(equals.type(torch.FloatTensor))
running_accuracy += accuracy.item()
model.train()
print(
"Epoch:",
e + 1,
"Training loss:{:.3f}..".format(running_loss /
len(trainloader)),
"Validation loss:{:.3f}..".format(running_valid_loss /
len(validloader)),
#"Running Accuracy:{:.3f}..".format(running_accuracy),
"Validation accuracy:{:.3f}%..".format(running_accuracy * 100 /
len(validloader)))
running_losses.append(running_loss / len(trainloader))
running_valid_losses.append(running_valid_loss / len(validloader))
print('Train Losses:', running_losses, 'Validation Losses:',
running_valid_losses)
return running_losses, running_valid_losses, model
