options = yaml.load(f)
f.close()
num_train = options['num_train']
num_val = options['num_val']
num_filters, filter_size = options['filters'], options['filter_size']
weight = float(options['weight_scale'])
hidden_dim = options['hidden_dim']
reg = float(options['reg'])
num_epochs = options['num_epochs']
batch_size = options['batch_size']
learning_rate = float(options['learning_rate'])
lr_decay = float(options['lr_decay'])
# read in data
X_train, y_train, X_test, y_test, X_val, y_val = process_data.read_faces_csv(
csv_file)
N_train, _, H, W = X_train.shape
N_val, _, _, _ = X_val.shape
data = {'X_train' : X_train[:num_train,:,:,:], 'y_train' : y_train[:num_train], \
"X_val" : X_val[:num_val,:,:,:], "y_val" : y_val[:num_val] }
print "Done loading data"
model = ThreeLayerConvNet(input_dim=(1, H, W), num_classes=7, num_filters=num_filters, filter_size=filter_size, \
weight_scale=weight, hidden_dim=hidden_dim, reg=reg)
solver = Solver(model,
data,
num_epochs=num_epochs,
num_val = 100 num_total = num_val + num_train num_filters = 32 num_filters2 = 64 filter_size = 5 num_classes = 7 weight_scale = 0.002088691359 hidden_dim = 500 reg = 0.000511417045343 num_epochs = 5 batch_size = 64 learning_rate = 0.000135139484703 * 100000000 lr_decay = 0.95 # read in data X_train, y_train, X_test, y_test, X_val, y_val = process_data.read_faces_csv(csv_file, num_total, use_tensorflow = True) _, H, W, _ = X_train.shape # Split data into training and validation X_val = X_train[-num_val:,:,:,:] y_val = y_train[-num_val:] X_train = X_train[:num_train,:,:,:] y_train = y_train[:num_train] # Convert y to one hot y_train_one_hot = np.zeros((num_train, num_classes)) y_train_one_hot[np.arange(num_train), y_train.astype(int)] = 1 y_val_one_hot = np.zeros((num_val, num_classes))
options = yaml.load(f)
f.close()
num_train = options['num_train']
num_val = options['num_val']
num_filters, filter_size = options['filters'], options['filter_size']
weight = float(options['weight_scale'])
hidden_dim = options['hidden_dim']
reg = float(options['reg'])
num_epochs = options['num_epochs']
batch_size = options['batch_size']
learning_rate = float(options['learning_rate'])
lr_decay = float(options['lr_decay'])
# read in data
X_train, y_train, X_test, y_test, X_val, y_val = process_data.read_faces_csv(csv_file)
# create mirrored training image, and append on X_train
N,_, row, col = X_train.shape
X_train_mirrored = np.zeros(X_train.shape)
y_train_mirrored = y_train
print X_train[0]
for i in xrange(N):
for r in xrange(row):
for c in xrange(col/2):
X_train_mirrored[i,:,r,c] = X_train[i,:,r,col-1-c]
X_train_mirrored[i,:,r,col-1-c] = X_train[i,:,r,c]
