def __init__(self, db_path, db_type, kmeans_file): # initialize the renderedImages dataset first super().__init__(db_path, db_type) # add the kmeans part self.kmeans = pickle.load(open(kmeans_file, 'rb')) self.num_clusters = self.kmeans.n_clusters self.rotations_dict = np.stack([get_R(self.kmeans.cluster_centers_[i]) for i in range(self.num_clusters)])
def testing():
model.eval()
ypred = []
ytrue = []
labels = []
for i, sample in enumerate(test_loader):
xdata = Variable(sample['xdata'].cuda())
label = Variable(sample['label'].cuda())
output = model(xdata, label)
ypred_bin = np.argmax(output[0].data.cpu().numpy(), axis=1)
ypred_res = output[1].data.cpu().numpy()
y = [
get_y(np.dot(rotations_dict[ypred_bin[j]], get_R(ypred_res[j])))
for j in range(ypred_bin.shape[0])
]
ypred.append(y)
ytrue.append(sample['ydata'].numpy())
labels.append(sample['label'].numpy())
del xdata, label, output, sample
gc.collect()
ypred = np.concatenate(ypred)
ytrue = np.concatenate(ytrue)
labels = np.concatenate(labels)
model.train()
return ytrue, ypred, labels
def get_residuals(ydata, key_rotations):
ydata_res = np.zeros((ydata.shape[0], len(key_rotations), 3))
for i in range(ydata.shape[0]):
for j in range(len(key_rotations)):
ydata_res[i,
j, :] = get_y(np.dot(key_rotations[j].T,
get_R(ydata[i])))
return ydata_res
def __init__(self, db_path, db_type, problem_type, kmeans_file): # initialize the renderedImages dataset first super().__init__(db_path, db_type) self.problem_type = problem_type # add the kmeans part self.kmeans = pickle.load(open(kmeans_file, 'rb')) self.num_clusters = self.kmeans.n_clusters #pdb.set_trace() if self.problem_type == 'm2': self.key_rotations = [get_R(y) for y in self.kmeans.cluster_centers_]
def __getitem__(self, idx): # run the item handler of the renderedImages dataset sample = super().__getitem__(idx) # update the ydata target using kmeans dictionary ydata = sample['ydata'].numpy() # rotation matrix ydata_rot = np.stack([get_R(ydata[i]) for i in range(ydata.shape[0])]) sample['ydata_rot'] = torch.from_numpy(ydata_rot).float() # bin part ydata_bin = self.kmeans.predict(ydata) sample['ydata_bin'] = torch.from_numpy(ydata_bin).long() # residual part ydata_res = np.stack([get_y(np.dot(self.rotations_dict[ydata_bin[i]].T, ydata_rot[i])) for i in range(ydata.shape[0])]) sample['ydata_res'] = torch.from_numpy(ydata_res).float() return sample
def get_residuals(ydata, ydata_bin):
ydata_res = np.zeros((ydata.shape[0], 3))
for i in range(ydata.shape[0]):
ydata_res[i, :] = get_y(
np.dot(rotations_dict[ydata_bin[i]].T, get_R(ydata[i])))
return ydata_res
# assign GPU
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
# save stuff here
results_file = os.path.join('results', args.save_str)
model_file = os.path.join('models', args.save_str + '.tar')
plots_file = os.path.join('plots', args.save_str)
log_dir = os.path.join('logs', args.save_str)
# kmeans data
kmeans_file = 'data/kmeans_dictionary_axis_angle_' + str(
args.dict_size) + '.pkl'
kmeans = pickle.load(open(kmeans_file, 'rb'))
num_clusters = kmeans.n_clusters
rotations_dict = np.stack(
[get_R(kmeans.cluster_centers_[i]) for i in range(kmeans.n_clusters)])
# relevant variables
ndim = 3
num_classes = len(classes)
# loss
mse_loss = nn.MSELoss().cuda()
ce_loss = nn.CrossEntropyLoss().cuda()
# DATA
# datasets
real_data = GBDGenerator(args.augmented_path, 'real', kmeans_file)
render_data = GBDGenerator(args.render_path, 'render', kmeans_file)
test_data = TestImages(args.pascal3d_path)
# setup data loaders
