def __init__(self, filename_obj, filename_ref=None):
super(Model, self).__init__()
# load .obj
vertices, faces = nr.load_obj(filename_obj)
self.vertices = vertices.unsqueeze(0).float32().stop_grad()
self.faces = faces.unsqueeze(0).float32().stop_grad()
# create textures
texture_size = 2
self.textures = jt.ones((1, self.faces.shape[1], texture_size, texture_size, texture_size, 3)).float32().stop_grad()
# load reference image
self.image_ref = jt.array((imread(filename_ref).max(-1) != 0).astype(np.float32)).stop_grad()
# camera parameters
self.camera_position = jt.array([6,10,-14]).float32()
# setup renderer
renderer = nr.Renderer(camera_mode='look_at')
renderer.eye = self.camera_position
self.renderer = renderer
def test_data(self):
a = jt.array([1, 2, 3])
assert (a.data == [1, 2, 3]).all()
d = a.data
a.data[1] = -2
assert (a.data == [1, -2, 3]).all()
assert (a.fetch_sync() == [1, -2, 3]).all()
li = jt.liveness_info()
del a
assert li == jt.liveness_info()
del d
assert li != jt.liveness_info()
def __init__(self, requires_grad=False):
super(PerceptualLoss, self).__init__()
self.mean_rgb = jt.array([0.485, 0.456, 0.406])
self.std_rgb = jt.array([0.229, 0.224, 0.225])
vgg_pretrained_features = models.vgg.vgg16().features
vgg_pretrained_features.load('init_models/vgg_pretrained_features.pkl')
self.slice1 = nn.Sequential()
self.slice2 = nn.Sequential()
self.slice3 = nn.Sequential()
self.slice4 = nn.Sequential()
for x in range(4):
self.slice1.append(vgg_pretrained_features[x])
for x in range(4, 9):
self.slice2.append(vgg_pretrained_features[x])
for x in range(9, 16):
self.slice3.append(vgg_pretrained_features[x])
for x in range(16, 23):
self.slice4.append(vgg_pretrained_features[x])
if (not requires_grad):
for param in self.parameters():
param = param.stop_grad()
def get_latent(self, input_image):
input_image = (input_image - 127.5) / 127.5
input_image = np.expand_dims(input_image, axis=2)
input_image = input_image.transpose(2, 0, 1)
input_image = np.expand_dims(input_image, axis=0)
input_image = input_image.astype('float32')
input_image = transform.to_tensor(jt.array(input_image))
# print(input_image.shape)
mus_mouth = self.net_encoder(input_image)
return mus_mouth
def test_avg_pool2d(self):
from torch.nn.functional import avg_pool2d as t_avg_pool2d
arr = np.random.random((2, 16, 33, 33))
jt_model = avg_pool2d(jt.array(arr), 3, 1, 1, ceil_mode=True)
torch_model = t_avg_pool2d(torch.Tensor(arr), 3, 1, 1, ceil_mode=True)
assert np.allclose(jt_model.numpy(), torch_model.numpy())
jt_model = avg_pool2d(jt.array(arr),
3,
1,
1,
ceil_mode=True,
count_include_pad=False)
torch_model = t_avg_pool2d(torch.Tensor(arr),
3,
1,
1,
ceil_mode=True,
count_include_pad=False)
assert np.allclose(jt_model.numpy(), torch_model.numpy())
print('finish')
def test_forward(self):
@contextlib.contextmanager
def check(bop_num):
jt.clean()
yield
graph = jt.dump_all_graphs()
bop = [
node for node in graph.nodes_info
if node.startswith("Op") and "broadcast_to" in node
]
assert len(bop) == bop_num, (len(bop), bop_num)
with check(1):
a = jt.array([1, 2, 3])
b = a + 1
assert (b.data == [2, 3, 4]).all()
del a, b
with check(0):
a = jt.array([1, 2, 3])
b = a + a
assert (b.data == [2, 4, 6]).all()
del a, b
def test_shape(shape1, shape2, bop_num):
with check(bop_num):
a = jt.random(shape1)
b = jt.random(shape2)
c = a + b
test_shape([3, 3, 3], [3, 3, 3], 0)
test_shape([3, 3, 3], [3, 3, 1], 1)
test_shape([3, 3, 3], [3, 1, 1], 1)
test_shape([3, 3, 3], [1, 1, 1], 1)
test_shape([3, 3, 3], [1, 1, 3], 1)
test_shape([3, 3, 3], [1, 3, 3], 1)
test_shape([3, 3, 1], [1, 3, 3], 2)
test_shape([3, 1, 3], [1, 3, 3], 2)
test_shape([3, 3], [1, 3, 3], 1)
test_shape([3, 3], [1, 3, 1], 2)
def test_unpool_diff_kernel_stride(self):
from jittor import nn
pool = nn.MaxPool2d(3, stride=2, return_indices=True)
unpool = nn.MaxUnpool2d(3, stride=2)
input = jt.array([[[[1., 2, 3, 4, 0], [5, 6, 7, 8, 0],
[9, 10, 11, 12, 0], [13, 14, 16, 15, 0],
[0, 0, 0, 0, 0]]]])
output, indices = pool(input)
out = unpool(output, indices, output_size=input.shape)
assert (out == jt.array([[[[
0.,
0.,
0.,
0.,
0.,
], [
0.,
0.,
0.,
0.,
0.,
], [
0.,
0.,
11.,
12.,
0.,
], [
0.,
0.,
32.,
0.,
0.,
], [
0.,
0.,
0.,
0.,
0.,
]]]])).all()
def farthest_point_sample(points, num_point):
"""
Input:
points: pointcloud data, [B, N, C]
num_point: number of samples
Return:
centroids: sampled pointcloud index, [B, num_point]
"""
B, N, C = points.shape
centroids = jt.zeros((B, num_point))
distance = jt.ones((B, N)) * 1e10
farthest = np.random.randint(0, N, B, dtype='l')
batch_indices = np.arange(B, dtype='l')
farthest = jt.array(farthest)
batch_indices = jt.array(batch_indices)
# jt.sync_all(True)
for i in range(num_point):
centroids[:, i] = farthest
centroid = points[batch_indices, farthest, :]
centroid = centroid.view(B, 1, 3)
dist = jt.sum((points - centroid.repeat(1, N, 1))**2, 2)
mask = dist < distance
# distance = mask.ternary(distance, dist)
# print (mask.size())
if mask.sum().data[0] > 0:
distance[mask] = dist[mask] # bug if mask.sum() == 0
farthest = jt.argmax(distance, 1)[0]
# print (farthest)
# print (farthest.shape)
# B, N, C = xyz.size()
# sample_list = random.sample(range(0, N), npoint)
# centroids = jt.zeros((1, npoint))
# centroids[0,:] = jt.array(sample_list)
# centroids = centroids.view(1, -1).repeat(B, 1)
# x_center = x[:,sample_list, :]
return centroids
def detect(original_image, min_score, max_overlap, top_k, suppress=None):
""" Detect objects in an image with a trained SSD300, and visualize the results.
Args:
original_image: image, a PIL Image
min_score: minimum threshold for a detected box to be considered a match for a certain class
max_overlap: maximum overlap two boxes can have so that the one with the lower score is not suppressed via Non-Maximum Suppression (NMS)
top_k: if there are a lot of resulting detection across all classes, keep only the top 'k'
suppress: classes that you know for sure cannot be in the image or you do not want in the image, a list
Return: annotated image, a PIL Image
"""
image = np.array(original_image).astype('float32')
H, W, C = image.shape
image = transform(image)
image = jt.array(image[np.newaxis, :]).float32()
predicted_locs, predicted_scores = model(image)
det_boxes, det_labels, det_scores = model.detect_objects(
predicted_locs,
predicted_scores,
min_score=min_score,
max_overlap=max_overlap,
top_k=top_k)
det_boxes = det_boxes[0]
original_dims = np.array([[W, H, W, H]])
det_boxes = det_boxes * original_dims
det_labels = [rev_label_map[l] for l in det_labels[0]]
if det_labels == ['background']:
return original_image
annotated_image = original_image
draw = ImageDraw.Draw(annotated_image)
font = ImageFont.truetype("ahronbd.ttf", 15)
for i in range(det_boxes.shape[0]):
if suppress is not None:
if det_labels[i] in suppress:
continue
box_location = det_boxes[i].tolist()
draw.rectangle(xy=box_location, outline=label_color_map[det_labels[i]])
draw.rectangle(xy=[l + 1. for l in box_location],
outline=label_color_map[det_labels[i]])
text_size = font.getsize(det_labels[i].upper())
text_location = [box_location[0] + 2., box_location[1] - text_size[1]]
textbox_location = [
box_location[0], box_location[1] - text_size[1],
box_location[0] + text_size[0] + 4., box_location[1]
]
draw.rectangle(xy=textbox_location,
fill=label_color_map[det_labels[i]])
draw.text(xy=text_location,
text=det_labels[i].upper(),
fill='white',
font=font)
del draw
return annotated_image
def test_ring_buffer():
buffer = jt.RingBuffer(1000)
def test_send_recv(data):
print("test send recv", type(data))
buffer.push(data)
recv = buffer.pop()
if isinstance(data, (np.ndarray, jt.Var)):
assert (recv == data).all()
else:
assert data == recv
n_byte = 0
test_send_recv(1)
n_byte += 1 + 8
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv(100000000000)
n_byte += 1 + 8
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv(1e-5)
n_byte += 1 + 8
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv(100000000000.0)
n_byte += 1 + 8
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv("float32")
n_byte += 1 + 8 + 7
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv("")
n_byte += 1 + 8 + 0
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv("xxxxxxxxxx")
n_byte += 1 + 8 + 10
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv([1, 0.2])
n_byte += 1 + 8 + 1 + 8 + 1 + 8
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv({'asd': 1})
n_byte += 1 + 8 + 1 + 8 + 3 + 1 + 8
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv(np.random.rand(10, 10))
n_byte += 1 + 16 + 2 + 10 * 10 * 8
assert n_byte == buffer.total_pop() and n_byte == buffer.total_push()
test_send_recv(test_ring_buffer)
test_send_recv(jt.array(np.random.rand(10, 10)))
expect_error(lambda: test_send_recv(np.random.rand(10, 1000)))
def sample_image(n_row, batches_done):
"""Saves a grid of generated digits ranging from 0 to n_classes"""
# Static sample
z = jt.array(np.random.normal(0, 1, (n_row**2, opt.latent_dim))).float32()
static_sample = generator(z, static_label, static_code)
save_image(static_sample.numpy(),
"images/static/%d.png" % batches_done,
nrow=n_row)
# Get varied c1 and c2
zeros = np.zeros((n_row**2, 1))
c_varied = np.repeat(np.linspace(-1, 1, n_row)[:, np.newaxis], n_row, 0)
c1 = jt.array(np.concatenate((c_varied, zeros), -1)).float32()
c2 = jt.array(np.concatenate((zeros, c_varied), -1)).float32()
sample1 = generator(static_z, static_label, c1)
sample2 = generator(static_z, static_label, c2)
save_image(sample1.numpy(),
"images/varying_c1/%d.png" % batches_done,
nrow=n_row)
save_image(sample2.numpy(),
"images/varying_c2/%d.png" % batches_done,
nrow=n_row)
def compute_gradient_penalty(D, real_samples, fake_samples):
"""Calculates the gradient penalty loss for WGAN GP"""
# Random weight term for interpolation between real and fake samples
alpha = jt.array(
np.random.random((real_samples.size(0), 1, 1, 1)).astype(np.float32))
# Get random interpolation between real and fake samples
interpolates = alpha * real_samples + ((1 - alpha) * fake_samples)
d_interpolates, _ = D(interpolates)
# Get gradient w.r.t. interpolates
gradients = jt.grad(d_interpolates, interpolates)
gradients = gradients.view(gradients.size(0), -1)
gradient_penalty = ((gradients.norm(2, dim=1) - 1)**2).mean()
return gradient_penalty
def test_wrong_fuse2(self):
a = jt.array([1])
b = jt.random([
10,
])
c = jt.random([
100,
])
bb = a * b
cc = a * c
jt.sync([bb, cc])
np.testing.assert_allclose(b.data, bb.data)
np.testing.assert_allclose(c.data, cc.data)
def run_models(self):
def to_cuda(x):
if jt.has_cuda:
return x.cuda()
return x
threshold = 1e-2
# Define numpy input image
bs = 1
test_img = np.random.random((bs,3,224,224)).astype('float32')
# Define pytorch & jittor input image
pytorch_test_img = to_cuda(torch.Tensor(test_img))
jittor_test_img = jt.array(test_img)
for test_model in self.models:
print("test model", test_model)
if test_model == "inception_v3":
test_img = np.random.random((bs,3,300,300)).astype('float32')
pytorch_test_img = to_cuda(torch.Tensor(test_img))
jittor_test_img = jt.array(test_img)
# Define pytorch & jittor model
pytorch_model = to_cuda(tcmodels.__dict__[test_model]())
jittor_model = jtmodels.__dict__[test_model]()
# Set eval to avoid dropout layer
pytorch_model.eval()
jittor_model.eval()
# Jittor loads pytorch parameters to ensure forward alignment
jittor_model.load_parameters(pytorch_model.state_dict())
# Judge pytorch & jittor forward relative error. If the differece is lower than threshold, this test passes.
pytorch_result = pytorch_model(pytorch_test_img)
jittor_result = jittor_model(jittor_test_img)
x = pytorch_result.detach().cpu().numpy() + 1
y = jittor_result.data + 1
relative_error = abs(x - y) / abs(y)
diff = relative_error.mean()
assert diff < threshold, f"[*] {test_model} forward fails..., Relative Error: {diff}"
print(f"[*] {test_model} forword passes with Relative Error {diff}")
jt.clean()
jt.gc()
torch.cuda.empty_cache()
print('all models pass test.')
def test_batchnorm_backward(self):
mpi = jt.compile_extern.mpi
data = np.random.rand(30, 3, 10, 10).astype("float32")
global_x = jt.array(data)
x = jt.array(data[mpi.world_rank() * 10:(mpi.world_rank() + 1) * 10,
...])
bn1 = nn.BatchNorm(3, sync=True)
bn2 = FakeMpiBatchNorm(3)
y1 = bn1(x)
y2 = bn2(x, global_x)
gs1 = jt.grad(y1, bn1.parameters())
gs2 = jt.grad(y2, bn2.parameters())
assert np.allclose(y1.data, y2.data,
atol=1e-5), (mpi.world_rank(), y1.data, y2.data,
y1.data - y2.data)
for i in range(len(gs1)):
assert np.allclose(gs1[i].data, gs2[i].data,
rtol=1e-3), (mpi.world_rank(), gs1[i].data,
gs2[i].data,
gs1[i].data - gs2[i].data)
def test_nll_loss(self):
tc_loss = tnn.functional.nll_loss
jt_loss = jnn.nll_loss
output = np.random.randn(10, 10).astype(np.float32)
target = np.random.randint(10, size=(10))
jt_y = jt_loss(jt.array(output), jt.array(target), reduction='mean')
tc_y = tc_loss(torch.from_numpy(output),
torch.from_numpy(target),
reduction='mean')
assert np.allclose(jt_y.numpy(), tc_y.numpy())
output = np.random.randn(10, 10).astype(np.float32)
target = np.random.randint(10, size=(10))
weight = np.random.randn(10, ).astype(np.float32)
jt_y = jt_loss(jt.array(output),
jt.array(target),
jt.array(weight),
reduction='mean')
tc_y = tc_loss(torch.from_numpy(output),
torch.from_numpy(target),
torch.from_numpy(weight),
reduction='mean')
assert np.allclose(jt_y.numpy(), tc_y.numpy())
def test_forward(self):
a = np.random.rand(1,3,224,224).astype(np.float32)
b = np.random.rand(64,3,7,7).astype(np.float32)
c = jt.mkl_ops.mkl_conv(a,b,2,3).data
a_jt = jt.array(a)
b_jt = jt.array(b)
with jt.flag_scope(enable_tuner=0,compile_options={"test_mkl_conv":1}):
c_jt = conv(a_jt, b_jt, 3, 2).data
with jt.log_capture_scope(
enable_tuner=1,
compile_options={"test_mkl_conv":2},
log_v=0, log_vprefix="tuner_manager=100,conv_tuner=1000",
) as raw_logs:
c_jt_tune = conv(a_jt, b_jt, 3, 2).data
assert np.max(c_jt-c)<1e-4 and np.max(c_jt_tune-c)<1e-4
logs = find_log_with_re(raw_logs,
"Run tuner conv: confidence\\((.*)\\) candidates\\((.*)\\)$")
assert len(logs)==1
assert logs[0][0] == '20'
assert simple_parser(logs[0][1]) == {'relay0':[1,0]}
def __getitem__(self, index):
A_path = self.files_A[index % len(self.files_A)]
B_path = self.files_B[random.randint(0, len(self.files_B) - 1)]
basenA = os.path.basename(A_path)
basenB = os.path.basename(B_path)
image_A = Image.open(A_path).convert('RGB')
image_B = Image.open(B_path).convert('RGB')
mask_A_ln = Image.open(os.path.join(self.auxdir_A+'_nose',basenA))
mask_A_le = Image.open(os.path.join(self.auxdir_A+'_eyes',basenA))
mask_A_ll = Image.open(os.path.join(self.auxdir_A+'_lips',basenA))
mask_B_ln = Image.open(os.path.join(self.auxdir_B+'_nose',basenB))
mask_B_le = Image.open(os.path.join(self.auxdir_B+'_eyes',basenB))
mask_B_ll = Image.open(os.path.join(self.auxdir_B+'_lips',basenB))
# Image transformations
params_A = get_params(self.load_h, self.load_w, self.crop_h, self.crop_w)
params_B = get_params(self.load_h, self.load_w, self.crop_h, self.crop_w)
transform_A = get_transform(params_A)
transform_A_mask = get_transform(params_A, gray=True, mask=True)
transform_B = get_transform(params_B, gray=True)
transform_B_mask = get_transform(params_B, gray=True, mask=True)
item_A = transform_A(image_A)
item_A_mask_ln = transform_A_mask(mask_A_ln)
item_A_mask_le = transform_A_mask(mask_A_le)
item_A_mask_ll = transform_A_mask(mask_A_ll)
item_B = transform_B(image_B)
item_B_mask_ln = transform_B_mask(mask_B_ln)
item_B_mask_le = transform_B_mask(mask_B_le)
item_B_mask_ll = transform_B_mask(mask_B_ll)
B_feat = np.load(os.path.join(self.auxdir_B+'_feat',basenB[:-4]+'.npy'))
item_B_label = jt.array(np.argmax(B_feat))
item_B_style = jt.array(B_feat).view(3, 1, 1).repeat(1, 128, 128)
item_B_style0 = jt.array(B_feat)
return item_A, item_B, item_A_mask_ln, item_A_mask_le, item_A_mask_ll, item_B_mask_ln, item_B_mask_le, item_B_mask_ll, item_B_label, item_B_style, item_B_style0
def test_default_var(self):
a = jt.array((2, 3, 3), np.float32)
b = a * 2.0
assert str(b.dtype) == "float32"
b = a * 2
assert str(b.dtype) == "float32"
a = jt.array((2, 3, 3), np.int32)
b = a * 2.0
assert str(b.dtype) == "float32"
b = a * 2
assert str(b.dtype) == "int32"
a = jt.array((2, 3, 3), np.float64)
b = a * 2.0
assert str(b.dtype) == "float64"
b = a * 2
assert str(b.dtype) == "float64"
a = jt.array((2, 3, 3), np.int64)
b = a * 2.0
assert str(b.dtype) == "float64"
b = a * 2
assert str(b.dtype) == "int64"
def __getitem__(self, index):
index = (index % self.dataset_size)
data_path = self.train_paths[index]
try:
data = sio.loadmat(data_path,
verify_compressed_data_integrity=False)
except Exception as e:
print(data_path, e)
return None
sample = data['surfaceSamples']
voxel = data['Volume']
cp = data['closestPoints']
voxel = jt.transform.to_tensor(jt.array(voxel)).float().unsqueeze(0)
sample = jt.transform.to_tensor(jt.array(sample)).float().t()
cp = jt.transform.to_tensor(jt.array(cp)).float().view(((-1), 3))
input_dict = {
'voxel': voxel,
'sample': sample,
'cp': cp,
'path': data_path
}
return input_dict
def test_cumprod_cpu(self):
for i in range(1,6):
for j in range(i):
x = np.random.rand(*((10,)*i))
x_jt = jt.array(x)
y_jt = jt.cumprod(x_jt, j).sqr()
g_jt = jt.grad(y_jt.sum(), x_jt)
x_tc = Variable(torch.from_numpy(x), requires_grad=True)
y_tc = torch.cumprod(x_tc, j)**2
y_tc.sum().backward()
g_tc = x_tc.grad
assert np.allclose(y_jt.numpy(), y_tc.data)
assert np.allclose(g_jt.numpy(), g_tc.data)
def look_at(vertices, eye, at=[0, 0, 0], up=[0, 1, 0]):
""""Look at" transformation of vertices. The z axis is changed to (at - eye). Original vertices are transformed to the new axis.
"""
if len(vertices.shape) != 3:
raise ValueError('vertices Tensor should have 3 dimensions')
at = jt.array(at).float32()
up = jt.array(up).float32()
if isinstance(eye, tuple):
eye = jt.array(list(eye)).float32()
else:
eye = jt.array(eye).float32()
batch_size = vertices.shape[0]
if len(eye.shape) == 1:
eye = eye.broadcast([batch_size] + eye.shape)
if len(at.shape) == 1:
at = at.broadcast([batch_size] + at.shape)
if len(up.shape) == 1:
up = up.broadcast([batch_size] + up.shape)
# create new axes
# eps is chosen as 0.5 to match the chainer version
z_axis = jt.normalize(at - eye, eps=1e-5)
x_axis = jt.normalize(jt.cross(up, z_axis), eps=1e-5)
y_axis = jt.normalize(jt.cross(z_axis, x_axis), eps=1e-5)
# create rotation matrix: [bs, 3, 3]
r = jt.contrib.concat(
(x_axis.unsqueeze(1), y_axis.unsqueeze(1), z_axis.unsqueeze(1)), dim=1)
# apply
# [bs, nv, 3] -> [bs, nv, 3] -> [bs, nv, 3]
if vertices.shape != eye.shape:
eye = eye.unsqueeze(1)
vertices = vertices - eye
vertices = jt.matmul(vertices, r.transpose(0, 2, 1)[0])
return vertices
def load_weights_from_keras(self, weights):
assert self.use_viewdirs, "Not implemented if use_viewdirs=False"
# Load pts_linears
for i in range(self.D):
idx_pts_linears = 2 * i
self.pts_linears[i].weight.data = jt.array(
np.transpose(weights[idx_pts_linears]))
self.pts_linears[i].bias.data = jt.array(
np.transpose(weights[idx_pts_linears + 1]))
# Load feature_linear
idx_feature_linear = 2 * self.D
self.feature_linear.weight.data = jt.array(
np.transpose(weights[idx_feature_linear]))
self.feature_linear.bias.data = jt.array(
np.transpose(weights[idx_feature_linear + 1]))
# Load views_linears
idx_views_linears = 2 * self.D + 2
self.views_linears[0].weight.data = jt.array(
np.transpose(weights[idx_views_linears]))
self.views_linears[0].bias.data = jt.array(
np.transpose(weights[idx_views_linears + 1]))
# Load rgb_linear
idx_rbg_linear = 2 * self.D + 4
self.rgb_linear.weight.data = jt.array(
np.transpose(weights[idx_rbg_linear]))
self.rgb_linear.bias.data = jt.array(
np.transpose(weights[idx_rbg_linear + 1]))
# Load alpha_linear
idx_alpha_linear = 2 * self.D + 6
self.alpha_linear.weight.data = jt.array(
np.transpose(weights[idx_alpha_linear]))
self.alpha_linear.bias.data = jt.array(
np.transpose(weights[idx_alpha_linear + 1]))
def look(vertices, eye, direction=[0, 1, 0], up=None):
"""
"Look" transformation of vertices.
"""
if len(vertices.shape) != 3:
raise ValueError('vertices Tensor should have 3 dimensions')
direction = jt.array(direction).float32()
if isinstance(eye, tuple):
eye = jt.array(list(eye)).float32()
else:
eye = jt.array(eye).float32()
if up is None:
up = jt.array([0, 1, 0]).float32()
if len(eye.shape) == 1:
eye = eye.unsqueeze(0)
if len(direction.shape) == 1:
direction = direction.unsqueeze(0)
if len(up.shape) == 1:
up = up.unsqueeze(0)
# create new axes
z_axis = jt.normalize(direction, eps=1e-5)
x_axis = jt.normalize(jt.cross(up, z_axis), eps=1e-5)
y_axis = jt.normalize(jt.cross(z_axis, x_axis), eps=1e-5)
# create rotation matrix: [bs, 3, 3]
r = jt.contrib.concat(
(x_axis.unsqueeze(1), y_axis.unsqueeze(1), z_axis.unsqueeze(1)), dim=1)
# apply
# [bs, nv, 3] -> [bs, nv, 3] -> [bs, nv, 3]
if vertices.shape != eye.shape:
eye = eye.unsqueeze(1)
vertices = vertices - eye
vertices = jt.matmul(vertices, r.transpose(0, 2, 1))
return vertices
def R1Penalty(self, real_img, height, alpha):
# TODO: use_loss_scaling, for fp16
# apply_loss_scaling = lambda x: x * torch.exp(x * torch.Tensor([np.float32(np.log(2.0))]).to(real_img.device))
apply_loss_scaling = lambda x: x * jt.exp(x * jt.array(
[np.float32(np.log(2.0))]))
# undo_loss_scaling = lambda x: x * torch.exp(-x * torch.Tensor([np.float32(np.log(2.0))]).to(real_img.device))
undo_loss_scaling = lambda x: x * jt.exp(-x * jt.array(
[np.float32(np.log(2.0))]))
# real_img = torch.autograd.Variable(real_img, requires_grad=True)
real_img = init.constant(real_img.shape, 'float32', real_img)
assert not real_img.is_stop_grad()
real_logit = self.dis(real_img, height, alpha)
# real_logit = apply_loss_scaling(torch.sum(real_logit))
# real_grads = torch.autograd.grad(outputs=real_logit, inputs=real_img,
# grad_outputs=torch.ones(real_logit.size()).to(real_img.device),
# create_graph=True, retain_graph=True)[0].view(real_img.size(0), -1)
real_grads = jt.grad(real_logit, real_img).view(real_img.size(0), -1)
# real_grads = undo_loss_scaling(real_grads)
# r1_penalty = torch.sum(torch.mul(real_grads, real_grads))
r1_penalty = jt.sum(jt.multiply(real_grads, real_grads))
return r1_penalty
def test_lived(self):
jt.clean()
check(0,0,0)
a = jt.array(1.0).stop_fuse()
a.name('a')
b = jt.array(1.0).stop_fuse()
b.name('b')
check(2,2,2)
c = a * b
c.name('c')
check(3,3,3)
vc = c.numpy()
check(3,3,1)
da, db = jt.grad(c, [a, b])
da.name('da')
db.name('db')
check(5,6,4) # dc, 3, da, 1, db, 1
del a, b, c
check(2,5,3)
da.sync(), db.sync()
check(2,2,0)
del da, db
check(0,0,0)
def read_image(input_path, *, sidelength=256, channels=1):
image = Image.open(input_path).resize((sidelength, sidelength))
if channels == 1:
image = image.convert('L')
elif channels == 3:
image = image.convert('RGB')
else:
raise ValueError()
image_arr = jittor.array(np.array(image))
if channels == 1:
image_arr = image_arr.unsqueeze(-1)
image_arr /= 255
image_arr = 2 * image_arr - 1
return image_arr
def test_cross(self):
arr1 = np.random.randn(16, 3, 224, 224, 3)
arr2 = np.random.randn(16, 3, 224, 224, 3)
check_equal(
torch.Tensor(arr1).cross(torch.Tensor(arr2), dim=1),
jt.array(arr1).cross(jt.array(arr2), dim=1), 1e-1)
check_equal(
torch.Tensor(arr1).cross(torch.Tensor(arr2), dim=-4),
jt.array(arr1).cross(jt.array(arr2), dim=-4), 1e-1)
check_equal(
torch.Tensor(arr1).cross(torch.Tensor(arr2), dim=-1),
jt.array(arr1).cross(jt.array(arr2), dim=-1), 1e-1)
check_equal(
torch.Tensor(arr1).cross(torch.Tensor(arr2), dim=4),
jt.array(arr1).cross(jt.array(arr2), dim=4), 1e-1)
print('pass cross test ...')
def test(self):
a = jt.array([1,2,3])
a.sync()
assert a.compile_options=={}
a.compile_options = {"compile_shapes":1}
assert a.compile_options=={"compile_shapes":1}
b = a+a
assert b.compile_options=={}
with jt.flag_scope(compile_options={"compile_shapes":1}):
c = a+b
assert c.compile_options=={"compile_shapes":1}
with jt.profile_scope() as report:
c.sync()
assert len(report)==2 and "compile_shapes:1" in report[1][0]
def _calcLoss(self, netOutput):
mask_loss_func = CrossEntropyLoss(ignore_index=255)
gts = []
for masks, Matrixs in zip(self.batchmasks, self.maskAlignMatrixs):
for mask, matrix in zip(masks, Matrixs):
gts.append(
cv2.warpAffine(mask, matrix[0:2],
(self.size_output, self.size_output)))
gts = jt.array(np.array(gts)).int32()
#print(gts)
loss = mask_loss_func(netOutput, gts)
return loss