def sample_cgan_concat_given_labels(netG,
given_labels,
batch_size=100,
denorm=True,
to_numpy=True,
verbose=True):
'''
netG: pretrained generator network
given_labels: float. unnormalized labels. we need to convert them to values in [-1,1].
'''
## num of fake images will be generated
nfake = len(given_labels)
## normalize regression
labels = given_labels / max_label
## generate images
if batch_size > nfake:
batch_size = nfake
fake_images = []
## concat to avoid out of index errors
labels = np.concatenate((labels, labels[0:batch_size]), axis=0)
netG = netG.cuda()
netG.eval()
with torch.no_grad():
if verbose:
pb = SimpleProgressBar()
tmp = 0
while tmp < nfake:
z = torch.randn(batch_size, dim_gan, dtype=torch.float).cuda()
c = torch.from_numpy(labels[tmp:(tmp + batch_size)]).type(
torch.float).cuda()
batch_fake_images = netG(z, c)
if denorm: #denorm imgs to save memory
assert batch_fake_images.max().item(
) <= 1.0 and batch_fake_images.min().item() >= -1.0
batch_fake_images = batch_fake_images * 0.5 + 0.5
batch_fake_images = batch_fake_images * 255.0
batch_fake_images = batch_fake_images.type(torch.uint8)
fake_images.append(batch_fake_images.detach().cpu())
tmp += batch_size
if verbose:
pb.update(min(float(tmp) / nfake, 1) * 100)
fake_images = torch.cat(fake_images, dim=0)
#remove extra entries
fake_images = fake_images[0:nfake]
if to_numpy:
fake_images = fake_images.numpy()
return fake_images, given_labels
def cal_labelscore(PreNet, images, labels_assi, min_label_before_shift, max_label_after_shift, batch_size = 200, resize = None, norm_img = False, num_workers=0):
'''
PreNet: pre-trained CNN
images: fake images
labels_assi: assigned labels
resize: if None, do not resize; if resize = (H,W), resize images to 3 x H x W
'''
PreNet.eval()
# assume images are nxncximg_sizeximg_size
n = images.shape[0]
nc = images.shape[1] #number of channels
img_size = images.shape[2]
labels_assi = labels_assi.reshape(-1)
eval_trainset = IMGs_dataset(images, labels_assi, normalize=False)
eval_dataloader = torch.utils.data.DataLoader(eval_trainset, batch_size=batch_size, shuffle=False, num_workers=num_workers)
labels_pred = np.zeros(n+batch_size)
nimgs_got = 0
pb = SimpleProgressBar()
for batch_idx, (batch_images, batch_labels) in enumerate(eval_dataloader):
batch_images = batch_images.type(torch.float).cuda()
batch_labels = batch_labels.type(torch.float).cuda()
batch_size_curr = len(batch_labels)
if norm_img:
batch_images = normalize_images(batch_images)
batch_labels_pred, _ = PreNet(batch_images)
labels_pred[nimgs_got:(nimgs_got+batch_size_curr)] = batch_labels_pred.detach().cpu().numpy().reshape(-1)
nimgs_got += batch_size_curr
pb.update((float(nimgs_got)/n)*100)
del batch_images; gc.collect()
torch.cuda.empty_cache()
#end for batch_idx
labels_pred = labels_pred[0:n]
labels_pred = (labels_pred*max_label_after_shift)-np.abs(min_label_before_shift)
labels_assi = (labels_assi*max_label_after_shift)-np.abs(min_label_before_shift)
ls_mean = np.mean(np.abs(labels_pred-labels_assi))
ls_std = np.std(np.abs(labels_pred-labels_assi))
return ls_mean, ls_std
def sample_cgan_given_labels(netG, given_labels, batch_size=500):
'''
netG: pretrained generator network
given_labels: float. unnormalized labels. we need to convert them to values in [-1,1].
'''
## num of fake images will be generated
nfake = len(given_labels)
## normalize regression
labels = given_labels / max_label
## generate images
if batch_size > nfake:
batch_size = nfake
netG = netG.cuda()
netG.eval()
## concat to avoid out of index errors
labels = np.concatenate((labels, labels[0:batch_size]), axis=0)
fake_images = []
with torch.no_grad():
pb = SimpleProgressBar()
tmp = 0
while tmp < nfake:
z = torch.randn(batch_size, dim_z, dtype=torch.float).cuda()
c = torch.from_numpy(labels[tmp:(tmp + batch_size)]).type(
torch.float).cuda()
batch_fake_images = netG(z, c)
fake_images.append(batch_fake_images.detach().cpu().numpy())
tmp += batch_size
pb.update(min(float(tmp) / nfake, 1) * 100)
fake_images = np.concatenate(fake_images, axis=0)
#remove extra images
fake_images = fake_images[0:nfake]
#denomarlized fake images
if fake_images.max() <= 1.0:
fake_images = fake_images * 0.5 + 0.5
fake_images = (fake_images * 255.0).astype(np.uint8)
return fake_images, given_labels
def sample_ccgan_given_labels(netG, net_y2h, labels, batch_size = 500, to_numpy=True, denorm=True, verbose=True):
'''
netG: pretrained generator network
labels: float. normalized labels.
'''
nfake = len(labels)
if batch_size>nfake:
batch_size=nfake
fake_images = []
fake_labels = np.concatenate((labels, labels[0:batch_size]))
netG=netG.cuda()
netG.eval()
net_y2h = net_y2h.cuda()
net_y2h.eval()
with torch.no_grad():
if verbose:
pb = SimpleProgressBar()
n_img_got = 0
while n_img_got < nfake:
z = torch.randn(batch_size, dim_gan, dtype=torch.float).cuda()
y = torch.from_numpy(fake_labels[n_img_got:(n_img_got+batch_size)]).type(torch.float).view(-1,1).cuda()
batch_fake_images = netG(z, net_y2h(y))
if denorm: #denorm imgs to save memory
assert batch_fake_images.max().item()<=1.0 and batch_fake_images.min().item()>=-1.0
batch_fake_images = batch_fake_images*0.5+0.5
batch_fake_images = batch_fake_images*255.0
batch_fake_images = batch_fake_images.type(torch.uint8)
# assert batch_fake_images.max().item()>1
fake_images.append(batch_fake_images.cpu())
n_img_got += batch_size
if verbose:
pb.update(min(float(n_img_got)/nfake, 1)*100)
##end while
fake_images = torch.cat(fake_images, dim=0)
#remove extra entries
fake_images = fake_images[0:nfake]
fake_labels = fake_labels[0:nfake]
if to_numpy:
fake_images = fake_images.numpy()
return fake_images, fake_labels
def cal_labelscore(PreNet, images, labels_assi, min_label_before_shift, max_label_after_shift, batch_size = 500, resize = None):
'''
PreNet: pre-trained CNN
images: fake images
labels_assi: assigned labels
resize: if None, do not resize; if resize = (H,W), resize images to 3 x H x W
'''
PreNet.eval()
# assume images are nxncximg_sizeximg_size
n = images.shape[0]
nc = images.shape[1] #number of channels
img_size = images.shape[2]
labels_assi = labels_assi.reshape(-1)
# predict labels
labels_pred = np.zeros(n)
with torch.no_grad():
tmp = 0
pb = SimpleProgressBar()
for i in range(n//batch_size):
pb.update(float(i)*100/(n//batch_size))
image_tensor = torch.from_numpy(images[tmp:(tmp+batch_size)]).type(torch.float).cuda()
if resize is not None:
image_tensor = nn.functional.interpolate(image_tensor, size = resize, scale_factor=None, mode='bilinear', align_corners=False)
labels_batch, _ = PreNet(image_tensor)
labels_pred[tmp:(tmp+batch_size)] = labels_batch.detach().cpu().numpy().reshape(-1)
tmp+=batch_size
del image_tensor; gc.collect()
torch.cuda.empty_cache()
labels_pred = (labels_pred*max_label_after_shift)-np.abs(min_label_before_shift)
labels_assi = (labels_assi*max_label_after_shift)-np.abs(min_label_before_shift)
ls_mean = np.mean(np.abs(labels_pred-labels_assi))
ls_std = np.std(np.abs(labels_pred-labels_assi))
return ls_mean, ls_std
def sample_cgan_given_labels(netG,
given_labels,
class_cutoff_points,
batch_size=200,
denorm=True,
to_numpy=True,
verbose=True):
'''
given_labels: a numpy array; raw label without any normalization; not class label
class_cutoff_points: the cutoff points to determine the membership of a give label
'''
class_cutoff_points = np.array(class_cutoff_points)
num_classes = len(class_cutoff_points) - 1
nfake = len(given_labels)
given_class_labels = np.zeros(nfake)
for i in range(nfake):
curr_given_label = given_labels[i]
diff_tmp = class_cutoff_points - curr_given_label
indx_nonneg = np.where(diff_tmp >= 0)[0]
if len(indx_nonneg
) == 1: #the last element of diff_tmp is non-negative
curr_given_class_label = num_classes - 1
assert indx_nonneg[0] == num_classes
elif len(indx_nonneg) > 1:
if diff_tmp[indx_nonneg[0]] > 0:
curr_given_class_label = indx_nonneg[0] - 1
else:
curr_given_class_label = indx_nonneg[0]
given_class_labels[i] = curr_given_class_label
given_class_labels = np.concatenate(
(given_class_labels, given_class_labels[0:batch_size]))
if batch_size > nfake:
batch_size = nfake
fake_images = []
netG = netG.cuda()
netG.eval()
with torch.no_grad():
if verbose:
pb = SimpleProgressBar()
tmp = 0
while tmp < nfake:
z = torch.randn(batch_size, dim_gan, dtype=torch.float).cuda()
labels = torch.from_numpy(
given_class_labels[tmp:(tmp + batch_size)]).type(
torch.long).cuda()
if labels.max().item() > num_classes:
print("Error: max label {}".format(labels.max().item()))
batch_fake_images = netG(z, labels)
if denorm: #denorm imgs to save memory
assert batch_fake_images.max().item(
) <= 1.0 and batch_fake_images.min().item() >= -1.0
batch_fake_images = batch_fake_images * 0.5 + 0.5
batch_fake_images = batch_fake_images * 255.0
batch_fake_images = batch_fake_images.type(torch.uint8)
# assert batch_fake_images.max().item()>1
fake_images.append(batch_fake_images.detach().cpu())
tmp += batch_size
if verbose:
pb.update(min(float(tmp) / nfake, 1) * 100)
fake_images = torch.cat(fake_images, dim=0)
#remove extra entries
fake_images = fake_images[0:nfake]
if to_numpy:
fake_images = fake_images.numpy()
return fake_images, given_labels
def cal_FID(PreNetFID,
IMGSr,
IMGSg,
batch_size=500,
resize=None,
norm_img=False):
#resize: if None, do not resize; if resize = (H,W), resize images to 3 x H x W
PreNetFID.eval()
nr = IMGSr.shape[0]
ng = IMGSg.shape[0]
nc = IMGSr.shape[1] #IMGSr is nrxNCxIMG_SIExIMG_SIZE
img_size = IMGSr.shape[2]
if batch_size > min(nr, ng):
batch_size = min(nr, ng)
# print("FID: recude batch size to {}".format(batch_size))
#compute the length of extracted features
with torch.no_grad():
test_img = torch.from_numpy(IMGSr[0].reshape(
(1, nc, img_size, img_size))).type(torch.float).cuda()
if resize is not None:
test_img = nn.functional.interpolate(test_img,
size=resize,
scale_factor=None,
mode='bilinear',
align_corners=False)
if norm_img:
test_img = normalize_images(test_img)
# _, test_features = PreNetFID(test_img)
test_features = PreNetFID(test_img)
d = test_features.shape[1] #length of extracted features
Xr = np.zeros((nr, d))
Xg = np.zeros((ng, d))
#batch_size = 500
with torch.no_grad():
tmp = 0
pb1 = SimpleProgressBar()
for i in range(nr // batch_size):
imgr_tensor = torch.from_numpy(IMGSr[tmp:(tmp + batch_size)]).type(
torch.float).cuda()
if resize is not None:
imgr_tensor = nn.functional.interpolate(imgr_tensor,
size=resize,
scale_factor=None,
mode='bilinear',
align_corners=False)
if norm_img:
imgr_tensor = normalize_images(imgr_tensor)
# _, Xr_tmp = PreNetFID(imgr_tensor)
Xr_tmp = PreNetFID(imgr_tensor)
Xr[tmp:(tmp + batch_size)] = Xr_tmp.detach().cpu().numpy()
tmp += batch_size
# pb1.update(min(float(i)*100/(nr//batch_size), 100))
pb1.update(min(max(tmp / nr * 100, 100), 100))
del Xr_tmp, imgr_tensor
gc.collect()
torch.cuda.empty_cache()
tmp = 0
pb2 = SimpleProgressBar()
for j in range(ng // batch_size):
imgg_tensor = torch.from_numpy(IMGSg[tmp:(tmp + batch_size)]).type(
torch.float).cuda()
if resize is not None:
imgg_tensor = nn.functional.interpolate(imgg_tensor,
size=resize,
scale_factor=None,
mode='bilinear',
align_corners=False)
if norm_img:
imgg_tensor = normalize_images(imgg_tensor)
# _, Xg_tmp = PreNetFID(imgg_tensor)
Xg_tmp = PreNetFID(imgg_tensor)
Xg[tmp:(tmp + batch_size)] = Xg_tmp.detach().cpu().numpy()
tmp += batch_size
# pb2.update(min(float(j)*100/(ng//batch_size), 100))
pb2.update(min(max(tmp / ng * 100, 100), 100))
del Xg_tmp, imgg_tensor
gc.collect()
torch.cuda.empty_cache()
fid_score = FID(Xr, Xg, eps=1e-6)
return fid_score