def plot_residuals(pred,
data,
range=None,
variable_names=['pT', 'eta', 'phi', 'E'],
bins=1000,
save=None,
title=None):
alph = 0.8
residuals = (pred.numpy() - data.numpy()) / data.numpy()
for kk in np.arange(4):
plt.figure()
n_hist_pred, bin_edges, _ = plt.hist(residuals[:, kk],
label='Residuals',
alpha=alph,
bins=bins,
range=range)
if title is None:
plt.suptitle('Residuals of %s' % variable_names[kk])
else:
plt.suptitle(title)
plt.xlabel(
r'$(%s_{recon} - %s_{true}) / %s_{true}$' %
(variable_names[kk], variable_names[kk], variable_names[kk]))
plt.ylabel('Number of events')
sciy()
if save is not None:
plt.savefig(save + '_%s' % variable_names[kk])
def visualize_sample(self, path, number, shape, class_=None):
data, target = self.get_sample(number, shape)
# get sample in order from 0 to 9
target, order = target.sort()
data = data[order]
image_frame_dim = int(np.floor(np.sqrt(number)))
if shape[2] == 1:
data_np = data.numpy().reshape(number, shape[0], shape[1],
shape[2])
save_images(data_np[:image_frame_dim * image_frame_dim, :, :, :],
[image_frame_dim, image_frame_dim], path)
elif shape[2] == 3:
# data = data.numpy().reshape(number, shape[0], shape[1], shape[2])
# if self.dataset_name == 'cifar10':
data = data.numpy().reshape(number, shape[2], shape[1], shape[0])
# data = data.numpy().reshape(number, shape[0], shape[1], shape[2])
# remap between 0 and 1
# data = data - data.min()
# data = data / data.max()
data = data / 2 + 0.5 # unnormalize
make_samples_batche(data[:number], number, path)
else:
save_images(data[:image_frame_dim * image_frame_dim, :, :, :],
[image_frame_dim, image_frame_dim], path)
return data
def matplotlib_imshow(data, is_image):
if is_image: #for images
data = data / 4 + 0.5 # unnormalize
npimg = data.numpy()
plt.imshow(npimg, cmap="gray")
else: # for labels
nplbl = data.numpy()
plt.imshow(t_loader.decode_segmap(nplbl))
def train(epoch, prior):
prior = BayesianGaussianMixture(n_components=50,
covariance_type='diag',
n_init=5,
max_iter=1000)
tmp = []
for (data, _) in train_loader:
#print(data.numpy().shape)
tmp.append(data.numpy().reshape(data.numpy().shape[0], -1))
prior.fit(np.vstack(tmp))
return prior
def calculate_fvd_from_inference_result(gen_path,
ref_path='./Evaluation/ref',
num_of_video=16,
video_length=10):
VIDEO_LENGTH = video_length
print('{}'.format(video_length))
base_ref = VideoGenerateDataset(ref_path, min_len=VIDEO_LENGTH)
base_tar = VideoGenerateDataset(gen_path, min_len=VIDEO_LENGTH)
bs = num_of_video
assert bs % 16 == 0
videoloader_ref = torch.utils.data.DataLoader(
base_ref,
batch_size=bs, #len(videodataset),
drop_last=True,
shuffle=False)
videoloader_tar = torch.utils.data.DataLoader(
base_tar,
batch_size=bs, #len(videodataset),
drop_last=True,
shuffle=False)
with tqdm(total=len(videoloader_ref), dynamic_ncols=True) as pbar:
for i, data in enumerate(videoloader_ref):
images_ref = data.numpy()
break
for i, data in enumerate(videoloader_tar):
images_tar = data.numpy()
break
with tf.Graph().as_default():
ref_tf = tf.convert_to_tensor(images_ref, dtype=tf.uint8)
tar_tf = tf.convert_to_tensor(images_tar, dtype=tf.uint8)
first_set_of_videos = ref_tf #14592
second_set_of_videos = tar_tf
result = calculate_fvd(
create_id3_embedding(preprocess(first_set_of_videos, (224, 224)),
bs),
create_id3_embedding(preprocess(second_set_of_videos, (224, 224)),
bs))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
return sess.run(result)
def double_check_size(self, data, gt_boxes, im_info):
check = np.array(data.size(), dtype=np.int)
scale_list = list(cfg.TRAIN.SCALES)
scale_list.append(cfg.TRAIN.MAX_SIZE)
scale_list = np.unique(np.array(scale_list))
if not np.in1d(check, scale_list).any():
import cv2
data = data.numpy()
gt_boxes = gt_boxes.numpy()
im_info = im_info.numpy()
data = data.astype(np.float32, copy=False)
im_shape = data.shape
im_size_min = np.min(im_shape[0:2])
im_size_max = np.max(im_shape[0:2])
target_size = min(cfg.TRAIN.SCALES)
max_size = cfg.TRAIN.MAX_SIZE
im_scale = float(target_size) / float(im_size_min)
# Prevent the biggest axis from being more than MAX_SIZE
if im_scale * im_size_max > max_size:
im_scale = float(max_size) / float(im_size_max)
data = cv2.resize(data,
None,
None,
fx=im_scale,
fy=im_scale,
interpolation=cv2.INTER_LINEAR)
gt_boxes[:, 0:4] = gt_boxes[:, 0:4] * im_scale
im_info = np.array(
[[data.shape[0], data.shape[1], im_info[0][2] * im_scale]],
dtype=np.float32)
data = torch.from_numpy(data)
gt_boxes = torch.from_numpy(gt_boxes)
im_info = torch.from_numpy(im_info)
return data, gt_boxes, im_info
def get_mean_std(data_path, input_size, rgb):
tform = []
tform.append(transforms.Resize(size=input_size))
if not rgb:
tform.append(transforms.Grayscale())
tform.append(transforms.ToTensor())
tform = transforms.Compose(tform)
dset = datasets.ImageFolder(root=data_path, transform=tform)
train_loader = DataLoader(dataset=dset, batch_size=50)
scaler = StandardScaler(with_mean=True, with_std=True)
print('Computing pixel mean and stdev...')
for idx, (data, labels) in enumerate(train_loader):
if idx % 20 == 0:
print("Batch {:d} / {:d}".format(idx, len(train_loader)))
data = data.numpy()
n_channels = data.shape[1]
# reshape into [n_pixels x 3]
data = data.transpose((0, 2, 3, 1)).reshape((-1, n_channels))
# pass batch to incremental mean and stdev calculator
scaler.partial_fit(data)
print('Done, mean = ')
pixel_mean = scaler.mean_
pixel_std = scaler.scale_
print(pixel_mean)
print('std = ')
print(pixel_std)
return pixel_mean, pixel_std
def loadWaymoTestFrames(PATH):
#train_folders = ["training_0005","training_0004","training_0003","training_0002","training_0001","training_0000"]#["training_0001"]# ["training_0000", "training_0001"]
test_folders = [
"testing_0007", "testing_0006", "testing_0005", "testing_0004",
"testing_0003", "testing_0002", "testing_0001", "testing_0000"
]
data_files = [
path for x in test_folders
for path in glob(os.path.join(PATH, x, "*.tfrecord"))
]
print(data_files) #all TFRecord file list
print(len(data_files))
dataset = [
tf.data.TFRecordDataset(FILENAME, compression_type='')
for FILENAME in data_files
] #create a list of dataset for each TFRecord file
frames = [
] #store all frames = total number of TFrecord files * 40 frame(each TFrecord)
for i, data_file in enumerate(dataset):
print("Datafile: ", data_files[i]) #Each TFrecord file
for idx, data in enumerate(
data_file
): #Create frame based on Waymo API, 199 frames per TFrecord (20s, 10Hz)
# if idx % 5 != 0: #Downsample every 5 images, reduce to 2Hz, total around 40 frames
# continue
frame = open_dataset.Frame()
frame.ParseFromString(bytearray(data.numpy()))
frames.append(frame)
return frames
def train(epoch):
model.train()
train_loss = 0
for batch_idx, data in enumerate(train_loader):
noise = np.random.poisson(np.random.uniform(1, 5, 1),
data.numpy().shape)
noise = torch.from_numpy(noise.astype(float))
noise = noise.float()
noisy_data = data + noise
noisy_data = Variable(noisy_data)
noisy_data = noisy_data.cuda()
data = Variable(data)
data = data.cuda()
optimizer.zero_grad()
recon_batch = model(noisy_data)
loss = loss_function(recon_batch, data)
loss.backward()
train_loss += loss.data[0]
optimizer.step()
if batch_idx % log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader),
loss.data[0] / len(data)))
train_loss_log.append(train_loss)
print('====> Epoch: {} Average loss: {:.4f}'.format(
epoch, train_loss / len(train_loader.dataset)))
avg_train_loss_log.append(train_loss / len(train_loader.dataset))
def load_wav_to_torch(full_path, target_sr=None):
if full_path.endswith('wav') and sf is not None:
sampling_rate, data = read(
full_path) # scipy only supports .wav but reads faster...
else:
data, sampling_rate = sf.read(full_path,
always_2d=True)[:, 0] # than soundfile.
if np.issubdtype(data.dtype, np.integer): # if audio data is type int
max_mag = -np.iinfo(
data.dtype).min # maximum magnitude = min possible value of intXX
else: # if audio data is type fp32
max_mag = max(np.amax(data), -np.amin(data))
max_mag = (2**31) + 1 if max_mag > (2**15) else (
(2**15) + 1 if max_mag > 1.01 else 1.0
) # data should be either 16-bit INT, 32-bit INT or [-1 to 1] float32
data = torch.FloatTensor(data.astype(np.float32)) / max_mag
if target_sr is not None and sampling_rate != target_sr:
data = torch.from_numpy(
librosa.core.resample(data.numpy(), sampling_rate, target_sr))
sampling_rate = target_sr
data = data.clamp(min=-0.9999, max=0.9999)
return data, sampling_rate
def test(epoch, prior):
ans = np.zeros((50, 10))
for data, lab in test_loader:
C = prior.predict(data.numpy().reshape(data.numpy().shape[0], -1))
for i in xrange(len(lab)):
ans[C[i], lab[i]] += 1
print(ans)
s = np.sum(ans)
v = 0
for i in xrange(ans.shape[0]):
for j in xrange(ans.shape[1]):
if ans[i, j] > 0:
v += ans[i, j] / s * np.log(ans[i, j] / s /
(np.sum(ans[i, :]) / s) /
(np.sum(ans[:, j]) / s))
print("Mutual information: " + str(v))
def save_embedding_process(model, save_loader, feed_data, is_cuda):
fts1 = feed_data['fts1']
fts2 = feed_data['fts2']
user_embedding1_list = []
user_embedding2_list = []
model.eval()
for batch_idx, data in enumerate(save_loader):
data = data.reshape([-1])
val_user_arr = data.numpy()
v_item1 = fts1[val_user_arr]
v_item2 = fts2[val_user_arr]
if is_cuda:
v_user = torch.LongTensor(val_user_arr).cuda()
v_item1 = torch.FloatTensor(v_item1).cuda()
v_item2 = torch.FloatTensor(v_item2).cuda()
else:
v_user = torch.LongTensor(val_user_arr)
v_item1 = torch.FloatTensor(v_item1)
v_item2 = torch.FloatTensor(v_item2)
res = model.get_user_embedding(v_item1, v_item2)
user_embedding1 = res[0]
user_embedding2 = res[1]
if is_cuda:
user_embedding1 = user_embedding1.detach().cpu().numpy()
user_embedding2 = user_embedding2.detach().cpu().numpy()
else:
user_embedding1 = user_embedding1.detach().numpy()
user_embedding2 = user_embedding2.detach().numpy()
user_embedding1_list.append(user_embedding1)
user_embedding2_list.append(user_embedding2)
return np.concatenate(user_embedding1_list, axis=0), np.concatenate(user_embedding2_list, axis=0)
def do_transforms(self, data):
toPIL = vision.transforms.ToPILImage()
toTensor = vision.transforms.ToTensor()
data = toTensor(
self.transform(toPIL(data.astype('uint8').transpose(
(1, 2, 0))))) * 255
return data.numpy()
def idx2str(self, data):
'''translate a list of int back to string'''
if not hasattr(self, 'rev_vocab'):
self.rev_vocab = {v: k for k, v in self._vocab.items()}
if data.__class__ == torch.Tensor:
data = data.numpy()
return [self.rev_vocab[d] for d in data]
def test(epoch):
model.eval()
test_loss = 0
correct = 0
total = 0
for batch_idx, (data, z_class) in enumerate(test_loader):
y_class = np.eye(10)[z_class.numpy()]
data = Variable(data, volatile=True)
recon_batch, mu, logvar, y_pred = model(data)
test_loss += loss_function([recon_batch], data, mu, logvar, y_pred,
y_class, epoch)
z = y_pred.data.cpu().numpy()
for i, row in enumerate(z):
pred = np.argmax(row)
if pred == z_class[i]:
correct += 1
total += len(z_class)
test_loss /= len(test_loader.dataset)
print('Correct: ' + str(correct))
print('Total: ' + str(total))
print('====> Test set loss: ' + str(test_loss.data.cpu().numpy()[0]))
if epoch % args.eval_interval == 0:
imgs = []
# for z in zs:
# # for c in cs:
# model.eval()
# x = model.decode(z)
# imgFile = np.swapaxes((x.data).cpu().numpy()[0],0,2)
# print(imgFile.shape)
# imgs.append(imgFile)
for batch_idx, (data, y_class) in enumerate(test_loader2):
if batch_idx % 200 != 0:
continue
img = np.swapaxes(np.swapaxes(data.numpy()[0], 0, 1), 1, 2)
imgs.append(img)
data = Variable(data, volatile=True)
mu, _, _ = model.encode(data)
# print(batch_idx)
a = torch.from_numpy(np.eye(10)[y_class.numpy()]).type(
torch.FloatTensor)
img = model.decode(mu, Variable(a))
img = np.swapaxes(np.swapaxes(img.data.numpy()[0], 0, 1), 1, 2)
imgs.append(img)
imgFile = stack(imgs)
imgFile = imgFile * 255 #/ np.max(imgFile)
imgFileName = args.save_image + "_" + str(epoch) + ".png"
cv.imwrite(imgFileName, imgFile)
def MyPlotFuc(data):
import matplotlib.pyplot as plt
import numpy as np
data = data.cpu()
data = data.numpy()
data = np.where(data > 0)
plt.xlim(0, 128)
plt.ylim(0, 128)
plt.scatter(data[0], data[1], s=5)
plt.show()
def visualize_sample(self, path, number, shape):
data, target = self.get_sample(number, shape)
# get sample in order from 0 to 9
target, order = target.sort()
data = data[order]
#
# if self.transform is not None:
# tf_bacth = None
# for i in range(number):
# tf_data = TF.to_pil_image(data[i])
# tf_data = self.transform(tf_data)
# if i == 0:
# tf_bacth = torch.FloatTensor(number, tf_data.shape[0], tf_data.shape[1], tf_data.shape[2])
# tf_bacth[i] = tf_data
# data = tf_bacth
image_frame_dim = int(np.floor(np.sqrt(number)))
if shape[2] == 1:
data_np = data.numpy().reshape(number, shape[0], shape[1],
shape[2])
save_images(data_np[:image_frame_dim * image_frame_dim, :, :, :],
[image_frame_dim, image_frame_dim], path)
elif shape[2] == 3:
# data = data.numpy().reshape(number, shape[0], shape[1], shape[2])
# if self.dataset_name == 'cifar10':
data = data.numpy().reshape(number, shape[2], shape[1], shape[0])
# data = data.numpy().reshape(number, shape[0], shape[1], shape[2])
# remap between 0 and 1
# data = data - data.min()
# data = data / data.max()
data = data / 2 + 0.5 # unnormalize
make_samples_batche(data[:number], number, path)
else:
save_images(data[:image_frame_dim * image_frame_dim, :, :, :],
[image_frame_dim, image_frame_dim], path)
return data
def get_batch(self):
data = next(self.dataset_iter)
data = data.numpy().astype(self.input_data_type) / 255.0
num_files, num_data, height, width, channel = data.shape
num_batch_per_file = num_data // (self.seq_len + self.horizon)
batch = data[:, :num_batch_per_file *
(self.seq_len + self.horizon)].reshape(
num_files * num_batch_per_file,
(self.seq_len + self.horizon), height, width, -1)
return batch
def stack_examples(data):
data = data.numpy().squeeze()
sample = data[0]
n = len(data)
w, h = sample.shape
out = numpy.ones((w, h * n))
space = 1
for i in range(n):
out[space:-space,
i * h + space:(i + 1) * h - space] = data[i, space:-space,
space:-space]
scipy.misc.imsave("samples.png", out)
def generate_variant_dataset(self, variant):
"""Generate a dataset corresponding to the given MNIST variant.
The modified MNIST data will be saved in a similar fashion to
that of the original MNIST dataset. Also, presumably some randomness will be
involved, meaning the dataset will change every time this function is called.
"""
# process and save as torch files
print('Generating...')
if not os.path.exists(self.variant_folder):
os.makedirs(self.variant_folder)
def _rot(image_data):
"""Destructive rotation."""
for i in range(image_data.shape[0]):
rand_deg = np.random.random() * 360.0
image_data[i] = rotate(image_data[i], rand_deg, reshape=False)
def _bg_rand(image_data):
"""Destructive random background."""
noise = np.random.randint(0,
256,
image_data.shape,
dtype=image_data.dtype)
image_data[image_data == 0] = noise[image_data == 0]
for data_file in (self.training_file, self.test_file):
# load original MNIST data
data, targets = torch.load(
os.path.join(self.processed_folder, data_file))
modified_data = data.numpy() # shape: (n, 28, 28)
if variant == 'rot':
_rot(modified_data)
elif variant == 'bg_rand':
_bg_rand(modified_data)
elif variant == 'bg_rand_rot':
_rot(modified_data)
_bg_rand(modified_data)
with open(os.path.join(self.variant_folder, data_file), 'wb') as f:
torch.save((torch.from_numpy(modified_data), targets), f)
print('Done!')
print('Saved dataset to %s.' % self.variant_folder)
def testTorchKtop():
batch_size = 20
train_dataset = downloadData(True)
train_loader = loadData(train_dataset, batch_size, True)
images,labels = next(iter(train_loader))
data = images.view(images.shape[0],-1)
D = data.numpy()
W = cosine_similarity(D,D)
print(W)
W = torch.FloatTensor(W)
nn = torch.topk(W,batch_size)
qq = torch.randn(batch_size,batch_size)
nnInd = nn[1]
qq[nnInd] = 0
# nn = getKneibor(X=data,n_nbrs=batch_size)
print(nnInd)
print(qq)
show_img(images,batch_size)
def load_wav_to_torch(full_path,
target_sr=None,
return_empty_on_exception=False):
sampling_rate = None
try:
data, sampling_rate = sf.read(full_path,
always_2d=True) # than soundfile.
except Exception as ex:
print(f"'{full_path}' failed to load.\nException:")
print(ex)
if return_empty_on_exception:
return [], sampling_rate or target_sr or 48000
else:
raise Exception(ex)
if len(data.shape) > 1:
data = data[:, 0]
assert len(
data
) > 2 # check duration of audio file is > 2 samples (because otherwise the slice operation was on the wrong dimension)
if np.issubdtype(data.dtype, np.integer): # if audio data is type int
max_mag = -np.iinfo(
data.dtype).min # maximum magnitude = min possible value of intXX
else: # if audio data is type fp32
max_mag = max(np.amax(data), -np.amin(data))
max_mag = (2**31) + 1 if max_mag > (2**15) else (
(2**15) + 1 if max_mag > 1.01 else 1.0
) # data should be either 16-bit INT, 32-bit INT or [-1 to 1] float32
data = torch.FloatTensor(data.astype(np.float32)) / max_mag
if (torch.isinf(data) | torch.isnan(data)).any(
) and return_empty_on_exception: # resample will crash with inf/NaN inputs. return_empty_on_exception will return empty arr instead of except
return [], sampling_rate or target_sr or 48000
if target_sr is not None and sampling_rate != target_sr:
data = torch.from_numpy(
librosa.core.resample(data.numpy(), sampling_rate, target_sr))
sampling_rate = target_sr
return data, sampling_rate
def visualize_data(self):
n = len(self)
for i in range(n):
image, impulse, response, bbox, one_hot = [
data.numpy() for data in self[i]
]
image = np.moveaxis(image, 0, -1)
image *= self.config.STD_PIXEL
image += self.config.MEAN_PIXEL
image *= 255
impulse = impulse * 255
image[:, :, 0] = np.where(impulse[0] == 255, 255, image[:, :, 0])
response = np.squeeze(response) * 255
bbox = np.squeeze(bbox) * 255
Image.fromarray(image.astype(np.uint8), "RGB").show()
# for i in range(impulse.shape[0]):
# Image.fromarray(impulse[i].astype(np.uint8), "L").show()
Image.fromarray(response.astype(np.uint8), "L").show()
Image.fromarray(bbox.astype(np.uint8), "L").show()
print(self.config.CLASS_NAMES[np.argmax(one_hot)])
input()
def datasets_initialization_kcenter(data,
labels,
init_size,
init_weight,
pho_p=0,
pho_n=0):
if torch.is_tensor(data):
data = data.numpy()
if torch.is_tensor(labels):
labels = labels.numpy()
kcenter = KCenter(data, labels)
for i in range(init_size - 1):
kcenter.select_one()
unlabeled_set = torch.utils.data.TensorDataset(
torch.from_numpy(kcenter.pool).float(),
torch.from_numpy(kcenter.pool_y).float())
labeled_set = WeightedTensorDataset(
torch.from_numpy(kcenter.selected).float(),
torch.from_numpy(kcenter.selected_y).float(),
init_weight * torch.ones(init_size, 1), pho_p, pho_n)
return unlabeled_set, labeled_set
def plot_log_dist(data,
logger,
name="undefined",
save_path="./",
cumulative=False,
hist=True,
kde=False,
bins=60):
if not os.path.exists(save_dir):
os.makedirs(save_dir)
if (isinstance(data, torch.Tensor)):
data = data.numpy()
data = data.flatten()
data_log = []
count_0 = 0
count = data.shape[0]
for dat in data:
if (dat > 0.0):
data_log.append(math.log(dat, 10))
elif (dat < 0.0):
print("%s is not a positive data" % (name))
return False
else:
count_0 += 1
data_log = np.array(data_log)
note = "non-zero rate: %.4f(%d/%d)" % (
(count - count_0) / count, count - count_0, count)
if count_0 == count:
logger.write("%s is all-zero." % (name))
return True
plot_dist(data=data_log,
logger=logger,
name=name + "(log)",
save_dir=save_dir,
notes=note,
cumulative=cumulative,
hist=hist,
kde=kde,
bins=bins)
return True
def __getitem__(self, index):
data = self.data[index]
lbl = self.labels[index]
data = Image.fromarray(data.numpy(), mode='L')
if self.is_noisy:
opencvImage = np.expand_dims(np.array(data), 2)
opencvImage = addNoise(opencvImage, var=.00001)
opencvImage = np.squeeze(opencvImage, 2)
imgNoisy = Image.fromarray(opencvImage, mode='L')
if self.transform is not None:
imgNoisy = self.transform(imgNoisy)
if self.transform is not None:
data = self.transform(data)
if not self.is_noisy:
return data, lbl
else:
return data, lbl, imgNoisy
pass
def __init__(self,
label_file,
file_dir,
frame_qty,
keyframe_skip,
keyframe_interval,
cache_file_prefix,
file_ext='mp4'):
VideoFramesBase.__init__(self)
for line_num, file_name_prefix, label_str in get_labels(label_file):
lab = LABEL_DICT[label_str]
cache_file_data = VideoFramesTest.cache_file_name(
cache_file_prefix, file_name_prefix)
if os.path.exists(cache_file_data):
print(f'File {cache_file_data} is already generated!')
continue
gc.collect()
file_name = os.path.join(file_dir,
file_name_prefix + '.' + file_ext)
tensor_list = []
statinfo = os.stat(file_name)
if False and statinfo.st_size == 0:
print(f'Empty file {file_name}, generating zero tensor')
tensor_list.append(
torch.zeros(3, MIN_WIDTH_HEIGHT, MIN_WIDTH_HEIGHT))
else:
frame_iter = GetFrames(file_name)
while frame_iter.read_next():
frame = frame_iter.frame
if frame >= keyframe_skip:
if (frame - keyframe_skip) % keyframe_interval != 0:
continue
tran_img_tensor = self.get_image_tensor(
frame_iter.img, frame_iter.width,
frame_iter.height)
tensor_list.append(tran_img_tensor)
if len(tensor_list) >= frame_qty:
break
frame_iter.release()
if not tensor_list:
raise Exception(
f'Faulty video {file_name}, remove it from the list!')
data = torch.stack(tensor_list)
frame_qty = len(tensor_list)
print(f'Saving {frame_qty} frames {cache_file_data}')
np.save(cache_file_data, data.numpy())
def save_image(filename, data):
img = data.numpy()
img = np.clip(img, 0, 255)
img = img.transpose(1, 2, 0).astype("uint8")
cv2.imwrite(filename, img)
def extract_features(dtypes, opts, start_time=time.time()):
cnn = cnns.__dict__[opts.cnn](pretrained=True)
if opts.gpu:
cnn = torch.nn.DataParallel(cnn).cuda()
torch.backends.cudnn.benchmark = True
cnn.eval()
for dtype in dtypes:
path = 'datasets/{dataset}'.format(dataset=opts.dataset)
data_path = '{path}/{cnn}_{dtype}.h5'.format(path=path,
cnn=opts.cnn,
dtype=dtype)
if not os.path.isfile(data_path):
# dataset
dataset = folder.StaticImageFolder(
dataset=opts.dataset,
dtype=dtype,
transform=transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
]))
num_data = len(dataset)
labels_ = np.array([dataset.imgs[k][1] for k in range(num_data)])
# data loader
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=opts.batch_size,
shuffle=False,
num_workers=opts.workers,
pin_memory=True)
print('{dtype}; data loader; '.format(dtype=dtype), end='')
print('{time:8.3f} s'.format(time=time.time() - start_time))
# feature extraction
data_dim = 2048 # feature dimension before softmax
data = torch.zeros(num_data, data_dim)
labels = torch.zeros(num_data, out=torch.LongTensor())
for i, (inputs, targets) in enumerate(data_loader):
pos = i * opts.batch_size
inputs = Variable(inputs.cuda(),
volatile=True) if opts.gpu else Variable(
inputs, volatile=True)
data[pos:pos + inputs.size(0)] = cnn(inputs).data.cpu()
labels[pos:pos + targets.size(0)] = targets
print('{dtype}; '.format(dtype=dtype), end='')
print('{pos:7d}/{num_data:7d} i; '.format(pos=pos,
num_data=num_data),
end='')
print('{time:8.3f} s'.format(time=time.time() - start_time))
data = data.numpy()
labels = labels.numpy()
# sanity check
print('{dtype}; '.format(dtype=dtype), end='')
print('order of image path and data {consistency}consistent; ' \
.format(consistency='' if (labels_ == labels).all() else 'not '), end='')
print('{time:8.3f} s'.format(time=time.time() - start_time))
# save
with h5py.File(data_path, 'w') as f:
f.create_dataset('data',
data=data,
compression='gzip',
compression_opts=9)
f.create_dataset('labels',
data=labels,
compression='gzip',
compression_opts=9)
print('{dtype}; {time:8.3f} s'.format(dtype=dtype,
time=time.time() - start_time))
def show(self, samples=8, plot=True, rand=False):
if self.num > 0 and plot:
if self.samples['score'].size()[0] < samples:
samples = self.samples['score'].size()[0]
image_num = 0
scalar_list = []
for key, data in self.samples.items():
if len(data.size()) == 1:
scalar_list.append(key)
else:
image_num += 1
scalar_num = np.size(scalar_list)
pindex = np.arange(samples)
if rand:
pindex = random.sample(
np.arange(self.samples['score'].size()[0]).tolist(),
samples)
plt.figure(figsize=(15., 20. * samples / 8))
k = 1
col = max(2, image_num)
for p in pindex:
for key, data in self.samples.items():
data = data.numpy()
if data.ndim > 1:
plt.subplot(samples, col, k)
k += 1
if data.shape[1] == 1:
plt.imshow(data[p, 0, :, :],
cm.Greys_r,
interpolation='nearest',
vmin=0.,
vmax=1.)
else:
plt.imshow(np.array(
np.clip(data[p, :3, :, :].transpose([1, 2, 0]),
0., 1.) * 255.,
dtype=np.uint8),
interpolation='nearest',
vmin=0.,
vmax=1.)
title = ''
if image_num > 1:
title = str(key) + ' - '
for scalar in scalar_list:
title += str(scalar) + '=' + \
str(self.samples[scalar][p])[:4] + ' - '
plt.title(title[:-3])
plt.subplots_adjust(hspace=0.3)
plt.show()
k = 1
plt.figure(figsize=(10., 2.))
for scalar in scalar_list:
plt.subplot(1, scalar_num, k)
k += 1
hist, bins = np.histogram(self.samples[scalar].numpy(),
bins=50)
width = 0.7 * (bins[1] - bins[0])
center = (bins[:-1] + bins[1:]) / 2
plt.bar(center, hist, align='center', width=width)
plt.title("Histogram " + scalar)
plt.show()
