def get_loss(self, image_a_pred, image_b_pred, mask_a, mask_b):
loss = 0
# get the nonzero indices
mask_a_indices_flat = torch.nonzero(mask_a)
mask_b_indices_flat = torch.nonzero(mask_b)
if len(mask_a_indices_flat) == 0:
return Variable(torch.cuda.LongTensor([0]), requires_grad=True)
if len(mask_b_indices_flat) == 0:
return Variable(torch.cuda.LongTensor([0]), requires_grad=True)
# take 5000 random pixel samples of the object, using the mask
num_samples = 10000
rand_numbers_a = (torch.rand(num_samples)*len(mask_a_indices_flat)).cuda()
rand_indices_a = Variable(torch.floor(rand_numbers_a).type(torch.cuda.LongTensor), requires_grad=False)
randomized_mask_a_indices_flat = torch.index_select(mask_a_indices_flat, 0, rand_indices_a).squeeze(1)
rand_numbers_b = (torch.rand(num_samples)*len(mask_b_indices_flat)).cuda()
rand_indices_b = Variable(torch.floor(rand_numbers_b).type(torch.cuda.LongTensor), requires_grad=False)
randomized_mask_b_indices_flat = torch.index_select(mask_b_indices_flat, 0, rand_indices_b).squeeze(1)
# index into the image and get descriptors
M_margin = 0.5 # margin parameter
random_img_a_object_descriptors = torch.index_select(image_a_pred, 1, randomized_mask_a_indices_flat)
random_img_b_object_descriptors = torch.index_select(image_b_pred, 1, randomized_mask_b_indices_flat)
pixel_wise_loss = (random_img_a_object_descriptors - random_img_b_object_descriptors).pow(2).sum(dim=2)
pixel_wise_loss = torch.add(pixel_wise_loss, -2*M_margin)
zeros_vec = torch.zeros_like(pixel_wise_loss)
loss += torch.max(zeros_vec, pixel_wise_loss).sum()
return loss
def __call__(self, spec_f):
spec_f, is_variable = _check_is_variable(spec_f)
n_fft = spec_f.size(2)
m_min = 0. if self.f_min == 0 else 2595 * np.log10(1. + (self.f_min / 700))
m_max = 2595 * np.log10(1. + (self.f_max / 700))
m_pts = torch.linspace(m_min, m_max, self.n_mels + 2)
f_pts = (700 * (10**(m_pts / 2595) - 1))
bins = torch.floor(((n_fft - 1) * 2) * f_pts / self.sr).long()
fb = torch.zeros(n_fft, self.n_mels)
for m in range(1, self.n_mels + 1):
f_m_minus = bins[m - 1].item()
f_m = bins[m].item()
f_m_plus = bins[m + 1].item()
if f_m_minus != f_m:
fb[f_m_minus:f_m, m - 1] = (torch.arange(f_m_minus, f_m) - f_m_minus) / (f_m - f_m_minus)
if f_m != f_m_plus:
fb[f_m:f_m_plus, m - 1] = (f_m_plus - torch.arange(f_m, f_m_plus)) / (f_m_plus - f_m)
fb = Variable(fb)
spec_m = torch.matmul(spec_f, fb) # (c, l, n_fft) dot (n_fft, n_mels) -> (c, l, n_mels)
return spec_m if is_variable else spec_m.data
def pixelcnn_generate(self, z1, z2):
# Sampling from PixelCNN
x_zeros = torch.zeros(
(z1.size(0), self.args.input_size[0], self.args.input_size[1], self.args.input_size[2]))
if self.args.cuda:
x_zeros = x_zeros.cuda()
for i in range(self.args.input_size[1]):
for j in range(self.args.input_size[2]):
samples_mean, samples_logvar = self.p_x(Variable(x_zeros, volatile=True), z1, z2)
samples_mean = samples_mean.view(samples_mean.size(0), self.args.input_size[0], self.args.input_size[1],
self.args.input_size[2])
if self.args.input_type == 'binary':
probs = samples_mean[:, :, i, j].data
x_zeros[:, :, i, j] = torch.bernoulli(probs).float()
samples_gen = samples_mean
elif self.args.input_type == 'gray' or self.args.input_type == 'continuous':
binsize = 1. / 256.
samples_logvar = samples_logvar.view(samples_mean.size(0), self.args.input_size[0],
self.args.input_size[1], self.args.input_size[2])
means = samples_mean[:, :, i, j].data
logvar = samples_logvar[:, :, i, j].data
# sample from logistic distribution
u = torch.rand(means.size()).cuda()
y = torch.log(u) - torch.log(1. - u)
sample = means + torch.exp(logvar) * y
x_zeros[:, :, i, j] = torch.floor(sample / binsize) * binsize
samples_gen = samples_mean
return samples_gen
def random_sample_from_masked_image_torch(img_mask, num_samples):
"""
:param img_mask: Numpy array [H,W] or torch.Tensor with shape [H,W]
:type img_mask:
:param num_samples: an integer
:type num_samples:
:return: tuple of torch.LongTensor in (u,v) format. Each torch.LongTensor has shape
[num_samples]
:rtype:
"""
image_height, image_width = img_mask.shape
if isinstance(img_mask, np.ndarray):
img_mask_torch = torch.from_numpy(img_mask).float()
else:
img_mask_torch = img_mask
# This code would randomly subsample from the mask
mask = img_mask_torch.view(image_width*image_height,1).squeeze(1)
mask_indices_flat = torch.nonzero(mask)
if len(mask_indices_flat) == 0:
return (None, None)
rand_numbers = torch.rand(num_samples)*len(mask_indices_flat)
rand_indices = torch.floor(rand_numbers).long()
uv_vec_flattened = torch.index_select(mask_indices_flat, 0, rand_indices).squeeze(1)
uv_vec = utils.flattened_pixel_locations_to_u_v(uv_vec_flattened, image_width)
return uv_vec
def _PyramidRoI_Feat(self, feat_maps, rois, im_info):
''' roi pool on pyramid feature maps'''
# do roi pooling based on predicted rois
img_area = im_info[0][0] * im_info[0][1]
h = rois.data[:, 4] - rois.data[:, 2] + 1
w = rois.data[:, 3] - rois.data[:, 1] + 1
roi_level = torch.log(torch.sqrt(h * w) / 224.0) / np.log(2)
roi_level = torch.floor(roi_level + 4)
# --------
# roi_level = torch.log(torch.sqrt(h * w) / 224.0)
# roi_level = torch.round(roi_level + 4)
# ------
roi_level[roi_level < 2] = 2
roi_level[roi_level > 5] = 5
# roi_level.fill_(5)
if cfg.POOLING_MODE == 'crop':
# pdb.set_trace()
# pooled_feat_anchor = _crop_pool_layer(base_feat, rois.view(-1, 5))
# NOTE: need to add pyrmaid
grid_xy = _affine_grid_gen(rois, feat_maps.size()[2:], self.grid_size) ##
grid_yx = torch.stack([grid_xy.data[:,:,:,1], grid_xy.data[:,:,:,0]], 3).contiguous()
roi_pool_feat = self.RCNN_roi_crop(feat_maps, Variable(grid_yx).detach()) ##
if cfg.CROP_RESIZE_WITH_MAX_POOL:
roi_pool_feat = F.max_pool2d(roi_pool_feat, 2, 2)
elif cfg.POOLING_MODE == 'align':
roi_pool_feats = []
box_to_levels = []
for i, l in enumerate(range(2, 6)):
if (roi_level == l).sum() == 0:
continue
idx_l = (roi_level == l).nonzero().squeeze()
box_to_levels.append(idx_l)
scale = feat_maps[i].size(2) / im_info[0][0]
feat = self.RCNN_roi_align(feat_maps[i], rois[idx_l], scale)
roi_pool_feats.append(feat)
roi_pool_feat = torch.cat(roi_pool_feats, 0)
box_to_level = torch.cat(box_to_levels, 0)
idx_sorted, order = torch.sort(box_to_level)
roi_pool_feat = roi_pool_feat[order]
elif cfg.POOLING_MODE == 'pool':
roi_pool_feats = []
box_to_levels = []
for i, l in enumerate(range(2, 6)):
if (roi_level == l).sum() == 0:
continue
idx_l = (roi_level == l).nonzero().squeeze()
box_to_levels.append(idx_l)
scale = feat_maps[i].size(2) / im_info[0][0]
feat = self.RCNN_roi_pool(feat_maps[i], rois[idx_l], scale)
roi_pool_feats.append(feat)
roi_pool_feat = torch.cat(roi_pool_feats, 0)
box_to_level = torch.cat(box_to_levels, 0)
idx_sorted, order = torch.sort(box_to_level)
roi_pool_feat = roi_pool_feat[order]
return roi_pool_feat
def _create_dummy_data(filename):
data = torch.rand(num_examples * maxlen)
data = 97 + torch.floor(26 * data).int()
with open(os.path.join(data_dir, filename), 'w') as h:
offset = 0
for _ in range(num_examples):
ex_len = random.randint(1, maxlen)
ex_str = ' '.join(map(chr, data[offset:offset+ex_len]))
print(ex_str, file=h)
offset += ex_len
def th_random_choice(a, n_samples=1, replace=True, p=None):
"""
Parameters
-----------
a : 1-D array-like
If a th.Tensor, a random sample is generated from its elements.
If an int, the random sample is generated as if a was th.range(n)
n_samples : int, optional
Number of samples to draw. Default is None, in which case a
single value is returned.
replace : boolean, optional
Whether the sample is with or without replacement
p : 1-D array-like, optional
The probabilities associated with each entry in a.
If not given the sample assumes a uniform distribution over all
entries in a.
Returns
--------
samples : 1-D ndarray, shape (size,)
The generated random samples
"""
if isinstance(a, int):
a = th.arange(0, a)
if p is None:
if replace:
idx = th.floor(th.rand(n_samples)*a.size(0)).long()
else:
idx = th.randperm(len(a))[:n_samples]
else:
if abs(1.0-sum(p)) > 1e-3:
raise ValueError('p must sum to 1.0')
if not replace:
raise ValueError('replace must equal true if probabilities given')
idx_vec = th.cat([th.zeros(round(p[i]*1000))+i for i in range(len(p))])
idx = (th.floor(th.rand(n_samples)*999)).long()
idx = idx_vec[idx].long()
selection = a[idx]
if n_samples == 1:
selection = selection[0]
return selection
def tanh_quantize(input, bits):
assert bits >= 1, bits
if bits == 1:
return torch.sign(input)
input = torch.tanh(input) # [-1, 1]
input_rescale = (input + 1.0) / 2 #[0, 1]
n = math.pow(2.0, bits) - 1
v = torch.floor(input_rescale * n + 0.5) / n
v = 2 * v - 1 # [-1, 1]
v = 0.5 * torch.log((1 + v) / (1 - v)) # arctanh
return v
def linear_quantize(input, sf, bits):
assert bits >= 1, bits
if bits == 1:
return torch.sign(input) - 1
delta = math.pow(2.0, -sf)
bound = math.pow(2.0, bits-1)
min_val = - bound
max_val = bound - 1
rounded = torch.floor(input / delta + 0.5)
clipped_value = torch.clamp(rounded, min_val, max_val) * delta
return clipped_value
def __call__(self, boxlists):
"""
Arguments:
boxlists (list[BoxList])
"""
# Compute level ids
s = torch.sqrt(cat([boxlist.area() for boxlist in boxlists]))
# Eqn.(1) in FPN paper
target_lvls = torch.floor(self.lvl0 + torch.log2(s / self.s0 + self.eps))
target_lvls = torch.clamp(target_lvls, min=self.k_min, max=self.k_max)
return target_lvls.to(torch.int64) - self.k_min
def min_max_quantize(input, bits):
assert bits >= 1, bits
if bits == 1:
return torch.sign(input) - 1
min_val, max_val = input.min(), input.max()
if isinstance(min_val, Variable):
max_val = float(max_val.data.cpu().numpy()[0])
min_val = float(min_val.data.cpu().numpy()[0])
input_rescale = (input - min_val) / (max_val - min_val)
n = math.pow(2.0, bits) - 1
v = torch.floor(input_rescale * n + 0.5) / n
v = v * (max_val - min_val) + min_val
return v
def log_Logistic_256(x, mean, logvar, average=False, reduce=True, dim=None):
bin_size = 1. / 256.
# implementation like https://github.com/openai/iaf/blob/master/tf_utils/distributions.py#L28
scale = torch.exp(logvar)
x = (torch.floor(x / bin_size) * bin_size - mean) / scale
cdf_plus = torch.sigmoid(x + bin_size/scale)
cdf_minus = torch.sigmoid(x)
# calculate final log-likelihood for an image
log_logist_256 = - torch.log(cdf_plus - cdf_minus + 1.e-7)
if reduce:
if average:
return torch.mean(log_logist_256, dim)
else:
return torch.sum(log_logist_256, dim)
else:
return log_logist_256
def fit_positive(rows, cols, yx_min, yx_max, anchors):
device_id = anchors.get_device() if torch.cuda.is_available() else None
batch_size, num, _ = yx_min.size()
num_anchors, _ = anchors.size()
valid = torch.prod(yx_min < yx_max, -1)
center = (yx_min + yx_max) / 2
ij = torch.floor(center)
i, j = torch.unbind(ij.long(), -1)
index = i * cols + j
anchors2 = anchors / 2
iou_matrix = utils.iou.torch.iou_matrix((yx_min - center).view(-1, 2), (yx_max - center).view(-1, 2), -anchors2, anchors2).view(batch_size, -1, num_anchors)
iou, index_anchor = iou_matrix.max(-1)
_positive = []
cells = rows * cols
for valid, index, index_anchor in zip(torch.unbind(valid), torch.unbind(index), torch.unbind(index_anchor)):
index, index_anchor = (t[valid] for t in (index, index_anchor))
t = utils.ensure_device(torch.ByteTensor(cells, num_anchors).zero_(), device_id)
t[index, index_anchor] = 1
_positive.append(t)
return torch.stack(_positive)
def compute_frustum_bounds(self, world_to_grid, camera_to_world):
corner_points = camera_to_world.new(8, 4, 1).fill_(1)
# depth min
corner_points[0,:3,0] = self.depth_to_skeleton(0, 0, self.depth_min)
corner_points[1,:3,0] = self.depth_to_skeleton(self.image_dims[0] - 1, 0, self.depth_min)
corner_points[2,:3,0] = self.depth_to_skeleton(self.image_dims[0] - 1, self.image_dims[1] - 1, self.depth_min)
corner_points[3,:3,0] = self.depth_to_skeleton(0, self.image_dims[1] - 1, self.depth_min)
# depth max
corner_points[4,:3,0] = self.depth_to_skeleton(0, 0, self.depth_max)
corner_points[5,:3,0] = self.depth_to_skeleton(self.image_dims[0] - 1, 0, self.depth_max)
corner_points[6,:3,0] = self.depth_to_skeleton(self.image_dims[0] - 1, self.image_dims[1] - 1, self.depth_max)
corner_points[7,:3,0] = self.depth_to_skeleton(0, self.image_dims[1] - 1, self.depth_max)
p = torch.bmm(camera_to_world.repeat(8, 1, 1), corner_points)
pl = torch.round(torch.bmm(world_to_grid.repeat(8, 1, 1), torch.floor(p)))
pu = torch.round(torch.bmm(world_to_grid.repeat(8, 1, 1), torch.ceil(p)))
bbox_min0, _ = torch.min(pl[:, :3, 0], 0)
bbox_min1, _ = torch.min(pu[:, :3, 0], 0)
bbox_min = np.minimum(bbox_min0, bbox_min1)
bbox_max0, _ = torch.max(pl[:, :3, 0], 0)
bbox_max1, _ = torch.max(pu[:, :3, 0], 0)
bbox_max = np.maximum(bbox_max0, bbox_max1)
return bbox_min, bbox_max
def create_non_correspondences(uv_b_matches, img_b_shape, num_non_matches_per_match=100, img_b_mask=None):
"""
Takes in pixel matches (uv_b_matches) that correspond to matches in another image, and generates non-matches by just sampling in image space.
Optionally, the non-matches can be sampled from a mask for image b.
Returns non-matches as pixel positions in image b.
Please see 'coordinate_conventions.md' documentation for an explanation of pixel coordinate conventions.
## Note that arg uv_b_matches are the outputs of batch_find_pixel_correspondences()
:param uv_b_matches: tuple of torch.FloatTensors, where each FloatTensor is length n, i.e.:
(torch.FloatTensor, torch.FloatTensor)
:param img_b_shape: tuple of (H,W) which is the shape of the image
(optional)
:param num_non_matches_per_match: int
(optional)
:param img_b_mask: torch.FloatTensor (can be cuda or not)
- masked image, we will select from the non-zero entries
- shape is H x W
:return: tuple of torch.FloatTensors, i.e. (torch.FloatTensor, torch.FloatTensor).
- The first element of the tuple is all "u" pixel positions, and the right element of the tuple is all "v" positions
- Each torch.FloatTensor is of shape torch.Shape([num_matches, non_matches_per_match])
- This shape makes it so that each row of the non-matches corresponds to the row for the match in uv_a
"""
image_width = img_b_shape[1]
image_height = img_b_shape[0]
if uv_b_matches == None:
return None
num_matches = len(uv_b_matches[0])
def get_random_uv_b_non_matches():
return pytorch_rand_select_pixel(width=image_width,height=image_height,
num_samples=num_matches*num_non_matches_per_match)
if img_b_mask is not None:
img_b_mask_flat = img_b_mask.view(-1,1).squeeze(1)
mask_b_indices_flat = torch.nonzero(img_b_mask_flat)
if len(mask_b_indices_flat) == 0:
print "warning, empty mask b"
uv_b_non_matches = get_random_uv_b_non_matches()
else:
num_samples = num_matches*num_non_matches_per_match
rand_numbers_b = torch.rand(num_samples)*len(mask_b_indices_flat)
rand_indices_b = torch.floor(rand_numbers_b).long()
randomized_mask_b_indices_flat = torch.index_select(mask_b_indices_flat, 0, rand_indices_b).squeeze(1)
uv_b_non_matches = (randomized_mask_b_indices_flat%image_width, randomized_mask_b_indices_flat/image_width)
else:
uv_b_non_matches = get_random_uv_b_non_matches()
# for each in uv_a, we want non-matches
# first just randomly sample "non_matches"
# we will later move random samples that were too close to being matches
uv_b_non_matches = (uv_b_non_matches[0].view(num_matches,num_non_matches_per_match), uv_b_non_matches[1].view(num_matches,num_non_matches_per_match))
# uv_b_matches can now be used to make sure no "non_matches" are too close
# to preserve tensor size, rather than pruning, we can perturb these in pixel space
copied_uv_b_matches_0 = torch.t(uv_b_matches[0].repeat(num_non_matches_per_match, 1))
copied_uv_b_matches_1 = torch.t(uv_b_matches[1].repeat(num_non_matches_per_match, 1))
diffs_0 = copied_uv_b_matches_0 - uv_b_non_matches[0].type(dtype_float)
diffs_1 = copied_uv_b_matches_1 - uv_b_non_matches[1].type(dtype_float)
diffs_0_flattened = diffs_0.view(-1,1)
diffs_1_flattened = diffs_1.view(-1,1)
diffs_0_flattened = torch.abs(diffs_0_flattened).squeeze(1)
diffs_1_flattened = torch.abs(diffs_1_flattened).squeeze(1)
need_to_be_perturbed = torch.zeros_like(diffs_0_flattened)
ones = torch.zeros_like(diffs_0_flattened)
num_pixels_too_close = 1.0
threshold = torch.ones_like(diffs_0_flattened)*num_pixels_too_close
# determine which pixels are too close to being matches
need_to_be_perturbed = where(diffs_0_flattened < threshold, ones, need_to_be_perturbed)
need_to_be_perturbed = where(diffs_1_flattened < threshold, ones, need_to_be_perturbed)
minimal_perturb = num_pixels_too_close/2
minimal_perturb_vector = (torch.rand(len(need_to_be_perturbed))*2).floor()*(minimal_perturb*2)-minimal_perturb
std_dev = 10
random_vector = torch.randn(len(need_to_be_perturbed))*std_dev + minimal_perturb_vector
perturb_vector = need_to_be_perturbed*random_vector
uv_b_non_matches_0_flat = uv_b_non_matches[0].view(-1,1).type(dtype_float).squeeze(1)
uv_b_non_matches_1_flat = uv_b_non_matches[1].view(-1,1).type(dtype_float).squeeze(1)
uv_b_non_matches_0_flat = uv_b_non_matches_0_flat + perturb_vector
uv_b_non_matches_1_flat = uv_b_non_matches_1_flat + perturb_vector
# now just need to wrap around any that went out of bounds
# handle wrapping in width
lower_bound = 0.0
upper_bound = image_width*1.0 - 1
lower_bound_vec = torch.ones_like(uv_b_non_matches_0_flat) * lower_bound
upper_bound_vec = torch.ones_like(uv_b_non_matches_0_flat) * upper_bound
uv_b_non_matches_0_flat = where(uv_b_non_matches_0_flat > upper_bound_vec,
uv_b_non_matches_0_flat - upper_bound_vec,
uv_b_non_matches_0_flat)
uv_b_non_matches_0_flat = where(uv_b_non_matches_0_flat < lower_bound_vec,
uv_b_non_matches_0_flat + upper_bound_vec,
uv_b_non_matches_0_flat)
# handle wrapping in height
lower_bound = 0.0
upper_bound = image_height*1.0 - 1
lower_bound_vec = torch.ones_like(uv_b_non_matches_1_flat) * lower_bound
upper_bound_vec = torch.ones_like(uv_b_non_matches_1_flat) * upper_bound
uv_b_non_matches_1_flat = where(uv_b_non_matches_1_flat > upper_bound_vec,
uv_b_non_matches_1_flat - upper_bound_vec,
uv_b_non_matches_1_flat)
uv_b_non_matches_1_flat = where(uv_b_non_matches_1_flat < lower_bound_vec,
uv_b_non_matches_1_flat + upper_bound_vec,
uv_b_non_matches_1_flat)
return (uv_b_non_matches_0_flat.view(num_matches, num_non_matches_per_match),
uv_b_non_matches_1_flat.view(num_matches, num_non_matches_per_match))
def mod289(x):
return x - torch.floor(x * (1.0 / 289.0)) * 289.0
def mask2d(batch_size, dim, keep_prob, device):
mask = torch.floor(torch.rand(batch_size, dim, device=device) +
keep_prob) / keep_prob
return mask
def fill_norm(yx_min, yx_max, anchors):
center = (yx_min + yx_max) / 2
ij = torch.floor(center)
center_offset = center - ij
size = yx_max - yx_min
return center_offset, torch.log(size / anchors.view(1, -1, 2))
def generate_vcmr_predictions_from_res(eval_res, max_prop_per_query=300, query_bsz_in_sort=1000):
""" This function is for Video Corpus Moment Retrieval (VCMR).
Generate prediction file which could be evaluated using standalone_eval.eval.
Args:
eval_res: dict(
query_meta=query_meta_list, # N_q * dict(), each dict is {"desc_id": int, "desc": str}
video_meta=video_meta_list, # N_videos * dict(), {"vid_name": str, "duration": float, "proposals": ndarray}
video2idx=eval_dataset.video2idx, # dict {vid_name: index}
video_bsz_in_sort=[], # N_videos * (N_q, N_prop)
)
max_prop_per_query: int or None. If None, generate ranking for all possible moments, else generate top {}.
query_bsz_in_sort: int, only sort a subset of queries at a time, it will be too large to sort all queries.
return:
list(dicts): each dict is dict(desc=str, desc_id=int, predictions=list(sublist)),
each sublist is [vid_name (str), st (float), ed (float), score (float)], score is negative distance.
"""
# video2idx
video2idx = eval_res["video2idx"]
# (N_videos, N_prop, N_q), (N_videos, N_prop)
padded_dist, padded_mask = pad_sequences_1d([e.transpose(0, 1) for e in eval_res["query_prop_dist_vcmr"]],
dtype=eval_res["query_prop_dist_vcmr"][0].dtype,
device=eval_res["query_prop_dist_vcmr"][0].device)
# putting 'NaN' into the invalid bits, torch.sort considers 'NaN' as larger than any number!!!
padded_dist += (padded_mask.unsqueeze(2) == 0).float() * 1e10
n_videos, n_prop, n_q = padded_dist.shape
print("n_videos, n_prop, n_q {}".format((n_videos, n_prop, n_q)))
padded_dist = padded_dist.view(n_videos * n_prop, n_q).transpose(0, 1).contiguous() # (N_q, N_video*N_prop)
print("padded_dist, {}".format(padded_dist.shape))
sorted_distances, sorted_indices = torch.topk(padded_dist.to(torch.device("cuda:0"), non_blocking=True),
k=min(max_prop_per_query, n_videos * n_prop),
dim=1, largest=False, sorted=True) # (N_q, max_prop_per_query) * 2
sorted_distances = - sorted_distances.cpu().numpy()
# (N_q, max_prop_per_query) * 2, prop_indices: inside video indices.
video_meta_indices = torch.floor(sorted_indices.float() / n_prop).long().cpu().numpy()
prop_indices = torch.remainder(sorted_indices, n_prop).cpu().numpy()
vr_res = []
query_meta = eval_res["query_meta"]
for i in trange(n_q, desc="[VR] Loop over queries to generate predictions"):
row = video_meta_indices[i]
score_row = - sorted_distances[i]
cur_vr_redictions = []
for j, meta_idx in enumerate(row):
video_idx = video2idx[eval_res["video_meta"][meta_idx]["vid_name"]]
cur_vr_redictions.append([video_idx, 0, 0, float(score_row[j])])
cur_query_pred = dict(
desc_id=query_meta[i]["desc_id"],
desc=query_meta[i]["desc"],
predictions=cur_vr_redictions
)
vr_res.append(cur_query_pred)
vcmr_res = []
logger.debug("sorted_indices {}".format(sorted_indices.shape))
logger.debug("sorted_distances {}".format(sorted_distances.shape))
for idx, (vm_row_indices, p_row_indices) in tqdm(enumerate(zip(video_meta_indices, prop_indices)),
desc="[VCMR] Loop over queries to generate predictions",
total=n_q): # query
sorted_distances_row = - sorted_distances[idx] # converted to negative distance
# [video_idx(int), st(float), ed(float), score(float)]
cur_ranked_predictions = []
for col_idx, (v_col_idx, p_col_idx) in enumerate(zip(vm_row_indices, p_row_indices)):
cur_pred = []
cur_pred += [video2idx[eval_res["video_meta"][v_col_idx]["vid_name"]], ]
cur_pred += eval_res["video_meta"][v_col_idx]["proposals"][p_col_idx].tolist()
cur_pred += [float(sorted_distances_row[col_idx])]
cur_ranked_predictions.append(cur_pred)
cur_query_pred = dict(
desc_id=eval_res["query_meta"][idx]["desc_id"],
desc=eval_res["query_meta"][idx]["desc"],
predictions=cur_ranked_predictions
)
vcmr_res.append(cur_query_pred)
return vcmr_res, vr_res
def floor_ste(x: torch.Tensor) -> torch.Tensor:
y = torch.floor(x)
return _ste(x, y)
def regularize(f):
f_ = (f - PI) / TWO_PI
return TWO_PI * (f_ - torch.floor(f_) - 0.5)
def train(self):
if self.T - self.target_sync_T > self.args.target:
self.sync_target_network()
self.target_sync_T = self.T
info = {}
for _ in range(self.args.iters):
self.dqn.eval()
batch, indices, is_weights = self.replay.Sample_N(self.args.batch_size, self.args.n_step, self.args.gamma)
columns = list(zip(*batch))
states = Variable(torch.from_numpy(np.array(columns[0])).float().transpose_(1, 3))
actions = Variable(torch.LongTensor(columns[1]))
terminal_states = Variable(torch.FloatTensor(columns[5]))
rewards = Variable(torch.FloatTensor(columns[2]))
# Have to clip rewards for DQN
rewards = torch.clamp(rewards, -1, 1)
steps = Variable(torch.FloatTensor(columns[4]))
new_states = Variable(torch.from_numpy(np.array(columns[3])).float().transpose_(1, 3))
target_dqn_qvals = self.target_dqn(new_states).cpu()
# Make a new variable with those values so that these are treated as constants
target_dqn_qvals_data = Variable(target_dqn_qvals.data)
q_value_gammas = (Variable(torch.ones(terminal_states.size()[0])) - terminal_states)
inter = Variable(torch.ones(terminal_states.size()[0]) * self.args.gamma)
# print(steps)
q_value_gammas = q_value_gammas * torch.pow(inter, steps)
values = torch.linspace(self.args.v_min, self.args.v_max, steps=self.args.atoms)
values = Variable(values)
values = values.view(1, 1, self.args.atoms)
values = values.expand(self.args.batch_size, self.args.actions, self.args.atoms)
# print(values)
q_value_gammas = q_value_gammas.view(self.args.batch_size, 1, 1)
q_value_gammas = q_value_gammas.expand(self.args.batch_size, self.args.actions, self.args.atoms)
# print(q_value_gammas)
gamma_values = q_value_gammas * values
# print(gamma_values)
rewards = rewards.view(self.args.batch_size, 1, 1)
rewards = rewards.expand(self.args.batch_size, self.args.actions, self.args.atoms)
# print(rewards)
operator_q_values = rewards + gamma_values
# print(operator_q_values)
clipped_operator_q_values = torch.clamp(operator_q_values, self.args.v_min, self.args.v_max)
delta_z = (self.args.v_max - self.args.v_min) / (self.args.atoms - 1)
# Using the notation from the categorical paper
b_j = (clipped_operator_q_values - self.args.v_min) / delta_z
# print(b_j)
lower_bounds = torch.floor(b_j)
upper_bounds = torch.ceil(b_j)
# Work out the max action
atom_values = Variable(torch.linspace(self.args.v_min, self.args.v_max, steps=self.args.atoms))
atom_values = atom_values.view(1, 1, self.args.atoms)
atom_values = atom_values.expand(self.args.batch_size, self.args.actions, self.args.atoms)
# Sum over the atoms dimension
target_expected_qvalues = torch.sum(target_dqn_qvals_data * atom_values, dim=2)
# Get the maximum actions index across the batch size
max_actions = target_expected_qvalues.max(dim=1)[1].view(-1)
# Project back onto the original support for the max actions
q_value_distribution_targets = torch.zeros(self.args.batch_size, self.args.atoms)
# Distributions for the max actions
# print(target_dqn_qvals_data, max_actions)
q_value_max_actions_distribs = target_dqn_qvals_data.index_select(dim=1, index=max_actions)[:,0,:]
# print(q_value_max_actions_distribs)
# Lower_bounds_actions
lower_bounds_actions = lower_bounds.index_select(dim=1, index=max_actions)[:,0,:]
upper_bounds_actions = upper_bounds.index_select(dim=1, index=max_actions)[:,0,:]
b_j_actions = b_j.index_select(dim=1, index=max_actions)[:,0,:]
lower_bound_values_to_add = q_value_max_actions_distribs * (upper_bounds_actions - b_j_actions)
upper_bound_values_to_add = q_value_max_actions_distribs * (b_j_actions - lower_bounds_actions)
# print(lower_bounds_actions)
# print(lower_bound_values_to_add)
# Naive looping
for b in range(self.args.batch_size):
for l, pj in zip(lower_bounds_actions.data.type(torch.LongTensor)[b], lower_bound_values_to_add[b].data):
q_value_distribution_targets[b][l] += pj
for u, pj in zip(upper_bounds_actions.data.type(torch.LongTensor)[b], upper_bound_values_to_add[b].data):
q_value_distribution_targets[b][u] += pj
self.dqn.train()
if self.args.gpu:
actions = actions.cuda()
# q_value_targets = q_value_targets.cuda()
q_value_distribution_targets = q_value_distribution_targets.cuda()
model_predictions = self.dqn(states).index_select(1, actions.view(-1))[:,0,:]
q_value_distribution_targets = Variable(q_value_distribution_targets)
# print(q_value_distribution_targets)
# print(model_predictions)
# Cross entropy loss
ce_loss = -torch.sum(q_value_distribution_targets * torch.log(model_predictions), dim=1)
ce_batch_loss = ce_loss.mean()
info = {}
self.log("DQN/X_Entropy_Loss", ce_batch_loss.data[0], step=self.T)
# Update
self.optimizer.zero_grad()
ce_batch_loss.backward()
# Taken from pytorch clip_grad_norm
# Remove once the pip version it up to date with source
gradient_norm = clip_grad_norm(self.dqn.parameters(), self.args.clip_value)
if gradient_norm is not None:
info["Norm"] = gradient_norm
self.optimizer.step()
if "States" in info:
states_trained = info["States"]
info["States"] = states_trained + columns[0]
else:
info["States"] = columns[0]
# Pad out the states to be of size batch_size
if len(info["States"]) < self.args.batch_size:
old_states = info["States"]
new_states = old_states[0] * (self.args.batch_size - len(old_states))
info["States"] = new_states
return info
def _get_random_inputs(device):
return torch.floor(torch.rand(
(BATCH_SIZE, 2)) * EMB_SIZE).to(dtype=torch.long, device=device)
def _interpolate(im, x, y, out_size):
# constants
num_batch, height, width, channels = im.shape[0], im.shape[
1], im.shape[2], im.shape[3]
out_height = out_size[0]
out_width = out_size[1]
max_y = int(height - 1)
max_x = int(width - 1)
# scale indices from [-1, 1] to [0, width/height]
x = (x + 1.0) * (width - 1.0) / 2.0
y = (y + 1.0) * (height - 1.0) / 2.0
# do sampling
x0 = (torch.floor(x)).int()
x1 = x0 + 1
y0 = (torch.floor(y)).int()
y1 = y0 + 1
x0_c = torch.clamp(x0, 0, max_x)
x1_c = torch.clamp(x1, 0, max_x)
y0_c = torch.clamp(y0, 0, max_y)
y1_c = torch.clamp(y1, 0, max_y)
dim2 = width
dim1 = width * height
base = _repeat(
torch.arange(0, num_batch) * dim1,
out_height * out_width).to(im.get_device())
base_y0 = base + y0_c * dim2
base_y1 = base + y1_c * dim2
idx_a = base_y0 + x0_c
idx_b = base_y1 + x0_c
idx_c = base_y0 + x1_c
idx_d = base_y1 + x1_c
# use indices to lookup pixels in the flat image and restore
# channels dim
im_flat = im.view([-1, channels])
im_flat = im_flat.float()
# and finally calculate interpolated values
x0_f = x0.float()
x1_f = x1.float()
y0_f = y0.float()
y1_f = y1.float()
wa = ((x1_f - x) * (y1_f - y)).unsqueeze(1)
wb = ((x1_f - x) * (y - y0_f)).unsqueeze(1)
wc = ((x - x0_f) * (y1_f - y)).unsqueeze(1)
wd = ((x - x0_f) * (y - y0_f)).unsqueeze(1)
zerof = torch.zeros_like(wa)
wa = torch.where(
(torch.eq(x0_c, x0) & torch.eq(y0_c, y0)).unsqueeze(1), wa, zerof)
wb = torch.where(
(torch.eq(x0_c, x0) & torch.eq(y1_c, y1)).unsqueeze(1), wb, zerof)
wc = torch.where(
(torch.eq(x1_c, x1) & torch.eq(y0_c, y0)).unsqueeze(1), wc, zerof)
wd = torch.where(
(torch.eq(x1_c, x1) & torch.eq(y1_c, y1)).unsqueeze(1), wd, zerof)
zeros = torch.zeros(
size=[int(num_batch) * int(height) * int(width),
int(channels)],
dtype=torch.float)
output = zeros.to(im.get_device())
output = output.scatter_add(dim=0,
index=idx_a.long().unsqueeze(1).repeat(
1, channels),
src=im_flat * wa)
output = output.scatter_add(dim=0,
index=idx_b.long().unsqueeze(1).repeat(
1, channels),
src=im_flat * wb)
output = output.scatter_add(dim=0,
index=idx_c.long().unsqueeze(1).repeat(
1, channels),
src=im_flat * wc)
output = output.scatter_add(dim=0,
index=idx_d.long().unsqueeze(1).repeat(
1, channels),
src=im_flat * wd)
return output
def pytorch_rand_select_pixel(width,height,num_samples=1):
two_rand_numbers = torch.rand(2,num_samples)
two_rand_numbers[0,:] = two_rand_numbers[0,:]*width
two_rand_numbers[1,:] = two_rand_numbers[1,:]*height
two_rand_ints = torch.floor(two_rand_numbers).type(dtype_long)
return (two_rand_ints[0], two_rand_ints[1])
def random_int_tensor(seed, size, low=0, high=2**32, a=22695477, c=1, m=2**32):
""" Same as random_float_tensor but integers between [low, high)
"""
return (
torch.floor(random_float_tensor(seed, size, a, c, m) *
(high - low)) + low).to(torch.int64)
def _PyramidRoI_Feat(self, feat_maps, rois, im_info):
''' roi pool on pyramid feature maps'''
# do roi pooling based on predicted rois
img_area = im_info[0][0] * im_info[0][1]
h = rois.data[:, 4] - rois.data[:, 2] + 1
w = rois.data[:, 3] - rois.data[:, 1] + 1
roi_level = torch.log(torch.sqrt(h * w) / 224.0) / np.log(2)
roi_level = torch.floor(roi_level + 4)
# --------
# roi_level = torch.log(torch.sqrt(h * w) / 224.0)
# roi_level = torch.round(roi_level + 4)
# ------
roi_level[roi_level < 2] = 2
roi_level[roi_level > 5] = 5
# roi_level.fill_(5)
if cfg.POOLING_MODE == 'crop':
# pdb.set_trace()
# pooled_feat_anchor = _crop_pool_layer(base_feat, rois.view(-1, 5))
# NOTE: need to add pyrmaid
grid_xy = _affine_grid_gen(rois,
feat_maps.size()[2:],
self.grid_size) ##
grid_yx = torch.stack(
[grid_xy.data[:, :, :, 1], grid_xy.data[:, :, :, 0]],
3).contiguous()
roi_pool_feat = self.RCNN_roi_crop(feat_maps,
Variable(grid_yx).detach()) ##
if cfg.CROP_RESIZE_WITH_MAX_POOL:
roi_pool_feat = F.max_pool2d(roi_pool_feat, 2, 2)
elif cfg.POOLING_MODE == 'align':
roi_pool_feats = []
box_to_levels = []
for i, l in enumerate(range(2, 6)):
if (roi_level == l).sum() == 0:
continue
idx_l = (roi_level == l).nonzero().squeeze()
box_to_levels.append(idx_l)
scale = feat_maps[i].size(2) / im_info[0][0]
feat = self.RCNN_roi_align(feat_maps[i], rois[idx_l], scale)
roi_pool_feats.append(feat)
roi_pool_feat = torch.cat(roi_pool_feats, 0)
box_to_level = torch.cat(box_to_levels, 0)
idx_sorted, order = torch.sort(box_to_level)
roi_pool_feat = roi_pool_feat[order]
elif cfg.POOLING_MODE == 'pool':
roi_pool_feats = []
box_to_levels = []
for i, l in enumerate(range(2, 6)):
if (roi_level == l).sum() == 0:
continue
idx_l = (roi_level == l).nonzero().squeeze()
box_to_levels.append(idx_l)
scale = feat_maps[i].size(2) / im_info[0][0]
feat = self.RCNN_roi_pool(feat_maps[i], rois[idx_l], scale)
roi_pool_feats.append(feat)
roi_pool_feat = torch.cat(roi_pool_feats, 0)
box_to_level = torch.cat(box_to_levels, 0)
idx_sorted, order = torch.sort(box_to_level)
roi_pool_feat = roi_pool_feat[order]
return roi_pool_feat