def forward(self, x):
x = self.conv1(x)
for l, l2, l3 in zip(self.layers, self.layers2, self.layers3):
x = l3(l2(l(x)))
x = F.adaptive_max_pool2d(x, 1)
x = x.view(x.size(0), -1)
return F.log_softmax(self.out(x), dim=-1)
def resize2dmask(mask, size):
with torch.no_grad():
o_in = torch.zeros(mask.size()).cuda()
o_out = torch.zeros(size).cuda()
t = (o_out - F.adaptive_max_pool2d(Variable(o_in - mask), size)).data
# t = (F.adaptive_avg_pool2d(Variable(img), size)).data
return t
def pa_max_pool(in_dict):
"""Implement `local max pooling` as `masking + global max pooling`.
Args:
feat: pytorch tensor, with shape [N, C, H, W]
mask: pytorch tensor, with shape [N, pC, pH, pW]
Returns:
feat_list: a list (length = pC) of pytorch tensors with shape [N, C, 1, 1]
visible: pytorch tensor with shape [N, pC]
NOTE:
The implementation of `masking + global max pooling` is only equivalent
to `local max pooling` when feature values are non-negative, which holds
for ResNet that has ReLU as final operation of all blocks.
"""
feat = in_dict['feat']
mask = in_dict['pap_mask']
N, C, H, W = feat.size()
N, pC, pH, pW = mask.size()
feat_list = []
for i in range(pC):
# [N, C, pH, pW]
m = mask[:, i, :, :].unsqueeze(1).expand_as(feat)
# [N, C]
local_feat = F.adaptive_max_pool2d(feat * m, 1)
# local_feat = F.adaptive_max_pool2d(feat * m, 1).view(N, -1)
feat_list.append(local_feat)
# [N, pC]
visible = (mask.sum(-1).sum(-1) != 0).float()
out_dict = {'feat_list': feat_list, 'visible': visible}
return out_dict
def roi_pooling(input, rois, size=(7, 7), spatial_scale=1.0):
assert (rois.dim() == 2)
assert (rois.size(1) == 5)
output = []
rois = rois.data.float()
num_rois = rois.size(0)
rois[:, 1:].mul_(spatial_scale)
rois = rois.long()
for i in range(num_rois):
roi = rois[i]
im_idx = roi[0]
if roi[1] >= input.size(3) or roi[2] >= input.size(
2) or roi[1] < 0 or roi[2] < 0:
# print(f"Runtime Warning: roi top left corner out of range: {roi}", file=sys.stderr)
roi[1] = torch.clamp(roi[1], 0, input.size(3) - 1)
roi[2] = torch.clamp(roi[2], 0, input.size(2) - 1)
if roi[3] >= input.size(3) or roi[4] >= input.size(
2) or roi[3] < 0 or roi[4] < 0:
# print(f"Runtime Warning: roi bottom right corner out of range: {roi}", file=sys.stderr)
roi[3] = torch.clamp(roi[3], 0, input.size(3) - 1)
roi[4] = torch.clamp(roi[4], 0, input.size(2) - 1)
if (roi[3:5] - roi[1:3] < 0).any():
# print(f"Runtime Warning: invalid roi: {roi}", file=sys.stderr)
im = input.new_full((1, input.size(1), 1, 1), 0)
else:
im = input.narrow(0, im_idx, 1)[..., roi[2]:(roi[4] + 1),
roi[1]:(roi[3] + 1)]
output.append(F.adaptive_max_pool2d(im, size))
return torch.cat(output, 0)
def forward(self, x, ROI, roi_layer):
roiX = self.layer1(x)
x_batch = []
for k in range(batch_size):
x_set = []
for j in range(roi_layer):
for i in range(self.count):
y1, y2, x1, x2 = ROI[k][i].squeeze(0)
y1 = int(y1.float() * 0.9)
y2 = int(y2.float() * 0.9)
x2 = int(x2.float() * 0.9)
x1 = int(x1.float() * 0.9)
image = F.adaptive_max_pool2d(roiX[k,
j][y1:y2,
x1:x2].unsqueeze(0),
output_size=(125, 125))
x_set.append(image.cpu().detach().numpy())
x_batch = x_set + x_batch
x_batch = torch.FloatTensor(x_batch).cuda().squeeze(1)
x_batch = x_batch.view(batch_size, 15, 125, 125)
#得到roi_layer*count张图 都是一张图的特征图 要转成tensor
norX = self.layer2(x)
x = torch.cat((norX, x_batch), 1)
x = self.layer3(x)
x = self.layer4(x)
print(x.shape)
x = x.view(x.size(0), -1)
x = self.fc(x)
return F.softmax(x, dim=1)
def forward(self, feature):
x = func.adaptive_max_pool2d(feature, (self.R, self.R))
x = x.view(x.size(0), -1)
if self.norm:
x = func.normalize(x)
return x
def forward(self, x):
if len(x.shape) == 5:
x = x.squeeze(1)
x = self.base(x)
if not self.training and self.out_base_feat:
if self.lmp == 0:
out = self.avgpool(x).view(x.size(0), -1)
elif self.lmp >= 1:
out = F.adaptive_max_pool2d(x, output_size=(self.lmp, 1)).view(
x.size(0), -1)
else:
out = F.adaptive_avg_pool2d(x,
output_size=(-self.lmp, 1)).view(
x.size(0), -1)
return out
x = self.avgpool(x).view(x.size(0), -1)
feat = F.normalize(x)
if self.embed_dim > 0:
x = self.embed_fc(x)
if self.has_bn:
x = self.embed_bn(x)
x = F.normalize(x)
if self.dropout > 0:
x = self.drop(x)
# return feat, x
return x, feat
def forward(self, x, xr):
'''
the forward function
:param x: the first input
:param xr: the second input
:return: the output
'''
# convolution
x = self.Conv_classifer(x)
if self.maxpooling_dir:
# global average polling of cnn features map in direction of channels
x = F.adaptive_avg_pool2d(x, (x.size()[1], 1))
else:
# global max polling of cnn features map in direction of words
x = F.adaptive_max_pool2d(x, (x.size()[2], 1))
x = x.view(x.size()[0], -1)
if self.padding:
x = nn.ConstantPad1d((0, self.w_size - x.size()[1]), 0)(x)
if self.xr:
x = torch.cat((x, xr), 1)
x = self.FC_classifier(x)
return x
def forward(self, x):
feature_map = self.feature_map(x)
cls, reg = self.rpn(feature_map)
feature_map = feature_map.view(
(-1, self.OUTPUT_SIZE[0], self.OUTPUT_SIZE[1]))
if self.training:
proposals = self.rpn.get_proposals(reg, cls)
else:
proposals = self.rpn.get_proposals(reg, cls)
all_cls = []
all_reg = []
for roi in proposals.int():
roi[np.where(roi < 0)] = 0
roi = roi / self.OUTPUT_CELL_SIZE
roi_feature_map = feature_map[:, roi[0]:roi[2] + 1,
roi[1]:roi[3] + 1]
pooled_roi = F.adaptive_max_pool2d(roi_feature_map, (7, 7)).view(
(-1, 50176))
r = F.relu(self.fc(pooled_roi))
r_cls = self.cls_layer(r)
r_reg = self.reg_layer(r).view((self.n_classes, 4))
all_cls.append(r_cls)
all_reg.append(r_reg[torch.argmax(r_cls)])
# print(all_cls.shape, all_reg.shape)
return torch.stack(all_cls).view(
(-1, self.n_classes)), torch.stack(all_reg), proposals, cls, reg
def forward(self, x):
if self.version in ['v1', 'v2']:
x_compress = self.compress(x)
x_out = self.spatial(x_compress)
scale = F.sigmoid(x_out) # broadcasting
elif self.version in ['v3', 'v4', 'v7', 'v8']:
x_compress = self.compress(x)
x_out = self.spatial(x_compress)
if not self.version == 'v7':
scale = F.sigmoid(x_out) # broadcasting
elif self.version == 'v5':
x_compress = self.compress(x)
x_ = self.spatial(x_compress)
x_out = self.expand(x_)
scale = F.sigmoid(x_out) # broadcasting
elif self.version == 'v6':
x_pool = F.adaptive_max_pool2d(
x,
x.size(2) // self.size_reduction_ratio)
x_ = self.compress(x_pool)
x_ = self.spatial(x_)
x_ = self.expand(x_)
x_up = F.upsample(x_, x.size()[2:], mode='bilinear')
scale = F.sigmoid(x_up) # broadcasting
if self.version == 'v7':
return x_out.expand_as(x)
else:
return x * scale
def forward(self, x):
s0 = s1 = self.stem(x)
pre_layers = [s1]
for i, cell in enumerate(self.cells):
weights = []
n = 2
start = 0
for _ in range(self._steps):
end = start + n
for j in range(start, end):
# import pdb;pdb.set_trace()
weights.append(F.softmax(self.alphas_normal[j], dim=-1))
start = end
n += 1
selected_idxs = self.normal_selected_idxs
# import pdb;pdb.set_trace()
s0, s1 = s1, cell(s0, s1, weights, selected_idxs, self.att)
pre_layers.append(s1)
# import pdb;pdb.set_trace()
fusion = torch.cat(pre_layers, dim=1)
fusion = self.fusion_conv(fusion)
x1 = F.adaptive_max_pool2d(fusion, 1)
x2 = F.adaptive_avg_pool2d(fusion, 1)
logits = self.classifier(torch.cat((x1, x2), dim=1))
return logits.squeeze(-1).squeeze(-1)
def reversal_att_map(self, attention):
opp_att_map = attention.new_ones(attention.size())
max_ = F.adaptive_max_pool2d(attention, 1)
opp_att_map *= max_
opp_att_map -= attention
return opp_att_map
def apply(features: Tensor, proposal_bboxes: Tensor, proposal_batch_indices: Tensor, mode: Mode) -> Tensor:
_, _, feature_map_height, feature_map_width = features.shape
scale = 1 / 16
output_size = (7 * 2, 7 * 2)
if mode == Pooler.Mode.POOLING:
pool = []
for (proposal_bbox, proposal_batch_index) in zip(proposal_bboxes, proposal_batch_indices):
start_x = max(min(round(proposal_bbox[0].item() * scale), feature_map_width - 1), 0) # [0, feature_map_width)
start_y = max(min(round(proposal_bbox[1].item() * scale), feature_map_height - 1), 0) # (0, feature_map_height]
end_x = max(min(round(proposal_bbox[2].item() * scale) + 1, feature_map_width), 1) # [0, feature_map_width)
end_y = max(min(round(proposal_bbox[3].item() * scale) + 1, feature_map_height), 1) # (0, feature_map_height]
roi_feature_map = features[proposal_batch_index, :, start_y:end_y, start_x:end_x]
pool.append(F.adaptive_max_pool2d(input=roi_feature_map, output_size=output_size))
pool = torch.stack(pool, dim=0)
elif mode == Pooler.Mode.ALIGN:
pool = ROIAlign(output_size, spatial_scale=scale, sampling_ratio=0)(
features,
torch.cat([proposal_batch_indices.view(-1, 1).float(), proposal_bboxes], dim=1)
)
else:
raise ValueError
pool = F.max_pool2d(input=pool, kernel_size=2, stride=2)
return pool
def disp_loss(self, pred, targe, sparse=True):
assert len(targe.size()) == 3, 'expect targe size is b,h,w'
b, channel, h, w = pred.size()
loss = 0.
targe = targe.unsqueeze(1)
if sparse:
targe = F.adaptive_max_pool2d(targe, (h, w))
pred=pred.view(b,-1).contiguous()
targe=targe.view(b,-1).contiguous()
EPE_map = self.SmoothL1Loss(pred, targe)
#print("EPE_map mean:",EPE_map.mean())
positive = (targe > 0).long()
#ignore = (EPE_map*positive)
#print("EPE_map mean:",EPE_map.mean())
#EPE_map = EPE_map[positive]
loss = EPE_map.mean()
return loss
EPE_map = self.SmoothL1Loss(pred, targe)
positive = (targe > 0).long()
EPE_map = EPE_map[positive]
loss = torch.mean(EPE_map)
for ib in range(b):
EPE_map = self.SmoothL1Loss(pred, targe.unsqueeze(1))
#EPE_map = torch.abs(targe[ib] - pred[ib, 0])
positive = (targe[ib] > 0).long()
loss += EPE_map.mean_()
#smooth = self.hardtanh(EPE_map)
#EPE_map = 0.5 * smooth * EPE_map
return loss
def train(model, data_loader, optimizer):
model.train()
batch_time = 0
total_loss = {'mse': 0, 'pos': 0, 'neg': 0}
for batch_no, (img, ann) in tqdm(enumerate(data_loader)):
start_time = time.time()
if args.cuda:
img = img.cuda(async=True)
ann = ann.cuda(async=True)
preds = model(V(img))
loss = 0
mse_loss = 0
pos_loss = 0
neg_loss = 0
for pred in preds:
b, c, h, w = pred.size()
_ann = F.adaptive_max_pool2d(ann, (h, w)).float()
diff = ((_ann - pred)**2)
mask = (_ann > 0).float()
npos = int(mask.sum().data[0])
pos_loss = (mask * diff).sum() / npos
mask = 1 - mask
neg_loss = (mask * diff).view(-1).topk(k=min(
npos * 3, diff.numel()),
sorted=False)[0]
neg_loss = neg_loss.sum() / neg_loss.numel()
mse = F.mse_loss(pred, _ann)
#loss += pos_loss + neg_loss
loss += mse
total_loss['mse'] += mse.data[0]
total_loss['pos'] += pos_loss.data[0]
total_loss['neg'] += neg_loss.data[0]
model.zero_grad()
loss.backward()
optimizer.step()
batch_time += time.time() - start_time
if (batch_no + 1) % args.log_interval == 0:
avg_time = batch_time * 1e3 / args.log_interval
for k in total_loss:
total_loss[k] /= args.log_interval
print(
'train | {:4d} /{:4d} batch | {:7.2f} ms/batch | mse loss {:.2e} | pos loss {:.2e} | neg loss {:.2e}'
.format(batch_no + 1, args.num_iter, avg_time,
total_loss['mse'], total_loss['pos'],
total_loss['neg']))
for k in total_loss:
total_loss[k] = 0
batch_time = 0
if (batch_no + 1) >= args.num_iter: return
def forward(self, x, get_features=False):
x0 = F.pad(x, (0, 0, 0, 0, 0, self.n_filters - self.input_shape[0]))
x0 = quantifier(x0, self.n_bits_activ)
hist = [None for _ in range(self.n_history - 1)] + [x0]
for t in range(self.n_iter):
a = self.Lconv(hist[-1])
a = self.Lactiv(a)
a = self.alpha[t, 0] * a
for i, x in enumerate(hist):
if x is not None:
a = a + self.alpha[t, i + 1] * x
a = self.Lnormalization[t](a)
a = quantifier(a, self.n_bits_activ)
for i in range(1, self.n_history - 1):
hist[i] = hist[i + 1]
hist[self.n_history - 1] = a
if self.pool_strategy[t]:
for i in range(len(hist)):
if hist[i] is not None:
hist[i] = F.max_pool2d(hist[i], 2)
out = F.adaptive_max_pool2d(hist[-1], (1, 1))[:, :, 0, 0]
if get_features:
return out, self.LOutput(out)
else:
return self.LOutput(out)
def max_pool(in_dict):
"""Implement `local max pooling` as `masking + global max pooling`.
Args:
feat: pytorch tensor, with shape [N, C, H, W]
mask: pytorch tensor, with shape [N, pC, pH, pW]
Returns:
feat_list: a list (length = pC) of pytorch tensors with shape [N, C]
visible: pytorch tensor with shape [N, pC]
NOTE:
The implementation of `masking + global max pooling` is only equivalent
to `local max pooling` when feature values are non-negative, which holds
for ResNet that has ReLU as final operation of all blocks.
"""
assert len(in_dict['feat']) == len(in_dict['ps_pred'])
N = in_dict['feat'][0].shape[0]
num_parts = len(in_dict['ps_pred'])
feat_list = []
visible_list = []
for i in range(num_parts):
feat = in_dict['feat'][i]
# [N, 1, pH, pW]
m = (in_dict['ps_pred'][i] > 0.5).float()
visible_list.append((m.sum(-1).sum(-1) > 0).float())
# [N, C, pH, pW]
m = m.expand_as(feat)
# [N, C]
local_feat = F.adaptive_max_pool2d(feat * m, 1).view(N, -1)
feat_list.append(local_feat)
# [N, pC]
visible = torch.cat(visible_list, 1)
out_dict = {'feat_list': feat_list, 'visible': visible}
return out_dict
def forward(self, data):
"""
Forward function.
:param data:
:return: tensor
"""
fm = self.conv1(data)
if self.channel_att:
# fm_pool = F.adaptive_avg_pool2d(fm, (1, 1)) + F.adaptive_max_pool2d(fm, (1, 1))
fm_pool = torch.cat([
F.adaptive_avg_pool2d(fm, (1, 1)),
F.adaptive_max_pool2d(fm, (1, 1))
],
dim=1)
att = self.att_c(fm_pool)
fm = fm * att
if self.spatial_att:
fm_pool = torch.cat([
torch.mean(fm, dim=1, keepdim=True),
torch.max(fm, dim=1, keepdim=True)[0]
],
dim=1)
att = self.att_s(fm_pool)
fm = fm * att
return fm
def forward(self, data):
"""
Forward function.
:param data:
:return: tensor
"""
fm = self.conv1(data) # (bs,out_ch,w,h)
if self.channel_att:
# adaptive_avg_pool2d 将(w,h)大小变成任意大小，如下面变成了(1,1),则经过cat之后为 (bs,2*out_ch,1,1)
fm_pool = torch.cat([
F.adaptive_avg_pool2d(fm, (1, 1)),
F.adaptive_max_pool2d(fm, (1, 1))
],
dim=1)
att = self.att_c(fm_pool) # (bs,out_ch,1,1)
fm = fm * att # (bs,out_ch,w,h)*(bs,out_ch,1,1) -> (bs,out_ch,w,h)
if self.spatial_att:
# (bs,1,w,h) + (bs,1,w,h) -> (bs,2,w,h) channel上一个是mean，一个是max
fm_pool = torch.cat([
torch.mean(fm, dim=1, keepdim=True),
torch.max(fm, dim=1, keepdim=True)[0]
],
dim=1)
att = self.att_s(fm_pool) # (bs,1,w,h)
fm = fm * att # (bs,out_ch,w,h)*(bs,1,w,h) -> (bs,out_ch,w,h)
return fm
def forward(self, x, xr):
conv1 = self.Conv_classifer(x)
# x = F.adaptive_avg_pool2d(x, (x.size()[1], 1))
x1 = F.adaptive_max_pool2d(conv1, (conv1.size()[1], 1))
x1 = x1.view(x1.size()[0], -1)
conv2 = self.Conv_classifer2(conv1)
x2 = F.adaptive_max_pool2d(conv2, (conv2.size()[1], 1))
x2 = x2.view(x2.size()[0], -1)
x = torch.cat((x1, x2, xr), 1)
x = self.FC_classifier(x)
return x
def forward(self, features, proposal_bboxes):
_, _, feature_map_height, feature_map_width = features.shape
pool = []
for proposal_bbox in proposal_bboxes:
start_x = max(
min(round(proposal_bbox[0].item() / 16),
feature_map_width - 1), 0)
start_y = max(
min(round(proposal_bbox[1].item() / 16),
feature_map_height - 1), 0)
end_x = max(
min(
round(proposal_bbox[2].item() / 16) + 1,
feature_map_width), 1)
end_y = max(
min(
round(proposal_bbox[3].item() / 16) + 1,
feature_map_height), 1)
roi_feature_map = features[..., start_y:end_y, start_x:end_x]
pool.append(F.adaptive_max_pool2d(roi_feature_map, 7))
pool = torch.cat(pool, dim=0)
pool = pool.view(pool.shape[0], -1)
h = self.fcs(pool)
classes = self._class(h)
transformers = self._transformer(h)
return classes, transformers
def forward(self, x):
x = self.down_sampling(x)
gap = F.adaptive_avg_pool2d(x, 1)
gap_logits = self.gap_fc(gap.view(x.shape[0], -1))
gap_weight = list(self.gap_fc.parameters())[0]
gap = x * gap_weight.unsqueeze(dim=2).unsqueeze(dim=3)
gmp = F.adaptive_max_pool2d(x, 1)
gmp_logits = self.gmp_fc(gmp.view(x.shape[0], -1))
gmp_weight = list(self.gmp_fc.parameters())[0]
gmp = x * gmp_weight.unsqueeze(dim=2).unsqueeze(dim=3)
cam_logit = torch.cat([gap_logits, gmp_logits], dim=1)
x = torch.cat([gap, gmp], dim=1)
x = self.conv(x)
heatmap = torch.sum(x, dim=1, keepdim=True)
out = self.fc(x.view(x.shape[0], -1))
gamma, beta = self.gamma(out), self.beta(out)
for i in range(self.num_blocks):
x = getattr(self, "UpBlock" + str(i + 1))(x, gamma, beta)
out = self.up_sampling(x)
return out, cam_logit, heatmap
def forward(self, x):
batch_size, C, H, W = x.shape
x0 = self.conv1(x)
x1 = self.encode2(x0)
x2 = self.encode3(x1)
x3 = self.encode4(x2)
x4 = self.encode5(x3)
##----
#segment
t0 = self.lateral0(x4)
t1 = upsize_add(t0, self.lateral1(x3)) #16x16
t2 = upsize_add(t1, self.lateral2(x2)) #32x32
t3 = upsize_add(t2, self.lateral3(x1)) #64x64
t1 = self.top1(t1) #; print(t1.shape)
t2 = self.top2(t2) #; print(t2.shape)
t3 = self.top3(t3) #; print(t3.shape)
x = fuse([t1, t2, t3], "cat")
logit = self.logit(x)
#---
probability_mask = torch.sigmoid(logit)
probability_label = F.adaptive_max_pool2d(probability_mask, 1).view(batch_size,-1)
if self.infer:
return logit
elif self.fp16:
# amp doesn't support regular BCELoss; only BCEWithLogitsLoss
return (probability_label, logit)
else:
return (probability_label, probability_mask)
def forward(self, pred, pred_semseg, label_info, label_i=0, sparse=True, shape=(30,30)):
targe = label_info['disp_label_0']
bs, num_class, h,w = pred.shape
if sparse:
targe = F.adaptive_max_pool2d(targe, shape)
#calucate disp output
#pred_scans = label_info['semseg_scans'].repeat(bs//cfg.TRAIN.IMS_PER_BATCH, 1, h, w)
#pred_scans = (pred_semseg == pred_scans.long()).float()
#pred_scans = pred_scans.unsqueeze(2)
#pred_scans.requires_grad = False
#pred_disp = pred*pred_scans
#pred_disp = torch.sum(pred_disp, 1)
disp_scans = label_info['disp_scans'].repeat(bs//cfg.TRAIN.IMS_PER_BATCH, 1, 1, 1)
pred_semseg, _ = torch.max(pred_semseg, dim=1)
pred_semseg = pred_semseg.unsqueeze(1)
pred = pred + pred_semseg
pred = F.softmax(pred, dim=1)
disp_pred = torch.sum(pred*disp_scans, dim=1)
#loss and pred
EPE_map = self.SmoothL1Loss(disp_pred, targe)
epe_pred = torch.abs(disp_pred - targe)
#ignore false disp values
positive = targe.ge(0)
EPE_map = torch.masked_select(EPE_map, positive)
epe_pred = torch.masked_select(epe_pred, positive)
#normlization
loss = EPE_map.mean()
epe_pred = epe_pred.mean()
return loss, epe_pred
def forward(self, x):
x = F.pad(x, (0, 0, 0, 0, 0, self.n_filters - self.input_shape[0]))
hist = [None for _ in range(self.n_history - 1)] + [x]
for t in range(self.n_iter):
a = self.Lconv(hist[-1])
fm_height = hist[-1].size(-1)
a = self.Lactiv(a)
a = self.alpha[t, 0] * a
for i, x in enumerate(hist):
if x is not None:
a = a + self.alpha[t, i + 1] * x
a = self.Lnormalization[t](a)
for i in range(1, self.n_history - 1):
hist[i] = hist[i + 1]
hist[self.n_history - 1] = a
if self.pool_strategy[t]:
for i in range(len(hist)):
if hist[i] is not None:
hist[i] = F.max_pool2d(hist[i], 2)
if self.out_mode == "pool" and hist[-1].size()[-1] > 1:
out = F.adaptive_max_pool2d(hist[-1], (1, 1))[:, :, 0, 0]
elif self.out_mode == "flatten":
out = hist[-1].view(hist[-1].size()[0], -1)
else:
out = hist[-1][:, :, 0, 0]
return self.LOutput(out)
def roi_pooling_2d_pytorch(input, rois, output_size=(7, 7), spatial_scale=1.0):
"""Spatial Region of Interest (ROI) pooling function in pure pytorch/python
This function acts similarly to `~roi_pooling_2d`, but performs a python
loop over ROI. Note that this is not a direct replacement of
`~roi_pooling_2d` (viceversa).
See :function:`~roi_pooling_2d` for details and output shape.
Args:
output_size (int or tuple): the target output size of the image of the
form H x W. Can be a tuple (H, W) or a single number H for a square
image H x H.
spatial_scale (float): scale of the rois if resized.
"""
assert rois.dim() == 2
assert rois.size(1) == 5
output = []
rois = rois.data.float()
num_rois = rois.size(0)
rois[:, 1:].mul_(spatial_scale)
rois = rois.long()
for i in range(num_rois):
roi = rois[i]
im_idx = roi[0]
im = input.narrow(0, im_idx, 1)[..., roi[2]:(roi[4] + 1),
roi[1]:(roi[3] + 1)]
output.append(F.adaptive_max_pool2d(im, output_size))
return torch.cat(output, 0)
def forward(self, x, blockID=None, ratio=None):
self.stochastic_downsampling(blockID, ratio)
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
if self.downsampling_ratio < 1:
if self.size_after_maxpool is None:
self.size_after_maxpool = self.maxpool(x).size(2)
x = F.adaptive_max_pool2d(
x,
int(round(self.size_after_maxpool * self.downsampling_ratio)))
else:
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x
def forward(self, features: Tensor,
proposal_bboxes: Tensor) -> Tuple[Tensor, Tensor]:
_, _, feature_map_height, feature_map_width = features.shape
pool = []
for proposal_bbox in proposal_bboxes:
start_x = max(
min(round(proposal_bbox[0].item() / 16),
feature_map_width - 1), 0) # [0, feature_map_width)
start_y = max(
min(round(proposal_bbox[1].item() / 16),
feature_map_height - 1), 0) # (0, feature_map_height]
end_x = max(
min(
round(proposal_bbox[2].item() / 16) + 1,
feature_map_width), 1) # [0, feature_map_width)
end_y = max(
min(
round(proposal_bbox[3].item() / 16) + 1,
feature_map_height), 1) # (0, feature_map_height]
roi_feature_map = features[..., start_y:end_y, start_x:end_x]
pool.append(F.adaptive_max_pool2d(roi_feature_map, 7))
print(pool)
pool = torch.cat(pool, dim=0) # pool has shape (128, 512, 7, 7)
pool = pool.view(pool.shape[0], -1)
h = self.fcs(pool)
classes = self._class(h)
transformers = self._transformer(h)
return classes, transformers
def simple_test_bboxes(self,
x,
img_meta,
mask_feats,
proposals,
rcnn_test_cfg,
rescale=False):
rois = bbox2roi(proposals)
roi_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
roi_feats = self.shared_head(roi_feats)
_, _, H, W = roi_feats.size()
mask_feats = F.adaptive_max_pool2d(mask_feats, (H, W))
roi_feats = torch.cat([roi_feats, mask_feats], dim=1)
cls_score, bbox_pred = self.bbox_head(roi_feats)
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
det_bboxes, det_labels = self.bbox_head.get_det_bboxes(
rois,
cls_score,
bbox_pred,
img_shape,
scale_factor,
rescale=rescale,
cfg=rcnn_test_cfg)
return det_bboxes, det_labels
def forward(self, x, visual=False):
x = self.cnn(x)
# G
x_g = self.embedding(x)
b, c, h, w = x_g.size()
x_g = torch.mean(x_g.contiguous().view(b, c, -1), dim=2)
out1 = self.fc(x_g) # [b, nc]
# L
x_l = self.special(x)
y_l = self.cls1(x_l) # [b, 14, 1, w]
y_l0 = F.softmax(y_l, dim=1)
y_l = F.adaptive_max_pool2d(y_l0, (1, 1)).contiguous().view(b, -1) # [b, 14]
out2 = self.cls2(y_l)
# Visualization
if visual:
nc1 = y_l0.size(1)
out2 = F.softmax(out2, dim=1)
return y_l0.contiguous().view(nc1, -1), y_l.contiguous().view(nc1, -1), out2.contiguous().view(nc1-1, -1)
# Dynamic weighting
weight = self.fusion(F.relu(out1)) # [b, 1]
weight = torch.cat((weight, 1 - weight), dim=1) # [b, 2]
y_all = torch.stack((out1, out2), dim=2)
out = y_all.matmul(weight.unsqueeze(2)).squeeze(2)
if self.training:
ally = y_l0.permute(0, 2, 3, 1).contiguous().view(b*w, -1) # [b*w, 14], it is p.
return ally, out1, out2, out
else:
return out
def forward(self, x):
out = self.conv1(x)
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = self.layer4(out)
out = F.adaptive_max_pool2d(out, 1)
out = out.view(out.size(0), -1)
return F.log_softmax(self.linear(out))
def forward(self, x):
x = self.conv_layers(x)
mean_pool = F.adaptive_avg_pool2d(x, 1).squeeze()
max_pool = F.adaptive_max_pool2d(x, 1).squeeze()
x = torch.cat([mean_pool, max_pool], dim=1)
x = self.fc(x)
return x
def _pyramid_pooling(self, input_x, output_sizes):
pyramid_level_tensors = []
for tsize in output_sizes:
if self.pool_type == 'max_pool':
pyramid_level_tensor = F.adaptive_max_pool2d(input_x, tsize)
if self.pool_type == 'avg_pool':
pyramid_level_tensor = F.adaptive_avg_pool2d(input_x, tsize)
pyramid_level_tensor = pyramid_level_tensor.view(input_x.size(0), -1)
pyramid_level_tensors.append(pyramid_level_tensor)
return torch.cat(pyramid_level_tensors, dim=1)
def forward(self, x):
x = F.adaptive_max_pool2d(x, output_size=(1, 1))
x = x.view(-1, x.size(1))
x = self.linear(x)
return x
def logits(self, features):
x = F.adaptive_max_pool2d(features, 1)
x = x.view(x.size(0), -1)
return self.classifier(x)
def classifier(self, x):
x = F.adaptive_max_pool2d(x, 1)
x = x.view(x.size(0), -1)
x = self.dropout(x)
return F.log_softmax(self.linear(x))
