def forward(self, out: Dict[str, torch.tensor],
inp: Dict[str, torch.tensor]) -> Dict[str, torch.tensor]:
annot = inp['annot']
att_box = out['att_out']
reg_box = out['bbx_out']
feat_sizes = out['feat_sizes']
num_f_out = out['num_f_out']
device = att_box.device
if len(num_f_out) > 1:
num_f_out = int(num_f_out[0].item())
else:
num_f_out = int(num_f_out.item())
feat_sizes = feat_sizes[:num_f_out, :]
if self.anchs is None:
feat_sizes = feat_sizes[:num_f_out, :]
anchs = self.get_anchors(feat_sizes)
anchs = anchs.to(device)
self.anchs = anchs
else:
anchs = self.anchs
att_box_sigmoid = torch.sigmoid(att_box).squeeze(-1)
att_box_best, att_box_best_ids = att_box_sigmoid.max(1)
# self.att_box_best = att_box_best
ious1 = IoU_values(annot, anchs)
gt_mask, expected_best_ids = ious1.max(1)
actual_bbox = reg_params_to_bbox(anchs, reg_box)
best_possible_result, _ = self.get_eval_result(actual_bbox, annot,
expected_best_ids)
msk = None
actual_result, pred_boxes = self.get_eval_result(
actual_bbox, annot, att_box_best_ids, msk)
out_dict = {}
out_dict['Acc'] = actual_result
out_dict['MaxPos'] = best_possible_result
out_dict['idxs'] = inp['idxs']
reshaped_boxes = x1y1x2y2_to_y1x1y2x2(
reshape((pred_boxes + 1) / 2, (inp['img_size'])))
out_dict['pred_boxes'] = reshaped_boxes
out_dict['pred_scores'] = att_box_best
# orig_annot = inp['orig_annot']
# Sanity check
# iou1 = (torch.diag(IoU_values(reshaped_boxes, orig_annot))
# >= self.acc_iou_threshold).float().mean()
# assert actual_result.item() == iou1.item()
return out_dict
def forward(self, out: Dict[str, torch.tensor],
inp: Dict[str, torch.tensor]) -> Dict[str, torch.tensor]:
"""
inp: att_box, reg_box, feat_sizes
annot: gt box (r1c1r2c2 form)
"""
annot = inp['annot']
att_box = out['att_out']
reg_box = out['bbx_out']
feat_sizes = out['feat_sizes']
num_f_out = out['num_f_out']
att_maps = out['att_maps']
if self.use_att_loss:
# self.loss_keys.append('att_ls')
iou_annot_stage_0 = inp['iou_annot_stage_0']
iou_annot_stage_1 = inp['iou_annot_stage_1']
iou_annot_stage_2 = inp['iou_annot_stage_2']
att_loss = self.att_losses(
att_maps[0], iou_annot_stage_0) + self.att_losses(
att_maps[1], iou_annot_stage_1) + self.att_losses(
att_maps[2], iou_annot_stage_2)
att_loss = att_loss / 3.
else:
att_loss = torch.zeros([1]).to(att_box.device)
device = att_box.device
# get the correct number of output features
# in the case of DataParallel
if len(num_f_out) > 1:
num_f_out = int(num_f_out[0].item())
else:
num_f_out = int(num_f_out.item())
# Computes Anchors only once since size is kept fixed
# Needs to be changed in case size is not fixed
if self.anchs is None:
feat_sizes = feat_sizes[:num_f_out, :]
anchs = self.get_anchors(feat_sizes)
anchs = anchs.to(device)
self.anchs = anchs
else:
anchs = self.anchs
matches = simple_match_anchors(
anchs, annot, match_thr=self.cfg['matching_threshold'])
#print(matches)
bbx_mask = (matches >= 0)
ious1 = IoU_values(annot, anchs)
_, msk = ious1.max(1)
bbx_mask2 = torch.eye(anchs.size(0))[msk]
bbx_mask2 = bbx_mask2 > 0
bbx_mask2 = bbx_mask2.to(device)
top1_mask = bbx_mask2
if not self.use_multi:
bbx_mask = bbx_mask2
else:
bbx_mask = bbx_mask | bbx_mask2
# all clear
gt_reg_params = bbox_to_reg_params(anchs, annot)
box_l = self.box_loss(reg_box, gt_reg_params)
# box_l_relv = box_l.sum(dim=2)[bbx_mask]
box_l_relv = box_l.sum(dim=2) * bbx_mask.float()
box_l_relv = box_l_relv.sum(dim=1) / bbx_mask.sum(dim=-1).float()
box_loss = box_l_relv.mean()
if box_loss.cpu() == torch.Tensor([float("Inf")]):
# There is a likely bug with annot box
# being very small
import pdb
pdb.set_trace()
att_box = att_box.squeeze(-1)
att_box_sigm = torch.sigmoid(att_box)
if self.use_softmax:
assert self.use_multi is False
gt_ids = msk
clas_loss = F.cross_entropy(att_box, gt_ids, reduction='none')
else:
if self.use_focal:
encoded_tgt = bbx_mask.float()
ps = att_box_sigm
weights = encoded_tgt * (1 - ps) + (1 - encoded_tgt) * ps
alphas = ((1 - encoded_tgt) * self.alpha + encoded_tgt *
(1 - self.alpha))
weights.pow_(self.gamma).mul_(alphas)
weights = weights.detach()
else:
weights = None
#print(att_box, att_box.shape, bbx_mask, bbx_mask.shape, weights, weights.shape)
clas_loss = F.binary_cross_entropy_with_logits(att_box,
bbx_mask.float(),
weight=weights,
reduction='none')
clas_loss = clas_loss.sum() / bbx_mask.sum()
# clas_loss = clas_loss.sum() / clas_loss.size(0)
if torch.isnan(box_loss) or torch.isnan(clas_loss):
# print('Nan Loss')
box_loss = box_loss.new_ones(box_loss.shape) * 0.01
box_loss.requires_grad = True
clas_loss = clas_loss.new_ones(clas_loss.shape)
clas_loss.requires_grad = True
if self.use_att_loss:
out_loss = self.lamb_reg * box_loss + clas_loss + att_loss
else:
out_loss = self.lamb_reg * box_loss + clas_loss
# + self.lamb_rel * rel_loss
out_dict = {}
out_dict['loss'] = out_loss
out_dict['cls_ls'] = clas_loss
out_dict['box_ls'] = box_loss
if self.use_att_loss:
out_dict['att_ls'] = att_loss
# out_dict['rel_ls'] = rel_loss
return out_dict
def forward(self, out: Dict[str, torch.tensor],
inp: Dict[str, torch.tensor]) -> Dict[str, torch.tensor]:
annot = inp['annot']
att_box = out['att_out']
reg_box = out['bbx_out']
feat_sizes = out['feat_sizes']
num_f_out = out['num_f_out']
device = att_box.device
if len(num_f_out) > 1:
num_f_out = int(num_f_out[0].item())
else:
num_f_out = int(num_f_out.item())
feat_sizes = feat_sizes[:num_f_out, :]
if self.anchs is None:
feat_sizes = feat_sizes[:num_f_out, :]
anchs = self.get_anchors(feat_sizes)
anchs = anchs.to(device)
self.anchs = anchs
else:
anchs = self.anchs
att_box_sigmoid = torch.sigmoid(att_box).squeeze(-1)
att_box_best, att_box_best_ids = att_box_sigmoid.max(1)
topk_box = torch.topk(att_box_sigmoid, k=100) #added by rishabh
att_box_best, att_box_best_ids = topk_box.values, topk_box.indices #added by rishabh
ious1 = IoU_values(annot, anchs)
#print("-->iou:", ious1.shape)
gt_mask, expected_best_ids = ious1.max(1)
actual_bbox = reg_params_to_bbox(anchs, reg_box)
#print("\n ->actual_bbox", actual_bbox[0][att_box_best_ids[0].item()])
best_possible_result, _ = self.get_eval_result(actual_bbox, annot,
expected_best_ids)
#print("\n--> att_box_best_ids:", att_box_best_ids)
msk = None
#'''
pred_box = self.get_eval_result(actual_bbox, annot,
att_box_best_ids[:, 0], msk)[1]
top_boxes = x1y1x2y2_to_y1x1y2x2(
reshape((pred_box + 1) / 2, (inp['img_size'])))
top_scores = att_box_best[:, 0]
for i in range(1, 100, 1):
#Rishabh--Break if att_box_best[:,i] < 0.5
if att_box_best[:, i] < 0.45:
break
actual_result, pred_boxes = self.get_eval_result(
actual_bbox, annot, att_box_best_ids[:, i], msk)
#'''
#actual_result, pred_boxes = self.get_eval_result(
# actual_bbox, annot, att_box_best_ids, msk)
#print("\n-> pred box", att_box_best_ids, pred_boxes)
'''
print("R-------In Eevaluator.py",
"\n annot", annot,
"\n att_box", att_box, att_box.shape,
"\n reg_box", reg_box, reg_box.shape,
"\n anchs", anchs, anchs.shape,
"\n actual_result", actual_result,
"\n pred_boxes", pred_boxes,
"\n best_possible_result", best_possible_result,
"\n actual_bbox", actual_bbox, actual_bbox.shape,
"\n att_box_best", att_box_best)
'''
#ris
out_dict = {}
out_dict['Acc'] = actual_result
out_dict['MaxPos'] = best_possible_result
out_dict['idxs'] = inp['idxs']
iou_sc = IoU_values(pred_boxes, anchs)
reshaped_boxes = x1y1x2y2_to_y1x1y2x2(
reshape((pred_boxes + 1) / 2, (inp['img_size'])))
out_dict['pred_boxes'] = reshaped_boxes
out_dict['pred_scores'] = att_box_best
#print("\n\n --> {}-th reshaped_box \n score:{} and \n box:{} iou:{}".format(i, att_box_best[:,i],reshaped_boxes, iou_sc.max(1)))
#print(IoU_values(top_boxes, reshaped_boxes) < 0.5)
if not False in (IoU_values(top_boxes, reshaped_boxes) < 0.5):
top_boxes = torch.cat([top_boxes, reshaped_boxes], axis=0)
top_scores = torch.cat([top_scores, att_box_best[:, i]],
axis=0)
#print("\n\n --> {}-th reshaped_box \n score:{} and \n box:{} iou:{}".format(i, att_box_best,reshaped_boxes, iou_sc.max(1)))
# orig_annot = inp['orig_annot']
# Sanity check
# iou1 = (torch.diag(IoU_values(reshaped_boxes, orig_annot))
# >= self.acc_iou_threshold).float().mean()
# assert actual_result.item() == iou1.item()
print("Best bounding boxes--------->\n\n", top_boxes, top_scores)
import cv2
import pandas as pd
#Written by Kritika
test_dat = pd.read_csv("data/referit/csv_dir/test.csv")
img_count = 0
for ind in range(len(test_dat)):
bb_data = top_boxes.cpu().numpy(
) #genfromtxt('bb.csv', delimiter=',')
for box_num in range(0, bb_data.shape[0]):
filename = test_dat.iloc[ind]['img_id']
#print(filename)
#Kritika
img_path = r'input/{}'.format('img1.jpeg')
#img_path=r'data/referit/saiapr_tc12_images/{}'.format('imageSend1.jpeg')
#img_path=r'data/referit/saiapr_tc12_images/{}'.format(filename)
img = cv2.imread(img_path)
img = cv2.cvtColor(img, 0)
#img = cv2.rectangle(img,(173, 182), (448, 360),(0,255,0),2)
#pred_box = test_dat.iloc[ind]['pred_box']
x1 = round(bb_data[box_num, 0])
y1 = round(bb_data[box_num, 1])
x2 = round(bb_data[box_num, 2])
y2 = round(bb_data[box_num, 3])
img = cv2.rectangle(img, (x1, y1), (x2, y2), (0, 0, 255), 20)
img_count = img_count + 1
nameToRename = "output/imgR" + str(img_count) + '.jpeg'
cv2.imwrite(nameToRename, img)
p = subprocess.Popen(["display", nameToRename])
if bb_data.shape[0] > 1:
found_obj = input("Do you mean this?: ")
'''
r = sr.Recognizer()
with sr.Microphone() as source:
print("Say something!")
audio = r.listen(source)
# Speech recognition using Google Speech Recognition
try:
# for testing purposes, we're just using the default API key
# to use another API key, use `r.recognize_google(audio, key="GOOGLE_SPEECH_RECOGNITION_API_KEY")`
# instead of `r.recognize_google(audio)`
#print("You said: " + r.recognize_google(audio))
found_obj = r.recognize_google(audio)
print(found_obj)
except sr.UnknownValueError:
print("Google Speech Recognition could not understand audio")
except sr.RequestError as e:
print("Could not request results from Google Speech Recognition service; {0}".format(e))
'''
p.kill()
if "y" == found_obj[0].lower():
break
else:
continue
else:
print("Thank you for your confirmation")
#print(top_boxes.cpu().numpy())
#np.savetxt("bb.csv", top_boxes.cpu().numpy(), delimiter=",")
return out_dict