def secure_anchor_iou_file(self):
"""
integrated function to secure ious' supply
"""
# check all the image files, if its .npy is not presented, gen it rightaway
logging.info('check if all the iou files are ready...')
for self.lineIdx in xrange(self.imgNum):
self.lineStr = self.fs[self.lineIdx]
if ';' not in self.lineStr:
continue
self.lg = self.lineStr.rsplit(';\t')
self.imgName = self.lg.pop(0)
ioufile = self.gen_ioufileDir(self.imgName, self.featName)
if '.npz' in ioufile: # already exists
continue
# if not, gen it...
logging.info('.npz missed for %s, gen it...' % self.imgName)
self.parseLine()
if self.stride is None:
self.stride = (self.im_info[0] / self.feat_shape[-2],
self.im_info[1] / self.feat_shape[-1])
anchor = tool.genAnchor(self.im_info,self.angleD, self.HoW, self.sideLength,\
self.feat_shape,self.stride) # anchor : N x 5
# iou_matrix = it_IoU(anchor, self.gdt[:,1:], \
# self.x_num) # BE CAREFUL ! [:,0] IS FOR LABEL !
iou_matrix = gpu_it_IoU(anchor, self.gdt[:,1:], \
self.x_num) # BE CAREFUL ! [:,0] IS FOR LABEL !
np.savez(ioufile, anchor=anchor, iou=iou_matrix)
logging.info('saved at %s' % ioufile)
logging.info('ok.')
def next(self): # one img per call
self.checkAndSet()
self.genIdx()
self.parseLine()
# havnt ended yet! longway to go
# assign gdt for each anchor, but 1st, u need generate anchors
# an_gdtbox IS NOT gdtbox FOR NOW !
anchor = tool.genAnchor(self.im_info,self.angleD, self.HoW, self.sideLength,\
self.feat_shape,self.stride) # anchor : N x 5
# logging.debug('anchor size:'+str(anchor.shape))
# logging.debug('anchor data:'+str(anchor))
if config.it.debug.debug:
self.anchor=anchor
# IoU ?
start=time.time() if config.debug.debug else None
ioufile = os.path.join(self.ioutdir,self.imgName[:-4])+'_'+str(self.imgHWC[:2])+'.npy' # encode the additional id
# if self.hasiou_flag and not os.path.isfile(ioufile): # need to notify the owner
# logging.info('ioufile does not exsists: '+ioufile)
if not self.hasiou_flag or not os.path.isfile(ioufile): # calculate and save it
iou_matrix = it_IoU(anchor, self.gdt[:,1:], \
self.x_num) # BE CAREFUL ! [:,0] IS FOR LABEL !
np.save(ioufile,iou_maxtrix)
else: # just load it
iou_matrix = np.load(ioufile)
if config.it.debug.debug:
logging.debug('[IoU] time elapse:'+str(time.time()-start))
logging.debug('[IoU] maximum:%f'%np.max(iou_matrix))
# logging.debug('iou matrix:'+str(iou_matrix))
# logging.debug('begin to calculate IoU...')
# logging.debug('anchor:'+str(anchor))
# logging.debug('gdt:'+str(self.gdt))
# display the img with anchors and gdt onit !
img=self.img.asnumpy()
img=tool.draw_angleBox(img,anchor,(0,250,0))
img=tool.draw_angleBox(img,self.gdt[:,1:],(0,0,250))
plt.imshow(img)
plt.show()
self.target_label,self.target_bbx, self.target_weight =post_iou(iou_matrix,anchor,self.gdt)
self.img=nd.transpose(self.img,axes=(2,0,1))
self.img=nd.expand_dims(self.img,axis=0)
return mx.io.DataBatch( [self.img,self.gdt,self.im_info],[self.target_label,self.target_bbx, self.target_weight])
s=raw_input('Press S to save raw featureMap\n')
if s=='S':
v.save_rawfeat('feature-%s'%imgname)#v.save_rawfeature('feature-%s'%imgname)
#s=raw_input('press any key to exit')
"""
##################################
# draw anchor
##################################
current_dir = os.getcwd()
os.chdir('..')
from tool import genAnchor, draw_angleBox
os.chdir(current_dir)
imgsize = 224
feat_size = 14 * 4
stride = imgsize / feat_size
stride = [imgsize, imgsize]
feat_shape = [feat_size, feat_size]
angleD = [0, 60, -60]
HoW = [2, 0.5, 1]
sideLength = [90, 60, 30]
img = cv2.imread(os.path.join(imgdir, imgname))[:, :, [2, 1, 0]]
anchor = genAnchor(None, angleD, HoW, sideLength, feat_shape, stride)
img_anchor = draw_angleBox(img, anchor, [0, 0, 255])
plt.imshow(img_anchor)
plt.show()
def predict(self,img_path):
os.path.isfile(img_path),'%s does not exist!'%img_path
self.imgname = os.path.basename(img_path)
self.img = mx.img.imdecode(open(img_path,'rb').read()).asnumpy()
img = mx.nd.array(self.img - self.m, cfg.predict.ctx)
img = mx.nd.transpose(img, axes=(2,0,1))
img = mx.nd.expand_dims( img, axis=0)
self.im_info = np.array(self.img.shape[:2])
_,feat_shape, _ = self.feature_symbol.infer_shape(data=img.shape)
feat_shape = feat_shape[0]
self.feat_shape = feat_shape
self.stride=(self.im_info[0]/self.feat_shape[-2],self.im_info[1]/self.feat_shape[-1])
self.anchor = tool.genAnchor(self.im_info,self.angleD, self.HoW, self.sideLength, self.feat_shape, self.stride)
logging.debug(img.shape)
# assert 0
d = mx.io.DataBatch([img],provide_data=[ ('data',img.shape),], provide_label=None)
self.mod.forward(d)
#self.raw_predict_bbox, self.raw_predict_score = [mx.nd.reshape(x,(-3,0)) for x in self.mod.get_outputs()[:2] ]
self.predict_bbox, self.predict_score = [mx.nd.reshape(x,(-3,0)).asnumpy() for x in self.mod.get_outputs()[:2] ]
self.predict_transfered_bbox = tool.bbox_inv_transfer(self.anchor, self.predict_bbox)
# filter who obviously out of boundary...
pick_idx = self.predict_transfered_bbox[:,0] < self.im_info[1]
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
pick_idx = self.predict_transfered_bbox[:,0] >0
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
# logging.debug('x:'+str(pick_idx.sum()))
pick_idx = self.predict_transfered_bbox[:,1] < self.im_info[0]
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
pick_idx = self.predict_transfered_bbox[:,1] > 0
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
# logging.debug('y:'+str(pick_idx.sum()))
pick_idx =self.predict_transfered_bbox[:,3] < max(self.im_info)
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
pick_idx =self.predict_transfered_bbox[:,3]> 0
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
# logging.debug('rh:'+str(pick_idx.sum()))
pick_idx =self.predict_transfered_bbox[:,4] < max(self.im_info)
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
pick_idx =self.predict_transfered_bbox[:,4] > 0
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
# logging.debug('rw:'+str(pick_idx.sum()))
pick_idx =self.predict_transfered_bbox[:,2] < np.pi
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
pick_idx =self.predict_transfered_bbox[:,2] > -np.pi
self.predict_transfered_bbox = self.predict_transfered_bbox[pick_idx,:]
self.predict_score = self.predict_score[pick_idx,:]
label_names=('target_label', ),
context=ctx)
# need an anchor generator...
# for raw image prediction use.
if os.path.isfile(imgPath):
feat_symbol = symbol.get_internals()[cfg.net.rpn_conv_name + '_output']
with open(imgPath, 'rb') as f:
img_mx_raw = mx.img.imdecode(f.read())
img = mx.nd.array(img_mx_raw.asnumpy() - np.array(cfg.it.mean))
H, W, C = img.shape
_, feat_shape, _ = feat_symbol.infer_shape(data=(1, C, H, W))
feat_shape = feat_shape[0]
im_info = np.array([H, W])
stride = (im_info[0] / feat_shape[-2], im_info[1] / feat_shape[-1])
anchor = tool.genAnchor( im_info,cfg.ANCHOR_angleD, cfg.ANCHOR_HoW, cfg.ANCHOR_sideLength,\
feat_shape, stride) # anchor : N x 5
img = mx.nd.transpose(img, axes=(2, 0, 1))
img_data = mx.nd.expand_dims(img, axis=0)
dataBatch = mx.io.DataBatch([
img_data,
])
mod.bind(data_shapes=[
('data', img_data.shape),
], for_training=False)
mod.init_params(None, arg_params=arg_params, aux_params= aux_params, allow_missing=False,\
allow_extra=True)
mod.forward(dataBatch)
viz_score_predict(img_mx_raw.asnumpy().astype(np.uint8), anchor, None,mod.get_outputs()[1].asnumpy(),\
mod.get_outputs()[0].asnumpy(), score_th=.1,show_num=150 )
mx.nd.waitall()