@@ -41,6 +41,39 @@ def _transforms(dataset):
4141 return transforms
4242
4343
44+ def find_images_with_bounding_size (eval_dataset : paddle .io .Dataset ):
45+ max_length_index = - 1
46+ max_width_index = - 1
47+ min_length_index = - 1
48+ min_width_index = - 1
49+
50+ max_length = float ('-inf' )
51+ max_width = float ('-inf' )
52+ min_length = float ('inf' )
53+ min_width = float ('inf' )
54+ for idx , data in enumerate (eval_dataset ):
55+ image = np .array (data ['img' ])
56+ h , w = image .shape [- 2 :]
57+ if h > max_length :
58+ max_length = h
59+ max_length_index = idx
60+ if w > max_width :
61+ max_width = w
62+ max_width_index = idx
63+ if h < min_length :
64+ min_length = h
65+ min_length_index = idx
66+ if w < min_width :
67+ min_width = w
68+ min_width_index = idx
69+ print (f"Found max image length: { max_length } { max_length_index } )
70+ print (f"Found max image width: { max_width } { max_width_index } )
71+ print (f"Found min image length: { min_length } { min_length_index } )
72+ print (f"Found min image width: { min_width } { min_width_index } )
73+ return paddle .io .Subset (eval_dataset , [max_width_index , max_length_index ,
74+ min_width_index , min_length_index ])
75+
76+
4477def load_predictor (args ):
4578 """
4679 load predictor func
@@ -109,7 +142,7 @@ def predict_image(args):
109142 data = transform ({'img' : args .image_file })
110143 data = data ['img' ][np .newaxis , :]
111144
112- # Step2: Prepare prdictor
145+ # Step2: Prepare predictor
113146 predictor , rerun_flag = load_predictor (args )
114147
115148 # Step3: Inference
@@ -167,6 +200,15 @@ def eval(args):
167200
168201 eval_dataset = builder .val_dataset
169202
203+ predictor , rerun_flag = load_predictor (args )
204+
205+ if rerun_flag and args .use_multi_img_for_dynamic_shape_collect :
206+ print (
207+ "***** Try to find the images with the largest and smallest length and width respectively in the ADE20K "
208+ "dataset for collecting dynamic shape. *****"
209+ )
210+ eval_dataset = find_images_with_bounding_size (eval_dataset )
211+
170212 batch_sampler = paddle .io .BatchSampler (
171213 eval_dataset , batch_size = 1 , shuffle = False , drop_last = False )
172214 loader = paddle .io .DataLoader (
@@ -175,8 +217,6 @@ def eval(args):
175217 num_workers = 0 ,
176218 return_list = True )
177219
178- predictor , rerun_flag = load_predictor (args )
179-
180220 intersect_area_all = 0
181221 pred_area_all = 0
182222 label_area_all = 0
@@ -207,14 +247,22 @@ def eval(args):
207247 time_max = max (time_max , timed )
208248 predict_time += timed
209249 if rerun_flag :
210- print (
211- "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
212- )
213- return
214-
250+ if args .use_multi_img_for_dynamic_shape_collect :
251+ if batch_id == sample_nums - 1 :
252+ print (
253+ "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
254+ )
255+ return
256+ else :
257+ continue
258+ else :
259+ print (
260+ "***** Collect dynamic shape done, Please rerun the program to get correct results. *****"
261+ )
262+ return
215263 logit = reverse_transform (
216- paddle .to_tensor (results ), data ['trans_info' ], mode = "bilinear" )
217- pred = paddle .to_tensor (logit )
264+ paddle .to_tensor (results ). unsqueeze ( 0 ) , data ['trans_info' ], mode = "bilinear" )
265+ pred = paddle .to_tensor (logit ). squeeze ( 0 )
218266 if len (
219267 pred .shape
220268 ) == 4 : # for humanseg model whose prediction is distribution but not class id
@@ -314,6 +362,12 @@ def eval(args):
314362 help = "Whether use mkldnn or not." )
315363 parser .add_argument (
316364 "--cpu_threads" , type = int , default = 1 , help = "Num of cpu threads." )
365+ parser .add_argument (
366+ "--use_multi_img_for_dynamic_shape_collect" ,
367+ type = bool ,
368+ default = False ,
369+ help = "Whether it is necessary to use multiple images to collect shape infomation,\
370+
317371 args = parser .parse_args ()
318372 if args .image_file :
319373 predict_image (args )
0 commit comments