fix(evaluator): compat bug and update requirment (#1416)

requirements.txt

@@ -3,14 +3,11 @@ numpy
 torch>=1.7
 opencv_python
 loguru
-scikit-image
 tqdm
 torchvision
-Pillow
 thop
 ninja
 tabulate
-tensorboard
 
 # verified versions
 # pycocotools corresponds to https://github.com/ppwwyyxx/cocoapi

setup.cfg

@@ -3,8 +3,8 @@ line_length = 100
 multi_line_output = 3
 balanced_wrapping = True
 known_standard_library = setuptools
-known_third_party = tqdm,loguru
-known_data_processing = cv2,numpy,scipy,PIL,matplotlib,scikit_image
+known_third_party = tqdm,loguru,tabulate
+known_data_processing = cv2,numpy,scipy,PIL,matplotlib
 known_datasets = pycocotools
 known_deeplearning = torch,torchvision,caffe2,onnx,apex,timm,thop,torch2trt,tensorrt,openvino,onnxruntime
 known_myself = yolox

yolox/evaluators/coco_evaluator.py

@@ -114,14 +114,8 @@ class COCOEvaluator:
         self.per_class_AR = per_class_AR
 
     def evaluate(
-        self,
-        model,
-        distributed=False,
-        half=False,
-        trt_file=None,
-        decoder=None,
-        test_size=None,
-        return_outputs=False
+        self, model, distributed=False, half=False, trt_file=None,
+        decoder=None, test_size=None, return_outputs=False
     ):
         """
         COCO average precision (AP) Evaluation. Iterate inference on the test dataset

yolox/evaluators/voc_eval.py

@@ -1,5 +1,4 @@
 #!/usr/bin/env python3
-# -*- coding:utf-8 -*-
 # Code are based on
 # https://github.com/rbgirshick/py-faster-rcnn/blob/master/lib/datasets/voc_eval.py
 # Copyright (c) Bharath Hariharan.
@@ -13,7 +12,7 @@ import numpy as np
 
 
 def parse_rec(filename):
-    """ Parse a PASCAL VOC xml file """
+    """Parse a PASCAL VOC xml file"""
     tree = ET.parse(filename)
     objects = []
     for obj in tree.findall("object"):
@@ -35,7 +34,7 @@ def parse_rec(filename):
 
 
 def voc_ap(rec, prec, use_07_metric=False):
-    """ap = voc_ap(rec, prec, [use_07_metric])
+    """
     Compute VOC AP given precision and recall.
     If use_07_metric is true, uses the
     VOC 07 11 point method (default:False).
@@ -92,9 +91,9 @@ def voc_eval(
         for i, imagename in enumerate(imagenames):
             recs[imagename] = parse_rec(annopath.format(imagename))
             if i % 100 == 0:
-                print("Reading annotation for {:d}/{:d}".format(i + 1, len(imagenames)))
+                print(f"Reading annotation for {i + 1}/{len(imagenames)}")
         # save
-        print("Saving cached annotations to {:s}".format(cachefile))
+        print(f"Saving cached annotations to {cachefile}")
         with open(cachefile, "wb") as f:
             pickle.dump(recs, f)
     else:
@@ -155,8 +154,7 @@ def voc_eval(
             # union
             uni = (
                 (bb[2] - bb[0] + 1.0) * (bb[3] - bb[1] + 1.0)
-                + (BBGT[:, 2] - BBGT[:, 0] + 1.0) * (BBGT[:, 3] - BBGT[:, 1] + 1.0)
-                - inters
+                + (BBGT[:, 2] - BBGT[:, 0] + 1.0) * (BBGT[:, 3] - BBGT[:, 1] + 1.0) - inters
             )
 
             overlaps = inters / uni

yolox/evaluators/voc_evaluator.py

@@ -21,14 +21,7 @@ class VOCEvaluator:
     VOC AP Evaluation class.
     """
 
-    def __init__(
-        self,
-        dataloader,
-        img_size,
-        confthre,
-        nmsthre,
-        num_classes,
-    ):
+    def __init__(self, dataloader, img_size, confthre, nmsthre, num_classes):
         """
         Args:
             dataloader (Dataloader): evaluate dataloader.
@@ -46,13 +39,8 @@ class VOCEvaluator:
         self.num_images = len(dataloader.dataset)
 
     def evaluate(
-        self,
-        model,
-        distributed=False,
-        half=False,
-        trt_file=None,
-        decoder=None,
-        test_size=None,
+        self, model, distributed=False, half=False, trt_file=None,
+        decoder=None, test_size=None, return_outputs=False,
     ):
         """
         VOC average precision (AP) Evaluation. Iterate inference on the test dataset
@@ -91,9 +79,7 @@ class VOCEvaluator:
             model(x)
             model = model_trt
 
-        for cur_iter, (imgs, _, info_imgs, ids) in enumerate(
-            progress_bar(self.dataloader)
-        ):
+        for cur_iter, (imgs, _, info_imgs, ids) in enumerate(progress_bar(self.dataloader)):
             with torch.no_grad():
                 imgs = imgs.type(tensor_type)
 
@@ -127,13 +113,13 @@ class VOCEvaluator:
 
         eval_results = self.evaluate_prediction(data_list, statistics)
         synchronize()
+        if return_outputs:
+            return eval_results, data_list
         return eval_results
 
     def convert_to_voc_format(self, outputs, info_imgs, ids):
         predictions = {}
-        for (output, img_h, img_w, img_id) in zip(
-            outputs, info_imgs[0], info_imgs[1], ids
-        ):
+        for output, img_h, img_w, img_id in zip(outputs, info_imgs[0], info_imgs[1], ids):
             if output is None:
                 predictions[int(img_id)] = (None, None, None)
                 continue
@@ -142,9 +128,7 @@ class VOCEvaluator:
             bboxes = output[:, 0:4]
 
             # preprocessing: resize
-            scale = min(
-                self.img_size[0] / float(img_h), self.img_size[1] / float(img_w)
-            )
+            scale = min(self.img_size[0] / float(img_h), self.img_size[1] / float(img_w))
             bboxes /= scale
 
             cls = output[:, 6]
@@ -175,7 +159,6 @@ class VOCEvaluator:
                 )
             ]
         )
-
         info = time_info + "\n"
 
         all_boxes = [
@@ -196,13 +179,9 @@ class VOCEvaluator:
                 c_dets = torch.cat((bboxes, scores.unsqueeze(1)), dim=1)
                 all_boxes[j][img_num] = c_dets[mask_c].numpy()
 
-            sys.stdout.write(
-                "im_eval: {:d}/{:d} \r".format(img_num + 1, self.num_images)
-            )
+            sys.stdout.write(f"im_eval: {img_num + 1}/{self.num_images} \r")
             sys.stdout.flush()
 
         with tempfile.TemporaryDirectory() as tempdir:
-            mAP50, mAP70 = self.dataloader.dataset.evaluate_detections(
-                all_boxes, tempdir
-            )
+            mAP50, mAP70 = self.dataloader.dataset.evaluate_detections(all_boxes, tempdir)
             return mAP50, mAP70, info

yolox/utils/model_utils.py

@@ -8,7 +8,6 @@ from typing import Sequence
 
 import torch
 import torch.nn as nn
-from thop import profile
 
 __all__ = [
     "fuse_conv_and_bn",
@@ -21,6 +20,8 @@ __all__ = [
 
 
 def get_model_info(model: nn.Module, tsize: Sequence[int]) -> str:
+    from thop import profile
+
     stride = 64
     img = torch.zeros((1, 3, stride, stride), device=next(model.parameters()).device)
     flops, params = profile(deepcopy(model), inputs=(img,), verbose=False)