代码收藏家 【YOLOv7】结合GradCAM热图可视化
文章目录
前言
本文使用类激活图（Class Activation Mapping, CAM）方法，对YOLOv7模型的检测结果进行特征可视化，便于我们观察网络在经过backbone和head之后，到底关注了图像的哪些区域才识别定位出了目标，也便于我们分析比较不同的YOLOv7模型，对同一个目标的关注程度。
为了方便各位小伙伴复现，这里放上了YOLOv7的论文和源码（不断更新中…），以及YOLOv7结合GradCAM的Gitee源码（参考：yolov5-gradcam）。
实现效果
实现细节
1、在YOLOv7源码的基础上进行修改和添加操作
---修改部分
---model/yolo.py(Detect类中的forward函数)
---model/yolo.py(Detect类中的_make_grid函数)
---model/experimental.py(attempt_load函数)
---添加部分
---model/gradcam.py
---model/yolov7_object_detector.py
---main_gradcam.py
特别说明：官方给出yolov7.pt预训练权重，最后一层是Detect，不是IDetect层，如果自己使用training/yolov7.yaml进行训练得到的权重，最后则是IDetect，可以仿照在Detect中进行的修改。
2、hook函数
基于Pytorch实现类激活图GradCAM方法的核心，是使用hook函数：在Pytorch的计算过程中，一些中间变量会释放掉，比如特征图、非叶子节点的梯度，在模型前向传播、反向传播的时候添加hook函数，提取一些释放掉而后面又需要用到的变量，也可以用hook函数来改变中间变量的梯度。
在YOLOv7中实现GradCAM时，具体会用到两个hook函数：register_forward_hook和register_full_backward_hook
class YOLOV7GradCAM:
# 初始化，得到target_layer层
def __init__(self, model, layer_name, img_size=(640, 640)):
self.model = model
self.gradients = dict()
self.activations = dict()
# 前向传播获取该层的输出
def forward_hook(module, input, output):
self.activations['value'] = output
return None
# 反向传播获取该层的梯度
def backward_hook(module, grad_input, grad_output):
self.gradients['value'] = grad_output[0]
return None
# 获取要进行反向传播求梯度的目标层
target_layer = find_yolo_layer(self.model, layer_name)
# 前向传播，获取目标层的输入和输出
target_layer.register_forward_hook(forward_hook)
# 反向传播，获取目标层梯度
target_layer.register_full_backward_hook(backward_hook)
3、GradCAM基本实现思路
ReLU函数以移除负值def forward(self, input_img, class_idx=True):
"""
Args:
input_img: input image with shape of (1, 3, H, W)
Return:
mask: saliency map of the same spatial dimension with input
logit: model output
preds: The object predictions
"""
saliency_maps = []
b, c, h, w = input_img.size()
preds, logits = self.model(input_img)
# 遍历每个目标
for logit, cls, cls_name in zip(logits[0], preds[1][0], preds[2][0]):
if class_idx:
score = logit[cls]
else:
score = logit.max()
self.model.zero_grad()
# 模型的对于某一类别的输出，进行反向传播
score.backward(retain_graph=True)
gradients = self.gradients['value']
activations = self.activations['value']
b, k, u, v = gradients.size()
# 对梯度的每个通道进行GAP(全局平均池化操作)
alpha = gradients.view(b, k, -1).mean(2)
# 维度调整，为了后续和目标层输出值逐点相乘
weights = alpha.view(b, k, 1, 1)
# GAP后的梯度值与目标层的输出值逐点相乘
saliency_map = (weights * activations).sum(1, keepdim=True)
# 剔除负值
saliency_map = F.relu(saliency_map)
saliency_map = F.interpolate(saliency_map, size=(h, w), mode='bilinear', align_corners=False)
saliency_map_min, saliency_map_max = saliency_map.min(), saliency_map.max()
saliency_map = (saliency_map - saliency_map_min).div(saliency_map_max - saliency_map_min).data
saliency_maps.append(saliency_map)
return saliency_maps, logits, preds
修改部分
1、Detect类中的forward函数
model/yolo.py文件中的Detect类中的forward函数logits_ = []logits = x[i][..., 5:]logits_.append(logits.view(bs, -1, self.no - 5))out = (torch.cat(z, 1), torch.cat(logits_, 1), x)def forward(self, x):
# x = x.copy() # for profiling
z = [] # inference output
# 1
logits_ = []
self.training |= self.export
for i in range(self.nl):
x[i] = self.m[i](x[i]) # conv
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
if not self.training: # inference
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
# 2
logits = x[i][..., 5:]
y = x[i].sigmoid()
if not torch.onnx.is_in_onnx_export():
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
else:
xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].data # wh
y = torch.cat((xy, wh, y[..., 4:]), -1)
z.append(y.view(bs, -1, self.no))
# 3
logits_.append(logits.view(bs, -1, self.no - 5))
if self.training:
out = x
elif self.end2end:
out = torch.cat(z, 1)
elif self.include_nms:
z = self.convert(z)
out = (z,)
else:
# out = (torch.cat(z, 1), x)
out = (torch.cat(z, 1), torch.cat(logits_, 1), x) # gradcam
return out
2、Detect类中的_make_grid函数
model/yolo.py文件中的Detect类中的_make_grid函数torch.meshgrid函数适配torch>=1.10.0版本@staticmethod
def _make_grid(nx=20, ny=20):
# yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
# torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)], indexing='ij')
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
3、attempt_load函数
inplace作为一个参数进行传入Silu等激活函数无法进行反向传播def attempt_load(weights, map_location=None, inplace=True):
# Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a
model = Ensemble()
for w in weights if isinstance(weights, list) else [weights]:
attempt_download(w)
ckpt = torch.load(w, map_location=map_location) # load
model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval()) # FP32 model
# Compatibility updates
for m in model.modules():
if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]:
# 修改这里
m.inplace = inplace # pytorch 1.7.0 compatibility
添加部分
1、gradcam.py
model文件夹中，添加gradcam.py文件：import time
import torch
import torch.nn.functional as F
def find_yolo_layer(model, layer_name):
"""Find yolov5 layer to calculate GradCAM and GradCAM++
Args:
model: yolov5 model.
layer_name (str): the name of layer with its hierarchical information.
Return:
target_layer: found layer
"""
hierarchy = layer_name.split('_')
target_layer = model.model.model._modules[hierarchy[0]]
for h in hierarchy[1:]:
target_layer = target_layer._modules[h]
return target_layer
class YOLOV7GradCAM:
# 初始化，得到target_layer层
def __init__(self, model, layer_name, img_size=(640, 640)):
self.model = model
self.gradients = dict()
self.activations = dict()
def backward_hook(module, grad_input, grad_output):
self.gradients['value'] = grad_output[0]
return None
def forward_hook(module, input, output):
self.activations['value'] = output
return None
target_layer = find_yolo_layer(self.model, layer_name)
# 获取forward过程中每层的输入和输出，用于对比hook是不是正确记录
target_layer.register_forward_hook(forward_hook)
target_layer.register_full_backward_hook(backward_hook)
device = 'cuda' if next(self.model.model.parameters()).is_cuda else 'cpu'
self.model(torch.zeros(1, 3, *img_size, device=device))
def forward(self, input_img, class_idx=True):
"""
Args:
input_img: input image with shape of (1, 3, H, W)
Return:
mask: saliency map of the same spatial dimension with input
logit: model output
preds: The object predictions
"""
saliency_maps = []
b, c, h, w = input_img.size()
preds, logits = self.model(input_img)
for logit, cls, cls_name in zip(logits[0], preds[1][0], preds[2][0]):
if class_idx:
score = logit[cls]
else:
score = logit.max()
self.model.zero_grad()
# 获取梯度
score.backward(retain_graph=True)
gradients = self.gradients['value']
activations = self.activations['value']
b, k, u, v = gradients.size()
alpha = gradients.view(b, k, -1).mean(2)
weights = alpha.view(b, k, 1, 1)
saliency_map = (weights * activations).sum(1, keepdim=True)
saliency_map = F.relu(saliency_map)
saliency_map = F.interpolate(saliency_map, size=(h, w), mode='bilinear', align_corners=False)
saliency_map_min, saliency_map_max = saliency_map.min(), saliency_map.max()
saliency_map = (saliency_map - saliency_map_min).div(saliency_map_max - saliency_map_min).data
saliency_maps.append(saliency_map)
return saliency_maps, logits, preds
def __call__(self, input_img):
return self.forward(input_img)
2、yolov7_object_detector.py
model文件夹中，添加yolov7_object_detector.py文件：import numpy as np
import torch
from models.experimental import attempt_load
from utils.general import xywh2xyxy
from utils.datasets import letterbox
import cv2
import time
import torchvision
import torch.nn as nn
from utils.general import box_iou
class YOLOV7TorchObjectDetector(nn.Module):
def __init__(self,
model_weight,
device,
img_size,
names=None,
mode='eval',
confidence=0.45,
iou_thresh=0.45,
agnostic_nms=False):
super(YOLOV7TorchObjectDetector, self).__init__()
self.device = device
self.model = None
self.img_size = img_size
self.mode = mode
self.confidence = confidence
self.iou_thresh = iou_thresh
self.agnostic = agnostic_nms
self.model = attempt_load(model_weight, map_location=device, inplace=False)
self.model.requires_grad_(True)
self.model.to(device)
if self.mode == 'train':
self.model.train()
else:
self.model.eval()
# fetch the names
if names is None:
self.names = ['your dataset classnames']
else:
self.names = names
# preventing cold start
img = torch.zeros((1, 3, *self.img_size), device=device)
self.model(img)
@staticmethod
def non_max_suppression(prediction, logits, conf_thres=0.3, iou_thres=0.45, classes=None, agnostic=False,
multi_label=False, labels=(), max_det=300):
"""Runs Non-Maximum Suppression (NMS) on inference and logits results
Returns:
list of detections, on (n,6) tensor per image [xyxy, conf, cls] and pruned input logits (n, number-classes)
"""
nc = prediction.shape[2] - 5 # number of classes
xc = prediction[..., 4] > conf_thres # candidates
# Checks
assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0'
assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0'
# Settings
min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height
max_nms = 30000 # maximum number of boxes into torchvision.ops.nms()
time_limit = 10.0 # seconds to quit after
redundant = True # require redundant detections
multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img)
merge = False # use merge-NMS
output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0]
logits_output = [torch.zeros((0, nc), device=logits.device)] * logits.shape[0]
# logits_output = [torch.zeros((0, 80), device=logits.device)] * logits.shape[0]
for xi, (x, log_) in enumerate(zip(prediction, logits)): # image index, image inference
# Apply constraints
# x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height
x = x[xc[xi]] # confidence
log_ = log_[xc[xi]]
# Cat apriori labels if autolabelling
if labels and len(labels[xi]):
l = labels[xi]
v = torch.zeros((len(l), nc + 5), device=x.device)
v[:, :4] = l[:, 1:5] # box
v[:, 4] = 1.0 # conf
v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls
x = torch.cat((x, v), 0)
# If none remain process next image
if not x.shape[0]:
continue
# Compute conf
x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf
# Box (center x, center y, width, height) to (x1, y1, x2, y2)
box = xywh2xyxy(x[:, :4])
# Detections matrix nx6 (xyxy, conf, cls)
if multi_label:
i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
else: # best class only
conf, j = x[:, 5:].max(1, keepdim=True)
x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
log_ = log_[conf.view(-1) > conf_thres]
# Filter by class
if classes is not None:
x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
# Check shape
n = x.shape[0] # number of boxes
if not n: # no boxes
continue
elif n > max_nms: # excess boxes
x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence
# Batched NMS
c = x[:, 5:6] * (0 if agnostic else max_wh) # classes
boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores
i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS
if i.shape[0] > max_det: # limit detections
i = i[:max_det]
if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean)
# update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix
weights = iou * scores[None] # box weights
x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes
if redundant:
i = i[iou.sum(1) > 1] # require redundancy
output[xi] = x[i]
logits_output[xi] = log_[i]
assert log_[i].shape[0] == x[i].shape[0]
return output, logits_output
@staticmethod
def yolo_resize(img, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True):
return letterbox(img, new_shape=new_shape, color=color, auto=auto, scaleFill=scaleFill, scaleup=scaleup)
def forward(self, img):
prediction, logits, _ = self.model(img, augment=False)
prediction, logits = self.non_max_suppression(prediction, logits, self.confidence, self.iou_thresh,
classes=None,
agnostic=self.agnostic)
self.boxes, self.class_names, self.classes, self.confidences = [[[] for _ in range(img.shape[0])] for _ in
range(4)]
for i, det in enumerate(prediction): # detections per image
if len(det):
for *xyxy, conf, cls in det:
# 返回整数
bbox = [int(b) for b in xyxy]
self.boxes[i].append(bbox)
self.confidences[i].append(round(conf.item(), 2))
cls = int(cls.item())
self.classes[i].append(cls)
if self.names is not None:
self.class_names[i].append(self.names[cls])
else:
self.class_names[i].append(cls)
return [self.boxes, self.classes, self.class_names, self.confidences], logits
def preprocessing(self, img):
if len(img.shape) != 4:
img = np.expand_dims(img, axis=0)
im0 = img.astype(np.uint8)
img = np.array([self.yolo_resize(im, new_shape=self.img_size)[0] for im in im0])
img = img.transpose((0, 3, 1, 2))
img = np.ascontiguousarray(img)
img = torch.from_numpy(img).to(self.device)
img = img / 255.0
return img
3、主函数main_gradcam.py
main_gradcam.py文件import os
import random
import time
import argparse
import numpy as np
from models.gradcam import YOLOV7GradCAM
from models.yolov7_object_detector import YOLOV7TorchObjectDetector
import cv2
# 数据集中的类别名(与标签数字相对应)
names = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
'hair drier', 'toothbrush'] # class names
# yolov7网络中，detect层前的三层输出
target_layers = ['102_act', '103_act', '104_act'] # yolov7
# target_layers = ['74_act', '75_act', '76_act'] # yolov7-tiny
# Arguments
parser = argparse.ArgumentParser()
parser.add_argument('--model-path', type=str, default="weights/yolov7.pt", help='Path to the model')
parser.add_argument('--img-path', type=str, default='figure/cam', help='input image path')
parser.add_argument('--output-dir', type=str, default='outputs/', help='output dir')
parser.add_argument('--img-size', type=int, default=640, help="input image size")
parser.add_argument('--target-layer', type=str, default='76_act',
help='The layer hierarchical address to which gradcam will applied,'
' the names should be separated by underline')
parser.add_argument('--method', type=str, default='gradcam', help='gradcam method: gradcam, gradcampp')
parser.add_argument('--device', type=str, default='cpu', help='cuda or cpu')
parser.add_argument('--names', type=str, default=None,
help='The name of the classes. The default is set to None and is set to coco classes. Provide your custom names as follow: object1,object2,object3')
parser.add_argument('--no_text_box', action='store_true',
help='do not show label and box on the heatmap')
args = parser.parse_args()
def get_res_img(bbox, mask, res_img):
mask = mask.squeeze(0).mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).detach().cpu().numpy().astype(
np.uint8)
heatmap = cv2.applyColorMap(mask, cv2.COLORMAP_JET)
# n_heatmat = (Box.fill_outer_box(heatmap, bbox) / 255).astype(np.float32)
n_heatmat = (heatmap / 255).astype(np.float32)
res_img = res_img / 255
res_img = cv2.add(res_img, n_heatmat)
res_img = (res_img / res_img.max())
return res_img, n_heatmat
def plot_one_box(x, img, color=None, label=None, line_thickness=3):
# this is a bug in cv2. It does not put box on a converted image from torch unless it's buffered and read again!
cv2.imwrite('temp.jpg', (img * 255).astype(np.uint8))
img = cv2.imread('temp.jpg')
# Plots one bounding box on image img
tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1 # line/font thickness
color = color or [random.randint(0, 255) for _ in range(3)]
c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)
if label:
tf = max(tl - 1, 1) # font thickness
t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
outside = c1[1] - t_size[1] - 3 >= 0 # label fits outside box up
c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3 if outside else c1[1] + t_size[1] + 3
outsize_right = c2[0] - img.shape[:2][1] > 0 # label fits outside box right
c1 = c1[0] - (c2[0] - img.shape[:2][1]) if outsize_right else c1[0], c1[1]
c2 = c2[0] - (c2[0] - img.shape[:2][1]) if outsize_right else c2[0], c2[1]
cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA) # filled
cv2.putText(img, label, (c1[0], c1[1] - 2 if outside else c2[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf,
lineType=cv2.LINE_AA)
return img
# 检测单个图片
def main(img_path):
colors = [[random.randint(0, 255) for _ in range(3)] for _ in names]
device = args.device
input_size = (args.img_size, args.img_size)
# 读入图片
img = cv2.imread(img_path) # 读取图像格式：BGR
print('[INFO] Loading the model')
# 实例化YOLOv7模型，得到检测结果
model = YOLOV7TorchObjectDetector(args.model_path, device, img_size=input_size, names=names)
# img[..., ::-1]: BGR --> RGB
# (480, 640, 3) --> (1, 3, 480, 640)
torch_img = model.preprocessing(img[..., ::-1])
tic = time.time()
# 遍历三层检测层
for target_layer in target_layers:
# 获取grad-cam方法
if args.method == 'gradcam':
saliency_method = YOLOV7GradCAM(model=model, layer_name=target_layer, img_size=input_size)
elif args.method == 'gradcampp':
saliency_method = YOLOV7GradCAMPP(model=model, layer_name=target_layer, img_size=input_size)
masks, logits, [boxes, _, class_names, conf] = saliency_method(torch_img) # 得到预测结果
result = torch_img.squeeze(0).mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).detach().cpu().numpy()
result = result[..., ::-1] # convert to bgr
# 保存设置
imgae_name = os.path.basename(img_path) # 获取图片名
save_path = f'{args.output_dir}{imgae_name[:-4]}/{args.method}'
if not os.path.exists(save_path):
os.makedirs(save_path)
print(f'[INFO] Saving the final image at {save_path}')
# 遍历每张图片中的每个目标
for i, mask in enumerate(masks):
# 遍历图片中的每个目标
res_img = result.copy()
# 获取目标的位置和类别信息
bbox, cls_name = boxes[0][i], class_names[0][i]
label = f'{cls_name} {conf[0][i]}' # 类别+置信分数
# 获取目标的热力图
res_img, heat_map = get_res_img(bbox, mask, res_img)
res_img = plot_one_box(bbox, res_img, label=label, color=colors[int(names.index(cls_name))],
line_thickness=3)
# 缩放到原图片大小
res_img = cv2.resize(res_img, dsize=(img.shape[:-1][::-1]))
output_path = f'{save_path}/{target_layer[:-4]}_{i}.jpg'
cv2.imwrite(output_path, res_img)
print(f'{imgae_name[:-4]}_{target_layer[:-4]}_{i}.jpg done!!')
print(f'Total time : {round(time.time() - tic, 4)} s')
if __name__ == '__main__':
# 图片路径为文件夹
if os.path.isdir(args.img_path):
img_list = os.listdir(args.img_path)
print(img_list)
for item in img_list:
# 依次获取文件夹中的图片名，组合成图片的路径
main(os.path.join(args.img_path, item))
# 单个图片
else:
main(args.img_path)
4、 GradCAM++(待完善…)
参考gradcam_plus_plus-pytorch初步在yolov7中实现了gradcam++可视化，但效果并不太好，比gradcam差很多，仍需要优化改进
目前的做法是：在model/gradcam.py文件中，加入如下函数：
class YOLOV7GradCAMPP(YOLOV7GradCAM):
def __init__(self, model, layer_name, img_size=(640, 640)):
super(YOLOV7GradCAMPP, self).__init__(model, layer_name, img_size)
def forward(self, input_img, class_idx=True):
saliency_maps = []
b, c, h, w = input_img.size()
preds, logits = self.model(input_img)
for logit, cls, cls_name in zip(logits[0], preds[1][0], preds[2][0]):
if class_idx:
score = logit[cls]
else:
score = logit.max()
self.model.zero_grad()
# 获取梯度
score.backward(retain_graph=True)
gradients = self.gradients['value'] # dS/dA
activations = self.activations['value'] # A
b, k, u, v = gradients.size()
alpha_num = gradients.pow(2)
alpha_denom = gradients.pow(2).mul(2) + \
activations.mul(gradients.pow(3)).view(b, k, u * v).sum(-1, keepdim=True).view(b, k, 1, 1)
# torch.where(condition, x, y) condition是条件，满足条件就返回x，不满足就返回y
alpha_denom = torch.where(alpha_denom != 0.0, alpha_denom, torch.ones_like(alpha_denom))
alpha = alpha_num.div(alpha_denom + 1e-7)
positive_gradients = F.relu(score.exp() * gradients) # ReLU(dY/dA) == ReLU(exp(S)*dS/dA))
weights = (alpha * positive_gradients).view(b, k, u * v).sum(-1).view(b, k, 1, 1)
saliency_map = (weights * activations).sum(1, keepdim=True)
saliency_map = F.relu(saliency_map)
saliency_map = F.interpolate(saliency_map, size=(h, w), mode='bilinear', align_corners=False)
saliency_map_min, saliency_map_max = saliency_map.min(), saliency_map.max()
saliency_map = (saliency_map - saliency_map_min).div(saliency_map_max - saliency_map_min).data
saliency_maps.append(saliency_map)
return saliency_maps, logits, preds
References
[1] yolov5-gradcam
[2] gradcam_plus_plus-pytorch
[3] pytorch学习笔记十五：Hook函数与CAM可视化
[4]GAP CAM Grad-CAM Grad-CAM++的解释
