def checkOneHot(self):
v = torch.LongTensor([1, 2, 1, 2, 0])
v_length = torch.LongTensor([2, 3])
v_onehot = utils.oneHot(v, v_length, 4)
target = torch.FloatTensor([[[0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 0]],
[[0, 1, 0, 0], [0, 0, 1, 0], [1, 0, 0, 0]]])
assert target.equal(v_onehot)
def checkOneHot(self):
v = torch.LongTensor([1, 2, 1, 2, 0])
v_length = torch.LongTensor([2, 3])
v_onehot = utils.oneHot(v, v_length, 4)
target = torch.FloatTensor([[[0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 0]],
[[0, 1, 0, 0], [0, 0, 1, 0], [1, 0, 0, 0]]])
assert target.equal(v_onehot)
def format_label_data(self, x, requires_grad=False, plabel=None, *args, **kwargs):
if x is None:
return
if plabel is None:
plabel = self.plabel
if issubclass(type(x), list):
new_x = []
for i, label in enumerate(x):
new_x.append(self.format_label_data(label, plabel=plabel[i]))
x = new_x
else:
if type(x)==list:
x = np.array(x)
if type(x)==np.ndarray:
if x.ndim == 1:
if issubclass(type(plabel), list):
label_dim = 0
for pl in plabel:
label_dim += pl['dim']
else:
label_dim = plabel['dim']
x = oneHot(x, label_dim)
if x.dtype!='float32':
x = x.astype('float32')
x = torch.from_numpy(x)
x = x.to(self.device, dtype=torch.float32)
for i in x:
i.requires_grad_(requires_grad)
return x
def selectMove(self, state, goal):
goalVec = utils.oneHot(goal)
if self.controllerEpsilon[goal] < random.random():
# predict action
dummyYtrue = np.zeros((1, 8))
dummyMask = np.zeros((1, 8))
return np.argmax(self.net.controllerNet.predict([np.reshape(state, (1, 84, 84, 4)), np.asarray([goalVec]), dummyYtrue, dummyMask], verbose=0)[1])
return random.choice(self.actionSet)
def __getitem__(self, index):
try:
lables={}
imgInfo=self.dataset[index]
imgName=imgInfo[0]
imgData= Image.open(os.path.join(self.imgPath,imgName))
imgData=trans.Resize((IMG_WIDTH,IMG_HEIGHT))(imgData)
imgData2=trans.ToTensor()(imgData)- 0.5
boxes=np.array([float(i) for i in imgInfo[1:]])
boxes2=np.split(boxes,len(boxes)//5) #标签为多分类，产生多个标记框
for featureSize,anchors in ANCHORS_GROUP.items():
#三种不同的特征图，4个偏移量+1个置信度+N个分类
lables[featureSize]=np.zeros(shape=(featureSize,featureSize,3, 5+CLASS_NUM))
for box in boxes2:
cls,x,y,w,h=box[0:5]
boxArea=w*h
# math.modf
xOffset,xIndex=math.modf(x*featureSize/IMG_WIDTH)
yOffset,yIndex=math.modf(y*featureSize/IMG_HEIGHT)
for i,anchor in enumerate(anchors):
anchorArea=ANCHORS_GROUP_AREA[featureSize][i]
wOffset, hOffset = w / anchor[0], h / anchor[1]
box_area = w * h
# 计算置信度(同心框的IOU(交并))
inter = np.minimum(w, anchor[0]) * np.minimum(h, anchor[1]) # 交集
conf = inter / (box_area + anchorArea - inter)
'''
加log函数方便求梯度
形状为(N,H,W,3,C),相当于增加了一个维度,
其中3相当于在做损失时与3个锚框对应,C维度存放偏移量,iou等数据
*oneHot:从数组中将值取出来
'''
lables[featureSize][int(yIndex),int(xIndex),i]=np.array(
[conf,xOffset,yOffset,np.log(wOffset),np.log(hOffset),*oneHot(CLASS_NUM,int(cls))]
)
return lables[13],lables[26],lables[52],imgData2
except Exception as e:
print("__getitem__",str(e))
def _update(self, stepCount):
batches = self.memory.sample(self.nSamples)
stateVector = []
goalVector = []
for batch in batches:
exp = batch[1]
stateVector.append(exp.state)
goalVector.append(utils.oneHot(exp.goal))
stateVector = np.asarray(stateVector)
goalVector = np.asarray(goalVector)
nextStateVector = []
for batch in batches:
exp = batch[1]
nextStateVector.append(exp.next_state)
nextStateVector = np.asarray(nextStateVector)
rewardVectors = self.net.controllerNet.predict([stateVector, goalVector, np.zeros((self.nSamples,8)), np.zeros((self.nSamples, 8 ))], verbose=0)[1]
rewardVectorsCopy = np.copy(rewardVectors)
rewardVectors = np.zeros((self.nSamples, 8))
nextStateRewardVectors = self.net.targetControllerNet.predict([nextStateVector, goalVector, np.zeros((self.nSamples,8)), np.zeros((self.nSamples, 8 ))], verbose=0)[1]
maskVector = np.zeros((self.nSamples, 8))
for i, batch in enumerate(batches):
exp = batch[1]
idx = batch[0]
maskVector[i, exp.action] = 1.
rewardVectors[i][exp.action] = exp.reward
if not exp.done:
rewardVectors[i][exp.action] += self.gamma * max(nextStateRewardVectors[i])
self.memory.update(idx, np.abs(rewardVectors[i][exp.action] - rewardVectorsCopy[i][exp.action]))
rewardVectors = np.asarray(rewardVectors)
loss = self.net.controllerNet.train_on_batch([stateVector, goalVector, rewardVectors, maskVector], [np.zeros(self.nSamples), rewardVectors])
#Update target network
controllerWeights = self.net.controllerNet.get_weights()
controllerTargetWeights = self.net.targetControllerNet.get_weights()
for i in range(len(controllerWeights)):
controllerTargetWeights[i] = self.targetTau * controllerWeights[i] + (1 - self.targetTau) * controllerTargetWeights[i]
self.net.targetControllerNet.set_weights(controllerTargetWeights)
return loss
def forward(self, x):
"""
This method defines the forward pass of the network.
Parameters:
x:
The input image tensor
Returns:
- A list of indices that forms a caption to the given image -
each element in the list is an index in vocab
- A list of the probabilities for each word in the generated caption
"""
batchSize = len(x)
outputSeq = [[] for _ in range(batchSize)]
inp = torch.zeros(batchSize,
1,
self._rnn.input_size,
requires_grad=False,
device=settings.device)
inp_ = torch.zeros(batchSize,
1,
self._rnn.input_size,
requires_grad=False,
device=settings.device)
first_hidden = self._cnn(x)
hidden_elems = []
for l in range(self._rnn.num_layers):
hidden_elems.append(first_hidden)
hidden = torch.stack(hidden_elems, 0)
assert hidden.size() == (self._rnn.num_layers, batchSize,
self._rnn.hidden_size)
if self._twoStates:
cell = torch.stack(hidden_elems, 0)
assert hidden.size() == (self._rnn.num_layers, batchSize,
self._rnn.hidden_size)
outputLogProbs_elems = []
for step in range(self._maxOutputLen +
1): # plus 1 for the termination char
if self._twoStates:
_, (hidden, cell) = self._rnn(inp, (hidden, cell))
else:
_, hidden = self._rnn(inp, hidden)
output = self._outputNet(hidden[-1])
outputLogProbs_elems.append(
nn.functional.log_softmax(output, dim=1))
for b in range(batchSize):
w = utils.randSamp(output[b], Hyperparams().topk)
outputSeq[b].append(w)
inp_[b, 0] = utils.oneHot(self._vocabLen + 1, w)
inp = inp_.detach().add(0)
outputLogProbs = torch.stack(outputLogProbs_elems, dim=1)
assert outputLogProbs.size() == (batchSize, self._maxOutputLen + 1,
self._vocabLen + 1)
return outputSeq, outputLogProbs