def finish_episode():
R = 0
policy_loss = []
value_loss = []
returns = []
for r in policy.rewards[::-1]:
R = r + gamma * R
returns.insert(0, R)
returns = paddle.to_tensor(returns)
returns = (returns - returns.mean()) / (returns.std() + eps)
for (log_prob, value), R in zip(policy.saved_log_probs, returns):
advantage = R - value
policy_loss.append(-log_prob * advantage)
value_loss.append(F.smooth_l1_loss(value.reshape([-1]), R.reshape([-1])))
optimizer.clear_grad()
policy_loss = paddle.concat(policy_loss).sum()
value_loss = paddle.concat(value_loss).sum()
loss = policy_loss + value_loss
loss.backward()
optimizer.step()
del policy.rewards[:]
del policy.saved_log_probs[:]
def forward(self, features, im_info, boxes=None):
# prediction
pred_cls_score_list = []
pred_bbox_offsets_list = []
for x in features:
t = F.relu(self.rpn_conv(x))
pred_cls_score_list.append(self.rpn_cls_score(t))
pred_bbox_offsets_list.append(self.rpn_bbox_offsets(t))
# get anchors
all_anchors_list = []
# stride: 64,32,16,8,4 p6->p2
base_stride = 4
off_stride = 2**(len(features) - 1) # 16
for fm in features:
layer_anchors = self.anchors_generator(fm, base_stride, off_stride)
off_stride = off_stride // 2
all_anchors_list.append(layer_anchors)
# sample from the predictions
rpn_rois = find_top_rpn_proposals(self.training,
pred_bbox_offsets_list,
pred_cls_score_list,
all_anchors_list, im_info)
rpn_rois = rpn_rois.cast('float32')
if self.training:
rpn_labels, rpn_bbox_targets = fpn_anchor_target(
boxes, im_info, all_anchors_list)
#rpn_labels = rpn_labels.astype(np.int32)
pred_cls_score, pred_bbox_offsets = fpn_rpn_reshape(
pred_cls_score_list, pred_bbox_offsets_list)
# rpn loss
valid_masks = rpn_labels >= 0
# objectness_loss = softmax_loss(
# torch.gather(pred_cls_score,torch.nonzero(valid_masks)),
# torch.gather(rpn_labels,torch.nonzero(valid_masks)))
objectness_loss = F.binary_cross_entropy(
F.softmax(
torch.gather(pred_cls_score, torch.nonzero(valid_masks))),
torch.gather(
torch.eye(2),
torch.gather(rpn_labels, torch.nonzero(valid_masks))))
pos_masks = rpn_labels > 0
# localization_loss = smooth_l1_loss(
# pred_bbox_offsets[pos_masks],
# rpn_bbox_targets[pos_masks],
# config.rpn_smooth_l1_beta)
localization_loss = \
F.smooth_l1_loss(torch.gather(pred_bbox_offsets, torch.nonzero(pos_masks)),
torch.gather(rpn_bbox_targets, torch.nonzero(pos_masks)),delta=config.rcnn_smooth_l1_beta)
normalizer = 1 / valid_masks.cast('float32').sum()
loss_rpn_cls = objectness_loss.sum() * normalizer
loss_rpn_loc = localization_loss.sum() * normalizer
loss_dict = {}
loss_dict['loss_rpn_cls'] = loss_rpn_cls
loss_dict['loss_rpn_loc'] = loss_rpn_loc
return rpn_rois, loss_dict
else:
return rpn_rois
def forward(self, student, teacher):
bs = student.shape[0]
student = student.reshape([bs, -1])
teacher = teacher.reshape([bs, -1])
t_d = pdist(teacher, squared=False, eps=self.eps)
mean_td = t_d.mean()
t_d = t_d / (mean_td + self.eps)
d = pdist(student, squared=False, eps=self.eps)
mean_d = d.mean()
d = d / (mean_d + self.eps)
loss = F.smooth_l1_loss(d, t_d, reduction="mean")
return loss
def forward(self, confidence, predicted_locations, labels, gt_locations):
"""Compute classification loss and smooth l1 loss.
Args:
confidence (batch_size, num_priors, num_classes): class predictions.
locations (batch_size, num_priors, 4): predicted locations.
labels (batch_size, num_priors): real labels of all the priors.
boxes (batch_size, num_priors, 4): real boxes corresponding all the priors.
"""
num_classes = confidence.shape[2]
with paddle.no_grad():
# derived from cross_entropy=sum(log(p))
loss = -F.log_softmax(confidence, 2)[:, :, 0]
mask = box_utils.hard_negative_mining(loss, labels,
self.neg_pos_ratio)
confidence = paddle.concat([
confidence[:, :, 0].masked_select(mask).reshape([-1, 1]),
confidence[:, :, 1].masked_select(mask).reshape([-1, 1])
],
axis=1)
classification_loss = F.cross_entropy(confidence.reshape(
[-1, num_classes]),
labels.masked_select(mask),
reduction='sum')
pos_mask = labels > 0
predicted_locations = predicted_locations.masked_select(
paddle.concat([
pos_mask.reshape(pos_mask.shape + [1]),
pos_mask.reshape(pos_mask.shape + [1]),
pos_mask.reshape(pos_mask.shape + [1]),
pos_mask.reshape(pos_mask.shape + [1])
],
axis=2)).reshape([-1, 4])
gt_locations = gt_locations.masked_select(
paddle.concat([
pos_mask.reshape(pos_mask.shape + [1]),
pos_mask.reshape(pos_mask.shape + [1]),
pos_mask.reshape(pos_mask.shape + [1]),
pos_mask.reshape(pos_mask.shape + [1])
],
axis=2)).reshape([-1, 4])
smooth_l1_loss = F.smooth_l1_loss(predicted_locations,
gt_locations.cast('float32'),
reduction='sum') # smooth_l1_loss
# smooth_l1_loss = F.mse_loss(predicted_locations, gt_locations, reduction='sum') #l2 loss
num_pos = gt_locations.shape[0]
return smooth_l1_loss / num_pos, classification_loss / num_pos
def train(model, data_loader, optimizer, lr_scheduler, epoch, LOG):
stages = 4
losses = [AverageMeter() for _ in range(stages)]
length_loader = len(data_loader)
model.train()
for batch_id, data in enumerate(data_loader()):
left_img, right_img, gt = data
mask = paddle.to_tensor(gt.numpy() > 0)
gt_mask = paddle.masked_select(gt, mask)
outputs = model(left_img, right_img)
outputs = [paddle.squeeze(output) for output in outputs]
tem_stage_loss = []
for index in range(stages):
temp_loss = args.loss_weights[index] * F.smooth_l1_loss(
paddle.masked_select(outputs[index], mask),
gt_mask,
reduction='mean')
tem_stage_loss.append(temp_loss)
losses[index].update(
float(temp_loss.numpy() / args.loss_weights[index]))
sum_loss = paddle.add_n(tem_stage_loss)
sum_loss.backward()
optimizer.step()
optimizer.clear_grad()
if batch_id % 5 == 0:
info_str = [
'Stage {} = {:.2f}({:.2f})'.format(x, losses[x].val,
losses[x].avg)
for x in range(stages)
]
info_str = '\t'.join(info_str)
info_str = 'Train Epoch{} [{}/{}] lr:{:.5f}\t{}'.format(
epoch, batch_id, length_loader, optimizer.get_lr(), info_str)
LOG.info(info_str)
lr_scheduler.step()
info_str = '\t'.join(
['Stage {} = {:.2f}'.format(x, losses[x].avg) for x in range(stages)])
LOG.info('Average train loss: ' + info_str)
def forward(self, student, teacher):
# reshape for feature map distillation
bs = student.shape[0]
student = student.reshape([bs, -1])
teacher = teacher.reshape([bs, -1])
td = (teacher.unsqueeze(0) - teacher.unsqueeze(1))
norm_td = F.normalize(td, p=2, axis=2)
t_angle = paddle.bmm(norm_td, norm_td.transpose([0, 2, 1])).reshape(
[-1, 1])
sd = (student.unsqueeze(0) - student.unsqueeze(1))
norm_sd = F.normalize(sd, p=2, axis=2)
s_angle = paddle.bmm(norm_sd, norm_sd.transpose([0, 2, 1])).reshape(
[-1, 1])
loss = F.smooth_l1_loss(s_angle, t_angle, reduction='mean')
return loss
def det_loss(self, p_det, anchor, t_conf, t_box):
pshape = paddle.shape(p_det)
pshape.stop_gradient = True
nB, nGh, nGw = pshape[0], pshape[-2], pshape[-1]
nA = len(anchor)
p_det = paddle.reshape(
p_det, [nB, nA, self.num_classes + 5, nGh, nGw]).transpose(
(0, 1, 3, 4, 2))
# 1. loss_conf: cross_entropy
p_conf = p_det[:, :, :, :, 4:6]
p_conf_flatten = paddle.reshape(p_conf, [-1, 2])
t_conf_flatten = t_conf.flatten()
t_conf_flatten = paddle.cast(t_conf_flatten, dtype="int64")
t_conf_flatten.stop_gradient = True
loss_conf = F.cross_entropy(p_conf_flatten,
t_conf_flatten,
ignore_index=-1,
reduction='mean')
loss_conf.stop_gradient = False
# 2. loss_box: smooth_l1_loss
p_box = p_det[:, :, :, :, :4]
p_box_flatten = paddle.reshape(p_box, [-1, 4])
t_box_flatten = paddle.reshape(t_box, [-1, 4])
fg_inds = paddle.nonzero(t_conf_flatten > 0).flatten()
if fg_inds.numel() > 0:
reg_delta = paddle.gather(p_box_flatten, fg_inds)
reg_target = paddle.gather(t_box_flatten, fg_inds)
else:
reg_delta = paddle.to_tensor([0, 0, 0, 0], dtype='float32')
reg_delta.stop_gradient = False
reg_target = paddle.to_tensor([0, 0, 0, 0], dtype='float32')
reg_target.stop_gradient = True
loss_box = F.smooth_l1_loss(reg_delta,
reg_target,
reduction='mean',
delta=1.0)
loss_box.stop_gradient = False
return loss_conf, loss_box
def forward(self, student, teacher):
# GAP to reduce memory
if self.avgpool is not None:
# NxC1xH1xW1 -> NxC1x1x1
student = self.avgpool(student)
# NxC2xH2xW2 -> NxC2x1x1
teacher = self.avgpool(teacher)
bs = student.shape[0]
student = student.reshape([bs, -1])
teacher = teacher.reshape([bs, -1])
t_d = pdist(teacher, squared=False)
mean_td = t_d.mean()
t_d = t_d / (mean_td + self.eps)
d = pdist(student, squared=False)
mean_d = d.mean()
d = d / (mean_d + self.eps)
loss = F.smooth_l1_loss(d, t_d, reduction="mean")
return loss
def forward(self, input, label, conf):
x_emb = self.embedding(input)
fc = self.lin_a(x_emb)
mask = conf > 0
mask = paddle.cast(mask, dtype="int64")
mask.stop_gradient = True
emb_mask = mask.max(1).flatten()
emb_mask_inds = paddle.nonzero(emb_mask > 0).flatten()
emb_mask_inds.stop_gradient = True
if emb_mask_inds.numel() == 0:
loss_box = self.phony * 0
else:
projection = self.lin_b(fc)
projection = paddle.reshape(projection, shape=[-1, 1])
output = paddle.gather(projection, emb_mask_inds)
target = paddle.gather(label, emb_mask_inds)
loss_box = F.smooth_l1_loss(output,
target,
reduction='sum',
delta=1.0)
loss_box = loss_box / len(conf)
return loss_box
def forward(self, student, teacher):
# GAP to reduce memory
if self.avgpool is not None:
# NxC1xH1xW1 -> NxC1x1x1
student = self.avgpool(student)
# NxC2xH2xW2 -> NxC2x1x1
teacher = self.avgpool(teacher)
# reshape for feature map distillation
bs = student.shape[0]
student = student.reshape([bs, -1])
teacher = teacher.reshape([bs, -1])
td = (teacher.unsqueeze(0) - teacher.unsqueeze(1))
norm_td = F.normalize(td, p=2, axis=2)
t_angle = paddle.bmm(norm_td, norm_td.transpose([0, 2,
1])).reshape([-1, 1])
sd = (student.unsqueeze(0) - student.unsqueeze(1))
norm_sd = F.normalize(sd, p=2, axis=2)
s_angle = paddle.bmm(norm_sd, norm_sd.transpose([0, 2,
1])).reshape([-1, 1])
loss = F.smooth_l1_loss(s_angle, t_angle, reduction='mean')
return loss
def forward(self, boxes, scores, gt_bbox, gt_label, prior_boxes):
boxes = paddle.concat(boxes, axis=1)
scores = paddle.concat(scores, axis=1)
gt_label = gt_label.unsqueeze(-1).astype('int64')
prior_boxes = paddle.concat(prior_boxes, axis=0)
bg_index = scores.shape[-1] - 1
# Match bbox and get targets.
targets_bbox, targets_label = \
self._bipartite_match_for_batch(gt_bbox, gt_label, prior_boxes, bg_index)
targets_bbox.stop_gradient = True
targets_label.stop_gradient = True
# Compute regression loss.
# Select positive samples.
bbox_mask = (targets_label != bg_index).astype(boxes.dtype)
loc_loss = bbox_mask * F.smooth_l1_loss(
boxes, targets_bbox, reduction='none')
loc_loss = loc_loss.sum() * self.loc_loss_weight
# Compute confidence loss.
conf_loss = F.softmax_with_cross_entropy(scores, targets_label)
# Mining hard examples.
label_mask = self._mine_hard_example(conf_loss.squeeze(-1),
targets_label.squeeze(-1),
bg_index)
conf_loss = conf_loss * label_mask.unsqueeze(-1).astype(
conf_loss.dtype)
conf_loss = conf_loss.sum() * self.conf_loss_weight
# Compute overall weighted loss.
normalizer = (targets_label != bg_index).astype('float32').sum().clip(
min=1)
loss = (conf_loss + loc_loss) / (normalizer + 1e-9)
return loss
def train(args):
# 使用 GPU训练
if paddle.is_compiled_with_cuda():
paddle.set_device("gpu:0")
# 创建多进程的游戏环境
envs = MultipleEnvironments(args.game, args.num_processes)
# 固定初始化状态
paddle.seed(123)
# 创建模型
model = Model(envs.num_states, envs.num_actions)
# 加载预训练模型
if args.trained_model is not None:
model.load_dict(paddle.load(args.trained_model))
# 创建保存模型的文件夹
if not os.path.isdir(args.saved_path):
os.makedirs(args.saved_path)
paddle.save(model.state_dict(),
"{}/model_{}.pdparams".format(args.saved_path, args.game))
# 为游戏评估单独开一个进程
mp = _mp.get_context("spawn")
process = mp.Process(target=eval,
args=(args, envs.num_states, envs.num_actions))
process.start()
# 创建优化方法
clip_grad = paddle.nn.ClipGradByNorm(clip_norm=0.5)
optimizer = paddle.optimizer.Adam(parameters=model.parameters(),
learning_rate=args.lr,
grad_clip=clip_grad)
# 刚开始给每个进程的游戏执行初始化
[agent_conn.send(("reset", None)) for agent_conn in envs.agent_conns]
# 获取游戏初始的界面
curr_states = [agent_conn.recv() for agent_conn in envs.agent_conns]
curr_states = paddle.to_tensor(np.concatenate(curr_states, 0),
dtype='float32')
curr_episode = 0
while True:
curr_episode += 1
old_log_policies, actions, values, states, rewards, dones = [], [], [], [], [], []
for _ in range(args.num_local_steps):
states.append(curr_states)
# 执行预测
logits, value = model(curr_states)
# 计算每个动作的概率值
policy = F.softmax(logits)
# 根据每个标签的概率随机生成符合概率的标签
old_m = Categorical(policy)
action = old_m.sample([1]).squeeze()
# 记录预测数据
actions.append(action)
values.append(value.squeeze())
# 计算类别的概率的对数
old_log_policy = old_m.log_prob(paddle.unsqueeze(action, axis=1))
old_log_policy = paddle.squeeze(old_log_policy)
old_log_policies.append(old_log_policy)
# 向各个进程游戏发送动作
[
agent_conn.send(("step", int(act[0])))
for agent_conn, act in zip(envs.agent_conns, action)
]
# 将多进程的游戏数据打包
state, reward, done, info = zip(
*[agent_conn.recv() for agent_conn in envs.agent_conns])
# 进行数据转换
state = paddle.to_tensor(np.concatenate(state, 0), dtype='float32')
# 转换为tensor数据
reward = paddle.to_tensor(reward, dtype='float32')
done = paddle.to_tensor(done, dtype='float32')
# 记录预测数据
rewards.append(reward)
dones.append(done)
curr_states = state
# 根据上面最后的图像预测
_, next_value, = model(curr_states)
next_value = next_value.squeeze()
old_log_policies = paddle.concat(old_log_policies).detach().squeeze()
actions = paddle.concat(actions).squeeze()
values = paddle.concat(values).squeeze().detach()
states = paddle.concat(states).squeeze()
gae = 0.0
R = []
for value, reward, done in list(zip(values, rewards, dones))[::-1]:
gae = gae * args.gamma * args.tau
gae = gae + reward + args.gamma * next_value.detach() * (
1.0 - done) - value.detach()
next_value = value
R.append(gae + value)
R = R[::-1]
R = paddle.concat(R).detach()
advantages = R - values
for i in range(args.num_epochs):
indice = paddle.randperm(args.num_local_steps * args.num_processes)
for j in range(args.batch_size):
batch_indices = indice[int(j * (
args.num_local_steps * args.num_processes / args.batch_size
)):int((j + 1) * (args.num_local_steps * args.num_processes /
args.batch_size))]
# 根据拿到的图像执行预测
logits, value = model(paddle.gather(states, batch_indices))
# 计算每个动作的概率值
new_policy = F.softmax(logits)
# 计算类别的概率的对数
new_m = Categorical(new_policy)
new_log_policy = new_m.log_prob(
paddle.unsqueeze(paddle.gather(actions, batch_indices),
axis=1))
new_log_policy = paddle.squeeze(new_log_policy)
# 计算actor损失
ratio = paddle.exp(
new_log_policy -
paddle.gather(old_log_policies, batch_indices))
advantage = paddle.gather(advantages, batch_indices)
actor_loss = paddle.clip(ratio, 1.0 - args.epsilon,
1.0 + args.epsilon) * advantage
actor_loss = paddle.concat([
paddle.unsqueeze(ratio * advantage, axis=0),
paddle.unsqueeze(actor_loss, axis=0)
])
actor_loss = -paddle.mean(paddle.min(actor_loss, axis=0))
# 计算critic损失
critic_loss = F.smooth_l1_loss(paddle.gather(R, batch_indices),
value.squeeze())
entropy_loss = paddle.mean(new_m.entropy())
# 计算全部损失
total_loss = actor_loss + critic_loss - args.beta * entropy_loss
# 计算梯度
total_loss.backward()
optimizer.step()
optimizer.clear_grad()
paddle.save(
model.state_dict(),
"{}/model_{}.pdparams".format(args.saved_path, args.game))
print("Episode: {}. Total loss: {:.4f}".format(curr_episode,
total_loss.numpy()[0]))
def forward(self, boxes, scores, gt_box, gt_class, anchors):
boxes = paddle.concat(boxes, axis=1)
scores = paddle.concat(scores, axis=1)
prior_boxes = paddle.concat(anchors, axis=0)
gt_label = gt_class.unsqueeze(-1)
batch_size, num_priors, num_classes = scores.shape
def _reshape_to_2d(x):
return paddle.flatten(x, start_axis=2)
# 1. Find matched bounding box by prior box.
# 1.1 Compute IOU similarity between ground-truth boxes and prior boxes.
# 1.2 Compute matched bounding box by bipartite matching algorithm.
matched_indices = []
matched_dist = []
for i in range(gt_box.shape[0]):
iou = iou_similarity(gt_box[i], prior_boxes)
matched_indice, matched_d = bipartite_match(
iou, self.match_type, self.overlap_threshold)
matched_indices.append(matched_indice)
matched_dist.append(matched_d)
matched_indices = paddle.concat(matched_indices, axis=0)
matched_indices.stop_gradient = True
matched_dist = paddle.concat(matched_dist, axis=0)
matched_dist.stop_gradient = True
# 2. Compute confidence for mining hard examples
# 2.1. Get the target label based on matched indices
target_label, _ = self._label_target_assign(gt_label, matched_indices)
confidence = _reshape_to_2d(scores)
# 2.2. Compute confidence loss.
# Reshape confidence to 2D tensor.
target_label = _reshape_to_2d(target_label).astype('int64')
conf_loss = F.softmax_with_cross_entropy(confidence, target_label)
conf_loss = paddle.reshape(conf_loss, [batch_size, num_priors])
# 3. Mining hard examples
neg_mask = self._mine_hard_example(conf_loss,
matched_indices,
matched_dist,
neg_pos_ratio=self.neg_pos_ratio,
neg_overlap=self.neg_overlap)
# 4. Assign classification and regression targets
# 4.1. Encoded bbox according to the prior boxes.
prior_box_var = paddle.to_tensor(
np.array([0.1, 0.1, 0.2, 0.2],
dtype='float32')).reshape([1, 4]).expand_as(prior_boxes)
encoded_bbox = []
for i in range(gt_box.shape[0]):
encoded_bbox.append(
box_coder(prior_box=prior_boxes,
prior_box_var=prior_box_var,
target_box=gt_box[i],
code_type='encode_center_size'))
encoded_bbox = paddle.stack(encoded_bbox, axis=0)
# 4.2. Assign regression targets
target_bbox, target_loc_weight = self._bbox_target_assign(
encoded_bbox, matched_indices)
# 4.3. Assign classification targets
target_label, target_conf_weight = self._label_target_assign(
gt_label, matched_indices, neg_mask=neg_mask)
# 5. Compute loss.
# 5.1 Compute confidence loss.
target_label = _reshape_to_2d(target_label).astype('int64')
conf_loss = F.softmax_with_cross_entropy(confidence, target_label)
target_conf_weight = _reshape_to_2d(target_conf_weight)
conf_loss = conf_loss * target_conf_weight * self.conf_loss_weight
# 5.2 Compute regression loss.
location = _reshape_to_2d(boxes)
target_bbox = _reshape_to_2d(target_bbox)
loc_loss = F.smooth_l1_loss(location, target_bbox, reduction='none')
loc_loss = paddle.sum(loc_loss, axis=-1, keepdim=True)
target_loc_weight = _reshape_to_2d(target_loc_weight)
loc_loss = loc_loss * target_loc_weight * self.loc_loss_weight
# 5.3 Compute overall weighted loss.
loss = conf_loss + loc_loss
loss = paddle.reshape(loss, [batch_size, num_priors])
loss = paddle.sum(loss, axis=1, keepdim=True)
normalizer = paddle.sum(target_loc_weight)
loss = paddle.sum(loss / normalizer)
return loss
def compute_res_loss(output, target):
# return F.smooth_l1_loss(output, target, reduction='elementwise_mean')
return F.smooth_l1_loss(output, target, reduction='mean')