AncTe/ginka-generator
1
0
Fork 0
mirror of https://github.com/unanmed/ginka-generator.git synced 2026-05-14 04:41:12 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

chore: 调整损失值计算

Browse Source
This commit is contained in:
unanmed 2026-05-06 22:47:08 +08:00
parent f0025df1ec
commit 7535ecc9fe
3 changed files with 10 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ginka/train_pretrain_split.py
View File

@ -37,7 +37,7 @@ from .utils import masked_focal
BATCH_SIZE = 64 BATCH_SIZE = 64
NUM_CLASSES = 7 NUM_CLASSES = 7
MAP_SIZE = 13 * 13 MAP_SIZE = 13 * 13
FOCAL_GAMMA = 2.0 FOCAL_GAMMA = 1.0
# 通道 1：空间骨架（floor+wall） # 通道 1：空间骨架（floor+wall）
CH1_KEEP = {0, 1} # 编码器输入保留的 tile CH1_KEEP = {0, 1} # 编码器输入保留的 tile
@ -47,13 +47,13 @@ CH1_NHEAD = 8
# 通道 2：关卡门控 # 通道 2：关卡门控
CH2_KEEP = {0, 1, 2, 4, 5} CH2_KEEP = {0, 1, 2, 4, 5}
CH2_LOSS = {0, 1, 2, 4, 5} CH2_LOSS = {2, 4, 5}
CH2_D_MODEL = 64 CH2_D_MODEL = 64
CH2_NHEAD = 8 CH2_NHEAD = 8
# 通道 3：收集资源 # 通道 3：收集资源
CH3_KEEP = None # 完整地图，无需切片 CH3_KEEP = None # 完整地图，无需切片
CH3_LOSS = {0, 1, 2, 3, 4, 5} CH3_LOSS = {3}
CH3_D_MODEL = 64 CH3_D_MODEL = 64
CH3_NHEAD = 8 CH3_NHEAD = 8

4
ginka/utils.py
View File

@ -75,9 +75,9 @@ def masked_focal(
# count[0]，导致 weight[0] 趋近于 0、非专属位置损失被消除的问题 # count[0]，导致 weight[0] 趋近于 0、非专属位置损失被消除的问题
class_weight = None class_weight = None
if balance: if balance:
flat = target.view(-1) # [B*S] 原始标签 flat = corrected.view(-1) # [B*S] 原始标签
counts = torch.bincount(flat, minlength=C).float() # [C] counts = torch.bincount(flat, minlength=C).float() # [C]
class_weight = flat.numel() / (counts.clamp(min=1.0) * C) class_weight = torch.sqrt(flat.numel() / (counts.clamp(min=1.0) * C))
class_weight[counts == 0] = 0.0 # 未出现类别不参与 class_weight[counts == 0] = 0.0 # 未出现类别不参与
ce = F.cross_entropy( ce = F.cross_entropy(

10
train_full.sh
View File

@ -24,17 +24,17 @@ TRAIN_DATA="ginka-dataset.json"
EVAL_DATA="ginka-eval.json" EVAL_DATA="ginka-eval.json"
# 阶段 0：三通道分拆预训练 # 阶段 0：三通道分拆预训练
P0_EPOCHS=10 P0_EPOCHS=20
P0_CHECKPOINT=5 P0_CHECKPOINT=10
P0_FINAL="result/pretrain_split/split_final.pth" P0_FINAL="result/pretrain_split/split_final.pth"
# 阶段 1：冻结编码器热身 # 阶段 1：冻结编码器热身
P1_EPOCHS=10 P1_EPOCHS=30
P1_CHECKPOINT=5 P1_CHECKPOINT=10
P1_FINAL="result/joint/warmup_final.pth" P1_FINAL="result/joint/warmup_final.pth"
# 阶段 2：完整联合训练 # 阶段 2：完整联合训练
P2_EPOCHS=400 P2_EPOCHS=470
P2_CHECKPOINT=20 P2_CHECKPOINT=20
# 从哪个阶段开始（0 = 从头）；命令行 --skip N 可覆盖此值 # 从哪个阶段开始（0 = 从头）；命令行 --skip N 可覆盖此值