训练出错，使用configs/dcn/cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms配置文件训练跑不起来 #74
Comments
|
看着是reshape的时候指定多个维度是-1没法推断具体shape，是否修改了这块相关代码？ |
|
配置文件和训练代码都没有修改，数据集采用coco |
|
你使用的paddle版本是多少呢，我本地在1.6.1上跑是能跑通的 |
|
您好，我的用的paddlepaddle-gpu 1.6.1.post107 |
|
我也是提示reshape这个问题。。。用的是paddlepaddle-gpu 1.6.2.post97 |
|
这是因为新的reader与之前reshape op的写法不兼容造成的（涉及到编译时shape），master分支已经修复了这个问题，现在使用master分支代码或者release0.1分支代码应该都是可以正常运行的~ @hoodpan @dragon515 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
具体的复现信息如下：
/home/user/anaconda3/envs/mmlab/bin/python /home/user/PaddleDetection-release-0.1/tools/train.py -c ../configs/dcn/cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms.yml --use_tb=True --tb_log_dir=../tb_1206_1452_cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms/scalar --eval
CascadeBBoxAssigner:
[32mclass_aware[0m: true
batch_size_per_im: 512
bbox_reg_weights:
bg_thresh_hi:
bg_thresh_lo:
fg_fraction: 0.25
fg_thresh:
num_classes: 81
shuffle_before_sample: true
CascadeBBoxHead:
[32mhead[0m: CascadeTwoFCHead
[32mnms[0m: MultiClassSoftNMS
num_classes: 81
CascadeRCNNClsAware:
[32mbackbone[0m: ResNet
[32mrpn_head[0m: FPNRPNHead
bbox_assigner: CascadeBBoxAssigner
bbox_head: CascadeBBoxHead
fpn: FPN
roi_extractor: FPNRoIAlign
CascadeTwoFCHead:
[32mmlp_dim[0m: 1024
FPN:
freeze_norm: false
has_extra_convs: false
max_level: 6
min_level: 2
norm_type: null
num_chan: 256
spatial_scale:
FPNRPNHead:
[32mrpn_target_assign[0m:
rpn_batch_size_per_im: 256
rpn_fg_fraction: 0.5
rpn_negative_overlap: 0.3
rpn_positive_overlap: 0.7
rpn_straddle_thresh: 0.0
[32mtest_proposal[0m:
min_size: 0.0
nms_thresh: 0.7
post_nms_top_n: 1000
pre_nms_top_n: 1000
[32mtrain_proposal[0m:
min_size: 0.0
nms_thresh: 0.7
post_nms_top_n: 2000
pre_nms_top_n: 2000
anchor_generator:
anchor_sizes:
aspect_ratios:
stride:
variance:
anchor_start_size: 32
max_level: 6
min_level: 2
num_chan: 256
num_classes: 1
FPNRoIAlign:
[32mbox_resolution[0m: 14
[32msampling_ratio[0m: 2
canconical_level: 4
canonical_size: 224
mask_resolution: 14
max_level: 5
min_level: 2
FasterRCNNEvalFeed:
[32mbatch_transforms[0m:
pad_to_stride: 32
[32mdataset[0m:
annotation: annotations/instances_val2017.json
dataset_dir: dataset/coco
image_dir: val2017
[32msample_transforms[0m:
to_rgb: true
with_mixup: false
is_channel_first: false
is_scale: true
mean:
std:
interp: 1
max_size: 2000
target_size:
use_cv2: true
channel_first: true
to_bgr: false
batch_size: 1
drop_last: false
enable_aug_flip: false
enable_multiscale: false
fields:
image_shape:
num_scale: 1
num_workers: 2
samples: -1
shuffle: false
use_padded_im_info: true
FasterRCNNTestFeed:
[32mbatch_transforms[0m:
pad_to_stride: 32
[32mdataset[0m:
annotation: dataset/coco/annotations/instances_val2017.json
batch_size: 1
drop_last: false
fields:
image_shape:
num_workers: 2
sample_transforms:
to_rgb: true
with_mixup: false
is_channel_first: false
is_scale: true
mean:
std:
interp: 1
max_size: 1333
target_size: 800
use_cv2: true
channel_first: true
to_bgr: false
samples: -1
shuffle: false
use_padded_im_info: true
FasterRCNNTrainFeed:
[32mbatch_transforms[0m:
pad_to_stride: 32
[32mdataset[0m:
annotation: annotations/instances_train2017.json
dataset_dir: dataset/coco
image_dir: train2017
[32msample_transforms[0m:
to_rgb: true
with_mixup: false
is_mask_flip: false
is_normalized: false
prob: 0.5
is_channel_first: false
is_scale: true
mean:
std:
interp: 1
max_size: 1800
target_size:
use_cv2: true
channel_first: true
to_bgr: false
batch_size: 1
bufsize: 10
class_aware_sampling: false
drop_last: false
fields:
image_shape:
memsize: null
num_workers: 2
samples: -1
shuffle: true
use_process: false
LearningRate:
[32mschedulers[0m:
gamma: 0.1
milestones:
values: null
start_factor: 0.1
steps: 1000
base_lr: 0.01
MultiClassSoftNMS:
background_label: 0
keep_top_k: 300
normalized: false
score_threshold: 0.01
softnms_sigma: 0.5
OptimizerBuilder:
optimizer:
momentum: 0.9
type: Momentum
regularizer:
factor: 0.0001
type: L2
ResNet:
[32mdcn_v2_stages[0m:
[32mdepth[0m: 200
[32mnonlocal_stages[0m:
[32mnorm_type[0m: bn
[32mvariant[0m: d
feature_maps:
freeze_at: 2
freeze_norm: true
norm_decay: 0.0
weight_prefix_name: ''
architecture: CascadeRCNNClsAware
eval_feed: FasterRCNNEvalFeed
log_smooth_window: 20
max_iters: 460000
metric: COCO
num_classes: 81
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet200_vd_pretrained.tar
save_dir: output
snapshot_iter: 10000
test_feed: FasterRCNNTestFeed
train_feed: FasterRCNNTrainFeed
use_gpu: true
weights: output/cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms/model_final
Traceback (most recent call last):
File "/home/user/PaddleDetection-release-0.1/tools/train.py", line 341, in
main()
File "/home/user/PaddleDetection-release-0.1/tools/train.py", line 129, in main
train_fetches = model.train(feed_vars)
File "/home/user/PaddleDetection-release-0.1/ppdet/modeling/architectures/cascade_rcnn_cls_aware.py", line 178, in train
return self.build(feed_vars, 'train')
File "/home/user/PaddleDetection-release-0.1/ppdet/modeling/architectures/cascade_rcnn_cls_aware.py", line 88, in build
body_feats = self.backbone(im)
File "/home/user/PaddleDetection-release-0.1/ppdet/modeling/backbones/resnet.py", line 432, in call
res = self.layer_warp(res, i)
File "/home/user/PaddleDetection-release-0.1/ppdet/modeling/backbones/resnet.py", line 383, in layer_warp
nonlocal_name + '_{}'.format(i), int(dim_in / 2) )
File "/home/user/PaddleDetection-release-0.1/ppdet/modeling/backbones/nonlocal_helper.py", line 152, in add_space_nonlocal
conv = space_nonlocal(input, dim_in, dim_out, prefix, dim_inner)
File "/home/user/PaddleDetection-release-0.1/ppdet/modeling/backbones/nonlocal_helper.py", line 101, in space_nonlocal
t_re = fluid.layers.reshape(t, shape=list(theta_shape), actual_shape=theta_shape_op )
File "/home/user/anaconda3/envs/paddle/lib/python3.7/site-packages/paddle/fluid/layers/nn.py", line 8976, in reshape
attrs["shape"] = get_attr_shape(shape)
File "/home/user/anaconda3/envs/paddle/lib/python3.7/site-packages/paddle/fluid/layers/nn.py", line 8949, in get_attr_shape
"be -1. But received shape[%d] is also -1." % dim_idx)
AssertionError: Only one dimension value of 'shape' in reshape can be -1. But received shape[2] is also -1.
Process finished with exit code 1
The text was updated successfully, but these errors were encountered: