Size don't match when augument

[Holmes2002]

Hi, when I get ['pre'] from model and augument them.
RuntimeError: The size of tensor a (57) must match the size of tensor b (224) at non-singleton dimension 1
Dimension of out don't match with real images
Why ? Error come from decode or encoder of model ?

[Holmes2002]

It come from encoder. I print out and get shape 57 57. Shoud I adjust somethings in encoder or decoder to get suitable size ?

[Haochen-Wang409]

Sorry, I cannot fully understand what you have encountered. What does ['pre'] exactly mean?

[Holmes2002]

It is outs ['pred'] of yours

[Holmes2002]

u use interpolate to enlarge outs["pred"] will match with shape of inputs but when augument you didn't use it. Is I miss some things when I use u code
Because I just have one gpu so I have to recode because syn_bn can't use in one gpu

I hope you explain for me !

[Haochen-Wang409]

Supervised loss is computed based on pred_l_large here.
Unsupervised loss is computed based on pred_u_large here.

large indicates the predictions are upsampled to HxW.

[Holmes2002]

Tks for reply sir !

[Holmes2002]

1.How many used images for extract unreliable and for contrastive loss ?
May be all for unreliable and 50 for contrastive(negative in your code) ?
2. Between Low rank and high rank are classes we are concerned ?
Hope u answer for me !

[Haochen-Wang409]

1. All features of unreliable pixels (served as negative samples) are stored in a memory bank, and thus there is no exact number. For contrastive loss, we select queries in the current mini-batch, which means they come from batch_size images.
2. For i-th unlabeled image, a qualified negative sample for class c should: (a) be unreliable; (b) probably not belongs to class c; (c) not belongs to most unlikely classes. We adopt low_rank and high_rank to achieve (b) and (c), respectively.

[Holmes2002]

U said that All features of Unreliable pixel were stored in a memory bank . But I search all find don't see u and them to list memory bank. Do u get from dist ?
I run the program but contras lost rezo. I find that my memory bank is still the same as the original ( all are (0,256))
Because I can run multi gps. Can u give me an advice ?

[Haochen-Wang409]

Provide your script and log, please.

[Holmes2002]

'''
dataset: # Required.
type: customs_semi
train:
data_root: /home/congvu/Semi_Segment/U2PL/data/customs
data_list: labeled.txt
data_LIP : LIPs.txt
flip: True
GaussianBlur: False
rand_resize: [0.5, 2.0]
#rand_rotation: [-10.0, 10.0]
crop:
type: rand
size: [224, 224] # crop image with HxW size
val:
data_root: /home/congvu/Semi_Segment/U2PL/data/customs
data_list: val.txt
crop:
type: center
size: [224, 224] # crop image with HxW size
batch_size: 4
n_sup: 59549
noise_std: 0.1
workers: 2
mean: [123.675, 116.28, 103.53]
std: [58.395, 57.12, 57.375]
ignore_label: 255

trainer: # Required.
epochs: 200
eval_on: True
sup_only_epoch: 0
optimizer:
type: SGD
kwargs:
lr: 0.01 # 8GPUs
momentum: 0.9
weight_decay: 0.0005
lr_scheduler:
mode: poly
kwargs:
power: 0.9
unsupervised:
TTA: False
drop_percent: 80
apply_aug: cutmix
contrastive:
negative_high_entropy: True
low_rank: 1
high_rank: 9
current_class_threshold: 0.3
current_class_negative_threshold: 1
unsupervised_entropy_ignore: 80
low_entropy_threshold: 20
num_negatives: 200
num_queries: 256
temperature: 0.5

saver:
auto_resume: True
snapshot_dir: checkpoints
pretrain: ''

criterion:
type: ohem
kwargs:
thresh: 0.7
min_kept: 100000

net: # Required.
num_classes: 9
sync_bn: False
ema_decay: 0.99
aux_loss:
aux_plane: 1024
loss_weight: 0.4
encoder:
type: model.resnet.resnet101
kwargs:
multi_grid: True
zero_init_residual: True
fpn: True
replace_stride_with_dilation: [False, True, True] #layer0...1 is fixed, layer2...4
decoder:
type: model.decoder.dec_deeplabv3_plus
kwargs:
inner_planes: 256
dilations: [12, 24, 36] '''
Is it ?

[Holmes2002]

what are results of dequeue_and_queue and gather_together, sir ?

[Haochen-Wang409]

It seems you run our code on your own dataset. I am not sure there are any issues in the dataset.
Maybe you could try our code on PASCAL VOC 2012 first.

[Holmes2002]

I tried with my own and succesed about 6 months ago. I want to know the memorybank and dequue_andqueue of u

[Haochen-Wang409]

Please refer to PyTorch documentation for understanding dequeue_and_enqueue.

[Holmes2002]

is high_valid_pixel_seg I put to memorybank same :
'''
# positive sample: center of the class
seg_proto_list.append(
torch.mean(
rep_teacher[low_valid_pixel_seg.bool()].detach(), dim=0, keepdim=True
)
)
'''
can use for negative_feat ?

[Haochen-Wang409]

high_valid_pixel_seg is a binary mask to filter unreliable pixels (i.e., negative samples), while low_valid_pixel_seg is to filter reliable pixels (i.e., positive samples).

[Holmes2002]

how can I extract negative samples sir ?

[Haochen-Wang409]

https://github.com/Haochen-Wang409/U2PL/blob/main/u2pl/utils/loss_helper.py#L190
Please check the code compute_contra_memobank_loss carefully.

[Holmes2002]

Not use contrastive loss will significantly affect to training model ?

[Haochen-Wang409]

Without contrastive loss, our method is equivalent to CutMix, which is expected to cause performance degradation.
I am not sure what significantly means.