RuntimeError: The size of tensor a (256) must match the size of tensor b (255) at non-singleton dimension 3

[HuXinzhi1004]

Hello! Thanks for your great job.
When I run

time CUDA_VISIBLE_DEVICES=0,4 python3 -m butterflydetector.train --lr=1e-3 --momentum=0.95 --epochs=150 --lr-decay 120 140 --batch-size=16 --basenet=hrnetw32det --head-quad=1 --headnets butterfly10 --square-edge=512 --lambdas 1 1 1 1 --dataset uavdt

I got the error.

(cluster) hxz@ubuntu16:/home/data/hxz/butterflydetector$ time CUDA_VISIBLE_DEVICES=0,4 python3 -m butterflydetector.train --lr=1e-3 --momentum=0.95 --epochs=150 --lr-decay 120 140 --batch-size=16 --basenet=hrnetw32det --head-quad=1 --headnets butterfly10 --square-edge=512 --lambdas 1 1 1 1 --dataset uavdt
INFO:butterflydetector.logs:{'type': 'process', 'argv': ['/home/data/hxz/butterflydetector/butterflydetector/train.py', '--lr=1e-3', '--momentum=0.95', '--epochs=150', '--lr-decay', '120', '140', '--batch-size=16', '--basenet=hrnetw32det', '--head-quad=1', '--headnets', 'butterfly10', '--square-edge=512', '--lambdas', '1', '1', '1', '1', '--dataset', 'uavdt'], 'args': {'debug': False, 'checkpoint': None, 'basenet': 'hrnetw32det', 'headnets': ['butterfly10'], 'pretrained': True, 'cross_talk': 0.0, 'head_dropout': 0.0, 'head_quad': 1, 'lambdas': [1.0, 1.0, 1.0, 1.0], 'r_smooth': 0.0, 'regression_loss': 'laplace', 'background_weight': 1.0, 'margin_loss': False, 'auto_tune_mtl': False, 'butterfly_side_length': 1, 'momentum': 0.95, 'beta2': 0.999, 'adam_eps': 1e-06, 'nesterov': True, 'weight_decay': 0.0, 'adam': False, 'amsgrad': False, 'lr': 0.001, 'lr_decay': [120, 140], 'lr_burn_in_epochs': 2, 'lr_burn_in_factor': 0.001, 'lr_gamma': 0.1, 'dataset': 'uavdt', 'train_annotations': None, 'train_image_dir': None, 'val_annotations': None, 'val_image_dir': None, 'pre_n_images': 8000, 'n_images': None, 'duplicate_data': None, 'pre_duplicate_data': None, 'loader_workers': 2, 'batch_size': 16, 'output': 'outputs/hrnetw32det-butterfly10-edge512-211210-152433.pkl', 'stride_apply': 1, 'epochs': 150, 'freeze_base': 0, 'pre_lr': 0.0001, 'rescale_images': 1.0, 'orientation_invariant': False, 'update_batchnorm_runningstatistics': False, 'square_edge': 512, 'ema': 0.001, 'disable_cuda': False, 'augmentation': True, 'debug_fields_indices': [], 'profile': None, 'device': device(type='cuda'), 'pin_memory': True}, 'version': '0.0.1', 'hostname': 'ubuntu16'}
INFO:butterflydetector.network.hrnet:=> init weights from normal distribution
INFO:butterflydetector.network.hrnet:=> loading pretrained model pretrained/imagenet/hrnet_w32-36af842e.pth
INFO:butterflydetector.network.basenetworks:stride = 4
INFO:butterflydetector.network.basenetworks:output features = 512
INFO:butterflydetector.network.heads:selected head CompositeField for butterfly10
Using multiple GPUs: 2
INFO:butterflydetector.network.losses:multihead loss: ['butterfly10.c', 'butterfly10.vec1', 'butterfly10.scales1', 'butterfly10.scales2'], [1.0, 1.0, 1.0, 1.0]
/home/data/hxz/butterflydetector/butterflydetector/data_manager/uavdt.py:66: VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray.
self.targets = np.asarray(self.targets)
/home/data/hxz/butterflydetector/butterflydetector/data_manager/uavdt.py:67: VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray.
self.targets_ignore = np.asarray(self.targets_ignore)
Images: 40409
Images: 16580
Images: 8000
INFO:butterflydetector.optimize:SGD optimizer
INFO:butterflydetector.network.trainer:{'type': 'config', 'field_names': ['butterfly10.c', 'butterfly10.vec1', 'butterfly10.scales1', 'butterfly10.scales2']}
/home/hxz/anaconda3/envs/cluster/lib/python3.7/site-packages/torch/nn/functional.py:3635: UserWarning: Default upsampling behavior when mode=bilinear is changed to align_corners=False since 0.4.0. Please specify align_corners=True if the old behavior is desired. See the documentation of nn.Upsample for details.
"See the documentation of nn.Upsample for details.".format(mode)
Traceback (most recent call last):
File "/home/hxz/anaconda3/envs/cluster/lib/python3.7/runpy.py", line 193, in _run_module_as_main
"main", mod_spec)
File "/home/hxz/anaconda3/envs/cluster/lib/python3.7/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/home/data/hxz/butterflydetector/butterflydetector/train.py", line 200, in
main()
File "/home/data/hxz/butterflydetector/butterflydetector/train.py", line 196, in main
trainer.loop(train_loader, val_loader, args.epochs, start_epoch=start_epoch)
File "/home/data/hxz/butterflydetector/butterflydetector/network/trainer.py", line 99, in loop
self.train(train_scenes, epoch)
File "/home/data/hxz/butterflydetector/butterflydetector/network/trainer.py", line 173, in train
loss, head_losses = self.train_batch(data, target, meta, apply_gradients)
File "/home/data/hxz/butterflydetector/butterflydetector/network/trainer.py", line 116, in train_batch
loss, head_losses = self.loss(outputs, targets)
File "/home/hxz/anaconda3/envs/cluster/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/home/data/hxz/butterflydetector/butterflydetector/network/losses.py", line 176, in forward
for l, f, t in zip(self.losses, head_fields, head_targets)
File "/home/data/hxz/butterflydetector/butterflydetector/network/losses.py", line 177, in
for ll in l(f, t)]
File "/home/hxz/anaconda3/envs/cluster/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
return forward_call(*input, **kwargs)
File "/home/data/hxz/butterflydetector/butterflydetector/network/losses.py", line 457, in forward
) / 100.0 / batch_size
RuntimeError: The size of tensor a (256) must match the size of tensor b (255) at non-singleton dimension 3

How to fix it ?

[TonyzBi]

@HuXinzhi1004 I got the same error with dataset Viddrone2019, Did you fix it ?

[HuXinzhi1004]

just change the image input size, but I don’t know if it’s the right thing to do 发自我的iPhone

…

------------------ Original ------------------ From: TonyzBi ***@***.***> Date: Wed,Dec 22,2021 10:39 AM To: vita-epfl/butterflydetector ***@***.***> Cc: HuXinzhi1004 ***@***.***>, Mention ***@***.***> Subject: Re: [vita-epfl/butterflydetector] RuntimeError: The size of tensor a (256) must match the size of tensor b (255) at non-singleton dimension 3 (Issue #5) @HuXinzhi1004 I got the the same error with dataset of Visdrone2019, Did you fixed it ? — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you were mentioned.Message ID: ***@***.***>

[TonyzBi]

@HuXinzhi1004 Thanks, I have tried other size of input, Just report other size error, it seems no matching the target size when running loss.

[TonyzBi]

@HuXinzhi1004 I change the size of output feature by Heads module, It worked, But I am not sure if this modification has change the purpose of the author, I don't know if it’s the right thing to do too.

+++ b/butterflydetector/network/heads.py
@@ -190,15 +190,26 @@ class CompositeField(torch.nn.Module):
         scales_x = [torch.nn.functional.relu(scale_x) for scale_x in scales_x]

         # upscale
+        # for _ in range(self._quad):
+        #     classes_x = [self.dequad_op(class_x)[:, :, :-1, :-1]
+        #                  for class_x in classes_x]
+        #     regs_x = [self.dequad_op(reg_x)[:, :, :-1, :-1]
+        #               for reg_x in regs_x]
+        #     regs_x_spread = [self.dequad_op(reg_x_spread)[:, :, :-1, :-1]
+        #                      for reg_x_spread in regs_x_spread]
+        #     scales_x = [self.dequad_op(scale_x)[:, :, :-1, :-1]
+        #                 for scale_x in scales_x]
+
         for _ in range(self._quad):
-            classes_x = [self.dequad_op(class_x)[:, :, :-1, :-1]
+            classes_x = [self.dequad_op(class_x)
                          for class_x in classes_x]
-            regs_x = [self.dequad_op(reg_x)[:, :, :-1, :-1]
+            regs_x = [self.dequad_op(reg_x)
                       for reg_x in regs_x]
-            regs_x_spread = [self.dequad_op(reg_x_spread)[:, :, :-1, :-1]
+            regs_x_spread = [self.dequad_op(reg_x_spread)
                              for reg_x_spread in regs_x_spread]
-            scales_x = [self.dequad_op(scale_x)[:, :, :-1, :-1]
+            scales_x = [self.dequad_op(scale_x)
                         for scale_x in scales_x]
+
         # reshape regressions