Comments (7)
Freddie, also the create_patches = false in the config natually increases the memory usage.
With create_patches = true; I can run batch-size 32 with min_l and n_views = 16, which matches your memory usage predictions table. So the below works on our tiny 16GB GPUs:
"training": {
"num_epochs": 400,
"batch_size": 32,
"min_L": 16,
"n_views": 16,
"n_workers": 4,
"crop": 3,
"lr": 0.0007,
"lr_step": 2,
"lr_decay": 0.97,
"load_lr_maps": false,
"beta": 50.0,
"create_patches": true,
"patch_size": 64,
"val_proportion": 0.10,
"lambda": 0.000001
}
}
from highres-net.
That's right.
-
min_L
acts as a lower-bound on the views always present. If the input has views fewer thanmin_L
then it will be padded tomin_L
with dummy views (0
-padding). -
n_views
: the number of LR views randomly* selected from the LR pool of a site.
For instance, what you want to avoid is n_views = 16
and min_L = 32
as that would always 0
-pad your views to 32.
* views with larger clearance are more likely to be selected
If I were you, I'd try min_L = n_views = 16
first, then min_L = n_views = 8
.
from highres-net.
Have you tried reducing the number of views?
Have a look at the table in the README
Increased batch size will buy you more performance, whereas 32 views has diminished returns compared to 16.
from highres-net.
Yes, already tried that :
"training": {
"num_epochs": 400,
"batch_size": 16,
"min_L": 32,
"n_views": 4,
"n_workers": 4,
"crop": 3,
"lr": 0.0007,
"lr_step": 2,
"lr_decay": 0.97,
"load_lr_maps": false,
"beta": 50.0,
"create_patches": false,
"patch_size": 64,
"val_proportion": 0.10,
"lambda": 0.000001
}
gives
root@8b530a18a524:/data# python src/train.py --config config/config.json
0%| | 0/400 [00:00<?, ?it/s]
Traceback (most recent call last): | 0/66 [00:00<?, ?it/s]
File "src/train.py", line 307, in <module>
main(config)
File "src/train.py", line 293, in main
trainAndGetBestModel(fusion_model, regis_model, optimizer, dataloaders, baseline_cpsnrs, config)
File "src/train.py", line 174, in trainAndGetBestModel
srs = fusion_model(lrs, alphas) # fuse multi frames (B, 1, 3*W, 3*H)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 493, in __call__
result = self.forward(*input, **kwargs)
File "/data/src/DeepNetworks/HRNet.py", line 207, in forward
layer1 = self.encode(stacked_input) # encode input tensor
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 493, in __call__
result = self.forward(*input, **kwargs)
File "/data/src/DeepNetworks/HRNet.py", line 74, in forward
x = self.res_layers(x)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 493, in __call__
result = self.forward(*input, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/container.py", line 92, in forward
input = module(input)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 493, in __call__
result = self.forward(*input, **kwargs)
File "/data/src/DeepNetworks/HRNet.py", line 34, in forward
residual = self.block(x)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 493, in __call__
result = self.forward(*input, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/container.py", line 92, in forward
input = module(input)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 493, in __call__
result = self.forward(*input, **kwargs)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/activation.py", line 961, in forward
return F.prelu(input, self.weight)
File "/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py", line 1121, in prelu
return torch.prelu(input, weight)
RuntimeError: CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 15.75 GiB total capacity; 14.27 GiB already allocated; 433.56 MiB free; 1.16 MiB cached)
while GPU mem was flushed just before the experiment
from highres-net.
Would you also decrease min_L (minimum number of views) to match n_views? On a 16GB P100 I can just about squeeze into memory: batch=8, min_L=16 n_views=16 (but not if min_L=32)
from highres-net.
Indeed. That flag flew under our radar!
from highres-net.
GG @ilkarman
from highres-net.
Related Issues (10)
- Shared representation in fusion block
- Expected 2D (unbatched) or 3D (batched) input to conv1d, but got input of size: [1, 4, 202, 202] HOT 6
- Calculation of val_score
- Lanczos code crashes HOT 3
- torch_mask 192 vs 384 HOT 3
- Pretrain-model HOT 1
- training problem because of the different sizes HOT 8
- ShiftNet loss HOT 1
- predic HOT 5
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from highres-net.