Comments (4)
Generally, you should use the provided init models for initialization instead of the models trained on UCF101/HMDB51. The high gradient norm is usually due to increase iter_size
, which is fine. You can increase the value of clip_gradient
in your solver to fit it.
from temporal-segment-networks.
test_iter
is computed by total_sample_num
/ batch_size
/ iter_size
/ num_gpu
.
However, the error you list is not related to test_iter
. If you look at your log, you will find that it is related to reshape
layer. So you may have some shape mismatch instead of anything to do with test_iter
.
from temporal-segment-networks.
One thing to add. This on-the-fly validation is just for monitoring the training. All experimental results should be reported with the provide testing scripts.
from temporal-segment-networks.
Thanks for your reply!! Now I know how to set the parameter test_iter
.
About the above errors, I have find the missing point. In the layer
layer { name: "reshape_fc" type: "Reshape" bottom: "fc" top: "reshape_fc" reshape_param { shape { dim: [-1, 1, 3, 101] } } }
,
the last parameter is 101, which I didn't change the corresponding classes 2.
Now the fine-tuned TSN could work, but it does not work well when I use the ucf101_split_1_tsn_rgb_reference_bn_inception.caffemodel
as the pretrained model, and shows as follows:
I0225 09:38:46.088831 16686 solver.cpp:631] Iteration 0, lr = 0.0001
I0225 09:38:46.588022 16686 solver.cpp:616] Gradient clipping: scaling down gradients (L2 norm 100.16 > 40) by scale factor 0.399362
I0225 09:38:47.023706 16686 solver.cpp:616] Gradient clipping: scaling down gradients (L2 norm 80.797 > 40) by scale factor 0.495068
I0225 09:38:47.458351 16686 solver.cpp:616] Gradient clipping: scaling down gradients (L2 norm 46.5655 > 40) by scale factor 0.859006
I0225 09:38:47.894577 16686 solver.cpp:616] Gradient clipping: scaling down gradients (L2 norm 51.9353 > 40) by scale factor 0.770189
I0225 09:38:48.331284 16686 solver.cpp:616] Gradient clipping: scaling down gradients (L2 norm 75.4552 > 40) by scale factor 0.530116
I0225 09:38:48.769134 16686 solver.cpp:616] Gradient clipping: scaling down gradients (L2 norm 63.1562 > 40) by scale factor 0.63335
......
What's wrong with it? Should I reduce the base_lr
or some other parameters ?? Could you give me some advice?
Thank you !!
from temporal-segment-networks.
Related Issues (20)
- What should I do if I want to change the main structure of the network to InceptionV3?
- how to change nem_segments?
- error when building dense_flow HOT 3
- could you supply the kinetics/labels/val_videofolder.txt ?
- calcDenseFlowGPU params
- extract_optical_flow.sh with single GPU HOT 1
- how to run build_all.sh on windows_10 system
- No registered converter was able to extract a C++ pointer to type char from this Python object of type bytes HOT 1
- how to dump_frames ?
- STOA vs SOTA
- Use RWF-2000 to train TSN
- Extract Frames and Optical Flow Images
- Docker extract flow: Only a part of videoes could be done, and the others were empty folders.
- cross modality gist
- For the problem 'Failed to build Caffe. '
- Failed to build OpenCV. Please check the logs above.
- How to use HMDB_ 51 data training
- how to detect? HOT 1
- Let me be clear: the code is non-functional, and video classification is not achievable with it.
- Unable to find 'extract_gpu' in Dense Flow during optical flow extraction
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 temporal-segment-networks.