caffe - a deep learning framework (ramin fahimi)
TRANSCRIPT
Caffe:Deep learningFramework
Ramin FahimiPyCon 2016 , IUST
Many contents has been taken from Caffe CVPR’15 tutorial and CS231n lectures, Stanford.
Caffe: Deep learning Framework Ramin Fahimi, #irpycon 2016
Took million years from nature to form effective visual recognition system.
It didn’t happened in one night!
Evolution.
Computer Vision1. Controlling processes: an industrial robot2. Navigation: an autonomous vehicle3. Detecting events: visual surveillance4. Organizing information: indexing databases of
images and image sequences5. Modeling objects or environments: medical
image analysis or topographical modeling6. Interaction: input to a device for computer-
human interaction7. Automatic inspection in manufacturing
applications.
Caffe: Deep learning Framework Ramin Fahimi, #irpycon 2016
What is deep learning? (DL)
Caffe: Deep learning Framework Ramin Fahimi, #irpycon 2016
Input image Weights Loss
The number of neurons in each layer is given by 253440, 186624, 64896, 64896, 43264, 4096, 4096, 1000
Why?What changed?
Caffe: Deep learning Framework Ramin Fahimi, #irpycon 2016
1. Improvements in hardware 2. Data Size3. Initialization4. Successfully applying back propagation5. Many other things
Use Cases
Caffe: Deep learning Framework Ramin Fahimi, #irpycon 2016
• Extract AlexNet or VGG features? Use Caffe• Fine-tune AlexNet for new classes? Use Caffe• Image Captioning with fine-tuning?
o Need pre-trained models (Caffe, Torch, Lasagne) o Need RNNs (Torch or Lasagne) o Use Torch or Lasagna
• Segmentation? (Classify every pixel) o Use Caffe If loss function exists in Caffe else Use Torch
• Object Detection? o Need pre-trained model (Torch, Caffe, Lasagne) o Need lots of custom imperative code (NOT Lasagne), Use Caffe or Torch
Use Cases – Cont.
Caffe: Deep learning Framework Ramin Fahimi, #irpycon 2016
• Feature extraction / fine-tuning existing models: Use Caffe • Complex uses of pre-trained models: Use Lasagne or Torch • Write your own layers: Use Torch • Crazy RNNs: Use Theano or Tensorflow • Huge model, need model parallelism: Use TensorFlow
Why Caffe?
Caffe: Deep learning Framework Ramin Fahimi, #irpycon 2016
§ Expression: models + optimizations are plaintext schemas, not code. § Speed: for state-of-the-art models and massive data. § Modularity: to extend to new tasks and architectures. § Openness: common code and reference models for reproducibility. § Community: joint discussion, development, and modeling
● Frameworks are more alike than different o All express deep models o All are nicely open-source o All include scripting for hacking and prototyping
● No strict winners – experiment and choose the framework that best fits your work
Open model collections
Caffe: Deep learning Framework Ramin Fahimi, #irpycon 2016
• The Caffe model zoo contains open collection of deep models
o VGG ILSVRC14 + Devil models in the zooo Network-in-Network / CCCP model in the zooo MIT Places scene recognition model in the zoo
• help disseminate and reproduce research• bundled tools for loading and publishing models • Share Your Models! with your citation + license of course