Human Protein Image Classification using PyTorch and fastai

Transcript

1. William Horton Human Protein Image Classification using PyTorch and fastai
2. Or, Beyond Dogs vs. Cats
3. https://github.com/fastai/fastai/blob/master/courses/dl1/lesson1.ipynb
4. Basic CNN architecture https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the- eli5-way-3bd2b1164a53
5. What happens when you start working with real(er) data?
6. What happens when Dogs vs. Cats doesn’t cut it anymore?
7. Who am I? • Backend Engineer on the Data Team at Compass (We’re hiring!) • Python, Spark, Kafka, Airflow
8. Deep Learning • fast.ai International Fellow (Spring 2018) • Open-source contributor (PyTorch and fastai) • Speaker, PyOhio 2018 -- “You Can Do Deep Learning” • Tweeting at @hortonhearsafoo
10. Structure of a Machine Learning Problem Inputs Models Loss Function Threshold- ing Outputs ? ?
11. What was different from Dogs vs. Cats?
12. Inputs Dogs vs. Cats - Regular (3-channel) JPGs - Even class distribution Human Protein Classification - 4 separate PNGs per sample - Wildly uneven class distribution
13. Outputs Dogs vs. Cats - Single (binary) class - Single label per image Human Protein Classification - Multi-class - Multi-label
14. Inputs
15. How do I even read this data? • 4 channels saved as separate PNGs • blue, green red, and yellow
16. Example https://www.kaggle.com/jschnab/exploring-the-human-protein-atlas-images
17. Two approaches • Custom open function • Preprocess into 4-channel PNG
18. Custom open function adapted from https://www.kaggle.com/iafoss/pretrained-resnet34-with-rgby-0-460-public-lb (fastai 0.7 version)
19. Using the custom open function then...
20. Trade-offs • Disk space • Precomputing vs. during training • Code complexity
21. "Extreme" augmentations Based on the way the data was gathered Rotation all the way up to 90 degrees Large random scaling (up to 1.4-1.5x)
22. Model
23. Problem Common pretrained models are designed for 3-channel inputs
24. How do we adapt the architecture while retaining the benefits of pretrained weights?
25. Possible solutions Why not create a new layer of the right shape? Why not create a new layer of the right shape, and copy over the weights for the first three layers? Why not create a new layer of the right shape, and copy over the weights for the first three layers, and re-use the first layer for the fourth layer as well?
26. Custom 4-Channel Resnet adapted from https://www.kaggle.com/iafoss/pretrained-resnet34-with-rgby-0-460-public-lb
27. What is the best architecture to use? Answer: try them all! My best submission was a combination of resnet18 + resnet50.
28. Loss function
29. What makes this data unique? Multi-class and multi-label Highly imbalanced classes
30. Multi-class and multi-label Multi-class: i.e. not binary Multi-label: each example could have multiple correct classes
31. Sigmoid and Softmax Sigmoid: "compresses" output to between 0 and 1 Softmax: "pushes" largest value closer to 1, other values closer to 0 Example: Output: [ 0.6590094 , -1.1441954 , 0.19337232, 1.57017275, -0.13396834] Sigmoid: [ 0.65903783, 0.24155091, 0.548193 , 0.82780823, 0.46655792] Softmax: [0.21131754, 0.03481879, 0.13265143, 0.52559203, 0.09562021]
32. Loss functions For Binary Classification (i.e. Dogs vs. Cats):
33. Loss functions For Multiclass, Single-label Classification (i.e. ImageNet):
34. What do we use for multi-class, multi-label data?
35. Loss functions It’s like we’re seeing each class as a binary decision
36. Class imbalance https://www.kaggle.com/jschnab/exploring-the-human-protein-atlas-images
37. Evaluation metric Macro F1 score Each class has the same impact on the final metric
38. Weighted Loss
39. Calculating weights Direct weights may be too extreme Log-dampened weights
40. Log-dampened weights https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/74065
41. From my training notebook
42. Focal loss From this object-detection paper (origin of RetinaNet): “The Focal Loss is designed to address the one-stage object detection scenario in which there is an extreme imbalance between foreground and background classes during training (e.g., 1:1000).”
43. Weighted Samplers
44. Standard Sampling: Weighted Random Sampling (maybe):
45. Thresholds
46. Calculating thresholds Model output for one example (after sigmoid): [0.9, 0.1, 0.2, 0.7, … C] How to decide which are 1s and which are 0s? Most naive: 0.5 Better: optimize threshold based on validation set Best: optimize separate threshold per class on validation set
47. Calculating thresholds all classes 0.5 optimize on validation set optimize per class on validation set
48. Calculating thresholds Optimization problem -- solve for C variables that minimize metric (in this case, F1) Combination of optimization methods L-BFGS-B and Basinhopping using sklearn (https://github.com/mratsim/Amazon-Forest-Computer-Vision)
49. (Bonus: Ensembling)
50. Ensembling with averaging Let’s say we have the outputs of 3 models for 1 example, with 5 possible classes: [0.8, 0.1, 0.3, 0.9, 0.5] [0.9, 0.3, 0.6, 0.1, 0.7] [0.9, 0.2, 0.4, 0.9, 0.7] Our output would be the average (then we’d apply thresholding post-averaging): [0.86666667, 0.2 , 0.43333333, 0.63333333, 0.63333333]
51. Ensembling with max [0.8, 0.1, 0.3, 0.9, 0.5] [0.9, 0.3, 0.6, 0.1, 0.7] [0.9, 0.2, 0.4, 0.9, 0.7] Our output would be the max: [0.9, 0.3, 0.6, 0.9, 0.7]
52. Ensembling with OR Let’s say we’ve already thresholded to 0s and 1s: [1, 0, 0, 1, 0] [1, 0, 1, 0, 1] [1, 0, 0, 1, 1] Our output would be the logical OR (basically if any model predicted that class): [1, 0, 1, 1, 1]
53. Conclusion
54. Structure of a Machine Learning Problem Inputs Models Loss Function Threshold- ing Outputs
55. Work with datasets that force you to explore the capabilities of your tools.
56. Further information My starter code: https://github.com/wdhorton/protein-atlas-fastai Kernel version: https://www.kaggle.com/hortonhearsafoo/fastai-v1-starter-pack-kernel-edition-lb-0-323 Kaggle discussion: https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/71039
57. References Kaggle is an amazing community! I tried to collect Kernels and Discussion threads that helped me in this competition (some already referenced in slides): https://www.kaggle.com/iafoss/pretrained-resnet34-with-rgby-0-460-public-lb (fastai v0.7) https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691 https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/73938 https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/74065 https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/69984#4 36319
58. Q&A (@hortonhearsafoo for anything that doesn’t get answered!)