Skip to content

Reproducibility challenge project for a candidate paper of ICLR2018: Incremental Learning through Deep Adaption

Notifications You must be signed in to change notification settings

jianing-sun/CNN-Incremental-Learning

Repository files navigation

cnn_incremental-learning

Reproductibility Challenge for ICLR 2018: Incremental Learning through Deep Adaption

cnn_v3: Dataset CIFAR10 trained with control module based on a regular 2-layer convolutional neural network.
cnn_vgg: Dataset CIFAR10 trained with control module based on vgg net
Model: details of regular vgg net and vgg with control module

Based on pytorch

A project from ECSE608: Machine Learning, worked with @harveyyan

Due on Dec 21

12/09/2017: Finally apply the control module with the original module with no error.

Problem: the accuracy didn't change no matter how many iterations. It seems the gradient descent didn't work. Still need to find the reason. But at least it didn't have error... a little relieved.

Solved:change newmodel = copy.deepcopy(model) to newmodel = model

Problem: which one is better with CIFAR10? old one or the new one?

Problem: Pre-train!!

Partial results 8pm 10/12/2017
Iteration: 100. Loss: 2.1288864612579346. Accuracy: 31.85.
Iteration: 200. Loss: 1.761412501335144. Accuracy: 31.19.
Iteration: 300. Loss: 1.2932252883911133. Accuracy: 48.96.
Iteration: 400. Loss: 1.426600694656372. Accuracy: 50.98.
Iteration: 500. Loss: 1.5066254138946533. Accuracy: 52.86.
Iteration: 600. Loss: 1.0674883127212524. Accuracy: 55.4.
Iteration: 700. Loss: 1.2660120725631714. Accuracy: 57.12.
Iteration: 800. Loss: 1.0268568992614746. Accuracy: 58.46.
Iteration: 900. Loss: 0.932781457901001. Accuracy: 62.76.
Iteration: 1000. Loss: 0.8914451003074646. Accuracy: 63.41.
Iteration: 1100. Loss: 0.8505222797393799. Accuracy: 66.0.
Iteration: 1200. Loss: 1.0242605209350586. Accuracy: 64.98.
Iteration: 1300. Loss: 0.9178219437599182. Accuracy: 67.42.
Iteration: 1400. Loss: 0.8666901588439941. Accuracy: 68.61.
Iteration: 1500. Loss: 0.892946720123291. Accuracy: 68.54.
Iteration: 1600. Loss: 0.8638588786125183. Accuracy: 68.35.
Iteration: 1700. Loss: 0.8628944158554077. Accuracy: 69.15.
Iteration: 1800. Loss: 1.0401527881622314. Accuracy: 68.27.
Iteration: 1900. Loss: 0.7258371710777283. Accuracy: 70.49.
Iteration: 2000. Loss: 0.6357018947601318. Accuracy: 70.98.
Iteration: 2100. Loss: 0.44135037064552307. Accuracy: 72.99.
Iteration: 2200. Loss: 0.7241382598876953. Accuracy: 71.54.
Iteration: 2300. Loss: 0.5983885526657104. Accuracy: 71.21.
Iteration: 2400. Loss: 0.3832913935184479. Accuracy: 73.9.
Iteration: 2500. Loss: 0.5777073502540588. Accuracy: 72.51.

About

Reproducibility challenge project for a candidate paper of ICLR2018: Incremental Learning through Deep Adaption

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages