Skip to main content

Dense and fastfood transform wrappers to reproduce "Intrinsic dimensionality of objective landscapes" by Li et al. (2018)

Project description

About

This package includes fastfood and dense transformation wrappers for pytorch modules, primarily to reproduce results from Li, Chunyuan, et al. "Measuring the intrinsic dimension of objective landscapes." arXiv preprint arXiv:1804.08838 (2018) - see below for info.

  • All contributions are welcome! Please raise an issue for a bug, feature or pull request!

  • Give this repo a star! :star:

Install

pip install intrinsic-dimensionality

Quick start on your classification task!

import os
os.environ["CUDA_VISIBLE_DEVICES"] = DEVICE_NUM
import torch
from torch import nn
import torchvision.models as models
from intrinsic import FastFoodWrap

class Classifier(nn.Module):
    def __init__(self, input_dim, n_classes):
        super(Classifier, self).__init__()
        self.fc = nn.Linear(input_dim, n_classes)
        self.maxpool = nn.AdaptiveMaxPool2d(1)

    def forward(self, x):
        x = self.maxpool(x)
        x = x.reshape(x.size(0), -1)
        x = self.fc(x)
        return x

def get_resnet(encoder_name, num_classes, pretrained=False):
    assert encoder_name in ["resnet18", "resnet50"], "{} is a wrong encoder name!".format(encoder_name)
    if encoder_name == "resnet18":
        model = models.resnet18(pretrained=pretrained)
        latent_dim = 512
    else:
        model = models.resnet50(pretrained=pretrained)
        latent_dim = 2048
    children = (list(model.children())[:-2] + [Classifier(latent_dim, num_classes)])
    model = torch.nn.Sequential(*children)
    return model

# Get model and wrap it in fastfood
model = get_resnet("resnet18", num_classes=YOUR_NUMBER_OF_CLASSES).cuda()
model = FastFoodWrap(model, intrinsic_dimension=100, device=DEVICE_NUM)

Reproducing experiments from the paper

Full thread about reproducibility results is available here. Note that some hyper-parameters were not listed in the paper - I raised issues on Uber's Github repo here.

I am able to reproduce their MNIST results with LR=0.0003, batch size 32 for both dense and fastfood transformations using FCN (fcn-dense, fcn-fastfood). However, not for LeNet (cnn-dense, cnn-fastfood).

For CIFAR-10, with far larger resnet (Resnet-18 11mil param) vs 280k 20-layer resnet used in the paper, results appear to be similar. FCN results in appendix (Fig S7) suggest some variation is to be expected.

Cite

@misc{jgamper2020intrinsic,
  title   = "Intrinsic-dimensionality Pytorch",
  author  = "Gamper, Jevgenij",
  year    = "2020",
  url     = "https://github.com/jgamper/intrinsic-dimensionality"
}

@article{li2018measuring,
  title={Measuring the intrinsic dimension of objective landscapes},
  author={Li, Chunyuan and Farkhoor, Heerad and Liu, Rosanne and Yosinski, Jason},
  journal={arXiv preprint arXiv:1804.08838},
  year={2018}
}

Project details


Release history Release notifications | RSS feed

This version

0.1

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

intrinsic-dimensionality-0.1.tar.gz (3.2 kB view hashes)

Uploaded Source

Built Distribution

intrinsic_dimensionality-0.1-py3-none-any.whl (4.0 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page