0.1.3 (#32)

* allow skipping model IDs in finalize scores

* allow subclassing of saver and score_computer directly from traker args

* default to BasicProjector if CudaProjector projeciton step errors out

* add another type of error that sometime occurs when fast_jl has issues

* update quickstart notebook

* Add link to colab with pre-computed trak scores to readme

* add dropbox links to quickstart nb

* update training code in quickstart tutorial

* bump version

---------

Co-authored-by: Joshua Vendrow

Author: kristian-georgiev

README.md

@@ -1,13 +1,12 @@
-[![PyPI version](https://badge.fury.io/py/traker.svg)](https://badge.fury.io/py/traker)
 [![arXiv](https://img.shields.io/badge/arXiv-2303.14186-b31b1b.svg?style=flat-square)](https://arxiv.org/abs/2303.14186)
+[![PyPI version](https://badge.fury.io/py/traker.svg)](https://badge.fury.io/py/traker)
+
+# TRAK: Attributing Model Behavior at Scale
 
 [[docs & tutorials]](https://trak.readthedocs.io/en/latest/)
-[[paper]](https://arxiv.org/abs/2303.14186)
 [[blog post]](https://gradientscience.org/trak/)
 [[website]](https://trak.csail.mit.edu)
 
-# TRAK: Attributing Model Behavior at Scale
-
 In our [paper](https://arxiv.org/abs/2303.14186), we introduce a new data attribution method called `TRAK` (Tracing with the
 Randomly-Projected After Kernel). Using `TRAK`, you can make  accurate
 counterfactual predictions (e.g., answers to questions of the form “what would
@@ -17,21 +16,10 @@ comparably effective methods, e.g., see our evaluation on:
 
 ![Main figure](/docs/assets/main_figure.png)
 
-## Citation
-If you use this code in your work, please cite using the following BibTeX entry:
-```
-@inproceedings{park2023trak,
-  title = {TRAK: Attributing Model Behavior at Scale},
-  author = {Sung Min Park and Kristian Georgiev and Andrew Ilyas and Guillaume Leclerc and Aleksander Madry},
-  booktitle = {Arxiv preprint arXiv:2303.14186},
-  year = {2023}
-}
-```
-
 ## Usage
 
-
-[[Quickstart]](https://trak.readthedocs.io/en/latest/quickstart.html)
+[[quickstart]](https://trak.readthedocs.io/en/latest/quickstart.html)
+[[pre-computed TRAK scores for CIFAR-10]](https://colab.research.google.com/drive/1Mlpzno97qpI3UC1jpOATXEHPD-lzn9Wg?usp=sharing)
 
 Check [our docs](https://trak.readthedocs.io/en/latest/) for more detailed examples and
 tutorials on how to use `TRAK`.  Below, we provide a brief blueprint of using `TRAK`'s API to compute attribution scores.
@@ -74,6 +62,17 @@ scores = traker.finalize_scores()
 ## Examples
 You can find several end-to-end examples in the `examples/` directory.
 
+## Citation
+If you use this code in your work, please cite using the following BibTeX entry:
+```
+@inproceedings{park2023trak,
+  title = {TRAK: Attributing Model Behavior at Scale},
+  author = {Sung Min Park and Kristian Georgiev and Andrew Ilyas and Guillaume Leclerc and Aleksander Madry},
+  booktitle = {Arxiv preprint arXiv:2303.14186},
+  year = {2023}
+}
+```
+
 ## Installation
 
 To install the version of our package which contains a fast, custom `CUDA`
@@ -93,9 +92,8 @@ pip install traker
 
 Please send an email to trak@mit.edu
 
-## Maintainers:
+## Maintainers
 
 [Kristian Georgiev](https://twitter.com/kris_georgiev1)<br>
 [Andrew Ilyas](https://twitter.com/andrew_ilyas)<br>
-[Guillaume Leclerc](https://twitter.com/gpoleclerc)<br>
 [Sung Min Park](https://twitter.com/smsampark)

docs/source/conf.py

@@ -22,8 +22,8 @@
 author = 'Kristian Georgiev'
 
 # The full version, including alpha/beta/rc tags
-release = '0.1.2'
-version = '0.1.2'
+release = '0.1.3'
+version = '0.1.3'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/quickstart.rst

@@ -92,21 +92,36 @@ classification task of your choice.)
         )
         return model
 
-    def get_dataloader(batch_size=256, num_workers=8, split='train'):
-
-        transforms = torchvision.transforms.Compose(
-                        [torchvision.transforms.RandomHorizontalFlip(),
-                        torchvision.transforms.RandomAffine(0),
-                        torchvision.transforms.ToTensor(),
-                        torchvision.transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.201))])
-
+    def get_dataloader(batch_size=256, num_workers=8, split='train', shuffle=False, augment=True):
+        if augment:
+            transforms = torchvision.transforms.Compose(
+                            [torchvision.transforms.RandomHorizontalFlip(),
+                             torchvision.transforms.RandomAffine(0),
+                             torchvision.transforms.ToTensor(),
+                             torchvision.transforms.Normalize((0.4914, 0.4822, 0.4465),
+                                                              (0.2023, 0.1994, 0.201))])
+        else:
+            transforms = torchvision.transforms.Compose([
+                             torchvision.transforms.ToTensor(),
+                             torchvision.transforms.Normalize((0.4914, 0.4822, 0.4465),
+                                                              (0.2023, 0.1994, 0.201))])
+            
         is_train = (split == 'train')
-        dataset = torchvision.datasets.CIFAR10(root='/tmp/cifar/', download=True, train=is_train, transform=transforms)
-        loader = torch.utils.data.DataLoader(dataset=dataset, shuffle=False, batch_size=batch_size, num_workers=num_workers)
-
+        dataset = torchvision.datasets.CIFAR10(root='/tmp/cifar/',
+                                               download=True,
+                                               train=is_train,
+                                               transform=transforms)
+
+        loader = torch.utils.data.DataLoader(dataset=dataset,
+                                             shuffle=shuffle,
+                                             batch_size=batch_size,
+                                             num_workers=num_workers)
+        
         return loader
 
-    def train(model, loader, lr=0.4, epochs=24, momentum=0.9, weight_decay=5e-4, lr_peak_epoch=5, label_smoothing=0.0):
+    def train(model, loader, lr=0.4, epochs=24, momentum=0.9,
+              weight_decay=5e-4, lr_peak_epoch=5, label_smoothing=0.0, model_id=0):
+        
         opt = SGD(model.parameters(), lr=lr, momentum=momentum, weight_decay=weight_decay)
         iters_per_epoch = len(loader)
         # Cyclic LR with single triangle
@@ -118,9 +133,8 @@ classification task of your choice.)
         loss_fn = CrossEntropyLoss(label_smoothing=label_smoothing)
 
         for ep in range(epochs):
-            model_count = 0
             for it, (ims, labs) in enumerate(loader):
-                ims = ims.float().cuda()
+                ims = ims.cuda()
                 labs = labs.cuda()
                 opt.zero_grad(set_to_none=True)
                 with autocast():
@@ -131,15 +145,19 @@ classification task of your choice.)
                 scaler.step(opt)
                 scaler.update()
                 scheduler.step()
+            if ep in [12, 15, 18, 21, 23]:
+                torch.save(model.state_dict(), f'./checkpoints/sd_{model_id}_epoch_{ep}.pt')
+            
+        return model
 
     os.makedirs('./checkpoints', exist_ok=True)
+    loader_for_training = get_dataloader(batch_size=512, split='train', shuffle=True)
 
-    for i in tqdm(range(3), desc='Training models..'):
+    # you can modify the for loop below to train more models
+    for i in tqdm(range(1), desc='Training models..'):
         model = construct_rn9().to(memory_format=torch.channels_last).cuda()
-        loader_train = get_dataloader(batch_size=512, split='train')
-        train(model, loader_train)
+        model = train(model, loader_for_training, model_id=i)
 
-        torch.save(model.state_dict(), f'./checkpoints/sd_{i}.pt')
 
 .. raw:: html
 
@@ -311,4 +329,4 @@ The final line above returns :code:`TRAK` scores as a :code:`numpy.array` from t
 
 That's it!
 Once you have your model(s) and your data, just a few API-calls to TRAK
-let's you compute data attribution scores.
+let's you compute data attribution scores.