Merge pull request #82 from lightly-ai/develop

Pre-release 1.0.7

.github/workflows/test.yml

@@ -0,0 +1,20 @@
+name: Unit Tests
+
+on: [push, pull_request]
+
+jobs:
+  test:
+    name: Test
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v2
+    - name: Set up Python 3.7
+      uses: actions/setup-python@v2
+      with:
+        python-version: 3.7
+    - name: Install Dependencies
+      run: pip install -e '.[all]'
+    - name: Run Pytest
+      run: python -m pytest -s -v

.pre-commit-config.yaml

@@ -7,9 +7,9 @@ repos:
     args: ['--maxkb=500']
 - repo: local
   hooks:
-  - id: tox                     # run all tests
-    name: Test with tox
-    entry: make tox
+  - id: pytest-check                     # run all tests
+    name: pytest-check
+    entry: make test
     language: system
     pass_filenames: false
     always_run: true

README.md

@@ -1,6 +1,8 @@
 
 ![Lightly Logo](docs/logos/lightly_logo_crop.png)
 
+![Unit Tests](https://github.com/lightly-ai/lightly/workflows/Unit%20Tests/badge.svg)
+
 Lightly is a computer vision framework for self-supervised learning.
 
 > We, at [Lightly](https://www.lightly.ai), are passionate engineers who want to make deep learning more efficient. We want to help popularize the use of self-supervised methods to understand and filter raw image data. Our solution can be applied before any data annotation step and the learned representations can be used to analyze and visualize datasets as well as for selecting a core set of samples.

docs/source/docker/advanced/meta_information.rst

@@ -89,7 +89,7 @@ additional samples that are enriching the existing selection.
           60 and want to sample 50, sampling would have no effect since there
           are already more than 50 samples selected.
 
-Custom Labels
+Custom Weak Labels
 -----------------------------------
 
 You can always add custom embeddings to the dataset by following the guide

docs/source/docker/examples/datasets_in_the_wild.rst

@@ -72,6 +72,36 @@ extract the frames with the following command:
     ffmpeg -i raw/video.mp4 -filter:v "fps=5" frames_ffmpeg/%d.png
 
 
+Extracting the frames without introducing compression artifacts is using lots of 
+storage. In this example, we have a small video of 6.4 MBytes. Once extracted, 
+the .png frames together with the video consume 553.4 MBytes. This is a 
+70x increase!
+
+.. list-table::
+   :widths: 50 50 50 30
+   :header-rows: 1
+
+   * - Metric
+     - ffmpeg extracted frames
+     - Lightly using video
+     - Reduction
+   * - Storage Consumption
+     - 447 MBytes + 6.4 MBytes
+     - 6.4 MBytes
+     - 70.84x
+
+.. note:: Why not extract the frames as compressed .jpg images? Extracting the 
+          frames as .jpg would indeed reduce storage consumption. The video from 
+          our example would end up using (14 MBytes + 6.4 MBytes). However, for 
+          critical applications where robustness and accuracy of the model are 
+          key, we have to think about the final system in production. Is your 
+          production system working with the raw camera stream (uncompressed) or 
+          with compressed frames (e.g. .jpg)? Very often we don’t have time to 
+          compress a frame in real-time systems or don’t want to introduce 
+          compression artifacts. You should also think about whether you want 
+          to train a model on compressed data whereas in production is runs 
+          using raw data.
+
 Now we want to do the same using Lightly Docker. Since the ffmpeg command
 extracted 99 frames let's extract 99 frames as well:
 
@@ -90,11 +120,9 @@ extracted 99 frames let's extract 99 frames as well:
 To perform a random selection we can simply replace "coreset" with "random" as
 our selected method. Note that if you don't specify any method coreset is used.
 
-To perform a random selection 
-
 Let's have a look at some statistics of the two obtained datasets:
 
-.. list-table:: video_dataset_statistics.csv
+.. list-table::
    :widths: 50 50 50 50 50
    :header-rows: 1
 

docs/source/getting_started/advanced.rst

@@ -1,3 +1,5 @@
+.. _lightly-advanced:
+
 Advanced
 ===================
 

docs/source/tutorials/package.rst

@@ -15,4 +15,5 @@ you might want to have a look at the two frameworks to understand basic concepts
    :maxdepth: 1
 
    structure_your_input.rst
-   package/tutorial_moco_memory_bank.rst
+   package/tutorial_moco_memory_bank.rst
+   package/tutorial_simclr_clothing.rst

docs/source/tutorials_source/package/tutorial_moco_memory_bank.py

@@ -54,12 +54,17 @@
 # 
 # We set some configuration parameters for our experiment.
 # Feel free to change them and analyze the effect.
+#
+# The default configuration uses a batch size of 512. This requires around 6.4GB
+# of GPU memory.
+# When training for 100 epochs you should achieve around 73% test set accuracy.
+# When training for 200 epochs accuracy increases to about 80%.
 
 num_workers = 8
-batch_size = 256
+batch_size = 512
 memory_bank_size = 4096
 seed = 1
-max_epochs = 50
+max_epochs = 100
 
 # %%
 # Replace the path with the location of your CIFAR-10 dataset.
@@ -243,8 +248,10 @@ def training_epoch_end(self, outputs):
 
 
     def configure_optimizers(self):
-        return torch.optim.SGD(self.resnet_moco.parameters(), lr=3e-2,
-                               momentum=0.9, weight_decay=5e-4)
+        optim = torch.optim.SGD(self.resnet_moco.parameters(), lr=6e-2,
+                                momentum=0.9, weight_decay=5e-4)
+        scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optim, max_epochs)
+        return [optim], [scheduler]
 
 
 # %%
@@ -301,7 +308,9 @@ def validation_step(self, batch, batch_idx):
                  on_epoch=True, prog_bar=True)
 
     def configure_optimizers(self):
-        return torch.optim.SGD(self.fc.parameters(), lr=10.0)
+        optim = torch.optim.SGD(self.fc.parameters(), lr=30.)
+        scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optim, max_epochs)
+        return [optim], [scheduler]
 
 
 # %%
@@ -324,6 +333,7 @@ def configure_optimizers(self):
 
 # %%
 # Train the Classifier
+model.eval()
 classifier = Classifier(model.resnet_moco)
 trainer = pl.Trainer(max_epochs=max_epochs, gpus=gpus,
                      progress_bar_refresh_rate=100)