Adding Data to Build a More Generic Model

Intro

You might be in the middle of training a model and then the business problem shifts. Now you have this model that has been going through the training process with a specific dataset and you need to make the model more generic.

There's likely something that your model learned that can be useful on this new dataset, so you might not have to restart the entire training process. We'll do an example of updating a pre-trained model to use a broader dataset with DVC. By the end of this, you should see how you can handle this quickly and start running new experiments to get a more generic model.

The original pre-trained model

For this post, we'll be making a more generic image classifier by taking the original dataset with bees and ants and adding cats and dogs to it. You can clone this GitHub repo to get the current bees and ants model and check out this post on how we experimented with both AlexNet and SqueezeNet to build this model.

So we're starting from our current bees and ants model and extending it to classify dogs and cats as well. We'll start by adding some cats and dogs data to our validation data and do some experiments with the current model to see how it performs on generic data.

Then we'll add the cats and dogs data to the training data and watch how the model improves as we run experiments.

Updating the dataset with DVC

To add the new cats and dogs dataset to the project, we'll use this DVC command.

$ dvc get https://github.com/iterative/dataset-registry blog/cats-dogs

This downloads a sample dataset with images of cats and dogs. You can use this command to download files or directories that are tracked by DVC or Git. This command can be used from anywhere in the file system, as long as DVC is installed.

This will make a new directory called ./cats-dogs/data/ that was downloaded from the DVC remote and it has images for cats and dogs. Now we can slowly add in the new data to the existing data.

We'll start by moving the val data for cats and dogs from the /cats-dogs/data/ directory to the corresponding directory in data/hymenoptera_data.

Just a quick note, cats and dogs don't really belong in the hymenoptera directory since that's specific to ants and bees, but it's the easiest and fastest way to add the data for this tutorial.

With this new data in place, we can start training our model.

Running new experiments with generic data

With the updated data, let's run an experiment on the model and see how good the results are. To run a new experiment, open your terminal and make sure you have a virtual environment enabled. Then run this command:

$ dvc exp run

Once the training epochs are finished, run the following command.

$ dvc exp show --no-timestamp \
--include-metrics step,acc,val_acc,loss,val_loss \
--include-params lr,momentum

The --no-timestamp hides the timestamps from table. The --includes-metrics option lets us choose which metrics we want to show in the table. The --includes-params option does the same for hyperparameters. This gives us a table that's easier to read quickly.

 ────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                metric:**step**       metric:**acc**   metric:**val_acc**      metric:**loss**   metric:**val_loss**   param:**lr**      param:**momentum**
 ────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                    **3**   **0.86885**      **0.46**   **0.31573**     **3.7067**   **0.001**   **0.09**
  **data-change**                  **-**         **-**         **-**         **-**          **-**   **0.001**   **0.09**
  │ ╓ 3b3a2a2 [exp-23593]      3   0.86885      0.46   0.31573     3.7067   0.001   0.09
  │ ╟ 93d015d                  2   0.83197   0.41333   0.36851     3.4259   0.001   0.09
  │ ╟ d474c42                  1   0.79918   0.43333   0.46612      3.286   0.001   0.09
  ├─╨ 1582b4b                  0   0.52869      0.39   0.94102     2.5967   0.001   0.09
 ────────────────────────────────────────────────────────────────────────────────────────────

You'll notice that the validation accuracy is really low. That's because the training metrics are based on bees and ants while the validation metrics are based on bees, ants, cats, and dogs. If we looked at the validation metrics by class, they'd likely be better for bees and ants than cats and dogs.

That means we should probably add more data to the training dataset.

Adding the cats data to the training dataset

Let's add the train data for cats to the corresponding directory in data/hymenoptera_data and go through another experiment run with a different learning rate. With this new data, we can run another experiment. One important thing to note here is that we're using checkpoints in our experiments. That's how we get the metrics for each training epoch.

If we want to run a fresh experiment that doesn't resume training from the last epoch, we need to reset our experiment. That's what we're going to do with this command.

$ dvc exp run --reset

This will reset all of the existing checkpoints and excute the training script. Once it's finished, let's take a look at the metrics table with this command. It's the same as the one we ran last time.

$ dvc exp show --no-timestamp \
    --include-metrics step,acc,val_acc,loss,val_loss \
    --include-params lr,momentum

Now you'll have a table that shows both experiments and you can see how much better the new one did with the cats data added.

 ────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                metric:**step**       metric:**acc**   metric:**val_acc**      metric:**loss**   metric:**val_loss**   param:**lr**      param:**momentum**
 ────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                    **3**   **0.91389**      **0.87**   **0.20506**    **0.66306**   **0.001**   **0.09**
  **data-change**                  **-**         **-**         **-**         **-**          **-**   **0.001**   **0.09**
  │ ╓ 9405575 [exp-54e8a]      3   0.91389      0.87   0.20506    0.66306   0.001   0.09
  │ ╟ 856d80f                  2   0.90215   0.87333   0.27204    0.61631   0.001   0.09
  │ ╟ 23dc98f                  1   0.87671      0.86   0.35964    0.61713   0.001   0.09
  ├─╨ 99a3c34                  0   0.71429      0.82   0.67674    0.62798   0.001   0.09
  │ ╓ 3b3a2a2 [exp-23593]      3   0.86885      0.46   0.31573     3.7067   0.001   0.09
  │ ╟ 93d015d                  2   0.83197   0.41333   0.36851     3.4259   0.001   0.09
  │ ╟ d474c42                  1   0.79918   0.43333   0.46612      3.286   0.001   0.09
  ├─╨ 1582b4b                  0   0.52869      0.39   0.94102     2.5967   0.001   0.09
 ────────────────────────────────────────────────────────────────────────────────────────────

There's another way you can look at the difference between the model before we added the cats data and after. If you run this in your terminal, you'll get a plot comparing the two experiments.

$ dvc plots diff exp-23593 exp-54e8a

The exp-23593 and exp-54e8a values are the ids for the experiments you want to compare. You'll see a new file gets generated in the dvc_plots directory in your project. That's where you'll find the index.html file you should open in your browser. You'll see something similar to this.

There's a huge difference in the accuracy of our model after we've added this additional data. Let's see if we can make it even better by adding the dogs data.

Adding the dogs data to the training dataset

We'll add the train data for dogs to the corresponding directory in data/hymenoptera_data just like we did for the cats data. Now we can run a new experiment with all of the new data included. We'll still need to reset the experiment like before, so run the following command.

$ dvc exp run --reset

Once the training epochs are finished, we can take one more look at that metrics table.

$ dvc exp show --no-timestamp \
--include-metrics step,acc,val_acc,loss,val_loss \
--include-params lr,momentum

Now we'll have all three experiments to compare.

 ────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                metric:**step**       metric:**acc**   metric:**val_acc**      metric:**loss**   metric:**val_loss**   param:**lr**      param:**momentum**
 ────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                    **3**    **0.8795**   **0.90667**   **0.29302**    **0.25752**   **0.001**   **0.09**
  **data-change**                  **-**         **-**         **-**         **-**          **-**   **0.001**   **0.09**
  │ ╓ c20220f [exp-82f70]      3    0.8795   0.90667   0.29302    0.25752   0.001   0.09
  │ ╟ fcb5a0b                  2   0.85915   0.92333   0.38274    0.25257   0.001   0.09
  │ ╟ 3768821                  1   0.80751   0.84667   0.47681    0.40228   0.001   0.09
  ├─╨ 7e1b8fb                  0   0.64632      0.84   0.87301    0.46744   0.001   0.09
  │ ╓ 9405575 [exp-54e8a]      3   0.91389      0.87   0.20506    0.66306   0.001   0.09
  │ ╟ 856d80f                  2   0.90215   0.87333   0.27204    0.61631   0.001   0.09
  │ ╟ 23dc98f                  1   0.87671      0.86   0.35964    0.61713   0.001   0.09
  ├─╨ 99a3c34                  0   0.71429      0.82   0.67674    0.62798   0.001   0.09
  │ ╓ 3b3a2a2 [exp-23593]      3   0.86885      0.46   0.31573     3.7067   0.001   0.09
  │ ╟ 93d015d                  2   0.83197   0.41333   0.36851     3.4259   0.001   0.09
  │ ╟ d474c42                  1   0.79918   0.43333   0.46612      3.286   0.001   0.09
  ├─╨ 1582b4b                  0   0.52869      0.39   0.94102     2.5967   0.001   0.09
 ────────────────────────────────────────────────────────────────────────────────────────────

These results make sense for the experiments we've run. We're paying attention to the validation accuracy here because this gives us a fair comparison of what's happening as we add more data.

The first experiment's training metrics are for bees and ants. The second experiment's training metrics are for bees, ants, and cats. And the third experiment's training metrics are for all four classes. So we can't really compare these metrics.

We can look at a comparison between the experiments with the cats data and both the cats and dogs data.

$ dvc plots diff exp-23593 exp-54e8a exp-82f70

The results you see line up with what is expected for the validation metrics based on how we added the data to the training set. Now you can keep running experiments until you get your model tuned like you need it!

Conclusion

When you want to change datasets quickly and start tracking how they affect our model, using a DVC remote makes it easy to do so on different computers. You'll be able to quickly upload and download GBs of data and see how changes affect individual experiments.

If you need help with anything DVC or CML, make sure to join our Discord community! We're always answering questions and having good conversations with everybody that shows up.