Using Experiments for Transfer Learning

DVC experiments help fine-tune models by tracking code and data changes.

Milecia McGregor
August 24, 2021 • 12 min read

Using Experiments to Improve Pre-trained Models

Intro

There are plenty of machine learning models available that have been trained to solve one problem and the knowledge gained from that can be applied to a new, yet related problem. For example, a model like AlexNet has been trained on millions of images so you could potentially use this to classify cars, animals, or even people. This is called transfer learning and it can save a lot of time on developing a model from scratch.

For us to take advantage of transfer learning, we can use fine-tuning to adopt the model to our new problem. In many cases, we start by replacing the last layer of the model. With the AlexNet example, this might mean the last layer was previously used to classify cars but our new problem is classifying animals.

Even though we already have the bulk of the model defined, we'll still have to do some experimentation to determine whether we need to replace more layers in the model or if any other changes need to be made.

In this post, we'll go through an example of fine-tuning AlexNet and SqueezeNet to classify bees and ants. We'll use DVC to handle experiments for us and we'll compare the results of both models at the end.

Initialize the pre-trained model

We'll be fine-tuning the AlexNet model and the SqueezeNet model to classify images of bees and ants. You can find the project we're working with in this repo, which is based on the tutorial over at this post.

In the pretrained_model_tuner.py file, you'll find the code that defines both the AlexNet and SqueezeNet models. We start by initializing these models so we can get the number of model features and the input size we need for fine-tuning.

Since the project has everything we need to initialize the models, we can start training and comparing the differences between them with the ants/bees dataset. Running experiments to get the best tuning for each model can make it difficult to see which changes led to a better result. That's why we will be using DVC to track changes in the code and the data.

Adding the train stage

Stages in DVC let us define individual data processes and can be used to build detailed machine learning pipelines. You have the ability to define the different steps of model creation like preprocessing, featurization, and training.

We currently have a train stage in the dvc.yaml file. If you take a look at it, you'll see something like:

stages:
  train:
    cmd: python pretrained_model_tuner.py
    deps:
      - data/hymenoptera_data
      - pretrained_model_tuner.py
    params:
      - lr
      - momentum
      - model_name
      - num_classes
      - batch_size
      - num_epochs
    outs:
      - model.pt:
          checkpoint: true
    live:
      results:
        summary: true
        html: true

The reason we need this dvc.yaml file is so DVC knows what to pay attention to in our workflow. It will start managing data, understand which metrics to pay attention to, and what the expected output for each step is.

You'll typically add stages to dvc.yaml using the dvc stage add command and this is one of the ways you can add new stages or update existing ones.

With the train stage defined, let's look at where the metrics actually come from in the code. If you open pretrained_model_tuner, you'll see a line where we dump the accuracy and loss for the training epochs into the results.json file. We're also saving the model on the epoch run and recording metrics for each epoch using dvclive logging.

if phase == 'train':
    torch.save(model.state_dict(), "model.pt")

    dvclive.log('acc', epoch_acc.item())
    dvclive.log('loss', epoch_loss)
    dvclive.log('training_time', epoch_time_elapsed)

if phase == 'val':
    dvclive.log('val_acc', epoch_acc.item())
    dvclive.log('val_loss', epoch_loss)

    val_acc_history.append(epoch_acc)

    dvclive.next_step()

This code is needed to let DVC access the metrics in the project because it will read the metrics from the dvclive.json file.

Since we have several hyperparameters set in the params.yaml, we need to use those values when we run the training stage. The following code makes the hyperparameter values accessible in the train function.

with open("params.yaml") as f:
    yaml=YAML(typ='safe')
    params = yaml.load(f)

With all of this in place, we can finally start running experiments to fine-tune the two models.

Fine-tuning AlexNet

You can find the code that initializes the AlexNet model in the initialize_model function in pretrained_model_tuner.py. Since we have DVC set up, we can jump straight into fine-tuning this model to see which hyperparameters give us the best accuracy.

We'll run the first experiment with the following command.

$ dvc exp run

This will execute the pretrained_model_tuner.py script and run for 5 epochs since that's what we defined in params.yaml. When this finishes, you can check out the metrics from this run with the current hyperparameter values.

$ dvc exp show

You'll see a table similar to this.

 ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                neutral:**Created**    metric:**step**       metric:**acc**      metric:**loss**   metric:**training_time**   metric:**val_acc**   metric:**val_loss**   param:**lr**      param:**momentum**   param:**model_name**   param:**num_classes**   param:**batch_size**   param:**num_epochs**
 ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                 -             **4**   **0.92623**   **0.19567**          **229.18**    **0.9085**    **0.25145**   **0.001**   **0.09**       **alexnet**      **2**             **8**            **5**
  **main**                      **01:58 PM**      -         -         -               -         -          -   **0.001**   **0.09**       **alexnet**      **2**             **8**            **5**
  │ ╓ bf81637 [exp-a1f53]   02:05 PM      4   0.92623   0.19567          229.18    0.9085    0.25145   0.001   0.09       alexnet      2             8            5
  │ ╟ 9ca3fb8               02:04 PM      3   0.89344   0.27423          178.34   0.90196    0.26965   0.001   0.09       alexnet      2             8            5
  │ ╟ a34ead1               02:03 PM      2   0.87295   0.29018          127.36    0.9085     0.2796   0.001   0.09       alexnet      2             8            5
  │ ╟ ae382c7               02:02 PM      1   0.89754   0.26993          76.419   0.89542    0.31113   0.001   0.09       alexnet      2             8            5
  ├─╨ a95260d               02:01 PM      0   0.73361    0.5271           25.71   0.86928    0.36408   0.001   0.09       alexnet      2             8            5
 ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

Now let's update the hyperparameters and run another experiment. There are several ways to do this with DVC:

Change the hyperparameter values directly in params.yaml
Update the values using the --set-param or the shorthand -S option on dvc exp run
Queue multiple experiments with different values using the --queue option on dvc exp run

We'll do an example of each of these throughout the rest of this article.

Let's start by updating the hyperparameter values in params.yaml. You should have these values in your file.

lr: 0.009
momentum: 0.017

Now run another experiment with dvc exp run. To make the table more readable, we're going to specify the parameters we want to show and take a look at the metrics with:

$ dvc exp show --no-timestamp --include-params lr,momentum,model_name

Your table should look something like this now. Since we're using checkpoints, note that we continue training additional epochs on top of your previous experiment. You'll see what it takes to start training from scratch later.

 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                metric:**step**       metric:**acc**      metric:**loss**   metric:**training_time**   metric:**val_acc**   metric:**val_loss**   param:**lr**      param:**momentum**   param:**model_name**
 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                    **9**   **0.91803**   **0.27989**          **228.59**   **0.82353**    **0.69077**   **0.009**   **0.017**      **alexnet**
  **main**                         **-**         **-**         **-**               **-**         **-**          **-**   **0.001**   **0.09**       **alexnet**
  │ ╓ 2361cff [exp-c0b11]      9   0.91803   0.27989          228.59   0.82353    0.69077   0.009   0.017      alexnet
  │ ╟ 7686d2f                  8   0.90984   0.23496          177.65   0.87582    0.50887   0.009   0.017      alexnet
  │ ╟ 671f8cd                  7   0.88934   0.39237           126.7   0.86928    0.47856   0.009   0.017      alexnet
  │ ╟ ea1bf61                  6   0.84836    0.4195          75.834   0.91503    0.30885   0.009   0.017      alexnet
  │ ╟ a9f8dab (bf81637)        5   0.79508   0.72891          25.219   0.66667     1.0311   0.009   0.017      alexnet
  │ ╓ bf81637 [exp-a1f53]      4   0.92623   0.19567          229.18    0.9085    0.25145   0.001   0.09       alexnet
  │ ╟ 9ca3fb8                  3   0.89344   0.27423          178.34   0.90196    0.26965   0.001   0.09       alexnet
  │ ╟ a34ead1                  2   0.87295   0.29018          127.36    0.9085     0.2796   0.001   0.09       alexnet
  │ ╟ ae382c7                  1   0.89754   0.26993          76.419   0.89542    0.31113   0.001   0.09       alexnet
  ├─╨ a95260d                  0   0.73361    0.5271           25.71   0.86928    0.36408   0.001   0.09       alexnet
 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

Finding good values for hyperparameters can take a few iterations, even when you're working with a pretrained model. So we'll run one more experiment to fine-tune this AlexNet model. This time we'll do it using the -S option.

$ dvc exp run -S lr=0.025 -S momentum=0.5 -S num_epochs=2

The updated table will have values similar to this.

 ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                metric:**step**       metric:**acc**      metric:**loss**    metric:**training_time**   metric:**val_acc**   metric:**val_loss**   param:**lr**      param:**momentum**   param:**model_name**
 ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                   **11**   **0.88525**    **1.1355**           **76.799**    **0.9085**     **1.7642**   **0.025**   **0.5**        **alexnet**
  **main**                         **-**         **-**          **-**               **-**         **-**          **-**   **0.001**   **0.09**       **alexnet**
  │ ╓ 54e87bc [exp-52406]     11   0.88525    1.1355           76.799    0.9085     1.7642   0.025   0.5        alexnet
  │ ╟ b2b9ad0 (2361cff)       10   0.79098    2.9427           25.715    0.8366     1.4148   0.025   0.5        alexnet
  │ ╓ 2361cff [exp-c0b11]      9   0.91803   0.27989           228.59   0.82353    0.69077   0.009   0.017      alexnet
  │ ╟ 7686d2f                  8   0.90984   0.23496           177.65   0.87582    0.50887   0.009   0.017      alexnet
  │ ╟ 671f8cd                  7   0.88934   0.39237            126.7   0.86928    0.47856   0.009   0.017      alexnet
  │ ╟ ea1bf61                  6   0.84836    0.4195           75.834   0.91503    0.30885   0.009   0.017      alexnet
  │ ╟ a9f8dab (bf81637)        5   0.79508   0.72891           25.219   0.66667     1.0311   0.009   0.017      alexnet
  │ ╓ bf81637 [exp-a1f53]      4   0.92623   0.19567           229.18    0.9085    0.25145   0.001   0.09       alexnet

If you take a look at the metrics and the corresponding hyperparameter values, you'll see which direction you should try next with your values. That's one way we can use DVC to fine-tune AlexNet for this particular dataset.

Fine-tuning SqueezeNet

We'll switch over to fine-tuning SqueezeNet now that you've seen how the process works in DVC. You'll need to update the model_name hyperparameter in params.yaml to squeezenet if you're following along. The other hyperparameter values can stay the same for now.

This is a good time to note that DVC is not only tracking the changes of your hyperparameters for each experiment, it also tracks any code changes and dataset changes as well.

Let's run one experiment with dvc exp run --reset just to show the difference in the metrics between the two models. Remember, since we're using checkpoints it continues training on top of the previous experiment. That's why we're using the --reset option here so that we can start a fresh experiment for the new model. You should see results similar to this in your table.

 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                metric:**step**       metric:**acc**      metric:**loss**   metric:**training_time**   metric:**val_acc**   metric:**val_loss**   param:**lr**      param:**momentum**   param:**model_name**
 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                    **1**   **0.85656**   **0.35667**          **83.414**   **0.87582**    **0.34273**   **0.025**   **0.5**        **squeezenet**
  **main**                         -         -         -               -         -          -   **0.001**   **0.09**       **squeezenet**
  │ ╓ 87ccd2e [exp-95f0f]      1   0.85656   0.35667          83.414   0.87582    0.34273   0.025   0.5        squeezenet
  ├─╨ 7d2fafc                  0   0.80328   0.50723          29.165   0.89542     0.3987   0.025   0.5        squeezenet
  │ ╓ 54e87bc [exp-52406]     11   0.88525    1.1355          76.799    0.9085     1.7642   0.025   0.5        alexnet
  │ ╟ b2b9ad0 (2361cff)       10   0.79098    2.9427          25.715    0.8366     1.4148   0.025   0.5        alexnet
  │ ╓ 2361cff [exp-c0b11]      9   0.91803   0.27989          228.59   0.82353    0.69077   0.009   0.017      alexnet

The newest experiment has an accuracy that's significantly different since we switched models. That tells us that the hyperparameter values that were good for AlexNet might not work the best for SqueezeNet.

So we'll need to run a few experiments to find the best hyperparameter values. This time, we'll take advantage of queues in DVC to set up the experiments and then run them at the same time. To set up a queue, we'll run this command.

$ dvc exp run --queue -S lr=0.0001 -S momentum=0.9 -S num_epochs=2

Running this sets up an experiment for future execution so we'll go ahead a run this command one more time with different values.

$ dvc exp run --queue -S lr=0.001 -S momentum=0.09 -S num_epochs=2

You can check out the details for the queues you have in place by looking at the experiments table with dvc exp show. You'll see something like this.

 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                metric:**step**       metric:**acc**      metric:**loss**   metric:**training_time**   metric:**val_acc**   metric:**val_loss**   param:**lr**      param:**momentum**   param:**model_name**
 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                    **1**   **0.85656**   **0.35667**          **83.414**   **0.87582**    **0.34273**   **0.025**   **0.5**        **squeezenet**
  **main**                         -         -         -               -         -          -   **0.001**   **0.09**       **squeezenet**
  │ ╓ 87ccd2e [exp-95f0f]      1   0.85656   0.35667          83.414   0.87582    0.34273   0.025   0.5        squeezenet
  ├─╨ 7d2fafc                  0   0.80328   0.50723          29.165   0.89542     0.3987   0.025   0.5        squeezenet
  │ ╓ 54e87bc [exp-52406]     11   0.88525    1.1355          76.799    0.9085     1.7642   0.025   0.5        alexnet
  │ ╟ b2b9ad0 (2361cff)       10   0.79098    2.9427          25.715    0.8366     1.4148   0.025   0.5        alexnet
  │ ╓ 2361cff [exp-c0b11]      9   0.91803   0.27989          228.59   0.82353    0.69077   0.009   0.017      alexnet
  │ ╟ 7686d2f                  8   0.90984   0.23496          177.65   0.87582    0.50887   0.009   0.017      alexnet
  │ ╟ 671f8cd                  7   0.88934   0.39237           126.7   0.86928    0.47856   0.009   0.017      alexnet
  │ ╟ ea1bf61                  6   0.84836    0.4195          75.834   0.91503    0.30885   0.009   0.017      alexnet
...
  ├── *2df7fa5                -          -          -         -         -               -   0.0001  0.9        squeezenet
  ├── *699dcae                -          -          -         -         -               -   0.001   0.09       squeezenet
 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────

Then you can execute all of the queues with this command.

$ dvc exp run --run-all

Now if you take a look at your table, you'll see the metrics from those 3 experiments.

 ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  neutral:**Experiment**                metric:**step**       metric:**acc**      metric:**loss**   metric:**training_time**   metric:**val_acc**   metric:**val_loss**   param:**lr**       param:**momentum**   param:**model_name**
 ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  **workspace**                    **5**   **0.76639**   **0.49865**          **85.705**   **0.81699**     **0.4518**   **0.001**    **0.09**       **squeezenet**
  **main**                         -         -         -               -         -          -   **0.001**    **0.09**       **squeezenet**
  │ ╓ 699dcae [exp-8322f]      5   0.76639   0.49865          85.705   0.81699     0.4518   0.001    0.09       squeezenet
  │ ╟ d26c25b (2df7fa5)        4   0.60246   0.68464          29.243   0.69935    0.55156   0.001    0.09       squeezenet
  │ ╓ 2df7fa5 [exp-d1c65]      3   0.78689     0.488          83.929   0.83007    0.41527   0.0001   0.9        squeezenet
  │ ╟ 05e1b41 (87ccd2e)        2   0.59016   0.76999          28.455   0.75163    0.49807   0.0001   0.9        squeezenet
  │ ╓ 87ccd2e [exp-95f0f]      1   0.85656   0.35667          83.414   0.87582    0.34273   0.025    0.5        squeezenet
  ├─╨ 7d2fafc                  0   0.80328   0.50723          29.165   0.89542     0.3987   0.025    0.5        squeezenet
  │ ╓ 54e87bc [exp-52406]     11   0.88525    1.1355          76.799    0.9085     1.7642   0.025    0.5        alexnet
  │ ╟ b2b9ad0 (2361cff)       10   0.79098    2.9427          25.715    0.8366     1.4148   0.025    0.5        alexnet
  │ ╓ 2361cff [exp-c0b11]      9   0.91803   0.27989          228.59   0.82353    0.69077   0.009    0.017      alexnet
  │ ╟ 7686d2f                  8   0.90984   0.23496          177.65   0.87582    0.50887   0.009    0.017      alexnet

Then you'll be able to make a decision on which way to go with your fine-tuning efforts and make a decision on which model works best for your project. In this case, it seems like SqueezeNet might be the winner!

You can take all of the DVC setup and apply this to your own custom fine-tuning use case.

Conclusion

When you're working with pretrained models, it can be hard to fine-tune them to give you the results you need. You might end up replacing the last layer of the model to fit your problem or you might need to dig deeper. Then you have to consider updating the hyperparameter values until you get the best model you can.

That's why it's important to research tools that make this process more efficient. Using DVC to help with this kind of experimentation will give you the ability to reproduce any experiment you run, making it easier to collaborate with others on a project. It will also help you keep track of what you've already tried in previous experiments.

← Back to blog