Helpful tips to debug ML systems¶

Always check your data
- Visualize outputs/labels and the inputs
- Visualize the data distribution and try to find patterns
- May also try if simple methods like Kernel SVM , Random Forest work reasonably. If not then some problem with the dataset
- Duplicated examples , bad data , corrupted labels
- Do you need to do Data processing ?
  - Mean centering , Whitening etc.
- Look for dataset imbalance and biases
  - Up-sample or Down-sample if you see imbalance
Start by overfitting on few samples by turning the regularization off. If the loss doesnot converge to zero then there might be a problem.
Start with simple model then gradually increase model complexity.
Is the model underfitting or overfitting? Simple checks can be done using k-Fold cross validation and seeing train and val losses. Run on small version or mini-validation data especially if you are running huge datasets. Use Tensorboard, weight&biases or other solutions to visualize the losses.
- Underfitting: both train and test errors are high.
- Solutions:
  - Increase model complexity
- Overfitting: train error low but test error high.
- Solutions:
  - Collect more data
  - Decrease model complexity eg. fewer units, reduce no. of hidden layers
  - Regularize:
    - Dropout
    - Batchnorm
    - Early stopping
    - L2 weight decay
    - Ensemble
See that scales of losses are OK
check for anomaly in gradient values
- If you find that gradient are blowing-up then you can perform gradient clipping.
check if there are NaNsor Infs or really weird values. Below, I describe a scenario that I faced.
- e.g. I was getting NaNs while running a Faster-RCNN model. I checked all the losses and found that my loss for classification was fine. But, I was getting NaNs just for co-ordinate regressors. Upon digging deep I found that something like below could be happening:

1. (x,y) of bounding boxes > H / W of the image. 
2. xmin = xmax or ymin = ymax
3. Small sized bounding boxes

Easy fixes could be something like not loading small boxes i.e ensuring |xmax -xmin| >= 10 in your dataloader.

Code to display warning if NaNs found in loss in pytorch:

loss =  ##
if torch.isnan(loss):
   print ("nans encountered", ids)
   import ipdb; ipdb.set_trace()

Additionally to debug during the backward pass you will need to add few hooks and print statements.

class MyModel(nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.model = models.resnet18(pretrained = True)
        self.gradients = []
        ## your layers
        self.model.avgpool.register_backward_hook(self.save_gradients)

    def forward(self, x):
        ## your forward pass
        return self.model(x)

    def save_gradients(self, module, grad_input, grad_output):
        name = module._get_name()
        self.gradients.append(grad_output)

model = MyModel()
x = torch.rand(1,3,224,224)
output = model(x)
output.backward()
print(model.gradients)

Code Sanity checks
- Check the conditional logic if/then/else in your code and be sure they are really doing intended things
- Deliver clear error/warning messages
  - eg. I made a silly mistake where I didnot initilialize a model if the checkpoint is not found. I forgot to add an error message to show the folder not found error. So, I was unaware if the model is being initialized or not.
- Have asserts in dataloader to prevent absolutely ridiculous things from happening
  - E.g. check if bounding box min and max are the same or if box coordinates are larger than the image
Learn how to use a debugger. ipdb or pdb for python to trace the points where you keep getting errors. Its better than using print debugging since you can examine each steps. Also, It might come handy to learn some pdb commands.

Helpful tips to debug ML systems¶

Resources¶