Vision Callbacks¶

Useful callbacks for vision models.

Note

We rely on the community to keep these updated and working. If something doesn’t work, we’d really appreciate a contribution to fix!

Confused Logit¶

Shows how the input would have to change to move the prediction from one logit to the other

Example outputs:

class pl_bolts.callbacks.vision.confused_logit.ConfusedLogitCallback(top_k, min_logit_value=5.0, logging_batch_interval=20, max_logit_difference=0.1)[source]

Bases: pytorch_lightning.callbacks.callback.Callback

Warning

The feature ConfusedLogitCallback is currently marked under review. The compatibility with other Lightning projects is not guaranteed and API may change at any time. The API and functionality may change without warning in future releases. More details: https://lightning-bolts.readthedocs.io/en/latest/stability.html

Takes the logit predictions of a model and when the probabilities of two classes are very close, the model doesn’t have high certainty that it should pick one vs the other class.

This callback shows how the input would have to change to swing the model from one label prediction to the other.

In this case, the network predicts a 5… but gives almost equal probability to an 8. The images show what about the original 5 would have to change to make it more like a 5 or more like an 8.

For each confused logit the confused images are generated by taking the gradient from a logit wrt an input for the top two closest logits.

Example:

from pl_bolts.callbacks.vision import ConfusedLogitCallback
trainer = Trainer(callbacks=[ConfusedLogitCallback()])

Note

Whenever called, this model will look for self.last_batch and self.last_logits in the LightningModule.

Note

This callback supports tensorboard only right now.

Authored by:

Alfredo Canziani

Parameters

top_k¶ (int) – How many “offending” images we should plot
projection_factor¶ – How much to multiply the input image to make it look more like this logit label
min_logit_value¶ (float) – Only consider logit values above this threshold
logging_batch_interval¶ (int) – How frequently to inspect/potentially plot something
max_logit_difference¶ (float) – When the top 2 logits are within this threshold we consider them confused

on_train_batch_end(trainer, pl_module, outputs, batch, batch_idx)[source]

Called when the train batch ends.

Note

The value outputs["loss"] here will be the normalized value w.r.t accumulate_grad_batches of the loss returned from training_step.

Return type: None

Tensorboard Image Generator¶

Generates images from a generative model and plots to tensorboard

class pl_bolts.callbacks.vision.image_generation.TensorboardGenerativeModelImageSampler(num_samples=3, nrow=8, padding=2, normalize=False, norm_range=None, scale_each=False, pad_value=0)[source]

Bases: pytorch_lightning.callbacks.callback.Callback

Warning

The feature TensorboardGenerativeModelImageSampler is currently marked under review. The compatibility with other Lightning projects is not guaranteed and API may change at any time. The API and functionality may change without warning in future releases. More details: https://lightning-bolts.readthedocs.io/en/latest/stability.html

Generates images and logs to tensorboard. Your model must implement the forward function for generation.

Requirements:

# model must have img_dim arg
model.img_dim = (1, 28, 28)

# model forward must work for sampling
z = torch.rand(batch_size, latent_dim)
img_samples = your_model(z)

Example:

from pl_bolts.callbacks import TensorboardGenerativeModelImageSampler

trainer = Trainer(callbacks=[TensorboardGenerativeModelImageSampler()])

Parameters

num_samples¶ (int) – Number of images displayed in the grid. Default: 3.
nrow¶ (int) – Number of images displayed in each row of the grid. The final grid size is (B / nrow, nrow). Default: 8.
padding¶ (int) – Amount of padding. Default: 2.
normalize¶ (bool) – If True, shift the image to the range (0, 1), by the min and max values specified by range. Default: False.
norm_range¶ (Optional[Tuple[int, int]]) – Tuple (min, max) where min and max are numbers, then these numbers are used to normalize the image. By default, min and max are computed from the tensor.
scale_each¶ (bool) – If True, scale each image in the batch of images separately rather than the (min, max) over all images. Default: False.
pad_value¶ (int) – Value for the padded pixels. Default: 0.

on_train_epoch_end(trainer, pl_module)[source]

Called when the train epoch ends.

To access all batch outputs at the end of the epoch, either:

Implement training_epoch_end in the LightningModule and access outputs via the module OR
Cache data across train batch hooks inside the callback implementation to post-process in this hook.

Return type: None