irisml-tasks-training 0.0.3

IrisML tasks for pytorch training

irisml-tasks-training

This is a package for IrisML training-related tasks.

See irisml repository for the detail of irisml framework.

Tasks

train

Train a pytorch model. A model object must have "criterion" and "predictor" property. See the documents for the detail. Returns a trained model.

predict

Run inference with a given pytorch model. Returns prediction results.

append_classifier

Append a classifier layer to an encoder model.

build_classification_prompt_dataset

Convert a multiclass classification Image Dataset into a dataset with text prompts.

build_zero_shot_classifier

Build a classifier FC layer from text features. See the CLIP repo for the detail.

create_classification_prompt_generator

Create a prompt generator for classification task.

export_onnx

Trace a pytorch model and export it as ONNX using torch.onnx.export(). Throws an exception if it couldn't export. Returns an exported onnx model.

evaluate_accuracy

Calculate top1 accuracy for given prediction results. It supports only image classification results.

evaluate_detection_average_precision

Calculate mAP for object detection results.

get_targets_from_dataset

Get a list or a tensor of targets from a Dataset.

make_feature_extractor_model

Make a new model to extract intermediate features from the given model. Use the predict task to run the extractor model.

make_image_text_contrastive_model

Make a new model to run image-text contrastive training like CLIP.

make_image_text_transform

Make a transform function that can be used for a contrastive training

split_image_text_model

Extract image_model and text_model from a image-text model.

Available plugins for train task.

• log_summary
• log_tensorboard
• progressbar

Interfaces for training and prediction

The tasks in this package expects the following interfaces

Notations

• input: An input object for a single example. For example, an image tensor.
• target: A ground truth for a single example.
• inputs_batch: A batch of input.
• targets_batch: A batch of target

Model

class Model(torch.nn.Module):
    def training_step(self, inputs_batch, targets_batch):  # Returns {'loss': loss_tensor}
        pass

A model for training must implement training_step() method. The trainer will provide inputs and targets to the method. It must return a dictionary containing 'loss' entry.

class Model(torch.nn.Module):
    def prediction_step(self, inputs_batch):  # Returns prediction results
        pass 

Similarily, a model for prediction must have 'prediction_step()' method. Inputs will be provided to this method and it must return prediction results.

For most of the case, a model implements both methods, training_step() and prediction_step().

Dataset

The trainer accepts an instance of torch.utils.data.Dataset class. For each index, it must return a tuple (raw_input, target). Curretly, raw_input must be a RGB PIL Image object.

Transform

A transform function must return (input, target) given (raw_inputs, target).

Inputs and targets formats

Multiclass Image classification

• input: A float tensor [3, H, W] that represents a RGB image. Its value range is [0-1].
• inputs_batch: A float tensor [N, 3, H, W] if all inputs have the same shape. Otherwise, a list of input.
• target: an integer tensor that represents a class index.
• targets: An integer tensor [N, 1].

Multilabel Image Classification

• inputs, inputs_batch: Same with above
• taget: An integer tensor [num_classes]. Its value is 0 (negative) or 1 (positive).
• targets_batch: An integer tensor [N, num_classes]

Object Detection

• inputs, inputs_batch: Same with above
• target: A float tensor [num_boxes, 5]. Each bounding box is represented as [class_index, x0, y0, x1, y1]. x0, y0, x1, y1 is relative coordinates of the left, top, right, bottom of the box. 0 <= x0 < x1 <= 1 and 0 <= y0 < y1 <= 1.
• targets_batch: A list of targets

Image Segmentation

• inputs, inputs_batch: Same with above
• target: A float tensor [num_classes, H, W]. Its value is 0 (negative) or 1 (positive) for each pixel on the sample.
• targets: A float tensor [N, num_classes, H, W]

CLIP Zero-shot classifier build

build_zero_shot_classifier task has a different interface. It doesn't require a Model instance. Instead, it requires two tensors, text_features and text_labels.

• text_features: A float tensor [N, feature_size].
• text_labels: An integer tensor[N, 1] that represents a class index for each text.