transformers/docs/source/en/tasks/semantic_segmentation.mdx

<!--Copyright 2022 The HuggingFace Team. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->

# Semantic segmentation

[[open-in-colab]]

<Youtube id="dKE8SIt9C-w"/>

Semantic segmentation assigns a label or class to each individual pixel of an image. There are several types of segmentation, and in the case of semantic segmentation, no distinction is made between unique instances of the same object. Both objects are given the same label (for example, "car" instead of "car-1" and "car-2"). Common real-world applications of semantic segmentation include training self-driving cars to identify pedestrians and important traffic information, identifying cells and abnormalities in medical imagery, and monitoring environmental changes from satellite imagery.

This guide will show you how to:

1. Finetune [SegFormer](https://huggingface.co/docs/transformers/main/en/model_doc/segformer#segformer) on the [SceneParse150](https://huggingface.co/datasets/scene_parse_150) dataset.
2. Use your finetuned model for inference.

<Tip>

See the image segmentation [task page](https://huggingface.co/tasks/image-segmentation) for more information about its associated models, datasets, and metrics.

</Tip>

Before you begin, make sure you have all the necessary libraries installed:

```bash
pip install -q datasets transformers evaluate
```

We encourage you to login to your Hugging Face account so you can upload and share your model with the community. When prompted, enter your token to login:

```py
>>> from huggingface_hub import notebook_login

>>> notebook_login()
```

## Load SceneParse150 dataset

Start by loading a smaller subset of the SceneParse150 dataset from the 🤗 Datasets library. This'll give you a chance to experiment and make sure everythings works before spending more time training on the full dataset.

```py
>>> from datasets import load_dataset

>>> ds = load_dataset("scene_parse_150", split="train[:50]")
```

Split the dataset's `train` split into a train and test set with the [`~datasets.Dataset.train_test_split`] method:

```py
>>> ds = ds.train_test_split(test_size=0.2)
>>> train_ds = ds["train"]
>>> test_ds = ds["test"]
```

Then take a look at an example:

```py
>>> train_ds[0]
{'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=512x683 at 0x7F9B0C201F90>,
 'annotation': <PIL.PngImagePlugin.PngImageFile image mode=L size=512x683 at 0x7F9B0C201DD0>,
 'scene_category': 368}
```

- `image`: a PIL image of the scene.
- `annotation`: a PIL image of the segmentation map, which is also the model's target.
- `scene_category`: a category id that describes the image scene like "kitchen" or "office". In this guide, you'll only need `image` and `annotation`, both of which are PIL images.

You'll also want to create a dictionary that maps a label id to a label class which will be useful when you set up the model later. Download the mappings from the Hub and create the `id2label` and `label2id` dictionaries:

```py
>>> import json
>>> from huggingface_hub import cached_download, hf_hub_url

>>> repo_id = "huggingface/label-files"
>>> filename = "ade20k-id2label.json"
>>> id2label = json.load(open(cached_download(hf_hub_url(repo_id, filename, repo_type="dataset")), "r"))
>>> id2label = {int(k): v for k, v in id2label.items()}
>>> label2id = {v: k for k, v in id2label.items()}
>>> num_labels = len(id2label)
```

## Preprocess

The next step is to load a SegFormer image processor to prepare the images and annotations for the model. Some datasets, like this one, use the zero-index as the background class. However, the background class isn't actually included in the 150 classes, so you'll need to set `reduce_labels=True` to subtract one from all the labels. The zero-index is replaced by `255` so it's ignored by SegFormer's loss function:

```py
>>> from transformers import AutoImageProcessor

>>> feature_extractor = AutoImageProcessor.from_pretrained("nvidia/mit-b0", reduce_labels=True)
```

It is common to apply some data augmentations to an image dataset to make a model more robust against overfitting. In this guide, you'll use the [`ColorJitter`](https://pytorch.org/vision/stable/generated/torchvision.transforms.ColorJitter.html) function from [torchvision](https://pytorch.org/vision/stable/index.html) to randomly change the color properties of an image, but you can also use any image library you like.

```py
>>> from torchvision.transforms import ColorJitter

>>> jitter = ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.1)
```

Now create two preprocessing functions to prepare the images and annotations for the model. These functions convert the images into `pixel_values` and annotations to `labels`. For the training set, `jitter` is applied before providing the images to the image processor. For the test set, the image processor crops and normalizes the `images`, and only crops the `labels` because no data augmentation is applied during testing.

```py
>>> def train_transforms(example_batch):
...     images = [jitter(x) for x in example_batch["image"]]
...     labels = [x for x in example_batch["annotation"]]
...     inputs = feature_extractor(images, labels)
...     return inputs


>>> def val_transforms(example_batch):
...     images = [x for x in example_batch["image"]]
...     labels = [x for x in example_batch["annotation"]]
...     inputs = feature_extractor(images, labels)
...     return inputs
```

To apply the `jitter` over the entire dataset, use the 🤗 Datasets [`~datasets.Dataset.set_transform`] function. The transform is applied on the fly which is faster and consumes less disk space:

```py
>>> train_ds.set_transform(train_transforms)
>>> test_ds.set_transform(val_transforms)
```

## Evaluate

Including a metric during training is often helpful for evaluating your model's performance. You can quickly load a evaluation method with the 🤗 [Evaluate](https://huggingface.co/docs/evaluate/index) library. For this task, load the [mean Intersection over Union](https://huggingface.co/spaces/evaluate-metric/accuracy) (IoU) metric (see the 🤗 Evaluate [quick tour](https://huggingface.co/docs/evaluate/a_quick_tour) to learn more about how to load and compute a metric):

```py
>>> import evaluate

>>> metric = evaluate.load("mean_iou")
```

Then create a function to [`~evaluate.EvaluationModule.compute`] the metrics. Your predictions need to be converted to logits first, and then reshaped to match the size of the labels before you can call [`~evaluate.EvaluationModule.compute`]:

```py
>>> def compute_metrics(eval_pred):
...     with torch.no_grad():
...         logits, labels = eval_pred
...         logits_tensor = torch.from_numpy(logits)
...         logits_tensor = nn.functional.interpolate(
...             logits_tensor,
...             size=labels.shape[-2:],
...             mode="bilinear",
...             align_corners=False,
...         ).argmax(dim=1)

...         pred_labels = logits_tensor.detach().cpu().numpy()
...         metrics = metric.compute(
...             predictions=pred_labels,
...             references=labels,
...             num_labels=num_labels,
...             ignore_index=255,
...             reduce_labels=False,
...         )
...         for key, value in metrics.items():
...             if type(value) is np.ndarray:
...                 metrics[key] = value.tolist()
...         return metrics
```

Your `compute_metrics` function is ready to go now, and you'll return to it when you setup your training.

## Train

<Tip>

If you aren't familiar with finetuning a model with the [`Trainer`], take a look at the basic tutorial [here](../training#finetune-with-trainer)!

</Tip>

You're ready to start training your model now! Load SegFormer with [`AutoModelForSemanticSegmentation`], and pass the model the mapping between label ids and label classes:

```py
>>> from transformers import AutoModelForSemanticSegmentation, TrainingArguments, Trainer

>>> pretrained_model_name = "nvidia/mit-b0"
>>> model = AutoModelForSemanticSegmentation.from_pretrained(
...     pretrained_model_name, id2label=id2label, label2id=label2id
... )
```

At this point, only three steps remain:

1. Define your training hyperparameters in [`TrainingArguments`]. It is important you don't remove unused columns because this'll drop the `image` column. Without the `image` column, you can't create `pixel_values`. Set `remove_unused_columns=False` to prevent this behavior! The only other required parameter is `output_dir` which specifies where to save your model. You'll push this model to the Hub by setting `push_to_hub=True` (you need to be signed in to Hugging Face to upload your model). At the end of each epoch, the [`Trainer`] will evaluate the IoU metric and save the training checkpoint.
2. Pass the training arguments to [`Trainer`] along with the model, dataset, tokenizer, data collator, and `compute_metrics` function.
3. Call [`~Trainer.train`] to finetune your model.

```py
>>> training_args = TrainingArguments(
...     output_dir="segformer-b0-scene-parse-150",
...     learning_rate=6e-5,
...     num_train_epochs=50,
...     per_device_train_batch_size=2,
...     per_device_eval_batch_size=2,
...     save_total_limit=3,
...     evaluation_strategy="steps",
...     save_strategy="steps",
...     save_steps=20,
...     eval_steps=20,
...     logging_steps=1,
...     eval_accumulation_steps=5,
...     remove_unused_columns=False,
...     push_to_hub=True,
... )

>>> trainer = Trainer(
...     model=model,
...     args=training_args,
...     train_dataset=train_ds,
...     eval_dataset=test_ds,
...     compute_metrics=compute_metrics,
... )

>>> trainer.train()
```

Once training is completed, share your model to the Hub with the [`~transformers.Trainer.push_to_hub`] method so everyone can use your model:

```py
>>> trainer.push_to_hub()
```

## Inference

Great, now that you've finetuned a model, you can use it for inference!

Load an image for inference:

```py
>>> image = ds[0]["image"]
>>> image
```

<div class="flex justify-center">
    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/semantic-seg-image.png" alt="Image of bedroom"/>
</div>

The simplest way to try out your finetuned model for inference is to use it in a [`pipeline`]. Instantiate a `pipeline` for image segmentation with your model, and pass your image to it:

```py
>>> from transformers import pipeline

>>> segmenter = pipeline("image-segmentation", model="my_awesome_seg_model")
>>> segmenter(image)
[{'score': None,
  'label': 'wall',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062690>},
 {'score': None,
  'label': 'sky',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062A50>},
 {'score': None,
  'label': 'floor',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062B50>},
 {'score': None,
  'label': 'ceiling',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062A10>},
 {'score': None,
  'label': 'bed ',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062E90>},
 {'score': None,
  'label': 'windowpane',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062390>},
 {'score': None,
  'label': 'cabinet',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062550>},
 {'score': None,
  'label': 'chair',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062D90>},
 {'score': None,
  'label': 'armchair',
  'mask': <PIL.Image.Image image mode=L size=640x427 at 0x7FD5B2062E10>}]
```

You can also manually replicate the results of the `pipeline` if you'd like. Process the image with an image processor and place the `pixel_values` on a GPU:

```py
>>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # use GPU if available, otherwise use a CPU
>>> encoding = feature_extractor(image, return_tensors="pt")
>>> pixel_values = encoding.pixel_values.to(device)
```

Pass your input to the model and return the `logits`:

```py
>>> outputs = model(pixel_values=pixel_values)
>>> logits = outputs.logits.cpu()
```

Next, rescale the logits to the original image size:

```py
>>> upsampled_logits = nn.functional.interpolate(
...     logits,
...     size=image.size[::-1],
...     mode="bilinear",
...     align_corners=False,
... )

>>> pred_seg = upsampled_logits.argmax(dim=1)[0]
```

To visualize the results, load the [dataset color palette](https://github.com/tensorflow/models/blob/3f1ca33afe3c1631b733ea7e40c294273b9e406d/research/deeplab/utils/get_dataset_colormap.py#L51) that maps each class to their RGB values. Then you can combine and plot your image and the predicted segmentation map:

```py
>>> import matplotlib.pyplot as plt

>>> color_seg = np.zeros((pred_seg.shape[0], pred_seg.shape[1], 3), dtype=np.uint8)
>>> palette = np.array(ade_palette())
>>> for label, color in enumerate(palette):
...     color_seg[pred_seg == label, :] = color
>>> color_seg = color_seg[..., ::-1]  # convert to BGR

>>> img = np.array(image) * 0.5 + color_seg * 0.5  # plot the image with the segmentation map
>>> img = img.astype(np.uint8)

>>> plt.figure(figsize=(15, 10))
>>> plt.imshow(img)
>>> plt.show()
```

<div class="flex justify-center">
    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/semantic-seg-preds.png" alt="Image of bedroom overlaid with segmentation map"/>
</div>