active-vision 0.0.2__py3-none-any.whl → 0.0.4__py3-none-any.whl

active_vision/__init__.py

@@ -1,3 +1,3 @@
-__version__ = "0.0.2"
+__version__ = "0.0.4"
 
 from .core import *

active_vision/core.py

@@ -87,7 +87,9 @@ class ActiveLearner:
         )
         return self.pred_df
 
-    def evaluate(self, df: pd.DataFrame, filepath_col: str, label_col: str, batch_size: int = 16):
+    def evaluate(
+        self, df: pd.DataFrame, filepath_col: str, label_col: str, batch_size: int = 16
+    ):
         """
         Evaluate on a labeled dataset. Returns a score.
         """
@@ -114,20 +116,158 @@ class ActiveLearner:
         """
         Sample top `num_samples` low confidence samples. Returns a df with filepaths and predicted labels, and confidence scores.
         """
-        uncertain_df = df.sort_values(
-            by="pred_conf", ascending=True
-        ).head(num_samples)
+        logger.info(f"Getting top {num_samples} low confidence samples")
+        uncertain_df = df.sort_values(by="pred_conf", ascending=True).head(num_samples)
         return uncertain_df
 
-    def add_to_train_set(self, df: pd.DataFrame):
+    def label(self, df: pd.DataFrame, output_filename: str = "labeled"):
+        """
+        Launch a labeling interface for the user to label the samples.
+        Input is a df with filepaths listing the files to be labeled. Output is a df with filepaths and labels.
+        """
+        import gradio as gr
+
+        shortcut_js = """
+        <script>
+        function shortcuts(e) {
+            // Only block shortcuts if we're in a text input or textarea
+            if (e.target.tagName.toLowerCase() === "textarea" || 
+                (e.target.tagName.toLowerCase() === "input" && e.target.type.toLowerCase() === "text")) {
+                return;
+            }
+            
+            if (e.key.toLowerCase() == "w") {
+                document.getElementById("submit_btn").click();
+            } else if (e.key.toLowerCase() == "d") {
+                document.getElementById("next_btn").click();
+            } else if (e.key.toLowerCase() == "a") {
+                document.getElementById("back_btn").click();
+            }
+        }
+        document.addEventListener('keypress', shortcuts, false);
+        </script>
+        """
+
+        logger.info(f"Launching labeling interface for {len(df)} samples")
+
+        filepaths = df["filepath"].tolist()
+
+        with gr.Blocks(head=shortcut_js) as demo:
+            current_index = gr.State(value=0)
+
+            filename = gr.Textbox(
+                label="Filename", value=filepaths[0], interactive=False
+            )
+
+            image = gr.Image(
+                type="filepath", label="Image", value=filepaths[0], height=500
+            )
+            category = gr.Radio(choices=self.class_names, label="Select Category")
+
+            with gr.Row():
+                back_btn = gr.Button("← Previous (A)", elem_id="back_btn")
+                submit_btn = gr.Button(
+                    "Submit (W)",
+                    variant="primary",
+                    elem_id="submit_btn",
+                    interactive=False,
+                )
+                next_btn = gr.Button("Next → (D)", elem_id="next_btn")
+
+            progress = gr.Slider(
+                minimum=0,
+                maximum=len(filepaths) - 1,
+                value=0,
+                label="Progress",
+                interactive=False,
+            )
+
+            finish_btn = gr.Button("Finish Labeling", variant="primary")
+
+            def update_submit_btn(choice):
+                return gr.Button(interactive=choice is not None)
+
+            category.change(
+                fn=update_submit_btn, inputs=[category], outputs=[submit_btn]
+            )
+
+            def navigate(current_idx, direction):
+                next_idx = current_idx + direction
+                if 0 <= next_idx < len(filepaths):
+                    return filepaths[next_idx], filepaths[next_idx], next_idx, next_idx
+                return (
+                    filepaths[current_idx],
+                    filepaths[current_idx],
+                    current_idx,
+                    current_idx,
+                )
+
+            def save_and_next(current_idx, selected_category):
+                if selected_category is None:
+                    return (
+                        filepaths[current_idx],
+                        filepaths[current_idx],
+                        current_idx,
+                        current_idx,
+                    )
+
+                # Save the current annotation
+                with open(f"{output_filename}.csv", "a") as f:
+                    f.write(f"{filepaths[current_idx]},{selected_category}\n")
+
+                # Move to next image if not at the end
+                next_idx = current_idx + 1
+                if next_idx >= len(filepaths):
+                    return (
+                        filepaths[current_idx],
+                        filepaths[current_idx],
+                        current_idx,
+                        current_idx,
+                    )
+                return filepaths[next_idx], filepaths[next_idx], next_idx, next_idx
+
+            def convert_csv_to_parquet():
+                try:
+                    df = pd.read_csv(f"{output_filename}.csv", header=None)
+                    df.columns = ["filepath", "label"]
+                    df = df.drop_duplicates(subset=["filepath"], keep="last")
+                    df.to_parquet(f"{output_filename}.parquet")
+                    gr.Info(f"Annotation saved to {output_filename}.parquet")
+                except Exception as e:
+                    logger.error(e)
+                    return
+
+            back_btn.click(
+                fn=lambda idx: navigate(idx, -1),
+                inputs=[current_index],
+                outputs=[filename, image, current_index, progress],
+            )
+
+            next_btn.click(
+                fn=lambda idx: navigate(idx, 1),
+                inputs=[current_index],
+                outputs=[filename, image, current_index, progress],
+            )
+
+            submit_btn.click(
+                fn=save_and_next,
+                inputs=[current_index, category],
+                outputs=[filename, image, current_index, progress],
+            )
+
+            finish_btn.click(fn=convert_csv_to_parquet)
+
+        demo.launch(height=1000)
+
+    def add_to_train_set(self, df: pd.DataFrame, output_filename: str):
         """
         Add samples to the training set.
         """
         new_train_set = df.copy()
-        new_train_set.drop(columns=["pred_conf"], inplace=True)
-        new_train_set.rename(columns={"pred_label": "label"}, inplace=True)
+        # new_train_set.drop(columns=["pred_conf"], inplace=True)
+        # new_train_set.rename(columns={"pred_label": "label"}, inplace=True)
 
-        len_old = len(self.train_set)
+        # len_old = len(self.train_set)
 
         logger.info(f"Adding {len(new_train_set)} samples to training set")
         self.train_set = pd.concat([self.train_set, new_train_set])
@@ -137,13 +277,15 @@ class ActiveLearner:
         )
         self.train_set.reset_index(drop=True, inplace=True)
 
+        self.train_set.to_parquet(f"{output_filename}.parquet")
+        logger.info(f"Saved training set to {output_filename}.parquet")
 
-        if len(self.train_set) == len_old:
-            logger.warning("No new samples added to training set")
+        # if len(self.train_set) == len_old:
+        #     logger.warning("No new samples added to training set")
 
-        elif len_old + len(new_train_set) < len(self.train_set):
-            logger.warning("Some samples were duplicates and removed from training set")
+        # elif len_old + len(new_train_set) < len(self.train_set):
+        #     logger.warning("Some samples were duplicates and removed from training set")
 
-        else:
-            logger.info("All new samples added to training set")
-            logger.info(f"Training set now has {len(self.train_set)} samples")
+        # else:
+        #     logger.info("All new samples added to training set")
+        #     logger.info(f"Training set now has {len(self.train_set)} samples")

{active_vision-0.0.2.dist-info → active_vision-0.0.4.dist-info}/METADATA

@@ -1,53 +1,77 @@
 Metadata-Version: 2.2
 Name: active-vision
-Version: 0.0.2
+Version: 0.0.4
 Summary: Active learning for edge vision.
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: datasets>=3.2.0
 Requires-Dist: fastai>=2.7.18
+Requires-Dist: gradio>=5.12.0
 Requires-Dist: ipykernel>=6.29.5
 Requires-Dist: ipywidgets>=8.1.5
 Requires-Dist: loguru>=0.7.3
 Requires-Dist: seaborn>=0.13.2
+Requires-Dist: timm>=1.0.13
 
 ![Python Version](https://img.shields.io/badge/python-3.10%2B-blue?style=for-the-badge)
 ![License](https://img.shields.io/badge/License-Apache%202.0-green.svg?style=for-the-badge)
-![PyPI](https://img.shields.io/pypi/v/active-vision?style=for-the-badge)
+[![PyPI](https://img.shields.io/pypi/v/active-vision?style=for-the-badge)](https://pypi.org/project/active-vision/)
 ![Downloads](https://img.shields.io/pepy/dt/active-vision?style=for-the-badge&logo=pypi&logoColor=white&label=Downloads&color=purple)   
 
 <p align="center">
-  <img src="https://github.com/dnth/active-vision/blob/main/assets/logo.png" alt="active-vision">
+  <img src="https://raw.githubusercontent.com/dnth/active-vision/main/assets/logo.png" alt="active-vision">
 </p>
 
 Active learning at the edge for computer vision.
 
-The goal of this project is to create a framework for active learning at the edge for computer vision. We should be able to train a model on a small dataset and then use active learning to iteratively improve the model all on a local machine.
+The goal of this project is to create a framework for the active learning loop for computer vision deployed on edge devices. 
 
-## Tech Stack
+Supported tasks:
+- [X] Image classification
+- [ ] Object detection
+- [ ] Segmentation
 
-- Training framework: fastai
-- User interface: streamlit
-- Database: sqlite
-- Experiment tracking: wandb
 
 ## Installation
 
-PyPI
+Get a release from PyPI
 ```bash
 pip install active-vision
 ```
 
-Local install
+Install from source
 ```bash
 git clone https://github.com/dnth/active-vision.git
 cd active-vision
 pip install -e .
 ```
 
+I recommend using [uv](https://docs.astral.sh/uv/) to set up a virtual environment and install the package. You can also use other virtual env of your choice.
+
+If you're using uv:
+
+```bash
+uv venv
+uv sync
+```
+Once the virtual environment is created, you can install the package using pip.
+
+> [!TIP]
+> If you're using uv add a uv before the pip install command to install into your virtual environment. Eg:
+> ```bash
+> uv pip install active-vision
+> ```
+
 ## Usage
-See the [notebook](./nbs/end-to-end.ipynb) for a complete example.
+See the [notebook](./nbs/04_relabel_loop.ipynb) for a complete example.
+
+Be sure to prepared 3 datasets:
+- [initial_samples](./nbs/initial_samples.parquet): A dataframe of an existing labeled training dataset to seed the training set.
+- [unlabeled](./nbs/unlabeled_samples.parquet): A dataframe of unlabeled data which we will sample from using active learning.
+- [eval](./nbs/evaluation_samples.parquet): A dataframe of labeled data which we will use to evaluate the performance of the model.
+
+As a toy example I created the above 3 datasets from the imagenette dataset.
 
 ```python
 from active_vision import ActiveLearner
@@ -56,29 +80,45 @@ import pandas as pd
 # Create an active learner instance with a model
 al = ActiveLearner("resnet18")
 
-# Load the dataset into the active learner
+# Load dataset 
 train_df = pd.read_parquet("training_samples.parquet")
-al.load_dataset(train_df, "filepath", "label")
+al.load_dataset(df, filepath_col="filepath", label_col="label")
 
-# Train the model
+# Train model
 al.train(epochs=3, lr=1e-3)
 
-# Load evaluation data
-eval_df = pd.read_parquet("evaluation_samples.parquet")
+# Evaluate the model on a *labeled* evaluation set
+accuracy = al.evaluate(eval_df, filepath_col="filepath", label_col="label")
 
-# Evaluate the model on a labeled evaluation set
-accuracy = al.evaluate(eval_df, "filepath", "label")
-
-# Get predictions from an unlabeled set
+# Get predictions from an *unlabeled* set
 pred_df = al.predict(filepaths)
 
-# Sample low confidence predictions
+# Sample low confidence predictions from unlabeled set
 uncertain_df = al.sample_uncertain(pred_df, num_samples=10)
 
-# Add newly labeled data to training set
-al.add_to_train_set(uncertain_df)
+# Launch a Gradio UI to label the low confidence samples
+al.label(uncertain_df, output_filename="uncertain")
 ```
 
+![Gradio UI](./assets/labeling_ui.png)
+
+Once complete, the labeled samples will be save into a new df.
+We can now add the newly labeled data to the training set.
+
+```python
+# Add newly labeled data to training set and save as a new file active_labeled
+al.add_to_train_set(labeled_df, output_filename="active_labeled")
+```
+
+Repeat the process until the model is good enough. Use the dataset to train a larger model and deploy.
+
+> [!TIP]
+> For the toy dataset, I got to about 93% accuracy on the evaluation set with 200+ labeled images. The best performing model on the [leaderboard](https://github.com/fastai/imagenette) got 95.11% accuracy training on all 9469 labeled images.
+> 
+> This took me about 6 iterations of relabeling. Each iteration took about 5 minutes to complete including labeling and model training (resnet18). See the [notebook](./nbs/04_relabel_loop.ipynb) for more details.
+>
+> But using the dataset of 200+ images, I trained a more capable model (convnext_small_in22k) and got 99.3% accuracy on the evaluation set. See the [notebook](./nbs/05_retrain_larger.ipynb) for more details.
+
 ## Workflow
 There are two workflows for active learning at the edge that we can use depending on the availability of labeled data.
 
@@ -86,10 +126,10 @@ There are two workflows for active learning at the edge that we can use dependin
 If we have no labeled data, we can use active learning to iteratively improve the model and build a labeled dataset.
 
 1. Load a small proxy model.
-2. Label an initial dataset.
+2. Label an initial dataset. If there is none, you'll have to label some images.
 3. Train the proxy model on the labeled dataset.
 4. Run inference on the unlabeled dataset.
-5. Evaluate the performance of the proxy model on the unlabeled dataset.
+5. Evaluate the performance of the proxy model.
 6. Is model good enough?    
     - Yes: Save the proxy model and the dataset.
     - No: Select the most informative images to label using active learning.
@@ -141,7 +181,7 @@ graph TD
 ```
 
 
-## Methodology
+<!-- ## Methodology
 To test out the workflows we will use the [imagenette dataset](https://huggingface.co/datasets/frgfm/imagenette). But this will be applicable to any dataset.
 
 Imagenette is a subset of the ImageNet dataset with 10 classes. We will use this dataset to test out the workflows. Additionally, Imagenette has an existing leaderboard which we can use to evaluate the performance of the models.
@@ -192,4 +232,4 @@ After the first iteration we got 94.57% accuracy on the validation set. See the
 > [!TIP]
 > | Train Epochs | Number of Images | Validation Accuracy |      Source      |
 > |--------------|-----------------|----------------------|------------------|
-> | 10           | 200             | 94.57%               | First relabeling [notebook](./nbs/03_retrain_model.ipynb) |
+> | 10           | 200             | 94.57%               | First relabeling [notebook](./nbs/03_retrain_model.ipynb) | -->

active_vision-0.0.4.dist-info/RECORD

@@ -0,0 +1,7 @@
+active_vision/__init__.py,sha256=XITukjUU49hPFzxCzmxqJAUWh3YE8sWQzmyZ5bVra88,43
+active_vision/core.py,sha256=0aXDI5Gpj0Spk7TSIxJf8aJDDBgZh0-jkWdYyZ1Zric,10713
+active_vision-0.0.4.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+active_vision-0.0.4.dist-info/METADATA,sha256=WlvtrzUy8m2nr8izUuTtysdQXO4ZjCO9vGWt2i_GMUI,10421
+active_vision-0.0.4.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
+active_vision-0.0.4.dist-info/top_level.txt,sha256=7qUQvccN2UU63z5S9vrgJmqK-8sFGrtpf1e9Z86nihE,14
+active_vision-0.0.4.dist-info/RECORD,,

active_vision-0.0.2.dist-info/RECORD

@@ -1,7 +0,0 @@
-active_vision/__init__.py,sha256=5VE_DRQ_Rgbo7NlPh3-rP2pUClK48jGxPqAcptBscZ8,43
-active_vision/core.py,sha256=RBVabC350wucYl7KJgIp3fc1pS9pxtG14iDb-ZyBJxI,5262
-active_vision-0.0.2.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-active_vision-0.0.2.dist-info/METADATA,sha256=7_eqZJnGeIPjb4LLZ-Bqu1AMJ_h77_0bNRyS_COEv5w,8350
-active_vision-0.0.2.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
-active_vision-0.0.2.dist-info/top_level.txt,sha256=7qUQvccN2UU63z5S9vrgJmqK-8sFGrtpf1e9Z86nihE,14
-active_vision-0.0.2.dist-info/RECORD,,