Using exported VL metadata

import time
from typing import Any, Dict

import requests


class DatasetExporter:
    def __init__(self, base_url: str):
        self.base_url = base_url.rstrip("/")
        self.headers = {"Accept": "application/json, text/plain, */*"}

    def initiate_export(
        self,
        dataset_id: str,
        file_name: str,
        export_format: str = "json",
        include_images: bool = False,
    ) -> Dict[str, Any]:
        """
        Initiate an export of a dataset.

        Args:
            dataset_id: The ID of the dataset to export
            file_name: Name of the export file
            export_format: Format of the export (default: 'json')
            include_images: Whether to include images in export (default: False)

        Returns:
            Dict containing the export task information
        """
        url = f"{self.base_url}/api/v1/dataset/{dataset_id}/export_context_async"
        params = {
            "file_name": file_name,
            "export_format": export_format,
            "include_images": str(include_images).lower(),
        }

        response = requests.get(url, headers=self.headers, params=params)
        response.raise_for_status()
        return response.json()

    def check_export_status(
        self, dataset_id: str, export_task_id: str
    ) -> Dict[str, Any]:
        """
        Check the status of an export task.

        Args:
            dataset_id: The ID of the dataset
            export_task_id: The ID of the export task to check

        Returns:
            Dict containing the status information
        """
        url = f"{self.base_url}/api/v1/dataset/{dataset_id}/export_status"
        params = {"export_task_id": export_task_id}

        response = requests.get(url, headers=self.headers, params=params)
        response.raise_for_status()
        return response.json()

    def wait_for_export(
        self,
        dataset_id: str,
        export_task_id: str,
        check_interval: int = 5,
        timeout: int = 300,
    ) -> Dict[str, Any]:
        """
        Wait for an export task to complete.

        Args:
            dataset_id: The ID of the dataset
            export_task_id: The ID of the export task to check
            check_interval: Time in seconds between status checks (default: 5)
            timeout: Maximum time to wait in seconds (default: 300)

        Returns:
            Dict containing the final status information

        Raises:
            TimeoutError: If the export doesn't complete within the timeout period
        """
        start_time = time.time()
        while True:
            status = self.check_export_status(dataset_id, export_task_id)

            if status["status"] == "COMPLETED":
                return status

            if status["status"] == "FAILED":
                raise Exception(f"Export failed: {status.get('result_message')}")

            if time.time() - start_time > timeout:
                raise TimeoutError("Export timed out")

            time.sleep(check_interval)

    def download_export(self, download_uri: str, output_path: str) -> None:
        """
        Download the exported file from the given URI.

        Args:
            download_uri: The URI path from the export status
            output_path: Local path where the file should be saved

        Returns:
            None
        """
        url = f"{self.base_url}{download_uri}"
        response = requests.get(url, headers=self.headers, stream=True)
        response.raise_for_status()

        with open(output_path, "wb") as f:
            for chunk in response.iter_content(chunk_size=8192):
                f.write(chunk)

import zipfile
import os

def main():
    # Replace with your own
    url = "http://3.129.25.115:2080"
    dataset_id = "4cd4d9a2-e793-11ef-80a1-0242ac13000a"
    file_name = "export.zip"

    exporter = DatasetExporter(url)

    try:
        print("Initiating export...")
        export_task = exporter.initiate_export(dataset_id, file_name)
        print(f"Export task created: {export_task}")

        print("Waiting for export to complete...")
        final_status = exporter.wait_for_export(dataset_id, export_task["id"])

        print("\nExport completed!")
        print(f"Download URI: {url}{final_status['download_uri']}")

        print("Downloading file...")
        exporter.download_export(final_status["download_uri"], file_name)
        print(f"File downloaded successfully to: {file_name}")

        print("Unzipping file...")
        output_dir = file_name.replace('.zip', '')
        os.makedirs(output_dir, exist_ok=True)
        with zipfile.ZipFile(file_name, "r") as zip_ref:
            zip_ref.extractall(output_dir)
        print(f"File unzipped successfully to: {output_dir}")

    except Exception as e:
        print(f"Error: {e}")


if __name__ == "__main__":
    main()

Initiating export...  
Export task created: {'id': 'cf8b3e78-cb20-4954-a782-436a03006dab', 'dataset_id': '4cd4d9a2-e793-11ef-80a1-0242ac13000a', 'created_at': '2025-02-10T14:41:14.390429', 'download_uri': None, 'progress': 0.0, 'status': 'INIT', 'entities_count': 4203, 'result_message': None}  
Waiting for export to complete...

Export completed!  
Download URI: http://3.129.25.115:2080/cdn/4cd4d9a2-e793-11ef-80a1-0242ac13000a/images/cf8b3e78-cb20-4954-a782-436a03006dab/cf8b3e78-cb20-4954-a782-436a03006dab.zip  
Downloading file...  
File downloaded successfully to: export.zip  
Unzipping file...  
File unzipped successfully to: export

import pandas as pd
import json

with open("food101/metadata.json") as f:
    data = json.load(f)

# This is the dataset level information
info_df = pd.DataFrame([data["info"]])

# This is the image level information
media_items_df = pd.json_normalize(data["media_items"])

info_df

media_items_df

UNIQUENESS_SCORE_THRESHOLD = 0.85

coreset_df = media_items_df[
    media_items_df["uniqueness_score"] > UNIQUENESS_SCORE_THRESHOLD
]
coreset_df

def has_duplicate_issue(metadata_items, confidence_threshold=0.8):
    if not isinstance(metadata_items, list):
        return False

    for item in metadata_items:
        if (
            item.get("type") == "issue"
            and item.get("properties", {}).get("issue_type") == "duplicates"
            and item.get("properties", {}).get("confidence", 0) > confidence_threshold
        ):
            return True
    return False

# Replace with your confidence threshold
CONFIDENCE_THRESHOLD = 0.8

duplicate_df = media_items_df[
    media_items_df["metadata_items"].apply(
        lambda x: has_duplicate_issue(x, confidence_threshold=CONFIDENCE_THRESHOLD)
    )
]

duplicate_df

def has_mislabel_issue(metadata_items, confidence_threshold=0.8):
    if not isinstance(metadata_items, list):
        return False

    for item in metadata_items:
        if (
            item.get("type") == "issue"
            and item.get("properties", {}).get("issue_type") == "mislabels"
            and item.get("properties", {}).get("confidence", 0) > confidence_threshold
        ):
            return True
    return False

# Replace with your confidence threshold
CONFIDENCE_THRESHOLD = 0.8

mislabel_df = media_items_df[
    media_items_df["metadata_items"].apply(
        lambda x: has_mislabel_issue(x, confidence_threshold=CONFIDENCE_THRESHOLD)
    )
]

mislabel_df

{
  "info": {
    "schema_version": "1.1",
    "dataset": "Extreme",
    "description": "Exported from Extreme at Visual Layer",
    "dataset_url": "http://app.visual-layer.com/dataset/42e3b7ae-dd5e-11ef-9fca-1a226f3de670/data",
    "export_time": "2025-02-15T18:23:34.164184",
    "dataset_creation_time": "2025-01-24T18:35:57.719308",
    "exported_by": "Danny Bickson",
    "total_media_items": 56
  },
  "media_items": [
    {
      "media_id": "8621054d-1d23-4ca7-a7ca-65feb2c83eb1",
      "media_type": "video_frame",
      "file_name": "output_000001.jpg",
      "file_path": "7370082532203988000_7450129782785756438_2024-12-19_14:28:36.mp4/output_000001.jpg",
      "file_size": "48.27KB",
      "uniqueness_score": 0.4989460029187489,
      "height": 720,
      "width": 576,
      "url": "http://app.visual-layer.com/dataset/42e3b7ae-dd5e-11ef-9fca-1a226f3de670/data/image/8621054d-1d23-4ca7-a7ca-65feb2c83eb1",
      "cluster_id": "49997382-b7c3-4c83-83a3-fa4e0a586f2b",
      "metadata_items": [
        {
          "type": "video_info",
          "properties": {
            "video_name": "7370082532203988000_7450129782785756438_2024-12-19_14:28:36.mp4",
            "frame_timestamp": 0.0
          }
        },
        {
          "type": "issue",
          "properties": {
            "issue_type": "duplicates",
            "confidence": 0.977624,
            "duplicate_group_id": "35395121-5cb1-4a92-906f-dd4cb9ff62d3",
            "duplicate_threshold": 0.96
          }
        }
      ]
    }
  ]
}

import json  
import pandas as pd  
import re

# Load JSON file
with open("metadata 8.json", "r") as f:  
    data = json.load(f)

# Prepare list for DataFrame
records = []

# Iterate over media items
for item in data.get("media_items", []):  
    video_name = None  
    frame_timestamp = None  
    categories = []  
    file_name = item.get("file_name", "")

    for metadata in item.get("metadata_items", []):
        if metadata["type"] == "video_info":
            video_name = metadata["properties"].get("video_name")
            frame_timestamp = metadata["properties"].get("frame_timestamp")
        elif metadata["type"] == "image_label":
            categories.append(metadata["properties"].get("category_name"))

    # Extract frame number from file name
    match = re.search(r'output_(\d+).jpg', file_name)
    frame_number = int(match.group(1)) if match else None

    # Append record with categories as a list
    records.append({
        "video_name": video_name,
        "file_name": file_name,
        "frame_number": frame_number,
        "time_in_video": frame_timestamp,
        "categories": categories
    })

# Convert to DataFrame
df = pd.DataFrame(records)  
print(df.tail())

{
  "info": {
    "schema_version": "1.1",
    "dataset": "General pics",
    "description": "Exported from General pics at Visual Layer",
    "total_media_items": 2
  },
  "media_items": [
    {
      "media_id": "01305347-b812-4717-99a3-aab54a206d4c",
      "file_name": "000000046252.jpg",
      "width": 640,
      "height": 480,
      "metadata_items": [
        {
          "type": "object_label",
          "properties": {
            "category_name": "person",
            "bbox": [148, 177, 127, 188],
            "metadata_items": [
              {
                "type": "user_tag",
                "properties": {"tag_name": "test-tags"}
              }
            ]
          }
        },
        {
          "type": "object_label",
          "properties": {
            "category_name": "shirt",
            "bbox": [192, 194, 75, 65]
          }
        }
      ]
    }
  ]
}

import json
import csv

# Load JSON file
input_file = "metadata_5.json"  # Change this to your actual file path
output_file = "output.csv"  # Output CSV file name

with open(input_file, "r") as f:
    data = json.load(f)

# Prepare CSV file
header = ["filename", "x", "y", "w", "h", "label", "user tag"]
rows = []

# Extract data
for media in data.get("media_items", []):
    filename = media.get("file_name", "")
    for item in media.get("metadata_items", []):
        if item.get("type") == "object_label":
            properties = item.get("properties", {})
            bbox = properties.get("bbox", [0, 0, 0, 0])
            label = properties.get("category_name", "")
            user_tag = ""

            # Extract user tag if available
            for meta in properties.get("metadata_items", []):
                if meta.get("type") == "user_tag":
                    user_tag = meta.get("properties", {}).get("tag_name", "")
                    break

            # Append row
            rows.append([filename, bbox[0], bbox[1], bbox[2], bbox[3], label, user_tag])

# Write to CSV
with open(output_file, "w", newline="") as f:
    writer = csv.writer(f)
    writer.writerow(header)
    writer.writerows(rows)

print(f"CSV file saved to {output_file}")

import json
import pandas as pd
from collections import defaultdict

# Load the exported metadata
with open("metadata.json", "r") as f:
    data = json.load(f)

# Extract video frame data with similarity information
frame_records = []

for item in data.get("media_items", []):
    if item.get("media_type") == "video_frame":
        # Extract basic frame info
        media_id = item.get("media_id")
        file_name = item.get("file_name")
        
        # Extract video info
        video_name = None
        frame_timestamp = None
        similar_frames = []
        
        for metadata in item.get("metadata_items", []):
            if metadata.get("type") == "video_info":
                video_name = metadata.get("properties", {}).get("video_name")
                frame_timestamp = metadata.get("properties", {}).get("frame_timestamp")
            elif metadata.get("type") == "issue" and metadata.get("properties", {}).get("issue_type") == "duplicates":
                # Extract similarity information
                confidence = metadata.get("properties", {}).get("confidence")
                duplicate_group_id = metadata.get("properties", {}).get("duplicate_group_id")
                duplicate_threshold = metadata.get("properties", {}).get("duplicate_threshold")
                
                similar_frames.append({
                    "confidence": confidence,
                    "group_id": duplicate_group_id,
                    "threshold": duplicate_threshold
                })
        
        frame_records.append({
            "media_id": media_id,
            "file_name": file_name,
            "video_name": video_name,
            "frame_timestamp": frame_timestamp,
            "similar_frames": similar_frames,
            "cluster_id": item.get("cluster_id")
        })

# Create DataFrame
frames_df = pd.DataFrame(frame_records)
print(f"Total video frames: {len(frames_df)}")
print(frames_df.head())

# Create a mapping of frames to their duplicate groups
frame_to_groups = defaultdict(set)
group_to_frames = defaultdict(set)

for idx, row in frames_df.iterrows():
    media_id = row['media_id']
    video_name = row['video_name']
    
    for similar_frame in row['similar_frames']:
        group_id = similar_frame['group_id']
        confidence = similar_frame['confidence']
        
        # Only consider high-confidence similarities (>= 0.95)
        if confidence >= 0.95:
            frame_to_groups[media_id].add(group_id)
            group_to_frames[group_id].add((media_id, video_name, confidence))

# Create video-to-video similarity matrix
video_similarity_records = []

# Get all unique videos
all_videos = frames_df['video_name'].unique()

for video_a in all_videos:
    for video_b in all_videos:
        if video_a != video_b:
            # Get frames from both videos
            frames_a = frames_df[frames_df['video_name'] == video_a]['media_id'].tolist()
            frames_b = frames_df[frames_df['video_name'] == video_b]['media_id'].tolist()
            
            # Find similarities between frames from different videos
            inter_video_similarities = []
            shared_groups = set()
            
            for frame_a in frames_a:
                for frame_b in frames_b:
                    groups_a = frame_to_groups.get(frame_a, set())
                    groups_b = frame_to_groups.get(frame_b, set())
                    
                    # Find groups that contain both frames (cross-video similarity)
                    common_groups = groups_a.intersection(groups_b)
                    if common_groups:
                        shared_groups.update(common_groups)
                        
                        # For each shared group, get the similarity confidence
                        # We know these frames are from different videos by construction
                        for group_id in common_groups:
                            frames_in_group = group_to_frames[group_id]
                            # Get confidence scores for frames from both videos in this group
                            confidences_in_group = [conf for _, _, conf in frames_in_group]
                            if confidences_in_group:
                                # Use the average confidence for this similarity group
                                inter_video_similarities.append(sum(confidences_in_group) / len(confidences_in_group))
            
            if inter_video_similarities:
                avg_similarity = sum(inter_video_similarities) / len(inter_video_similarities)
                num_similar_frames = len(shared_groups)
                
                video_similarity_records.append({
                    "video_a": video_a,
                    "video_b": video_b,
                    "average_similarity": round(avg_similarity, 4),
                    "number_of_similar_frames": num_similar_frames
                })

# Create the final similarity DataFrame
similarity_df = pd.DataFrame(video_similarity_records)

# Remove duplicates (keep only one direction of each pair)
similarity_df = similarity_df[similarity_df['video_a'] < similarity_df['video_b']]

# Sort by average similarity (descending)
similarity_df = similarity_df.sort_values('average_similarity', ascending=False)

print(f"Video pairs with similarities: {len(similarity_df)}")
print("\nTop 10 most similar video pairs:")
print(similarity_df.head(10))

# Create similarity table using cluster IDs
cluster_similarity_records = []

# Group frames by cluster
cluster_groups = frames_df.groupby('cluster_id')

for cluster_id, cluster_frames in cluster_groups:
    if len(cluster_frames) > 1:  # Only clusters with multiple frames
        videos_in_cluster = cluster_frames['video_name'].unique()
        
        # Create pairs of videos in the same cluster
        for i, video_a in enumerate(videos_in_cluster):
            for video_b in videos_in_cluster[i+1:]:
                frames_a_in_cluster = len(cluster_frames[cluster_frames['video_name'] == video_a])
                frames_b_in_cluster = len(cluster_frames[cluster_frames['video_name'] == video_b])
                
                cluster_similarity_records.append({
                    "video_a": video_a,
                    "video_b": video_b,
                    "cluster_id": cluster_id,
                    "frames_from_video_a": frames_a_in_cluster,
                    "frames_from_video_b": frames_b_in_cluster,
                    "total_frames_in_cluster": len(cluster_frames)
                })

# Create DataFrame
cluster_similarity_df = pd.DataFrame(cluster_similarity_records)

# Aggregate by video pair
video_cluster_summary = cluster_similarity_df.groupby(['video_a', 'video_b']).agg({
    'cluster_id': 'count',  # Number of shared clusters
    'frames_from_video_a': 'sum',
    'frames_from_video_b': 'sum'
}).rename(columns={'cluster_id': 'shared_clusters'}).reset_index()

print("\nVideo similarity based on shared clusters:")
print(video_cluster_summary.sort_values('shared_clusters', ascending=False).head())

# Export the main similarity table
similarity_df.to_csv("video_similarity_analysis.csv", index=False)

# Export the frame-level data
frames_df.to_csv("video_frames_with_similarity.csv", index=False)

# Export the cluster-based analysis
video_cluster_summary.to_csv("video_cluster_similarity.csv", index=False)

print("Analysis complete! Files exported:")
print("- video_similarity_analysis.csv")
print("- video_frames_with_similarity.csv") 
print("- video_cluster_similarity.csv")

{
  "type": "issue",
  "properties": {
    "issue_type": "mislabels",
    "issues_description": "3D-CONTAMINATION",
    "confidence": 0.8059
  }
}

import json
import pandas as pd
from collections import defaultdict, Counter

# Load the exported metadata
with open("metadata.json", "r") as f:
    data = json.load(f)

print(f"Dataset: {data['info']['dataset']}")
print(f"Total media items: {data['info']['total_media_items']}")

# Extract mislabel information
mislabel_records = []
total_objects = 0
objects_with_mislabels = 0

for item in data.get("media_items", []):
    media_id = item.get("media_id")
    file_name = item.get("file_name")
    
    # Look for object labels with mislabel issues
    for metadata in item.get("metadata_items", []):
        if metadata.get("type") == "object_label":
            total_objects += 1
            object_props = metadata.get("properties", {})
            original_category = object_props.get("category_name")
            bbox = object_props.get("bbox", [])
            
            # Check for mislabel issues in this object
            object_mislabels = []
            for obj_metadata in object_props.get("metadata_items", []):
                if (obj_metadata.get("type") == "issue" and 
                    obj_metadata.get("properties", {}).get("issue_type") == "mislabels"):
                    
                    mislabel_props = obj_metadata.get("properties", {})
                    suggested_label = mislabel_props.get("issues_description")
                    confidence = mislabel_props.get("confidence")
                    
                    object_mislabels.append({
                        "suggested_label": suggested_label,
                        "confidence": confidence
                    })
            
            if object_mislabels:
                objects_with_mislabels += 1
                for mislabel in object_mislabels:
                    mislabel_records.append({
                        "media_id": media_id,
                        "file_name": file_name,
                        "original_label": original_category,
                        "suggested_label": mislabel["suggested_label"],
                        "confidence": mislabel["confidence"],
                        "bbox": bbox
                    })

print(f"Total objects analyzed: {total_objects}")
print(f"Objects with mislabel issues: {objects_with_mislabels}")
print(f"Total mislabel records: {len(mislabel_records)}")

# Create DataFrame for analysis
mislabels_df = pd.DataFrame(mislabel_records)

# 1. How many mislabels total?
print(f"TOTAL MISLABELS: {len(mislabels_df)}")

# 2. Confidence distribution
print(f"\nCONFIDENCE DISTRIBUTION:")
print(f"Average confidence: {mislabels_df['confidence'].mean():.4f}")
print(f"Min confidence: {mislabels_df['confidence'].min():.4f}")
print(f"Max confidence: {mislabels_df['confidence'].max():.4f}")
print(f"High confidence (>0.8): {len(mislabels_df[mislabels_df['confidence'] > 0.8])}")
print(f"Medium confidence (0.7-0.8): {len(mislabels_df[(mislabels_df['confidence'] >= 0.7) & (mislabels_df['confidence'] <= 0.8)])}")
print(f"Lower confidence (<0.7): {len(mislabels_df[mislabels_df['confidence'] < 0.7])}")

# 3. From which classes are the mislabels?
print(f"\nMISLABELED CLASSES (Original Labels):")
original_label_counts = mislabels_df['original_label'].value_counts()
print(original_label_counts.head(10))

# 4. What are the suggested corrections?
print(f"\nSUGGESTED CORRECTIONS (Most Common):")
suggested_label_counts = mislabels_df['suggested_label'].value_counts()
print(suggested_label_counts.head(10))

# 5. Most common mislabel corrections (original -> suggested)
print(f"\nMOST COMMON MISLABEL CORRECTIONS:")
correction_pairs = mislabels_df.groupby(['original_label', 'suggested_label']).size().reset_index(name='count')
correction_pairs = correction_pairs.sort_values('count', ascending=False)
print("Original Label -> Suggested Label (Count)")
for _, row in correction_pairs.head(15).iterrows():
    print(f"{row['original_label']} -> {row['suggested_label']} ({row['count']})")

# Analyze high-confidence corrections for the most reliable suggestions
high_conf_corrections = correction_pairs[
    correction_pairs.apply(lambda x: mislabels_df[
        (mislabels_df['original_label'] == x['original_label']) & 
        (mislabels_df['suggested_label'] == x['suggested_label'])
    ]['confidence'].mean() > 0.8, axis=1)
]

print(f"High confidence corrections (>0.8): {len(high_conf_corrections)}")
print("\nTop high-confidence corrections:")
for _, row in high_conf_corrections.head(10).iterrows():
    avg_conf = mislabels_df[
        (mislabels_df['original_label'] == row['original_label']) & 
        (mislabels_df['suggested_label'] == row['suggested_label'])
    ]['confidence'].mean()
    print(f"{row['original_label']} -> {row['suggested_label']} ({row['count']} times, avg conf: {avg_conf:.3f})")

# Summary statistics
print(f"\nSUMMARY STATISTICS:")
print(f"Unique original labels with issues: {len(mislabels_df['original_label'].unique())}")
print(f"Unique suggested labels: {len(mislabels_df['suggested_label'].unique())}")
print(f"Unique correction pairs: {len(correction_pairs)}")
print(f"Images with mislabels: {len(mislabels_df['file_name'].unique())}")

# Export results
mislabels_df.to_csv("mislabel_analysis.csv", index=False)
correction_pairs.to_csv("mislabel_corrections.csv", index=False)

print("\nFiles exported:")
print("- mislabel_analysis.csv: Detailed mislabel data")
print("- mislabel_corrections.csv: Correction patterns summary")

TOTAL MISLABELS: 422

CONFIDENCE DISTRIBUTION:
Average confidence: 0.8059
High confidence (>0.8): 120
Medium confidence (0.7-0.8): 300

MISLABELED CLASSES (Original Labels):
3D-SHINYMARK         52
3D-SCRATCH           40
3D-CONTAMINATION     38
3D-SHINYLINE         28
3D-DENT              28

MOST COMMON MISLABEL CORRECTIONS:
Bandslot_discolor -> BS-DISCOLOR-B (20)
3D-SHINYMARK -> 3D-CONTAMINATION (18)
3D-SHINYLINE -> 3D-CONTAMINATION (12)
3D-SCRATCH -> 3D-CONTAMINATION (10)
3D-CHIP -> 3D2-CHIP (10)