This notebook demonstrates the use of the responsibleai API to assess an object detection pytorch model trained on the fridge dataset. It walks through the API calls necessary to create a widget with model analysis insights, then guides a visual analysis of the model.
The following section examines the code necessary to create datasets and a model. It then generates insights using the responsibleai API that can be visually analyzed.
import os
import sys
from zipfile import ZipFile
import numpy as np
import pandas as pd
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from responsibleai_vision.common.constants import ImageColumns
import json
from raiutils.common.retries import retry_function
import json
import os
import xml.etree.ElementTree as ET
try:
from urllib import urlretrieve
except ImportError:
from urllib.request import urlretrieve
EPOCHS = 10
LEARNING_RATE = 1e-4
IM_SIZE = 300
BATCH_SIZE = 16
FRIDGE_MODEL_NAME = 'fridge_od_model'
FRIDGE_MODEL_WINDOWS_NAME = 'fridge_model_windows'
WIN = 'win'
def load_fridge_object_detection_dataset_labels():
src_images = "./data/odFridgeObjects/"
# Path to the annotations
annotations_folder = os.path.join(src_images, "annotations")
labels = []
label_dict = {'can': 1, 'carton': 2, 'milk_bottle': 3, 'water_bottle': 4}
# Read each annotation
for i, filename in enumerate(os.listdir(annotations_folder)):
if filename.endswith(".xml"):
print("Parsing " + os.path.join(src_images, filename))
root = ET.parse(os.path.join(annotations_folder, filename)).getroot()
# use if needed
# width = int(root.find("size/width").text)
# height = int(root.find("size/height").text)
image_labels = []
for object in root.findall("object"):
name = object.find("name").text
xmin = object.find("bndbox/xmin").text
ymin = object.find("bndbox/ymin").text
xmax = object.find("bndbox/xmax").text
ymax = object.find("bndbox/ymax").text
isCrowd = int(object.find("difficult").text)
image_labels.append([
label_dict[name], # label
float(xmin), # topX. To normalize, divide by width.
float(ymin), # topY. To normalize, divide by height.
float(xmax), # bottomX. To normalize, divide by width
float(ymax), # bottomY. To normalize, divide by height
int(isCrowd)
])
labels.append(image_labels)
return labels
def load_fridge_object_detection_dataset():
# create data folder if it doesnt exist.
os.makedirs("data", exist_ok=True)
# download data
download_url = ("https://cvbp-secondary.z19.web.core.windows.net/" +
"datasets/object_detection/odFridgeObjects.zip")
data_file = "./odFridgeObjects.zip"
urlretrieve(download_url, filename=data_file)
# extract files
with ZipFile(data_file, "r") as zip:
print("extracting files...")
zip.extractall(path="./data")
print("done")
# delete zip file
os.remove(data_file)
labels = load_fridge_object_detection_dataset_labels()
# get all file names into a pandas dataframe with the labels
data = pd.DataFrame(columns=[ImageColumns.IMAGE.value,
ImageColumns.LABEL.value])
for i, file in enumerate(os.listdir("./data/odFridgeObjects/" + "images")):
image_path = "./data/odFridgeObjects/" + "images" + "/" + file
data = data.append({ImageColumns.IMAGE.value: image_path,
ImageColumns.LABEL.value: labels[i]}, # folder
ignore_index=True)
return data
import urllib.request as request_file
#download fine-tuned recycling model from url
def download_assets(filepath,force=False):
if force or not os.path.exists(filepath):
request_file.urlretrieve(
"https://publictestdatasets.blob.core.windows.net/models/fastrcnn.pt",
os.path.join(filepath))
else:
print('Found' + filepath)
return filepath
#Loading in our pretrained model
import torchvision
from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
import torch
import os
def get_instance_segmentation_model(num_classes):
# load an instance segmentation model pre-trained on COCO
model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
in_features = model.roi_heads.box_predictor.cls_score.in_features
# replace the pre-trained head with a new one
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)
return model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Device: ', str(device))
num_classes = 5
model = get_instance_segmentation_model(num_classes)
_ = download_assets('Recycling_finetuned_FastRCNN.pt')
model.load_state_dict(torch.load('Recycling_finetuned_FastRCNN.pt', map_location = device))
#if using the general torchvision pretrained model, comment above and uncomment below
# model = detection.fasterrcnn_resnet50_fpn(pretrained=True)
model.to(device)
data = load_fridge_object_detection_dataset()
train_data = data
test_data = data
class_names = np.array(['can', 'carton', 'milk_bottle', 'water_bottle'])
from raiwidgets import ResponsibleAIDashboard
from responsibleai_vision import ModelTask, RAIVisionInsights
To use Responsible AI Toolbox, initialize a RAIInsights object upon which different components can be loaded.
RAIInsights accepts the model, the full dataset, the test dataset, the target feature string and the task type string as its arguments.
rai_insights = RAIVisionInsights(model, test_data.sample(5, random_state=42),
"label",
task_type=ModelTask.OBJECT_DETECTION,
classes=class_names)
rai_insights
Add the components of the toolbox that are focused on model assessment.
# Interpretability
rai_insights.explainer.add()
# Error Analysis
rai_insights.error_analysis.add()
Once all the desired components have been loaded, compute insights on the test set.
rai_insights.compute()
Finally, visualize and explore the model insights. Use the resulting widget or follow the link to view this in a new tab.
ResponsibleAIDashboard(rai_insights)