We are using canned architectures with pre-trained weights provided by
TensorFlow Keras.
Here's a comparison between the SOTA ImageNet architectures.
MobileNetV2 | ResNet50 | VGG19 | InceptionV3 | Xception | |
---|---|---|---|---|---|
Input Size | 224 x 224 | 224 x 224 | 224 x 224 | 299 x 299 | 299 x 299 |
Size | 14 MB | 98 MB | 548 MB | 92 MB | 88 MB |
Depth | 88 | 168 | 26 | 159 | 126 |
Parameters | 3.5 M | 25.6 M | 143.6 M | 23.8 M | 22.9 M |
Top-1 Accuracy | 71.30% | 74.90% | 71.30% | 77.90% | 79.00% |
Top-5 Accuracy | 90.10% | 92.10% | 90.00% | 93.70% | 94.50% |
Inference Time | 1.15 | 1 | 2.45 | 2.35 | 4 |
# Provide `url` of the image to classify and ImageNet `model` architectures list import requests
url = "https://gramener.com/amle-image-recognition/classify" data = {"url": "https://i.imgur.com/WkomVeG.jpg", "model": ["MobileNetV2", "ResNet50", "VGG19", "InceptionV3", "Xception"]} requests.get(url, data=data).json()