app.py

import torch
import fastapi
import numpy as np
from PIL import Image

class TorchTensor(torch.Tensor):
    pass

class Prediction:
    prediction: TorchTensor

app = fastapi.FastAPI(docs_url="/")
from transformers import ViTForImageClassification

# Define the number of classes in your custom dataset
num_classes = 20

# Initialize the ViTForImageClassification model
model = ViTForImageClassification.from_pretrained(
    'google/vit-base-patch16-224-in21k',
    num_labels=num_classes  # Specify the number of classes
)


model.load_state_dict(torch.load('best_model.pth', map_location='cpu'))
# Define a function to preprocess the input image
def preprocess_input(input: fastapi.UploadFile):
    image = Image.open(input.file)
    image = image.resize((224, 224)).convert("RGB")
    input = np.array(image)
    input = np.transpose(input, (2, 0, 1))
    input = torch.from_numpy(input).float()
    input = input.unsqueeze(0)
    return input

# Define an endpoint to make predictions
@app.post("/predict")
async def predict_endpoint(input:fastapi.UploadFile):
    """Make a prediction on an image uploaded by the user."""

    # Preprocess the input image
    input = preprocess_input(input)

    # Make a prediction
    prediction = model(input)


    predicted_class = prediction.argmax(1).item()

    # Return the predicted class in JSON format
    return {"prediction": predicted_class}