predict.py

import torch
import torch.nn as nn
import torch.nn.init as init
from transformers import SwinForImageClassification
from huggingface_hub import hf_hub_download
from PIL import Image
import json
import os
import random
from torchvision import transforms

# Load labels
with open("labels.json", "r") as f:
    class_names = json.load(f)
print("class_names:", class_names)

MODEL_NAME = "microsoft/swin-large-patch4-window7-224"  

class SwinCustom(nn.Module):
    def __init__(self, model_name=MODEL_NAME, num_classes=40):
        super(SwinCustom, self).__init__()
        self.model = SwinForImageClassification.from_pretrained(model_name, num_labels=num_classes, ignore_mismatched_sizes=True)
        in_features = self.model.classifier.in_features
        self.model.classifier = nn.Sequential(
            nn.Linear(in_features, in_features),
            nn.LeakyReLU(),
            nn.Dropout(0.3),
            nn.Linear(in_features, num_classes)
        )
        # Weight initialization
        for m in self.model.classifier:
            if isinstance(m, nn.Linear):
                init.kaiming_uniform_(m.weight, a=0, mode='fan_in', nonlinearity='leaky_relu')

    def forward(self, images):
        outputs = self.model(images)
        return outputs.logits

model_path = hf_hub_download(repo_id="Noha90/AML_16", filename="large_swin_best_model.pth")
print("Model path:", model_path)
model = SwinCustom(model_name=MODEL_NAME, num_classes=40)
state_dict = torch.load(model_path, map_location="cpu")
if "model_state_dict" in state_dict:
    state_dict = state_dict["model_state_dict"]
model.load_state_dict(state_dict, strict=False)
model.eval()

# Preprocessing
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

def predict(image_path):
    image = Image.open(image_path).convert("RGB")
    x = transform(image).unsqueeze(0)
    with torch.no_grad():
        outputs = model(x)
        print("Logits:", outputs)
        probs = torch.nn.functional.softmax(outputs, dim=1)[0]
        print("Probs:", probs)
        print("Sum of probs:", probs.sum())
        top5 = torch.topk(probs, k=5)

    top1_idx = int(top5.indices[0])
    top1_label = class_names[top1_idx]

    # Select a random image from the class subfolder
    class_folder = f"sample_images/{str(top1_label).replace(' ', '_')}"
    reference_image = None
    if os.path.isdir(class_folder):
        image_files = [f for f in os.listdir(class_folder) if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp", ".gif", ".webp"))]
        if image_files:
            chosen_file = random.choice(image_files)
            ref_path = os.path.join(class_folder, chosen_file)
            print(f"[DEBUG] Randomly selected reference image: {ref_path}")
            reference_image = Image.open(ref_path).convert("RGB")
        else:
            print(f"[DEBUG] No images found in {class_folder}")
    else:
        print(f"[DEBUG] Class folder does not exist: {class_folder}")

    top5_probs = {class_names[int(idx)]: float(score) for idx, score in zip(top5.indices, top5.values)}
    print(f"image path: {image_path}")
    print(f"top1_label: {top1_label}")
    print(f"[DEBUG] Top-5 indices: {top5.indices}")
    print(f"[DEBUG] Top-5 labels: {[class_names[int(idx)] for idx in top5.indices]}")
    print(f"[DEBUG] Top-5 probs: {top5_probs}")

    return image, reference_image, top5_probs