import argparse import torch import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt from PIL import Image import torchvision.transforms as transforms import random import numpy as np from torchvision.datasets import CocoCaptions import matplotlib.gridspec as gridspec import os from tqdm import tqdm from model import EncoderCNN, CaptionTransformer, ImageCaptioningModel from data_utils import Vocabulary, get_transforms def main(): parser = argparse.ArgumentParser(description="Generate Image Captions") parser.add_argument("--model_path", type=str, default="caption_model.pt", help="Path to trained model (default: caption_model.pt)") parser.add_argument("--vocab_path", type=str, default="vocab.pkl", help="Path to vocabulary file (default: vocab.pkl)") parser.add_argument("--sample_img_dir", type=str, required=True, help="Path to sample images directory") parser.add_argument("--sample_ann_file", type=str, default="coco_sample/captions_val_sample.json", help="Path to sample annotations JSON file (only used if --sample_img_dir is coco_sample/images)") parser.add_argument("--num_samples", type=int, default=6, help="Maximum number of samples to display") parser.add_argument("--beam_size", type=int, default=5, help="Beam size for beam search") parser.add_argument("--temperature", type=float, default=1.0, help="Sampling temperature") parser.add_argument("--output", type=str, help="Path to save output image. If not provided, will use predictions.png") args = parser.parse_args() # Set default output filename if args.output is None: args.output = "predictions.png" device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(f"Using device: {device}") try: vocab = Vocabulary.load(args.vocab_path) print(f"Loaded vocabulary with {len(vocab)} tokens from {args.vocab_path}") except Exception as e: print(f"Error loading vocabulary: {e}") return encoder = EncoderCNN(embed_size=512) decoder = CaptionTransformer(vocab_size=len(vocab)) model = ImageCaptioningModel(encoder, decoder) try: model.load_state_dict(torch.load(args.model_path, map_location=device)) model = model.to(device) model.eval() print(f"Model loaded from {args.model_path}") except Exception as e: print(f"Error loading model: {e}") return model_transform = get_transforms(split="val") image_paths = [] gt_captions_map = {} use_coco_dataset = args.sample_img_dir.startswith("coco_sample") if use_coco_dataset: image_root = args.sample_img_dir if not image_root.endswith('images'): image_root = os.path.join(image_root, 'images') annotation_file = args.sample_ann_file print(f"Attempting to load COCO dataset: images from '{image_root}', annotations from '{annotation_file}'") try: if not os.path.isdir(image_root): print(f"Error: Image directory not found: {image_root}") return sample_dataset = CocoCaptions( root=image_root, annFile=annotation_file, transform=None ) print(f"Loaded COCO sample dataset with {len(sample_dataset)} images initially from annotation file.") available_img_ids = [] if not sample_dataset.ids: print("No image IDs found in the annotation file.") return print("Checking for available image files on disk...") for img_id in sample_dataset.ids: try: img_info = sample_dataset.coco.loadImgs(img_id)[0] img_file_path = os.path.join(image_root, img_info['file_name']) if os.path.exists(img_file_path): available_img_ids.append(img_id) except Exception as e: print(f"Warning: Could not process img_id {img_id} from annotations: {e}") continue if not available_img_ids: print("No valid image files found on disk corresponding to the annotations.") return print(f"Found {len(available_img_ids)} available images on disk.") num_to_sample = min(args.num_samples, 6, len(available_img_ids)) if num_to_sample == 0: print("No images available to sample after checking disk.") return selected_img_ids = random.sample(available_img_ids, num_to_sample) for img_id in selected_img_ids: img_info = sample_dataset.coco.loadImgs(img_id)[0] img_path = os.path.join(image_root, img_info['file_name']) image_paths.append(img_path) ann_ids = sample_dataset.coco.getAnnIds(imgIds=img_id) anns = sample_dataset.coco.loadAnns(ann_ids) gt_caps_for_img = [ann['caption'] for ann in anns] if gt_caps_for_img: gt_captions_map[img_path] = random.choice(gt_caps_for_img) else: gt_captions_map[img_path] = "N/A (no captions found for this image)" except FileNotFoundError: print(f"Error: Annotation file not found or critical issue during initial load: {annotation_file}") return except Exception as e: print(f"Error loading COCO sample dataset: {e}") return else: print(f"Loading images directly from custom directory: {args.sample_img_dir}") try: if not os.path.isdir(args.sample_img_dir): print(f"Error: Directory not found: {args.sample_img_dir}") return all_files = [os.path.join(args.sample_img_dir, f) for f in os.listdir(args.sample_img_dir)] image_files = [f for f in all_files if os.path.isfile(f) and f.lower().endswith(('.png', '.jpg', '.jpeg'))] if not image_files: print(f"No image files (.png, .jpg, .jpeg) found in {args.sample_img_dir}") return print(f"Found {len(image_files)} images.") num_available = len(image_files) num_samples = min(args.num_samples, 6, num_available) image_paths = random.sample(image_files, num_samples) except Exception as e: print(f"Error listing or sampling images from {args.sample_img_dir}: {e}") return if not image_paths: print("No images selected for processing.") return num_samples = len(image_paths) if num_samples == 0: print("No images to display.") return elif num_samples <= 3: nrows, ncols = 1, num_samples else: nrows = 2 ncols = (num_samples + 1) // 2 fig = plt.figure(figsize=(5*ncols, 5*nrows)) gs = gridspec.GridSpec(nrows, ncols, figure=fig) print(f"Generating captions for {num_samples} images...") # Wrap image_paths with tqdm for a progress bar for i, img_path in enumerate(tqdm(image_paths, desc="Generating Captions")): # Calculate grid position row, col = i // ncols, i % ncols ax = fig.add_subplot(gs[row, col]) try: # Load original image using PIL original_img = Image.open(img_path).convert("RGB") except Exception as e: print(f"Error opening image {img_path}: {e}") ax.set_title(f"Error loading image: {os.path.basename(img_path)}", fontsize=9) ax.axis('off') continue # Skip to next image # Apply transform only for model input model_input = model_transform(original_img).unsqueeze(0).to(device) # Generate caption with torch.no_grad(): # Ensure no gradients are calculated # model.generate now returns a list of (tokens_list, score) tuples, sorted by score beam_results = model.generate( model_input, vocab, max_len=30, beam_size=args.beam_size, temperature=args.temperature ) title_parts = [] # Ground Truth if use_coco_dataset: gt_caption = gt_captions_map.get(img_path, "./img") # Clean GT caption tokens gt_caption_cleaned = ' '.join([t for t in gt_caption.split() if t not in {"", "", "", ""}]) # Assuming gt_caption is a string title_parts.append(f"GT: {gt_caption_cleaned}") if beam_results: # Best predicted caption best_pred_tokens, best_pred_score = beam_results[0] best_pred_caption = ' '.join([t for t in best_pred_tokens if t not in {"", "", "", ""}]) title_parts.append(f"Best Pred (Score: {best_pred_score:.2f}): {best_pred_caption}") # Other beam search results if len(beam_results) > 1: title_parts.append("--- Other Beams ---") for i, (tokens, score) in enumerate(beam_results[1:]): if i < args.beam_size -1 : # Show up to beam_size-1 other beams other_caption = ' '.join([t for t in tokens if t not in {"", "", "", ""}]) title_parts.append(f"Beam {i+1} (Score: {score:.2f}): {other_caption}") else: break else: title_parts.append("Pred: No caption generated") # Display original image and set title ax.imshow(np.array(original_img)) # Use np.array to ensure correct display ax.set_title("\n".join(title_parts), fontsize=8) # Adjust fontsize if needed ax.axis('off') plt.tight_layout(h_pad=3.0) # Save to predictions.png plt.savefig(args.output, bbox_inches='tight', dpi=200) print(f"Output saved to {args.output}") if __name__ == "__main__": main()