Upload 96 files

app.py

@@ -0,0 +1,486 @@
+import os
+import sys
+import tempfile
+from pathlib import Path
+
+import gradio as gr
+import numpy as np
+import torch
+import yaml
+from PIL import Image
+from pytorch_grad_cam import GradCAM
+from pytorch_grad_cam.utils.image import show_cam_on_image
+
+
+WORKSPACE_ROOT = Path(
+    os.environ.get("WORKSPACE_ROOT", Path(__file__).resolve().parents[1])
+)
+SRC_DIR = WORKSPACE_ROOT / "src"
+sys.path.insert(0, str(SRC_DIR))
+
+from models.swin import EncoderSwinTiny
+from transforms.image_transform import get_classification_valid_transform
+from utils.captioning_inference import build_caption_runtime, decode_tokens
+from visualization.generate_gradcam import (
+    SwinClassifierWrapper,
+    reshape_transform,
+)
+
+
+CLASSIFICATION_STATE = None
+CAPTIONING_STATE = None
+
+
+def load_params():
+    """params.yaml을 읽어서 데모, 모델, 체크포인트 설정을 가져온다."""
+    with open(WORKSPACE_ROOT / "params.yaml", "r", encoding="utf-8") as f:
+        return yaml.safe_load(f)
+
+# params.yaml의 demo.class_names에서 학습 당시 클래스 목록을 가져온다.
+def load_class_names(params):
+    class_names = params.get("demo", {}).get("class_names", [])
+
+    if not isinstance(class_names, list) or not all(
+        isinstance(class_name, str)
+        for class_name in class_names
+    ):
+        raise ValueError("demo.class_names must be a list of class name strings.")
+
+    if not class_names:
+        raise ValueError("No class names found in params.yaml demo.class_names.")
+
+    return class_names
+
+# CUDA 사용 가능 여부를 기준으로 장치를 선택
+def get_device(params):
+    device_name = params.get("train", {}).get("device", "cuda")
+
+    # 설정이 cuda이고 실제 CUDA가 있으면 GPU를 사용한다.
+    if device_name == "cuda" and torch.cuda.is_available():
+        return torch.device("cuda")
+
+    return torch.device("cpu")
+
+
+def load_classification_checkpoint(model, checkpoint_path, device):
+    """분류 모델 체크포인트를 로드하고 model_state_dict 형식이면 내부 state_dict만 꺼낸다."""
+    checkpoint = torch.load(
+        checkpoint_path,
+        map_location=device,
+    )
+
+    # 저장 포맷이 {"model_state_dict": ...} 형태인 경우 실제 가중치만 사용한다.
+    if isinstance(checkpoint, dict) and "model_state_dict" in checkpoint:
+        checkpoint = checkpoint["model_state_dict"]
+
+    model.load_state_dict(checkpoint)
+
+
+def build_classification_runtime():
+    """분류 모델, transform, 클래스명, 체크포인트 경로를 묶은 런타임 상태를 만든다."""
+    params = load_params()
+    model_name = params["classification"]["model_name"]
+
+    # 현재 Grad-CAM wrapper와 모델 생성 로직은 Swin-T 전용이므로 다른 모델은 명시적으로 막는다.
+    if model_name != "swin_t":
+        raise ValueError(
+            "The combined Gradio demo currently supports only swin_t "
+            f"for classification, got: {model_name}"
+        )
+
+    class_names = load_class_names(params)
+    device = get_device(params)
+
+    model = EncoderSwinTiny(
+        num_classes=len(class_names)
+    ).to(device)
+
+    checkpoint_path = WORKSPACE_ROOT / params["classification"]["final_checkpoint"]
+    load_classification_checkpoint(
+        model,
+        checkpoint_path,
+        device,
+    )
+    model.eval()
+
+    return {
+        "params": params,
+        "model": model,
+        "model_name": model_name,
+        "device": device,
+        "class_names": class_names,
+        "transform": get_classification_valid_transform(),
+        "checkpoint_path": checkpoint_path,
+    }
+
+
+def get_classification_runtime():
+    """분류 런타임을 최초 요청 시 한 번만 만들고 이후에는 캐시된 상태를 재사용한다."""
+    global CLASSIFICATION_STATE
+
+    # 버튼 클릭 전에는 모델을 로드하지 않고, 첫 예측 시점에만 로드한다.
+    if CLASSIFICATION_STATE is None:
+        CLASSIFICATION_STATE = build_classification_runtime()
+
+    return CLASSIFICATION_STATE
+
+
+def get_caption_checkpoint_path(params):
+    """캡셔닝 체크포인트 경로를 params.yaml에서 우선 찾고, 없으면 기본 파일명 규칙으로 만든다."""
+    checkpoint_config = params["captioning"]["checkpoint"]
+    final_checkpoint = checkpoint_config.get("final_checkpoint")
+
+    # final_checkpoint가 명시되어 있으면 그 파일을 우선 사용한다.
+    if final_checkpoint:
+        return WORKSPACE_ROOT / checkpoint_config["save_dir"] / final_checkpoint
+
+    # 명시 경로가 없으면 학습 코드의 encoder-decoder_version_best.pt 규칙으로 fallback한다.
+    encoder_name = params["captioning"]["encoder"]
+    decoder_name = params["captioning"]["decoder"]
+    version = params["captioning"]["version"]
+    return (
+        WORKSPACE_ROOT
+        / checkpoint_config["save_dir"]
+        / f"{encoder_name}-{decoder_name}_{version}_best.pt"
+    )
+
+
+def get_captioning_runtime():
+    """캡셔닝 런타임을 최초 요청 시 한 번만 만들고 이후에는 캐시된 상태를 재사용한다."""
+    global CAPTIONING_STATE
+
+    # 캡셔닝 탭을 실제로 실행하기 전까지 encoder/decoder 로딩을 미룬다.
+    if CAPTIONING_STATE is None:
+        params = load_params()
+        CAPTIONING_STATE = build_caption_runtime(
+            WORKSPACE_ROOT,
+            checkpoint_path=get_caption_checkpoint_path(params),
+        )
+    return CAPTIONING_STATE
+
+
+def make_gradcam_overlay(model, image, tensor, device):
+    """분류 모델의 마지막 Swin block을 대상으로 Grad-CAM overlay 이미지를 생성한다."""
+    # Grad-CAM은 gradient가 필요하므로 frozen backbone/classifier도 일시적으로 gradient를 켠다.
+    for param in model.backbone.parameters():
+        param.requires_grad = True
+
+    for param in model.classifier.parameters():
+        param.requires_grad = True
+
+    gradcam_model = SwinClassifierWrapper(model).to(device)
+    gradcam_model.eval()
+
+    resized_image = image.resize((224, 224))
+    image_np = np.array(resized_image).astype(np.float32) / 255.0
+    target_layer = model.backbone.features[-1][-1].norm2
+
+    with GradCAM(
+        model=gradcam_model,
+        target_layers=[target_layer],
+        reshape_transform=reshape_transform,
+    ) as cam:
+        grayscale_cam = cam(input_tensor=tensor)[0]
+
+    overlay = show_cam_on_image(
+        image_np,
+        grayscale_cam,
+        use_rgb=True,
+    )
+
+    return Image.fromarray(overlay)
+
+
+def predict_classification(image, show_gradcam):
+    """업로드된 이미지를 분류하고, 선택 시 Grad-CAM 결과까지 함께 반환한다."""
+    # 이미지가 없으면 Gradio 출력 개수에 맞춰 빈 결과를 반환한다.
+    if image is None:
+        return None, "Please upload an image.", {}, []
+
+    runtime = get_classification_runtime()
+    params = runtime["params"]
+    model = runtime["model"]
+    device = runtime["device"]
+    class_names = runtime["class_names"]
+    transform = runtime["transform"]
+
+    image = image.convert("RGB")
+    tensor = transform(image).unsqueeze(0).to(device)
+
+    with torch.no_grad():
+        logits = model(tensor)
+        probs = torch.softmax(logits, dim=1)[0]
+
+    top_k = max(
+        1,
+        min(
+            int(params["demo"].get("top_k", 5)),
+            len(class_names),
+        ),
+    )
+    top_probs, top_indices = torch.topk(
+        probs,
+        k=top_k,
+    )
+
+    top_probs = top_probs.detach().cpu().tolist()
+    top_indices = top_indices.detach().cpu().tolist()
+
+    confidences = {
+        class_names[idx]: float(prob)
+        for idx, prob in zip(top_indices, top_probs)
+    }
+
+    predicted_idx = top_indices[0]
+    predicted_label = class_names[predicted_idx]
+    predicted_confidence = top_probs[0]
+    summary = (
+        f"Prediction: {predicted_label} "
+        f"({predicted_confidence * 100:.2f}%)"
+    )
+
+    table = [
+        [
+            rank,
+            class_names[idx],
+            f"{prob * 100:.2f}%",
+        ]
+        for rank, (idx, prob) in enumerate(
+            zip(top_indices, top_probs),
+            start=1,
+        )
+    ]
+
+    gradcam_image = None
+
+    # 사용자가 체크박스를 켠 경우에만 비용이 큰 Grad-CAM을 생성한다.
+    if show_gradcam:
+        gradcam_image = make_gradcam_overlay(
+            model,
+            image,
+            tensor,
+            device,
+        )
+
+    return gradcam_image, summary, confidences, table
+
+
+def caption_token_labels(generated_tokens, runtime, caption):
+    """attention heatmap 제목으로 사용할 생성 토큰 라벨을 만든다."""
+    special_ids = {
+        runtime["w2i"].get("<pad>"),
+        runtime["w2i"].get("<sos>"),
+        runtime["w2i"].get("<eos>"),
+    }
+    labels = [
+        runtime["i2w"].get(token, "<unk>")
+        for token in generated_tokens
+        if token not in special_ids
+    ]
+
+    # 토큰 id 기반 라벨이 있으면 attention 길이와 맞기 쉬운 이 라벨을 사용한다.
+    if labels:
+        return labels
+
+    # 예외적으로 라벨이 비어 있으면 문장 문자열을 단어 단위로 나눠 fallback한다.
+    return caption.split()
+
+
+@torch.no_grad()
+def predict_captioning(image):
+    """업로드된 이미지에 대해 캡션을 생성하고 cross-attention heatmap들을 반환한다."""
+    # 이미지가 없으면 Gradio 출력 개수에 맞춰 빈 결과를 반환한다.
+    if image is None:
+        return "Please upload an image.", []
+
+    runtime = get_captioning_runtime()
+    params = runtime["params"]
+    image = image.convert("RGB")
+    image_tensor = runtime["transform"](image)
+    image_tensor = image_tensor.unsqueeze(0).to(runtime["device"])
+
+    features = runtime["encoder"](
+        image_tensor,
+        return_features=True,
+    )
+    start_token = torch.full(
+        (features.size(0),),
+        runtime["w2i"]["<sos>"],
+        dtype=torch.long,
+        device=runtime["device"],
+    )
+
+    beam_config = params["captioning"]["beam_search"]
+    use_beam_search = beam_config.get("use_beam_search", True)
+    beam_size = beam_config.get("beam_size", 3)
+
+    # params.yaml에서 beam search를 켠 경우 여러 후보를 탐색해 캡션을 생성한다.
+    if use_beam_search:
+        generated_tokens, _, enc_dec_atten = runtime["decoder"].generate_beam(
+            features,
+            start_token,
+            runtime["w2i"]["<eos>"],
+            beam_size,
+        )
+    else:
+        # beam search를 끈 경우 매 step에서 가장 확률 높은 토큰을 선택하는 greedy 생성을 사용한다.
+        generated_tokens, _, enc_dec_atten = runtime["decoder"].generate(
+            features,
+            start_token,
+            runtime["w2i"]["<eos>"],
+        )
+
+    caption = decode_tokens(
+        generated_tokens[0],
+        runtime["w2i"],
+        runtime["i2w"],
+        params["captioning"]["tokenizer"]["use_subword"],
+        sp_model_path=runtime["sp_model_path"],
+    )
+
+    caption_tokens = caption_token_labels(
+        generated_tokens[0],
+        runtime,
+        caption,
+    )
+
+    tmp_dir = tempfile.mkdtemp(prefix="combined_captioning_gradio_")
+    heatmap_images = []
+    n_layers = len(runtime["decoder"].layers)
+
+    # 각 decoder layer별 cross-attention heatmap 이미지를 만들어 Gallery에 표시한다.
+    for layer in range(1, n_layers + 1):
+        cross_atten_path = Path(tmp_dir) / f"cross_attention_layer_{layer}.jpg"
+        runtime["decoder"].show_cross_atten(
+            enc_dec_atten[0],
+            caption_tokens,
+            layer,
+            image_tensor.squeeze(0).detach().cpu(),
+            str(cross_atten_path),
+        )
+        heatmap_images.append(
+            (
+                str(cross_atten_path),
+                f"Layer {layer}",
+            )
+        )
+
+    return caption, heatmap_images
+
+
+def create_demo():
+    """분류 탭과 캡셔닝 탭을 가진 하나의 Gradio Blocks 앱을 만든다."""
+    params = load_params()
+    top_k = max(1, int(params["demo"].get("top_k", 5)))
+    caption_checkpoint = get_caption_checkpoint_path(params)
+
+    with gr.Blocks(title="ImageNet Classification and Captioning Demo") as demo:
+        gr.Markdown("# ImageNet Classification and Captioning Demo")
+
+        with gr.Tabs():
+            with gr.Tab("Classification"):
+                gr.Markdown(
+                    "Upload an image and classify it with the final checkpoint."
+                )
+                gr.Markdown(
+                    f"checkpoint: {WORKSPACE_ROOT / params['classification']['final_checkpoint']}"
+                )
+
+                with gr.Row():
+                    with gr.Column():
+                        classification_image_input = gr.Image(
+                            type="pil",
+                            label="Input Image",
+                        )
+                        gradcam_checkbox = gr.Checkbox(
+                            value=bool(params["demo"].get("show_gradcam", True)),
+                            label="Show Grad-CAM",
+                        )
+                        classification_button = gr.Button(
+                            "Predict",
+                            variant="primary",
+                        )
+
+                    with gr.Column():
+                        gradcam_output = gr.Image(
+                            type="pil",
+                            label="Grad-CAM",
+                        )
+                        classification_summary_output = gr.Textbox(
+                            label="Prediction",
+                        )
+                        confidence_output = gr.Label(
+                            label="Top Prediction Scores",
+                            num_top_classes=top_k,
+                        )
+                        table_output = gr.Dataframe(
+                            headers=["Rank", "Class", "Confidence"],
+                            datatype=["number", "str", "str"],
+                            label=f"Top-{top_k}",
+                            interactive=False,
+                        )
+
+                classification_button.click(
+                    fn=predict_classification,
+                    inputs=[
+                        classification_image_input,
+                        gradcam_checkbox,
+                    ],
+                    outputs=[
+                        gradcam_output,
+                        classification_summary_output,
+                        confidence_output,
+                        table_output,
+                    ],
+                )
+
+            with gr.Tab("Captioning"):
+                gr.Markdown(
+                    "Upload an image and generate a caption with cross-attention heatmaps."
+                )
+                gr.Markdown(f"checkpoint: {caption_checkpoint}")
+
+                with gr.Row():
+                    with gr.Column():
+                        captioning_image_input = gr.Image(
+                            type="pil",
+                            label="Input Image",
+                        )
+                        captioning_button = gr.Button(
+                            "Generate Caption",
+                            variant="primary",
+                        )
+
+                    with gr.Column():
+                        caption_output = gr.Textbox(
+                            label="Generated Caption",
+                            lines=4,
+                        )
+                        cross_atten_output = gr.Gallery(
+                            label="Cross Attention Heatmaps",
+                            columns=2,
+                            object_fit="contain",
+                            height="auto",
+                        )
+
+                captioning_button.click(
+                    fn=predict_captioning,
+                    inputs=[captioning_image_input],
+                    outputs=[
+                        caption_output,
+                        cross_atten_output,
+                    ],
+                )
+
+    return demo
+
+
+if __name__ == "__main__":
+    params = load_params()
+
+    demo = create_demo()
+    demo.launch(
+        server_name=params["demo"]["host"],
+        server_port=params["demo"]["port"],
+        share=params["demo"]["share"],
+    )

outputs/captioning/swin-transformer_final_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9897bcdee87a9b20241c9c742c27feb0f1204cf2d65456f3d892300a23b59adc
+size 468449515

outputs/classification/cls_swin-t_base_cls_raw-20260525-v2_lr-0005_bs-32_adamw_none_wdc-0.05_ls-0.0_best.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec30dab609101cc6a58722968771d66a81d44cd3098e79975c4c4ab59141b1b2
+size 112163475

params.yaml

@@ -0,0 +1,242 @@
+project:
+  name: imagenet-project
+
+data:
+  raw_dir: ./data/raw/
+  captions_file: ./data/captioning/annotations/train.json
+  dataset_version: cls_raw-20260525-v2
+  # dataset_version: raw-20260509-v1
+
+split:
+  train_ratio: 0.7
+  val_ratio: 0.15
+  test_ratio: 0.15
+
+train:
+  seed: 42
+  # repeated experiment
+  # seed: 7
+  # seed: 21
+  epochs: 20
+  batch_size: 32
+  num_workers: 4
+  device: cuda
+  optimizer: adam
+
+preprocess:
+  image_size: 224
+  normalize: true
+
+loss:
+  name: cross_entropy
+  ignore_index: pad_token
+
+evaluate:
+  batch_size: 32
+  metrics:
+    - bleu
+    - rouge_l
+    - meteor
+
+logging:
+  use_wandb: true
+  project_name: imagenet-project
+  log_interval: 10
+
+outputs:
+  base_dir: outputs
+
+demo:
+  host: 0.0.0.0
+  port: 7860
+  share: false
+  top_k: 5
+  show_gradcam: true
+  class_names : [airplane, apple, aster, banana, bicycle, bracelet, bulldog, bus, butterfly, car, carrot, cucumber, cup-cake, daisy, dandelion, dumpling, earrings, elephant, glasses, golden-retriever, hamburger, horse, iris, lavender, lily, marigold, motorcycle, necklace, orange, orchid, pants, pasta, penguin, persian-cat, pizza, rose, salad, sandwich, sheep, siamese-cat, sneakers, squirrel, steak, strawberry, sunflower, sushi, tomato, t-shirt, tulip, waffle]
+
+cnn:
+  backbone: resnet18
+  pretrained: true
+  freeze: true
+  output_dim: 512
+  dropout: 0.3
+  pooling: avg
+
+captioning:
+
+  # encoder: resnet18
+  encoder: swin
+  # encoder: vit
+  decoder: transformer
+  # decoder: lstm
+  # decoder: gru
+  version: final
+
+  epochs: 25
+  learning_rate: 0.0001
+  batch_size: 32
+  optimizer: adamw
+  max_caption_length: 30
+  train_num_caption: 2
+
+  debug: False
+
+  lstm:
+    embed_dim: 256
+    hidden_dim: 512
+    num_layers: 1
+
+  gru:
+    embed_dim: 256
+    hidden_dim: 512
+    num_layers: 1
+
+  transformer:
+    n_layers: 6
+    nhead: 8
+    d_model: 512
+    drop_p: 0.3
+    label_smoothing: 0
+    weight_decay: 0.001
+
+  data:
+    dataset_version: cap_raw-20260524-v1
+    train_img: ./data/captioning/raw/train/
+    train_caption: ./data/captioning/annotations/train.json
+    val_img: ./data/captioning/raw/val/
+    val_caption: ./data/captioning/annotations/val.json
+    test_img: ./data/captioning/raw/test/
+    test_caption: ./data/captioning/annotations/test.json
+    
+  tokenizer:
+    min_freq: 3
+    max_vocab_size: 10000
+    sp_vocab_size: 2000
+    use_subword: False
+    sp_model_path: ./src/dataset/sub_tokenizer2000.model
+
+  checkpoint:
+    save_dir: ./outputs/captioning
+    final_checkpoint: swin-transformer_final_best.pt
+    resume: False
+
+  heatmap:
+    dec_atten_dir: /workspace/outputs/captioning/heatmap/
+    enc_dec_atten_dir: /workspace/outputs/captioning/heatmap/
+    layer: 6 # 몇번째 층
+    sample: [0, 410, 820, 1230, 1640] # caption & heatmap 몇번째 샘플(batch)
+
+  scheduler:
+    use_scheduler: False
+    warmup_step: 500
+    lr_scale: 0.5
+
+  beam_search:
+    use_beam_search: True
+    beam_size: 3
+    
+
+
+classification:
+
+  # model_name: resnet18
+  # model_name: efficientnet_b0
+  # model_name: convnext_tiny
+  # model_name: mobilenet_v3_small
+  # model_name: vit_b_16
+  model_name: swin_t
+  # model_name: deit_tiny_patch16_224
+
+  final_checkpoint: ./outputs/classification/cls_swin-t_base_cls_raw-20260525-v2_lr-0005_bs-32_adamw_none_wdc-0.05_ls-0.0_best.pth
+
+  epochs: 50
+
+  learning_rate:
+
+    # baseline
+    cnn: 0.001
+    transformer: 0.0005
+
+    # hyperparameter tuning
+    # cnn: 0.0005
+    # transformer: 0.0001
+
+  # optimizer: adam
+  # optimizer: sgd
+  optimizer: adamw
+
+  # default
+  # weight_decay: 0.01
+
+  # tuning
+  weight_decay: 0.05
+
+  scheduler:
+    use: false
+
+    # use: true
+    # name: cosineannealinglr
+
+  augmentation:
+
+    # baseline
+    use_aug: false
+    type: none
+
+    # mixup
+    # use_aug: true
+    # type: mixup
+
+    # cutmix
+    # use_aug: true
+    # type: cutmix
+
+  label_smoothing: 0.0
+
+  # label smoothing experiment
+  # label_smoothing: 0.05
+  # label_smoothing: 0.1
+
+
+  metrics:
+
+    train:
+      - loss
+      - accuracy
+
+    validation:
+      - loss
+      - accuracy
+      - macro_f1
+
+    final_test:
+      - accuracy
+      - macro_f1
+      - precision
+      - recall
+      - confusion_matrix
+
+  checkpoint:
+    save_dir: /workspace/outputs/classification
+
+
+latent_space:
+  data_dir: /workspace/data/raw
+  checkpoint: /workspace/outputs/classification/cls_swin-t_base_cls_raw-20260525-v2_lr-0005_bs-32_adamw_none_wdc-0.05_ls-0.0_best.pth
+  output_dir: /workspace/outputs/latent_space
+  output_umap_npy: cls_swin-t_best_umap_2d_test_nb10_md05
+  output_umap_png: cls_swin-t_best_umap_plt_test_nb10_md05
+  output_meta_csv: cls_swin-t_best_metadata_test_nb10_md05
+  split: test
+  batch_size: 32
+  num_workers: 4
+  device: cuda
+  seed: 42
+  save_meta: true
+  use_wandb: true
+  wandb_name: latent_space_umap
+
+  umap:
+    n_neighbors: 10
+    min_dist: 0.5
+    metric: cosine

requirements.txt

@@ -0,0 +1,342 @@
+absl-py==2.4.0
+accelerate==0.27.2
+aiobotocore==3.5.0
+aiohappyeyeballs==2.6.1
+aiohttp==3.13.5
+aiohttp-retry==2.9.1
+aioitertools==0.13.0
+aiosignal==1.4.0
+albumentations==1.4.7
+amqp==5.3.1
+annotated-doc==0.0.4
+annotated-types==0.7.0
+antlr4-python3-runtime==4.9.3
+anyio==4.13.0
+appdirs==1.4.4
+archspec @ file:///croot/archspec_1697725767277/work
+argon2-cffi==25.1.0
+argon2-cffi-bindings==25.1.0
+arrow==1.4.0
+asttokens @ file:///opt/conda/conda-bld/asttokens_1646925590279/work
+astunparse==1.6.3
+async-lru==2.3.0
+async-timeout==5.0.1
+asyncssh==2.22.0
+atpublic==7.0.0
+attrs==26.1.0
+babel==2.18.0
+backoff==2.2.1
+beautifulsoup4 @ file:///croot/beautifulsoup4-split_1681493039619/work
+billiard==4.2.4
+bleach==6.3.0
+boltons @ file:///croot/boltons_1677628692245/work
+boto3==1.42.91
+botocore==1.42.91
+Brotli @ file:///tmp/abs_ecyw11_7ze/croots/recipe/brotli-split_1659616059936/work
+celery==5.6.3
+certifi==2026.4.22
+cffi==2.0.0
+chardet @ file:///home/builder/ci_310/chardet_1640804867535/work
+charset-normalizer==3.4.7
+click==8.3.3
+click-didyoumean==0.3.1
+click-plugins==1.1.1.2
+click-repl==0.3.0
+colorama==0.4.6
+comm==0.2.3
+conda @ file:///croot/conda_1696257509808/work
+conda-build @ file:///croot/conda-build_1708025865815/work
+conda-content-trust @ file:///croot/conda-content-trust_1693490622020/work
+conda-libmamba-solver @ file:///croot/conda-libmamba-solver_1691418897561/work/src
+conda-package-handling @ file:///croot/conda-package-handling_1690999929514/work
+conda_index @ file:///croot/conda-index_1706633791028/work
+conda_package_streaming @ file:///croot/conda-package-streaming_1690987966409/work
+configobj==5.0.9
+contourpy==1.3.2
+cryptography @ file:///croot/cryptography_1707523700518/work
+cuda-bindings==12.9.4
+cuda-pathfinder==1.5.4
+cuda-toolkit==12.8.1
+cycler==0.12.1
+dacite==1.6.0
+dagshub==0.7.0
+dagshub-annotation-converter==0.2.0
+dataclasses-json==0.6.7
+datasets==2.18.0
+debugpy==1.8.20
+decorator @ file:///opt/conda/conda-bld/decorator_1643638310831/work
+defusedxml==0.7.1
+dictdiffer==0.9.0
+dill==0.3.8
+diskcache==5.6.3
+distro @ file:///croot/distro_1701455004953/work
+dnspython==2.6.1
+docker-pycreds==0.4.0
+dpath==2.2.0
+dulwich==1.2.0
+dvc==3.67.1
+dvc-data==3.18.3
+dvc-http==2.32.0
+dvc-objects==5.2.0
+dvc-render==1.0.2
+dvc-s3==3.3.0
+dvc-studio-client==0.22.0
+dvc-task==0.40.2
+entrypoints==0.4
+exceptiongroup==1.3.1
+executing @ file:///opt/conda/conda-bld/executing_1646925071911/work
+expecttest==0.2.1
+fastjsonschema==2.21.2
+filelock==3.29.0
+flatten-dict==0.4.2
+flufl.lock==9.0.0
+fonttools==4.62.1
+fqdn==1.5.1
+frozenlist==1.8.0
+fsspec==2024.2.0
+ftfy==6.3.1
+funcy==2.0
+gitdb==4.0.12
+GitPython==3.1.49
+gmpy2 @ file:///tmp/build/80754af9/gmpy2_1645455533097/work
+gql==4.0.0
+grandalf==0.8
+graphql-core==3.2.8
+gto==1.9.0
+h11==0.16.0
+hf-xet==1.4.3
+httpcore==1.0.9
+httpx==0.28.1
+huggingface-hub==0.20.3
+hydra-core==1.3.2
+hypothesis==6.98.10
+idna==3.13
+ImageHash==4.3.1
+ImageIO==2.37.3
+ipykernel==7.2.0
+ipython @ file:///croot/ipython_1704833016303/work
+isoduration==20.11.0
+iterative-telemetry==0.0.10
+jedi @ file:///tmp/build/80754af9/jedi_1644315229345/work
+Jinja2==3.1.6
+jmespath==1.1.0
+joblib==1.5.3
+json5==0.14.0
+jsonpatch @ file:///tmp/build/80754af9/jsonpatch_1615747632069/work
+jsonpointer==2.1
+jsonschema @ file:///croot/jsonschema_1699041609003/work
+jsonschema-specifications @ file:///croot/jsonschema-specifications_1699032386549/work
+jupyter-events==0.12.1
+jupyter-lsp==2.3.1
+jupyter_client==8.8.0
+jupyter_core==5.9.1
+jupyter_server==2.18.2
+jupyter_server_terminals==0.5.4
+jupyterlab==4.5.6
+jupyterlab_pygments==0.3.0
+jupyterlab_server==2.28.0
+kiwisolver==1.5.0
+kombu==5.6.2
+lazy-loader==0.5
+libarchive-c @ file:///tmp/build/80754af9/python-libarchive-c_1617780486945/work
+libmambapy @ file:///croot/mamba-split_1698782620632/work/libmambapy
+lightning-utilities==0.15.3
+lxml==6.1.0
+markdown-it-py==4.0.0
+MarkupSafe==3.0.3
+marshmallow==3.26.2
+matplotlib==3.8.3
+matplotlib-inline @ file:///opt/conda/conda-bld/matplotlib-inline_1662014470464/work
+mdurl==0.1.2
+menuinst @ file:///croot/menuinst_1706732933928/work
+mistune==3.2.1
+mkl-fft @ file:///croot/mkl_fft_1695058164594/work
+mkl-random @ file:///croot/mkl_random_1695059800811/work
+mkl-service==2.4.0
+more-itertools @ file:///croot/more-itertools_1700662129964/work
+mpmath==1.3.0
+multidict==6.7.1
+multiprocess==0.70.16
+mypy_extensions==1.1.0
+nbclient==0.10.4
+nbconvert==7.17.1
+nbformat==5.10.4
+nest-asyncio==1.6.0
+networkx==3.3
+nltk==3.8.1
+notebook_shim==0.2.4
+numpy==1.26.4
+nvidia-cublas==13.1.0.3
+nvidia-cublas-cu12==12.8.4.1
+nvidia-cuda-cupti==13.0.85
+nvidia-cuda-cupti-cu12==12.8.90
+nvidia-cuda-nvrtc==13.0.88
+nvidia-cuda-nvrtc-cu12==12.8.93
+nvidia-cuda-runtime==13.0.96
+nvidia-cuda-runtime-cu12==12.8.90
+nvidia-cudnn-cu12==9.19.0.56
+nvidia-cudnn-cu13==9.19.0.56
+nvidia-cufft==12.0.0.61
+nvidia-cufft-cu12==11.3.3.83
+nvidia-cufile==1.15.1.6
+nvidia-cufile-cu12==1.13.1.3
+nvidia-curand==10.4.0.35
+nvidia-curand-cu12==10.3.9.90
+nvidia-cusolver==12.0.4.66
+nvidia-cusolver-cu12==11.7.3.90
+nvidia-cusparse==12.6.3.3
+nvidia-cusparse-cu12==12.5.8.93
+nvidia-cusparselt-cu12==0.7.1
+nvidia-cusparselt-cu13==0.8.0
+nvidia-nccl-cu12==2.28.9
+nvidia-nccl-cu13==2.28.9
+nvidia-nvjitlink==13.0.88
+nvidia-nvjitlink-cu12==12.8.93
+nvidia-nvshmem-cu12==3.4.5
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvtx==13.0.85
+nvidia-nvtx-cu12==12.8.90
+omegaconf==2.3.0
+open-clip-torch==2.24.0
+opencv-python-headless==4.9.0.80
+optree==0.10.0
+orjson==3.11.8
+overrides==7.7.0
+packaging==26.2
+pandas==2.1.4
+pandocfilters==1.5.1
+parso @ file:///opt/conda/conda-bld/parso_1641458642106/work
+pathspec==1.1.1
+pathvalidate==3.3.1
+pexpect @ file:///tmp/build/80754af9/pexpect_1605563209008/work
+pillow==10.3.0
+pkginfo @ file:///croot/pkginfo_1679431160147/work
+platformdirs @ file:///croot/platformdirs_1692205439124/work
+pluggy @ file:///tmp/build/80754af9/pluggy_1648024709248/work
+portalocker==3.2.0
+prometheus_client==0.25.0
+prompt-toolkit @ file:///croot/prompt-toolkit_1704404351921/work
+propcache==0.4.1
+protobuf==4.25.9
+psutil==7.2.2
+ptyprocess @ file:///tmp/build/80754af9/ptyprocess_1609355006118/work/dist/ptyprocess-0.7.0-py2.py3-none-any.whl
+pure-eval @ file:///opt/conda/conda-bld/pure_eval_1646925070566/work
+pyarrow==24.0.0
+pyarrow-hotfix==0.7
+pycocoevalcap==1.2
+pycocotools==2.0.11
+pycosat @ file:///croot/pycosat_1696536503704/work
+pycparser @ file:///tmp/build/80754af9/pycparser_1636541352034/work
+pydantic==2.9.2
+pydantic-settings==2.14.0
+pydantic_core==2.23.4
+pydot==4.0.1
+pygit2==1.18.2
+Pygments @ file:///croot/pygments_1684279966437/work
+pygtrie==2.5.0
+pyOpenSSL @ file:///croot/pyopenssl_1708380408460/work
+pyparsing==3.3.2
+PySocks @ file:///home/builder/ci_310/pysocks_1640793678128/work
+python-dateutil==2.9.0.post0
+python-dotenv==1.2.2
+python-etcd==0.4.5
+python-json-logger==4.1.0
+pytz==2026.1.post1
+PyWavelets==1.8.0
+PyYAML==6.0.1
+pyzmq==27.1.0
+referencing @ file:///croot/referencing_1699012038513/work
+regex==2026.4.4
+requests==2.32.3
+requests-toolbelt==1.0.0
+rfc3339-validator==0.1.4
+rfc3986-validator==0.1.1
+rich==15.0.0
+rouge_score==0.1.2
+rpds-py @ file:///croot/rpds-py_1698945930462/work
+ruamel.yaml @ file:///croot/ruamel.yaml_1666304550667/work
+ruamel.yaml.clib @ file:///croot/ruamel.yaml.clib_1666302247304/work
+s3fs==2026.3.0
+s3transfer==0.16.1
+sacrebleu==2.4.0
+safetensors==0.4.2
+scikit-image==0.25.2
+scikit-learn==1.7.2
+scipy==1.15.3
+scmrepo==3.6.2
+semver==3.0.4
+Send2Trash==2.1.0
+sentence-transformers==2.7.0
+sentencepiece==0.2.0
+sentry-sdk==2.58.0
+setproctitle==1.3.7
+shellingham==1.5.4
+shortuuid==1.0.13
+shtab==1.8.0
+six==1.17.0
+smmap==5.0.3
+sortedcontainers==2.4.0
+soupsieve @ file:///croot/soupsieve_1696347547217/work
+sqltrie==0.11.2
+stack-data @ file:///opt/conda/conda-bld/stack_data_1646927590127/work
+sympy==1.14.0
+tabulate==0.10.0
+tenacity==9.1.4
+terminado==0.18.1
+threadpoolctl==3.6.0
+tifffile==2025.5.10
+timm==0.9.12
+tinycss2==1.4.0
+tokenizers==0.15.2
+tomli @ file:///opt/conda/conda-bld/tomli_1657175507142/work
+tomlkit==0.14.0
+toolz @ file:///croot/toolz_1667464077321/work
+torch==2.11.0+cu128
+torchaudio==2.11.0+cu128
+torchelastic==0.2.2
+torchmetrics==1.9.0
+torchvision==0.26.0+cu128
+tornado==6.5.5
+tqdm==4.66.2
+traitlets @ file:///croot/traitlets_1671143879854/work
+transformers==4.38.2
+treelib==1.8.0
+triton==3.6.0
+truststore @ file:///croot/truststore_1695244293384/work
+typer==0.25.0
+types-dataclasses==0.6.6
+typing-inspect==0.9.0
+typing-inspection==0.4.2
+typing_extensions==4.10.0
+tzdata==2026.2
+tzlocal==5.3.1
+uri-template==1.3.0
+urllib3==2.6.3
+vine==5.1.0
+voluptuous==0.16.0
+wandb==0.26.1
+wcwidth==0.7.0
+webcolors==25.10.0
+webencodings==0.5.1
+websocket-client==1.9.0
+wrapt==2.1.2
+xxhash==3.7.0
+yarl==1.23.0
+zc.lockfile==4.0
+zstandard @ file:///croot/zstandard_1677013143055/work
+
+# 5/25 추가필요
+notebook==7.5.0
+einops==0.8.2
+
+
+# 5/30 추가
+umap-learn==0.5.12
+
+# 5/31 가인 (추가 필요)
+grad-cam==1.5.5
+opencv-python==4.11.0.86
+ttach==0.0.3
+gradio==4.44.1
+fastapi==0.112.4
+starlette==0.38.6

src/caption/check_clip_score.py

@@ -0,0 +1,440 @@
+import json
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from PIL import Image
+from tqdm import tqdm
+from transformers import CLIPModel, CLIPProcessor
+import os
+from dotenv import load_dotenv
+
+# ============================================================
+# 설정값
+# ============================================================
+
+load_dotenv()
+
+# .env 안의 HF_TOKEN 읽기
+hf_token = os.getenv("HF_TOKEN")
+
+# 전체 클래스를 검수하려면 True
+# 특정 클래스만 검수하려면 False
+CHECK_ALL_CLASSES = True
+
+# 전체 클래스 검수 시 기준이 되는 raw 데이터 루트
+DATA_RAW_ROOT_DIR = Path("data/raw")
+
+# 특정 클래스만 검수할 때 사용할 클래스 폴더 경로
+# CHECK_ALL_CLASSES = False 일 때만 사용됨
+TARGET_CLASS_DIR = Path("data/raw")
+
+# 입력 JSON 파일
+INPUT_JSON_PATH = Path("data/annotations/captions_flo_all.json")
+
+# 출력 JSON 파일
+OUTPUT_JSON_PATH = Path("data/annotations/clip_checked_flo_all.json")
+
+# 사용할 CLIP 모델
+MODEL_NAME = "openai/clip-vit-base-patch32"
+
+# 한 번에 처리할 이미지-캡션 쌍 개수
+BATCH_SIZE = 32
+
+# 하위 몇 %를 fail / review로 볼지
+FAIL_BOTTOM_PERCENT = 10
+REVIEW_BOTTOM_PERCENT = 20
+
+print("경로 : " , INPUT_JSON_PATH)
+
+# ============================================================
+# JSON 입출력
+# ============================================================
+
+def load_json(path: Path) -> list[dict[str, Any]]:
+    with path.open("r", encoding="utf-8") as f:
+        data = json.load(f)
+
+    if not isinstance(data, list):
+        raise ValueError("입력 JSON은 반드시 배열 형태여야 합니다.")
+
+    return data
+
+
+def save_json(data: list[dict[str, Any]], path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    with path.open("w", encoding="utf-8") as f:
+        json.dump(data, f, ensure_ascii=False, indent=4)
+
+
+# ============================================================
+# 클래스 / 경로 처리
+# ============================================================
+
+def get_target_class_name() -> str:
+    """
+    TARGET_CLASS_DIR = data/raw/airplane 이면 airplane 반환
+    """
+    return TARGET_CLASS_DIR.name
+
+
+def get_class_name_from_image_value(image_value: str) -> str:
+    """
+    JSON의 image 값이 airplane/hf_airplane_001.jpg 라면 airplane 반환
+    """
+    image_value = image_value.replace("\\", "/")
+    image_path = Path(image_value)
+
+    if len(image_path.parts) < 2:
+        return ""
+
+    return image_path.parts[0]
+
+
+def is_target_item(item: dict[str, Any]) -> bool:
+    """
+    CHECK_ALL_CLASSES = True:
+        모든 item 처리
+
+    CHECK_ALL_CLASSES = False:
+        TARGET_CLASS_DIR.name과 JSON image의 첫 번째 폴더명이 같은 item만 처리
+    """
+    if CHECK_ALL_CLASSES:
+        return True
+
+    image_value = str(item.get("image", ""))
+    image_class_name = get_class_name_from_image_value(image_value)
+
+    return image_class_name == get_target_class_name()
+
+
+def resolve_image_path(image_value: str) -> Path:
+    """
+    JSON:
+        "image": "airplane/hf_airplane_001.jpg"
+
+    전체 클래스 검수:
+        DATA_RAW_ROOT_DIR / image
+        → data/raw/airplane/hf_airplane_001.jpg
+
+    특정 클래스 검수:
+        TARGET_CLASS_DIR / 파일명
+        → data/raw/airplane/hf_airplane_001.jpg
+    """
+    image_value = image_value.replace("\\", "/")
+    image_path = Path(image_value)
+
+    if CHECK_ALL_CLASSES:
+        return DATA_RAW_ROOT_DIR / image_path
+
+    return TARGET_CLASS_DIR / image_path.name
+
+
+def load_image(image_path: Path) -> Image.Image | None:
+    try:
+        with Image.open(image_path) as img:
+            return img.convert("RGB").copy()
+    except Exception:
+        return None
+
+
+# ============================================================
+# 캡션 펼치기
+# ============================================================
+
+def flatten_caption_items(data: list[dict[str, Any]]) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
+    """
+    이미지 1장에 caption 3개가 있으면
+    이미지-캡션 쌍 3개로 펼친다.
+    """
+    target_data = []
+    flat_items = []
+
+    for item in data:
+        if not is_target_item(item):
+            continue
+
+        target_item_index = len(target_data)
+        target_data.append(item)
+
+        image_value = str(item.get("image", ""))
+        captions = item.get("captions", [])
+
+        if not isinstance(captions, list):
+            captions = []
+
+        for caption_index, caption in enumerate(captions):
+            flat_items.append({
+                "item_index": target_item_index,
+                "caption_index": caption_index,
+                "image": image_value,
+                "class": item.get("class", ""),
+                "split": item.get("split", ""),
+                "caption": str(caption).strip()
+            })
+
+    return target_data, flat_items
+
+
+# ============================================================
+# CLIP Score 계산
+# ============================================================
+
+@torch.no_grad()
+def compute_clip_scores(
+    flat_items: list[dict[str, Any]],
+    model: CLIPModel,
+    processor: CLIPProcessor,
+    device: torch.device
+) -> list[dict[str, Any]]:
+
+    results = []
+
+    for start in tqdm(range(0, len(flat_items), BATCH_SIZE), desc="computing CLIP scores"):
+        batch_items = flat_items[start:start + BATCH_SIZE]
+
+        valid_items = []
+        images = []
+        texts = []
+
+        for item in batch_items:
+            image_path = resolve_image_path(item["image"])
+            image = load_image(image_path)
+
+            if image is None:
+                results.append({
+                    **item,
+                    "resolved_image_path": str(image_path).replace("\\", "/"),
+                    "clip_cosine": None,
+                    "clip_score": None,
+                    "clip_status": "missing_image",
+                    "clip_reason": f"image file could not be opened: {image_path}"
+                })
+                continue
+
+            caption = item["caption"]
+
+            if not caption:
+                results.append({
+                    **item,
+                    "resolved_image_path": str(image_path).replace("\\", "/"),
+                    "clip_cosine": None,
+                    "clip_score": None,
+                    "clip_status": "empty_caption",
+                    "clip_reason": "caption is empty"
+                })
+                continue
+
+            valid_items.append({
+                **item,
+                "resolved_image_path": str(image_path).replace("\\", "/")
+            })
+            images.append(image)
+            texts.append(caption)
+
+        if not valid_items:
+            continue
+
+        inputs = processor(
+            text=texts,
+            images=images,
+            return_tensors="pt",
+            padding=True,
+            truncation=True
+        )
+
+        inputs = {
+            key: value.to(device)
+            for key, value in inputs.items()
+        }
+
+        outputs = model(
+            input_ids=inputs["input_ids"],
+            attention_mask=inputs["attention_mask"],
+            pixel_values=inputs["pixel_values"]
+        )
+
+        image_features = outputs.image_embeds
+        text_features = outputs.text_embeds
+
+        image_features = F.normalize(image_features, p=2, dim=1)
+        text_features = F.normalize(text_features, p=2, dim=1)
+
+        cosine_scores = (image_features * text_features).sum(dim=1)
+
+        for item, cosine in zip(valid_items, cosine_scores):
+            cosine_value = float(cosine.detach().cpu().item())
+            clip_score = 2.5 * max(cosine_value, 0.0)
+
+            results.append({
+                **item,
+                "clip_cosine": round(cosine_value, 6),
+                "clip_score": round(clip_score, 6),
+                "clip_status": "pending",
+                "clip_reason": ""
+            })
+
+    return results
+
+
+# ============================================================
+# pass / review / fail 판정
+# ============================================================
+
+def assign_clip_status(results: list[dict[str, Any]]) -> None:
+    valid_scores = [
+        result["clip_score"]
+        for result in results
+        if isinstance(result.get("clip_score"), float)
+    ]
+
+    if not valid_scores:
+        return
+
+    fail_threshold = np.percentile(valid_scores, FAIL_BOTTOM_PERCENT)
+    review_threshold = np.percentile(valid_scores, REVIEW_BOTTOM_PERCENT)
+
+    for result in results:
+        clip_score = result.get("clip_score")
+
+        if clip_score is None:
+            continue
+
+        if clip_score <= fail_threshold:
+            result["clip_status"] = "fail"
+            result["clip_reason"] = f"clip score is in the bottom {FAIL_BOTTOM_PERCENT}%"
+        elif clip_score <= review_threshold:
+            result["clip_status"] = "review"
+            result["clip_reason"] = f"clip score is in the bottom {REVIEW_BOTTOM_PERCENT}%"
+        else:
+            result["clip_status"] = "pass"
+            result["clip_reason"] = "clip score is acceptable"
+
+
+# ============================================================
+# 결과를 원래 JSON 구조에 붙이기
+# ============================================================
+
+def attach_results_to_data(
+    target_data: list[dict[str, Any]],
+    results: list[dict[str, Any]]
+) -> list[dict[str, Any]]:
+
+    for item in target_data:
+        item["caption_checks"] = []
+
+    results = sorted(
+        results,
+        key=lambda x: (x["item_index"], x["caption_index"])
+    )
+
+    for result in results:
+        item_index = result["item_index"]
+
+        check = {
+            "caption_index": result["caption_index"],
+            "caption": result["caption"],
+            "resolved_image_path": result.get("resolved_image_path"),
+            "clip_cosine": result.get("clip_cosine"),
+            "clip_score": result.get("clip_score"),
+            "clip_status": result.get("clip_status"),
+            "clip_reason": result.get("clip_reason", "")
+        }
+
+        target_data[item_index]["caption_checks"].append(check)
+
+    return target_data
+
+
+# ============================================================
+# 요약 출력
+# ============================================================
+
+def print_summary(
+    target_data: list[dict[str, Any]],
+    flat_items: list[dict[str, Any]],
+    results: list[dict[str, Any]]
+) -> None:
+
+    status_count = {}
+    valid_scores = []
+
+    for result in results:
+        status = result.get("clip_status", "unknown")
+        status_count[status] = status_count.get(status, 0) + 1
+
+        if isinstance(result.get("clip_score"), float):
+            valid_scores.append(result["clip_score"])
+
+    print("\n===== CLIP Score Summary =====")
+    print(f"check all classes: {CHECK_ALL_CLASSES}")
+
+    if CHECK_ALL_CLASSES:
+        print(f"data raw root dir: {DATA_RAW_ROOT_DIR}")
+    else:
+        print(f"target class dir: {TARGET_CLASS_DIR}")
+        print(f"target class name: {get_target_class_name()}")
+
+    print(f"target images: {len(target_data)}")
+    print(f"target image-caption pairs: {len(flat_items)}")
+    print(f"status count: {status_count}")
+
+    if valid_scores:
+        print(f"min score: {min(valid_scores):.4f}")
+        print(f"max score: {max(valid_scores):.4f}")
+        print(f"mean score: {np.mean(valid_scores):.4f}")
+        print(f"bottom {FAIL_BOTTOM_PERCENT}% threshold: {np.percentile(valid_scores, FAIL_BOTTOM_PERCENT):.4f}")
+        print(f"bottom {REVIEW_BOTTOM_PERCENT}% threshold: {np.percentile(valid_scores, REVIEW_BOTTOM_PERCENT):.4f}")
+
+
+# ============================================================
+# 실행
+# ============================================================
+
+def main():
+    if not INPUT_JSON_PATH.exists():
+        raise FileNotFoundError(f"input file not found: {INPUT_JSON_PATH}")
+
+    if CHECK_ALL_CLASSES:
+        if not DATA_RAW_ROOT_DIR.exists():
+            raise FileNotFoundError(f"data raw root directory not found: {DATA_RAW_ROOT_DIR}")
+    else:
+        if not TARGET_CLASS_DIR.exists():
+            raise FileNotFoundError(f"target class directory not found: {TARGET_CLASS_DIR}")
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    print(f"device: {device}")
+
+    print(f"loading model: {MODEL_NAME}")
+    model = CLIPModel.from_pretrained(MODEL_NAME, token=hf_token).to(device)
+    processor = CLIPProcessor.from_pretrained(MODEL_NAME, token=hf_token)
+    model.eval()
+
+    data = load_json(INPUT_JSON_PATH)
+    target_data, flat_items = flatten_caption_items(data)
+
+    if not target_data:
+        raise ValueError("검수 대상 데이터가 없습니다. CHECK_ALL_CLASSES 또는 TARGET_CLASS_DIR 설정을 확인하세요.")
+
+    results = compute_clip_scores(
+        flat_items=flat_items,
+        model=model,
+        processor=processor,
+        device=device
+    )
+
+    assign_clip_status(results)
+
+    checked_data = attach_results_to_data(target_data, results)
+
+    save_json(checked_data, OUTPUT_JSON_PATH)
+
+    print_summary(target_data, flat_items, results)
+    print(f"\nsaved: {OUTPUT_JSON_PATH}")
+
+
+if __name__ == "__main__":
+    main()

src/caption/generate_captions_blip.py

@@ -0,0 +1,220 @@
+import os                      # 파일/폴더 탐색
+import json                    # JSON 저장
+import random                  # 데이터 섞기
+import torch                   # GPU 사용
+import re                      # 정규식 (문장 필터링)
+from collections import defaultdict  # 클래스별 그룹화
+from PIL import Image          # 이미지 로드
+from transformers import BlipProcessor, BlipForConditionalGeneration  # BLIP
+from sentence_transformers import SentenceTransformer, util           # SBERT
+
+# ----------------------
+# 1. 설정
+# ----------------------
+ROOT_DIR = "data/raw"                         # 이미지 루트 폴더 (raw/클래스/이미지)
+OUTPUT_JSON = "annotation.json"         # 결과 JSON 파일 이름
+
+TARGET_CAPTIONS = 3                        # 이미지당 캡션 개수 (3 또는 5 추천)
+SIM_THRESHOLD = 0.85                       # 문장 유사도 기준 (높을수록 엄격)
+MIN_WORDS = 3                              # 최소 단어 수 (짧은 문장 제거)
+MAX_ATTEMPTS = 10                          # 캡션 생성 최대 반복 횟수
+
+TRAIN_RATIO = 0.7                          # train 비율
+VAL_RATIO = 0.15                           # val 비율
+TEST_RATIO = 0.15                          # test 비율
+
+device = "cuda" if torch.cuda.is_available() else "cpu"  # GPU 사용 여부
+print("device : ", device)
+# ----------------------
+# 2. 모델 로드
+# ----------------------
+processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")  
+# 이미지 → 토큰 변환
+
+blip_model = BlipForConditionalGeneration.from_pretrained(
+    "Salesforce/blip-image-captioning-base"
+).to(device)  
+# 캡션 생성 모델
+
+embedder = SentenceTransformer("all-MiniLM-L6-v2", device=device)  
+# 문장 → 벡터 (유사도 계산용)
+
+# ----------------------
+# 3. 캡션 생성 함수
+# ----------------------
+def generate_captions(image, n):
+    inputs = processor(images=image, return_tensors="pt").to(device)  # 이미지 전처리
+
+    outputs = blip_model.generate(
+        **inputs,
+        do_sample=True,        # 다양성 확보 (샘플링)
+        top_k=50,
+        top_p=0.95,
+        temperature=0.9,
+        num_return_sequences=n,  # n개 생성
+        max_length=30
+    )
+
+    # 토큰 → 문자열 변환
+    return [
+        processor.decode(o, skip_special_tokens=True).strip().lower()
+        for o in outputs
+    ]
+
+# ----------------------
+# 4. 기본 품질 필터
+# ----------------------
+def basic_filter(captions):
+    filtered = []
+
+    for c in captions:
+        words = c.split()
+
+        if len(words) < MIN_WORDS:          # 너무 짧은 문장 제거
+            continue
+
+        if len(set(words)) < len(words) * 0.6:  # 반복 단어 많은 문장 제거
+            continue
+
+        if re.search(r"[^a-z0-9\s]", c):    # 이상한 문자 제거
+            continue
+
+        filtered.append(c)
+
+    return filtered
+
+# ----------------------
+# 5. 키워드 추출
+# ----------------------
+def extract_keywords(caption):
+    stopwords = {"a","the","on","in","at","with","and","of","to","is","are"}  # 불용어
+    return set([w for w in caption.split() if w not in stopwords])  # 핵심 단어만 추출
+
+# ----------------------
+# 6. 유사도 + 키워드 필터
+# ----------------------
+def advanced_filter(captions):
+    if not captions:
+        return []
+
+    embeddings = embedder.encode(captions, convert_to_tensor=True)  # 문장 → 벡터
+
+    selected = []
+    selected_idx = []
+
+    for i, cap in enumerate(captions):
+        keep = True
+        kw_i = extract_keywords(cap)
+
+        for j in selected_idx:
+            sim = util.cos_sim(embeddings[i], embeddings[j]).item()  # cosine similarity
+
+            if sim > SIM_THRESHOLD:   # 의미가 너무 비슷하면 제거
+                keep = False
+                break
+
+            kw_j = extract_keywords(captions[j])
+            overlap = len(kw_i & kw_j) / max(len(kw_i), 1)
+
+            if overlap > 0.7:         # 키워드 많이 겹치면 제거
+                keep = False
+                break
+
+        if keep:
+            selected.append(cap)
+            selected_idx.append(i)
+
+    return selected
+
+# ----------------------
+# 7. 캡션 생성 루프
+# ----------------------
+def get_captions(image):
+    final_caps = []
+    attempts = 0
+
+    while len(final_caps) < TARGET_CAPTIONS and attempts < MAX_ATTEMPTS:
+        needed = TARGET_CAPTIONS - len(final_caps)
+
+        new_caps = generate_captions(image, needed * 3)  # 부족분보다 넉넉히 생성
+        new_caps = basic_filter(new_caps)                # 1차 필터
+
+        combined = list(set(final_caps + new_caps))      # 중복 제거
+        filtered = advanced_filter(combined)             # 유사도 필터
+
+        final_caps = filtered[:TARGET_CAPTIONS]          # 목표 개수 맞춤
+        attempts += 1
+
+    return final_caps
+
+# ----------------------
+# 8. 데이터 수집
+# ----------------------
+dataset = []
+
+for class_name in os.listdir(ROOT_DIR):                  # 클래스 폴더 순회
+    class_path = os.path.join(ROOT_DIR, class_name)
+
+    if not os.path.isdir(class_path):
+        continue
+
+    for filename in os.listdir(class_path):              # 이미지 순회
+        if not filename.lower().endswith((".jpg", ".jpeg", ".png")):
+            continue
+
+        path = os.path.join(class_path, filename)
+        image = Image.open(path).convert("RGB")          # 이미지 로드
+
+        captions = get_captions(image)                   # 캡션 생성
+
+        dataset.append({
+            "image": f"{class_name}/{filename}",         # 상대 경로 저장
+            "class": class_name,                         # 클래스 라벨
+            "captions": captions                         # 캡션 리스트
+        })
+
+        print(f"\n{class_name}/{filename}")
+        for i, c in enumerate(captions):
+            print(f"{i+1}. {c}")
+
+# ----------------------
+# 9. Stratified Split
+# ----------------------
+class_groups = defaultdict(list)
+
+for item in dataset:
+    class_groups[item["class"]].append(item)             # 클래스별 묶기
+
+train_set, val_set, test_set = [], [], []
+
+for class_name, items in class_groups.items():
+    random.shuffle(items)                                # 클래스 내부 shuffle
+
+    total = len(items)
+
+    train_end = max(1, int(total * TRAIN_RATIO))         # 최소 1개 보장
+    val_end = train_end + max(1, int(total * VAL_RATIO))
+
+    train_set += items[:train_end]
+    val_set += items[train_end:val_end]
+    test_set += items[val_end:]
+
+# split 라벨 부여
+for item in train_set:
+    item["split"] = "train"
+
+for item in val_set:
+    item["split"] = "val"
+
+for item in test_set:
+    item["split"] = "test"
+
+dataset = train_set + val_set + test_set                 # 다시 하나로 합침
+
+# ----------------------
+# 10. JSON 저장
+# ----------------------
+with open(OUTPUT_JSON, "w", encoding="utf-8") as f:
+    json.dump(dataset, f, indent=4, ensure_ascii=False)
+
+print(f"\n완료: {OUTPUT_JSON} 생성됨")

src/caption/generate_captions_florence2.py

@@ -0,0 +1,345 @@
+import os
+import json
+import random
+from pathlib import Path
+
+import torch
+from PIL import Image
+from tqdm import tqdm
+from dotenv import load_dotenv
+from transformers import AutoProcessor, Florence2ForConditionalGeneration
+
+
+# =========================================================
+# 1. 설정값
+# =========================================================
+
+# 전체 클래스 캡셔닝: "data/raw"
+# 특정 클래스만 캡셔닝: "data/raw/apple"
+INPUT_IMAGE_DIR = "data/raw"
+
+# image 값을 "pizza/hf_pizza_001.jpg" 형태로 만들기 위한 기준 경로
+DATA_RAW_ROOT = "data/raw"
+
+# 결과 JSON 저장 경로
+OUTPUT_JSON_PATH = "data/annotations/captions_flo_all.json"
+
+# transformers 5.7.0에서는 florence-community 모델 사용 권장
+# base-ft: 가볍고 다운스트림 task에 fine-tuning된 모델
+# large-ft: 더 무겁지만 품질이 더 좋을 수 있음
+MODEL_ID = "florence-community/Florence-2-base-ft"
+# MODEL_ID = "florence-community/Florence-2-large-ft"
+
+# .env 파일에서 읽을 Hugging Face 토큰 이름
+# 공개 모델이면 없어도 동작할 수 있지만, 토큰을 넣어두는 편이 안정적입니다.
+HF_TOKEN_ENV_NAME = "HF_TOKEN"
+
+# split 비율: 기본 7 : 1.5 : 1.5
+TRAIN_RATIO = 0.7
+VAL_RATIO = 0.15
+TEST_RATIO = 0.15
+
+# split 재현을 위한 seed
+RANDOM_SEED = 42
+
+# 이미지당 캡션 3개 생성
+# Florence-2 문서에서 지원하는 caption task입니다.
+CAPTION_TASKS = [
+    "<CAPTION>",
+    "<DETAILED_CAPTION>",
+    "<MORE_DETAILED_CAPTION>",
+]
+
+# 생성 옵션
+NUM_BEAMS = 3
+MAX_NEW_TOKENS = 64
+
+# 몇 장마다 중간 저장할지
+SAVE_EVERY = 220
+
+# 이미 JSON에 있는 이미지는 건너뛸지 여부
+SKIP_ALREADY_DONE = True
+
+# 허용 이미지 확장자
+IMAGE_EXTENSIONS = [".jpg", ".jpeg", ".png", ".webp", ".bmp"]
+
+
+# =========================================================
+# 2. 이미지 목록 가져오기
+# =========================================================
+
+def get_image_list():
+    input_dir = Path(INPUT_IMAGE_DIR).resolve()
+    raw_root = Path(DATA_RAW_ROOT).resolve()
+
+    if not input_dir.exists():
+        raise FileNotFoundError(f"입력 경로가 없습니다: {input_dir}")
+
+    image_list = []
+
+    for image_path in sorted(input_dir.rglob("*")):
+        if image_path.suffix.lower() not in IMAGE_EXTENSIONS:
+            continue
+
+        # 예:
+        # /workspace/data/raw/pizza/hf_pizza_001.jpg
+        # -> pizza/hf_pizza_001.jpg
+        relative_image_path = image_path.resolve().relative_to(raw_root).as_posix()
+
+        # 예:
+        # pizza/hf_pizza_001.jpg
+        # -> pizza
+        class_name = relative_image_path.split("/")[0]
+
+        image_list.append({
+            "path": image_path,
+            "image": relative_image_path,
+            "class": class_name,
+        })
+
+    return image_list
+
+
+# =========================================================
+# 3. train / val / test 나누기
+# =========================================================
+
+def add_split(image_list):
+    random.seed(RANDOM_SEED)
+
+    total_ratio = TRAIN_RATIO + VAL_RATIO + TEST_RATIO
+
+    result = []
+
+    # 클래스별로 이미지 모으기
+    class_map = {}
+
+    for item in image_list:
+        class_name = item["class"]
+
+        if class_name not in class_map:
+            class_map[class_name] = []
+
+        class_map[class_name].append(item)
+
+    # 클래스별로 train / val / test 나누기
+    for class_name, items in class_map.items():
+        random.shuffle(items)
+
+        total_count = len(items)
+
+        train_count = round(total_count * TRAIN_RATIO / total_ratio)
+        val_count = round(total_count * VAL_RATIO / total_ratio)
+
+        for index, item in enumerate(items):
+            if index < train_count:
+                split = "train"
+            elif index < train_count + val_count:
+                split = "val"
+            else:
+                split = "test"
+
+            item["split"] = split
+            result.append(item)
+
+    return result
+
+
+# =========================================================
+# 4. Florence-2 모델 준비
+# =========================================================
+
+def load_model():
+    load_dotenv()
+
+    hf_token = os.getenv(HF_TOKEN_ENV_NAME)
+
+    if torch.cuda.is_available():
+        device = "cuda"
+
+        # GPU가 bfloat16을 지원하면 bfloat16 사용
+        # 아니면 float16 사용
+        if torch.cuda.is_bf16_supported():
+            torch_dtype = torch.bfloat16
+        else:
+            torch_dtype = torch.float16
+    else:
+        device = "cpu"
+        torch_dtype = torch.float32
+
+    print(f"device: {device}")
+    print(f"dtype: {torch_dtype}")
+    print(f"model: {MODEL_ID}")
+
+    processor = AutoProcessor.from_pretrained(
+        MODEL_ID,
+        token=hf_token,
+    )
+
+    model = Florence2ForConditionalGeneration.from_pretrained(
+        MODEL_ID,
+        dtype=torch_dtype,
+        token=hf_token,
+    ).to(device)
+
+    model.eval()
+
+    return model, processor, device, torch_dtype
+
+
+# =========================================================
+# 5. 이미지 1장 캡셔닝
+# =========================================================
+
+def make_caption(image, task, model, processor, device, torch_dtype):
+    inputs = processor(
+        text=task,
+        images=image,
+        return_tensors="pt",
+    )
+
+    inputs = inputs.to(device, torch_dtype)
+
+    with torch.no_grad():
+        generated_ids = model.generate(
+            **inputs,
+            max_new_tokens=MAX_NEW_TOKENS,
+            num_beams=NUM_BEAMS,
+            do_sample=False,
+        )
+
+    generated_text = processor.batch_decode(
+        generated_ids,
+        skip_special_tokens=False,
+    )[0]
+
+    parsed_result = processor.post_process_generation(
+        generated_text,
+        task=task,
+        image_size=image.size,
+    )
+
+    caption = parsed_result.get(task, "")
+
+    if not isinstance(caption, str):
+        caption = str(caption)
+
+    return caption.strip()
+
+
+def make_three_captions(image_path, model, processor, device, torch_dtype):
+    image = Image.open(image_path).convert("RGB")
+
+    captions = []
+
+    for task in CAPTION_TASKS:
+        caption = make_caption(
+            image=image,
+            task=task,
+            model=model,
+            processor=processor,
+            device=device,
+            torch_dtype=torch_dtype,
+        )
+
+        captions.append(caption)
+
+    return captions
+
+
+# =========================================================
+# 6. 기존 JSON 읽기 / 저장하기
+# =========================================================
+
+def load_existing_result():
+    output_path = Path(OUTPUT_JSON_PATH)
+
+    if not output_path.exists():
+        return {}
+
+    with output_path.open("r", encoding="utf-8") as f:
+        data = json.load(f)
+
+    result = {}
+
+    for item in data:
+        result[item["image"]] = item
+
+    return result
+
+
+def save_result(result_map):
+    output_path = Path(OUTPUT_JSON_PATH)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    result_list = list(result_map.values())
+    result_list.sort(key=lambda x: x["image"])
+
+    with output_path.open("w", encoding="utf-8") as f:
+        json.dump(result_list, f, ensure_ascii=False, indent=4)
+
+
+# =========================================================
+# 7. 실행
+# =========================================================
+
+def main():
+    print("이미지 목록을 읽는 중입니다.")
+
+    image_list = get_image_list()
+    image_list = add_split(image_list)
+
+    print(f"총 이미지 수: {len(image_list)}")
+
+    result_map = load_existing_result()
+
+    model, processor, device, torch_dtype = load_model()
+
+    new_count = 0
+    skip_count = 0
+    fail_count = 0
+
+    for item in tqdm(image_list):
+        image_key = item["image"]
+
+        if SKIP_ALREADY_DONE and image_key in result_map:
+            skip_count += 1
+            continue
+
+        try:
+            captions = make_three_captions(
+                image_path=item["path"],
+                model=model,
+                processor=processor,
+                device=device,
+                torch_dtype=torch_dtype,
+            )
+
+            result_map[image_key] = {
+                "image": item["image"],
+                "class": item["class"],
+                "captions": captions,
+                "split": item["split"],
+            }
+
+            new_count += 1
+
+            if new_count % SAVE_EVERY == 0:
+                save_result(result_map)
+
+        except Exception as e:
+            fail_count += 1
+            print(f"\n실패한 이미지: {item['path']}")
+            print(f"에러 내용: {e}")
+
+    save_result(result_map)
+
+    print("\n캡셔닝 완료")
+    print(f"새로 처리한 이미지 수: {new_count}")
+    print(f"건너뛴 이미지 수: {skip_count}")
+    print(f"실패한 이미지 수: {fail_count}")
+    print(f"저장 위치: {OUTPUT_JSON_PATH}")
+
+
+if __name__ == "__main__":
+    main()

src/caption/generate_captions_git.py

@@ -0,0 +1,600 @@
+import json
+import math
+import random
+import re
+from collections import defaultdict
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import os
+import torch
+from dotenv import load_dotenv
+from PIL import Image
+from tqdm import tqdm
+from transformers import AutoModelForCausalLM, AutoProcessor
+
+
+# ============================================================
+# 1. 설정값
+# ============================================================
+
+load_dotenv()
+
+# .env 안의 HF_TOKEN 읽기
+hf_token = os.getenv("HF_TOKEN")
+
+# 이미지 원본 루트 경로
+# 예:
+# - 전체 클래스 캡셔닝: "data/raw"
+# - 특정 클래스만 캡셔닝: "data/raw/apple"
+INPUT_IMAGE_PATH = "data/raw/airplane"
+
+# image 필드를 만들 때 기준이 되는 root
+# JSON에는 "pizza/hf_pizza_001.jpg" 형태로 저장됨
+DATA_RAW_ROOT = "data/raw"
+
+# 결과 저장 경로
+OUTPUT_JSON_PATH = "data/annotations/captions_git.json"
+
+# 에러 이미지 목록 저장 경로
+ERROR_JSON_PATH = "data/annotations/caption_git_errors.json"
+
+# GIT 모델
+# 기본 추천: microsoft/git-base-coco
+# 더 큰 모델을 쓰고 싶으면: microsoft/git-large-coco
+MODEL_NAME = "microsoft/git-large-coco"
+
+# 이미지당 생성할 캡션 개수
+CAPTIONS_PER_IMAGE = 3
+
+# split 비율
+# 기본 7 : 1.5 : 1.5
+SPLIT_RATIO = {
+    "train": 0.7,
+    "val": 0.15,
+    "test": 0.15,
+}
+
+# split 재현성을 위한 seed
+RANDOM_SEED = 42
+
+# 추론 배치 크기
+# GPU 메모리가 부족하면 8 -> 4 -> 2 -> 1 순서로 줄이기
+BATCH_SIZE = 8
+
+# 장치 설정
+# "auto": CUDA 가능하면 GPU, 아니면 CPU
+# 직접 지정 가능: "cuda", "cpu"
+DEVICE = "auto"
+
+# dtype 설정
+# "auto": CUDA면 float16, CPU면 float32
+# 직접 지정 가능: "float32", "float16", "bfloat16"
+TORCH_DTYPE = "auto"
+
+# 중간 저장 간격
+# 이미지가 많을 때 중간에 오류가 나도 일부 결과를 보존하기 위한 설정
+SAVE_EVERY_N_IMAGES = 100
+
+# 기존 OUTPUT_JSON_PATH가 있으면 이미 캡셔닝된 이미지는 건너뛸지 여부
+RESUME_FROM_EXISTING_JSON = True
+
+# 지원 이미지 확장자
+SUPPORTED_EXTENSIONS = {
+    ".jpg", ".jpeg", ".png", ".webp", ".bmp"
+}
+
+# 캡션 생성 설정
+# num_beams >= num_return_sequences 여야 함
+# num_beam_groups를 사용해서 서로 조금 다른 caption을 생성하도록 함
+GENERATION_CONFIG = {
+    "max_length": 40,
+    "num_beams": 5,
+    "num_return_sequences": CAPTIONS_PER_IMAGE,
+    "early_stopping": True,
+    "no_repeat_ngram_size": 2,
+}
+
+# beam search 결과가 너무 중복될 때 추가 샘플링으로 보완할지 여부
+ENABLE_SAMPLING_FALLBACK = True
+
+SAMPLING_FALLBACK_CONFIG = {
+    "max_length": 40,
+    "do_sample": True,
+    "top_p": 0.9,
+    "temperature": 0.8,
+    "num_return_sequences": CAPTIONS_PER_IMAGE,
+    "no_repeat_ngram_size": 2,
+}
+
+
+# ============================================================
+# 2. 데이터 구조
+# ============================================================
+
+@dataclass
+class ImageItem:
+    path: Path
+    image_field: str
+    class_name: str
+    split: str = ""
+
+
+# ============================================================
+# 3. 유틸 함수
+# ============================================================
+
+def resolve_device() -> torch.device:
+    if DEVICE == "auto":
+        return torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    return torch.device(DEVICE)
+
+
+def resolve_dtype(device: torch.device) -> torch.dtype:
+    if TORCH_DTYPE == "auto":
+        return torch.float16 if device.type == "cuda" else torch.float32
+
+    dtype_map = {
+        "float32": torch.float32,
+        "float16": torch.float16,
+        "bfloat16": torch.bfloat16,
+    }
+
+    if TORCH_DTYPE not in dtype_map:
+        raise ValueError(f"지원하지 않는 TORCH_DTYPE입니다: {TORCH_DTYPE}")
+
+    if device.type == "cpu" and TORCH_DTYPE in {"float16", "bfloat16"}:
+        print("[WARN] CPU에서는 float16/bfloat16이 불안정할 수 있어 float32로 변경합니다.")
+        return torch.float32
+
+    return dtype_map[TORCH_DTYPE]
+
+
+def normalize_caption(text: str) -> str:
+    text = text.strip()
+    text = re.sub(r"\s+", " ", text)
+    return text
+
+
+def deduplicate_captions(captions: List[str]) -> List[str]:
+    result = []
+    seen = set()
+
+    for caption in captions:
+        caption = normalize_caption(caption)
+
+        if not caption:
+            continue
+
+        key = caption.lower()
+
+        if key in seen:
+            continue
+
+        seen.add(key)
+        result.append(caption)
+
+    return result
+
+
+def ensure_caption_count(captions: List[str], target_count: int) -> List[str]:
+    captions = deduplicate_captions(captions)
+
+    if len(captions) >= target_count:
+        return captions[:target_count]
+
+    if len(captions) == 0:
+        return [""] * target_count
+
+    while len(captions) < target_count:
+        captions.append(captions[-1])
+
+    return captions
+
+
+def save_json(path: Path, data: List[dict]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    with path.open("w", encoding="utf-8") as f:
+        json.dump(data, f, ensure_ascii=False, indent=2)
+
+
+def load_existing_json(path: Path) -> Dict[str, dict]:
+    if not path.exists():
+        return {}
+
+    with path.open("r", encoding="utf-8") as f:
+        data = json.load(f)
+
+    result = {}
+
+    for item in data:
+        image_key = item.get("image")
+
+        if image_key:
+            result[image_key] = item
+
+    return result
+
+
+# ============================================================
+# 4. 이미지 수집
+# ============================================================
+
+def collect_images(input_path: Path, data_raw_root: Path) -> List[ImageItem]:
+    if not input_path.exists():
+        raise FileNotFoundError(f"입력 경로가 존재하지 않습니다: {input_path}")
+
+    image_paths = sorted([
+        path
+        for path in input_path.rglob("*")
+        if path.is_file() and path.suffix.lower() in SUPPORTED_EXTENSIONS
+    ])
+
+    if not image_paths:
+        raise RuntimeError(f"이미지를 찾지 못했습니다: {input_path}")
+
+    items = []
+
+    for image_path in image_paths:
+        try:
+            relative_path = image_path.relative_to(data_raw_root)
+        except ValueError:
+            raise ValueError(
+                f"이미지 경로가 DATA_RAW_ROOT 하위에 있어야 합니다.\n"
+                f"image_path={image_path}\n"
+                f"DATA_RAW_ROOT={data_raw_root}"
+            )
+
+        if len(relative_path.parts) < 2:
+            raise ValueError(
+                f"이미지는 클래스 폴더 하위에 있어야 합니다: {image_path}\n"
+                f"예: data/raw/pizza/hf_pizza_001.jpg"
+            )
+
+        class_name = relative_path.parts[0]
+        image_field = relative_path.as_posix()
+
+        items.append(
+            ImageItem(
+                path=image_path,
+                image_field=image_field,
+                class_name=class_name,
+            )
+        )
+
+    return items
+
+
+# ============================================================
+# 5. split 분리
+# ============================================================
+
+def calculate_split_counts(total_count: int) -> Dict[str, int]:
+    ratio_sum = sum(SPLIT_RATIO.values())
+
+    raw_counts = {
+        split_name: total_count * ratio / ratio_sum
+        for split_name, ratio in SPLIT_RATIO.items()
+    }
+
+    counts = {
+        split_name: int(math.floor(count))
+        for split_name, count in raw_counts.items()
+    }
+
+    remaining = total_count - sum(counts.values())
+
+    # 소수점이 큰 split부터 남은 개수 배분
+    sorted_splits = sorted(
+        raw_counts.keys(),
+        key=lambda split_name: raw_counts[split_name] - counts[split_name],
+        reverse=True,
+    )
+
+    for split_name in sorted_splits[:remaining]:
+        counts[split_name] += 1
+
+    return counts
+
+
+def assign_splits(items: List[ImageItem]) -> List[ImageItem]:
+    rng = random.Random(RANDOM_SEED)
+
+    class_map = defaultdict(list)
+
+    for item in items:
+        class_map[item.class_name].append(item)
+
+    for class_name, class_items in class_map.items():
+        rng.shuffle(class_items)
+
+        counts = calculate_split_counts(len(class_items))
+
+        start = 0
+
+        for split_name in ["train", "val", "test"]:
+            end = start + counts.get(split_name, 0)
+
+            for item in class_items[start:end]:
+                item.split = split_name
+
+            start = end
+
+    return items
+
+
+# ============================================================
+# 6. 모델 로드
+# ============================================================
+
+def load_model():
+    device = resolve_device()
+    torch_dtype = resolve_dtype(device)
+
+    print(f"[INFO] device={device}")
+    print(f"[INFO] dtype={torch_dtype}")
+    print(f"[INFO] model={MODEL_NAME}")
+
+    processor = AutoProcessor.from_pretrained(MODEL_NAME, token=hf_token)
+
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_NAME,
+        dtype=torch_dtype,
+        token=hf_token
+    )
+    model.to(device)
+    model.eval()
+
+    return model, processor, device, torch_dtype
+
+
+# ============================================================
+# 7. 캡셔닝
+# ============================================================
+
+def load_batch_images(batch_items: List[ImageItem]) -> Tuple[List[Image.Image], List[ImageItem], List[dict]]:
+    images = []
+    valid_items = []
+    errors = []
+
+    for item in batch_items:
+        try:
+            with Image.open(item.path) as img:
+                images.append(img.convert("RGB"))
+                valid_items.append(item)
+        except Exception as e:
+            errors.append({
+                "image": item.image_field,
+                "class": item.class_name,
+                "split": item.split,
+                "error": str(e),
+            })
+
+    return images, valid_items, errors
+
+
+@torch.inference_mode()
+def generate_batch_captions(
+    model,
+    processor,
+    device: torch.device,
+    torch_dtype: torch.dtype,
+    images: List[Image.Image],
+) -> List[List[str]]:
+    inputs = processor(images=images, return_tensors="pt")
+
+    inputs = {
+        key: value.to(device)
+        for key, value in inputs.items()
+    }
+
+    if "pixel_values" in inputs:
+        inputs["pixel_values"] = inputs["pixel_values"].to(dtype=torch_dtype)
+
+    generated_ids = model.generate(
+        **inputs,
+        **GENERATION_CONFIG,
+    )
+
+    decoded = processor.batch_decode(
+        generated_ids,
+        skip_special_tokens=True,
+    )
+
+    grouped_captions = []
+
+    for i in range(len(images)):
+        start = i * CAPTIONS_PER_IMAGE
+        end = start + CAPTIONS_PER_IMAGE
+
+        captions = decoded[start:end]
+        captions = deduplicate_captions(captions)
+
+        grouped_captions.append(captions)
+
+    return grouped_captions
+
+
+@torch.inference_mode()
+def generate_sampling_fallback_captions(
+    model,
+    processor,
+    device: torch.device,
+    torch_dtype: torch.dtype,
+    image: Image.Image,
+) -> List[str]:
+    inputs = processor(images=[image], return_tensors="pt")
+
+    inputs = {
+        key: value.to(device)
+        for key, value in inputs.items()
+    }
+
+    if "pixel_values" in inputs:
+        inputs["pixel_values"] = inputs["pixel_values"].to(dtype=torch_dtype)
+
+    generated_ids = model.generate(
+        **inputs,
+        **SAMPLING_FALLBACK_CONFIG,
+    )
+
+    decoded = processor.batch_decode(
+        generated_ids,
+        skip_special_tokens=True,
+    )
+
+    return deduplicate_captions(decoded)
+
+
+def make_result_item(item: ImageItem, captions: List[str]) -> dict:
+    return {
+        "image": item.image_field,
+        "class": item.class_name,
+        "captions": captions,
+        "split": item.split,
+    }
+
+
+def caption_images(
+    model,
+    processor,
+    device: torch.device,
+    torch_dtype: torch.dtype,
+    items: List[ImageItem],
+    existing_result_map: Dict[str, dict],
+) -> Tuple[Dict[str, dict], List[dict]]:
+    result_map = dict(existing_result_map)
+    error_list = []
+
+    target_items = [
+        item
+        for item in items
+        if item.image_field not in result_map
+    ]
+
+    print(f"[INFO] 전체 이미지 수: {len(items)}")
+    print(f"[INFO] 기존 결과 수: {len(existing_result_map)}")
+    print(f"[INFO] 새로 캡셔닝할 이미지 수: {len(target_items)}")
+
+    processed_count = 0
+
+    for batch_start in tqdm(range(0, len(target_items), BATCH_SIZE), desc="Captioning"):
+        batch_items = target_items[batch_start:batch_start + BATCH_SIZE]
+
+        images, valid_items, errors = load_batch_images(batch_items)
+        error_list.extend(errors)
+
+        if not images:
+            continue
+
+        try:
+            batch_captions = generate_batch_captions(
+                model=model,
+                processor=processor,
+                device=device,
+                torch_dtype=torch_dtype,
+                images=images,
+            )
+
+            for image, item, captions in zip(images, valid_items, batch_captions):
+                if ENABLE_SAMPLING_FALLBACK and len(captions) < CAPTIONS_PER_IMAGE:
+                    fallback_captions = generate_sampling_fallback_captions(
+                        model=model,
+                        processor=processor,
+                        device=device,
+                        torch_dtype=torch_dtype,
+                        image=image,
+                    )
+
+                    captions = deduplicate_captions(captions + fallback_captions)
+
+                captions = ensure_caption_count(
+                    captions=captions,
+                    target_count=CAPTIONS_PER_IMAGE,
+                )
+
+                result_map[item.image_field] = make_result_item(
+                    item=item,
+                    captions=captions,
+                )
+
+                processed_count += 1
+
+        except Exception as e:
+            print("[ERROR] 배치 캡셔닝 실패")
+            print(f"[ERROR] {type(e).__name__}: {e}")
+
+            for item in valid_items:
+                error_list.append({
+                    "image": item.image_field,
+                    "class": item.class_name,
+                    "split": item.split,
+                    "error": str(e),
+                })
+
+        if SAVE_EVERY_N_IMAGES > 0 and processed_count > 0:
+            if processed_count % SAVE_EVERY_N_IMAGES == 0:
+                current_results = [
+                    result_map[item.image_field]
+                    for item in items
+                    if item.image_field in result_map
+                ]
+                save_json(Path(OUTPUT_JSON_PATH), current_results)
+                save_json(Path(ERROR_JSON_PATH), error_list)
+
+    return result_map, error_list
+
+
+# ============================================================
+# 8. main
+# ============================================================
+
+def main():
+    input_path = Path(INPUT_IMAGE_PATH).resolve()
+    data_raw_root = Path(DATA_RAW_ROOT).resolve()
+    output_json_path = Path(OUTPUT_JSON_PATH)
+    error_json_path = Path(ERROR_JSON_PATH)
+
+    items = collect_images(
+        input_path=input_path,
+        data_raw_root=data_raw_root,
+    )
+
+    items = assign_splits(items)
+
+    existing_result_map = {}
+
+    if RESUME_FROM_EXISTING_JSON:
+        existing_result_map = load_existing_json(output_json_path)
+
+    model, processor, device, torch_dtype = load_model()
+
+    result_map, error_list = caption_images(
+        model=model,
+        processor=processor,
+        device=device,
+        torch_dtype=torch_dtype,
+        items=items,
+        existing_result_map=existing_result_map,
+    )
+
+    final_results = [
+        result_map[item.image_field]
+        for item in items
+        if item.image_field in result_map
+    ]
+
+    save_json(output_json_path, final_results)
+    save_json(error_json_path, error_list)
+
+    print("[DONE] 캡셔닝 완료")
+    print(f"[DONE] 결과 저장: {output_json_path}")
+    print(f"[DONE] 에러 저장: {error_json_path}")
+    print(f"[DONE] 정상 결과 수: {len(final_results)}")
+    print(f"[DONE] 에러 수: {len(error_list)}")
+
+
+if __name__ == "__main__":
+    main()

src/caption/generate_captions_vit_gpt2.py

@@ -0,0 +1,457 @@
+import json
+import random
+from pathlib import Path
+from collections import defaultdict
+
+import torch
+from PIL import Image, UnidentifiedImageError
+from tqdm import tqdm
+from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
+
+
+# =========================================================
+# 1. 설정값
+# =========================================================
+
+# 전체 클래스 캡셔닝:
+# INPUT_IMAGE_DIR = "/workspace/data/raw"
+#
+# 특정 클래스만 캡셔닝:
+# INPUT_IMAGE_DIR = "/workspace/data/raw/apple"
+INPUT_IMAGE_DIR = "/workspace/data/raw/airplane"
+
+OUTPUT_JSON_PATH = "/workspace/data/annotations/annotation.json"
+
+MODEL_NAME = "nlpconnect/vit-gpt2-image-captioning"
+
+CAPTIONS_PER_IMAGE = 3
+
+SPLIT_RATIO = {
+    "train": 0.7,
+    "val": 0.15,
+    "test": 0.15,
+}
+
+RANDOM_SEED = 42
+
+BATCH_SIZE = 8
+
+IMAGE_EXTENSIONS = [".jpg", ".jpeg", ".png", ".webp", ".bmp"]
+
+# "auto": data/raw 입력 시 전체 클래스, data/raw/apple 입력 시 apple 클래스만 자동 판단
+# "raw": INPUT_IMAGE_DIR 아래를 전체 raw 폴더로 간주
+# "class": INPUT_IMAGE_DIR 자체를 하나의 클래스 폴더로 간주
+INPUT_MODE = "auto"
+
+# 캡션 문장 끝의 마침표 제거 여부
+REMOVE_TRAILING_PERIOD = True
+
+# beam search 설정
+GENERATION_CONFIG = {
+    "max_new_tokens": 32,
+    "num_beams": 8,
+    "num_return_sequences": CAPTIONS_PER_IMAGE,
+    "early_stopping": True,
+    "no_repeat_ngram_size": 2,
+    "repetition_penalty": 1.1,
+    "length_penalty": 0.8,
+}
+
+# beam search 결과가 중복될 때 샘플링으로 보충
+ENABLE_SAMPLING_FALLBACK = True
+
+SAMPLING_FALLBACK_CONFIG = {
+    "max_new_tokens": 32,
+    "do_sample": True,
+    "top_p": 0.9,
+    "temperature": 0.8,
+    "num_return_sequences": CAPTIONS_PER_IMAGE * 2,
+    "no_repeat_ngram_size": 2,
+    "repetition_penalty": 1.1,
+}
+
+MAX_FALLBACK_ROUNDS = 3
+
+# 그래도 3개를 못 채우면 중복을 허용해서라도 3개를 맞출지 여부
+FILL_WITH_DUPLICATES_IF_NEEDED = True
+
+
+# =========================================================
+# 2. 기본 유틸 함수
+# =========================================================
+
+def validate_config():
+    total_ratio = sum(SPLIT_RATIO.values())
+
+    if abs(total_ratio - 1.0) > 1e-6:
+        raise ValueError(f"SPLIT_RATIO의 합은 1이어야 합니다. 현재 합: {total_ratio}")
+
+    if GENERATION_CONFIG["num_beams"] < CAPTIONS_PER_IMAGE:
+        raise ValueError("num_beams는 CAPTIONS_PER_IMAGE보다 크거나 같아야 합니다.")
+
+    if GENERATION_CONFIG["num_return_sequences"] != CAPTIONS_PER_IMAGE:
+        raise ValueError("GENERATION_CONFIG의 num_return_sequences는 CAPTIONS_PER_IMAGE와 같아야 합니다.")
+
+
+def is_image_file(path: Path) -> bool:
+    return path.suffix.lower() in IMAGE_EXTENSIONS
+
+
+def clean_caption(text: str) -> str:
+    caption = " ".join(text.strip().split())
+
+    if REMOVE_TRAILING_PERIOD:
+        caption = caption.rstrip(".")
+
+    return caption
+
+
+def unique_captions(captions):
+    result = []
+    seen = set()
+
+    for caption in captions:
+        caption = clean_caption(caption)
+        key = caption.lower()
+
+        if caption and key not in seen:
+            result.append(caption)
+            seen.add(key)
+
+    return result
+
+
+def load_image(image_path: Path):
+    try:
+        return Image.open(image_path).convert("RGB")
+    except (UnidentifiedImageError, OSError) as e:
+        print(f"[SKIP] 이미지를 열 수 없습니다: {image_path} / error: {e}")
+        return None
+
+
+# =========================================================
+# 3. 이미지 목록 수집
+# =========================================================
+
+def has_direct_images(input_dir: Path) -> bool:
+    for child in input_dir.iterdir():
+        if child.is_file() and is_image_file(child):
+            return True
+    return False
+
+
+def get_relative_base_dir(input_dir: Path) -> Path:
+    """
+    JSON의 image 값을 '클래스폴더/이미지명' 형태로 만들기 위한 기준 경로를 정한다.
+
+    예시 1)
+    INPUT_IMAGE_DIR = /workspace/data/raw
+    image file      = /workspace/data/raw/pizza/hf_pizza_001.jpg
+    relative base   = /workspace/data/raw
+    result          = pizza/hf_pizza_001.jpg
+
+    예시 2)
+    INPUT_IMAGE_DIR = /workspace/data/raw/apple
+    image file      = /workspace/data/raw/apple/hf_apple_001.jpg
+    relative base   = /workspace/data/raw
+    result          = apple/hf_apple_001.jpg
+    """
+
+    if INPUT_MODE == "raw":
+        return input_dir
+
+    if INPUT_MODE == "class":
+        return input_dir.parent
+
+    if INPUT_MODE == "auto":
+        if has_direct_images(input_dir):
+            return input_dir.parent
+        return input_dir
+
+    raise ValueError("INPUT_MODE은 'auto', 'raw', 'class' 중 하나여야 합니다.")
+
+
+def collect_image_records(input_dir: str):
+    input_path = Path(input_dir)
+
+    if not input_path.exists():
+        raise FileNotFoundError(f"��미지 경로가 존재하지 않습니다: {input_path}")
+
+    relative_base_dir = get_relative_base_dir(input_path)
+
+    records = []
+
+    for image_path in sorted(input_path.rglob("*")):
+        if not image_path.is_file():
+            continue
+
+        if not is_image_file(image_path):
+            continue
+
+        relative_path = image_path.relative_to(relative_base_dir)
+        relative_path_str = relative_path.as_posix()
+
+        # image 값이 apple/xxx.jpg 라면 class는 apple
+        class_name = relative_path.parts[0]
+
+        records.append({
+            "path": image_path,
+            "image": relative_path_str,
+            "class": class_name,
+        })
+
+    if not records:
+        raise ValueError(f"캡셔닝할 이미지가 없습니다: {input_path}")
+
+    return records
+
+
+# =========================================================
+# 4. train / val / test split 배정
+# =========================================================
+
+def assign_split(records):
+    random.seed(RANDOM_SEED)
+
+    class_map = defaultdict(list)
+
+    for record in records:
+        class_map[record["class"]].append(record)
+
+    result = []
+
+    for class_name, items in class_map.items():
+        random.shuffle(items)
+
+        total = len(items)
+        train_count = int(total * SPLIT_RATIO["train"])
+        val_count = int(total * SPLIT_RATIO["val"])
+
+        for idx, item in enumerate(items):
+            if idx < train_count:
+                item["split"] = "train"
+            elif idx < train_count + val_count:
+                item["split"] = "val"
+            else:
+                item["split"] = "test"
+
+            result.append(item)
+
+    result.sort(key=lambda x: x["image"])
+
+    return result
+
+
+# =========================================================
+# 5. 모델 로드
+# =========================================================
+
+def load_model():
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    print(f"[INFO] device: {device}")
+    print(f"[INFO] model: {MODEL_NAME}")
+
+    model = VisionEncoderDecoderModel.from_pretrained(MODEL_NAME)
+    processor = ViTImageProcessor.from_pretrained(MODEL_NAME)
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    model.config.pad_token_id = tokenizer.pad_token_id
+    model.to(device)
+    model.eval()
+
+    return model, processor, tokenizer, device
+
+
+# =========================================================
+# 6. 캡션 생성
+# =========================================================
+
+def decode_output_ids(output_ids, tokenizer):
+    captions = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
+    return [clean_caption(caption) for caption in captions]
+
+
+@torch.no_grad()
+def generate_by_beam_search(images, model, processor, tokenizer, device):
+    pixel_values = processor(
+        images=images,
+        return_tensors="pt"
+    ).pixel_values.to(device)
+
+    output_ids = model.generate(
+        pixel_values,
+        **GENERATION_CONFIG
+    )
+
+    captions = decode_output_ids(output_ids, tokenizer)
+
+    grouped = []
+    start = 0
+
+    for _ in images:
+        end = start + CAPTIONS_PER_IMAGE
+        grouped.append(captions[start:end])
+        start = end
+
+    return grouped
+
+
+@torch.no_grad()
+def generate_by_sampling(image, model, processor, tokenizer, device):
+    pixel_values = processor(
+        images=[image],
+        return_tensors="pt"
+    ).pixel_values.to(device)
+
+    output_ids = model.generate(
+        pixel_values,
+        **SAMPLING_FALLBACK_CONFIG
+    )
+
+    return decode_output_ids(output_ids, tokenizer)
+
+
+def complete_caption_count(captions, original_candidates):
+    """
+    기본 목표:
+    - 최대한 중복 없는 캡션 3개를 만든다.
+
+    단, 모델이 비슷한 문장만 계속 만들면 3개를 못 채울 수 있다.
+    이때 FILL_WITH_DUPLICATES_IF_NEEDED=True이면 중복을 허용해서 3개를 맞춘다.
+    """
+
+    captions = unique_captions(captions)
+
+    if len(captions) >= CAPTIONS_PER_IMAGE:
+        return captions[:CAPTIONS_PER_IMAGE]
+
+    if not FILL_WITH_DUPLICATES_IF_NEEDED:
+        return captions
+
+    for caption in original_candidates:
+        caption = clean_caption(caption)
+
+        if caption:
+            captions.append(caption)
+
+        if len(captions) >= CAPTIONS_PER_IMAGE:
+            break
+
+    return captions[:CAPTIONS_PER_IMAGE]
+
+
+def generate_captions_for_batch(batch_records, model, processor, tokenizer, device):
+    images = []
+    valid_records = []
+
+    for record in batch_records:
+        image = load_image(record["path"])
+
+        if image is None:
+            continue
+
+        images.append(image)
+        valid_records.append(record)
+
+    if not images:
+        return []
+
+    beam_caption_groups = generate_by_beam_search(
+        images=images,
+        model=model,
+        processor=processor,
+        tokenizer=tokenizer,
+        device=device
+    )
+
+    results = []
+
+    for record, image, beam_captions in zip(valid_records, images, beam_caption_groups):
+        all_candidates = list(beam_captions)
+        captions = unique_captions(beam_captions)
+
+        if ENABLE_SAMPLING_FALLBACK:
+            fallback_round = 0
+
+            while len(captions) < CAPTIONS_PER_IMAGE and fallback_round < MAX_FALLBACK_ROUNDS:
+                sampled_captions = generate_by_sampling(
+                    image=image,
+                    model=model,
+                    processor=processor,
+                    tokenizer=tokenizer,
+                    device=device
+                )
+
+                all_candidates.extend(sampled_captions)
+                captions = unique_captions(captions + sampled_captions)
+                fallback_round += 1
+
+        captions = complete_caption_count(
+            captions=captions,
+            original_candidates=all_candidates
+        )
+
+        results.append({
+            "image": record["image"],
+            "class": record["class"],
+            "captions": captions,
+            "split": record["split"],
+        })
+
+    return results
+
+
+# =========================================================
+# 7. JSON 저장
+# =========================================================
+
+def save_json(data, output_path: str):
+    output_path = Path(output_path)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    with open(output_path, "w", encoding="utf-8") as f:
+        json.dump(data, f, ensure_ascii=False, indent=4)
+
+    print(f"[DONE] JSON 저장 완료: {output_path}")
+    print(f"[DONE] 총 이미지 수: {len(data)}")
+
+
+# =========================================================
+# 8. 실행
+# =========================================================
+
+def main():
+    validate_config()
+
+    records = collect_image_records(INPUT_IMAGE_DIR)
+    records = assign_split(records)
+
+    print(f"[INFO] 캡셔닝 대상 이미지 수: {len(records)}")
+
+    model, processor, tokenizer, device = load_model()
+
+    results = []
+
+    for start in tqdm(range(0, len(records), BATCH_SIZE), desc="captioning"):
+        end = start + BATCH_SIZE
+        batch_records = records[start:end]
+
+        batch_results = generate_captions_for_batch(
+            batch_records=batch_records,
+            model=model,
+            processor=processor,
+            tokenizer=tokenizer,
+            device=device
+        )
+
+        results.extend(batch_results)
+
+    save_json(results, OUTPUT_JSON_PATH)
+
+
+if __name__ == "__main__":
+    main()

src/collection/check_class_counts.py

@@ -0,0 +1,81 @@
+import os
+
+# ================================
+# 0. 설정
+# ================================
+TARGET_COUNT = 60
+MIN_RES = 256  # 해상도 256
+PREFIX = "kg"
+BASE_DIR = "./data/raw"
+
+# ================================
+# 1. 경로
+# ================================
+HOME = os.path.expanduser("~")
+
+DATA_DIR = os.path.join(
+    HOME,
+    "Desktop",
+    "raw_kg"
+)
+
+THRESHOLD = TARGET_COUNT
+
+# ================================
+# 2. 클래스 목록
+# ================================
+CLASS_LIST = [
+    # 음식 및 식재료
+    "pizza","hamburger","sushi","pasta","salad",
+    "steak","cup_cake","sandwich","waffle","dumpling",
+
+    # 동물
+    "golden-retriever","bulldog","siamese_cat",
+    "persian_cat","elephant","sheep","horse",
+    "penguin","butterfly","squirrel",
+
+    # 꽃
+    "rose","sunflower","daisy","tulip","dandelion",
+    "lily","lavender","orchid","iris","marigold","aster",
+
+    # 과일
+    "apple","banana","strawberry","orange",
+    "carrot","tomato","cucumber",
+
+    # 탈것
+    "car","bicycle","motorcycle","airplane","bus",
+
+    # 패션 및 잡화
+    "t-shirt","sneakers","earrings","glasses",
+    "pants","bracelet","necklace"
+]
+
+print(f"{THRESHOLD}장 이하 클래스 목록 (0장 포함)\n")
+
+low_classes = []
+
+# ================================
+# 3. 클래스별 개수 체크
+# ================================
+for cls in sorted(CLASS_LIST):
+
+    cls_path = os.path.join(DATA_DIR, cls)
+
+    if not os.path.exists(cls_path):
+        count = 0
+
+    else:
+        count = len([
+            f for f in os.listdir(cls_path)
+            if os.path.isfile(os.path.join(cls_path, f))
+        ])
+
+    if count < THRESHOLD:
+        print(f"{cls}: {count}장")
+        low_classes.append((cls, count))
+
+# ================================
+# 4. 요약
+# ================================
+print("\n요약")
+print(f"{THRESHOLD}장 미만 클래스 수: {len(low_classes)}개")

src/collection/collect_filtering_images.py

@@ -0,0 +1,228 @@
+import os
+import hashlib
+from PIL import Image
+
+# ================================
+# 0. 설정
+# ================================
+TARGET_COUNT = 60
+MIN_RES = 128  # 해상도 128
+PREFIX = "kg"
+BASE_DIR = "./data/raw"
+
+# ================================
+# 1. 클래스 + 유사어 매핑
+# ================================
+CLASS_MAP = {
+    # 음식
+    "pizza": ["pizza"],
+    "hamburger": ["hamburger"],
+    "sushi": ["sushi"],
+    "pasta": ["pasta", "spaghetti"],
+    "salad": ["salad"],
+    "steak": ["steak"],
+    "cup_cake": ["cup_cake", "cup cake"],
+    "sandwich": ["sandwich"],
+    "waffle": ["waffle"],
+    "dumpling": ["dumpling"],
+
+    # 동물
+    "golden-retriever": ["golden retriever"],
+    "bulldog": ["bulldog"],
+    "siamese-cat": ["siamese"],
+    "persian-cat": ["persian"],
+    "elephant": ["elephant"],
+    "sheep": ["sheep"],
+    "horse": ["horse"],
+    "penguin": ["penguin"],
+    "butterfly": ["butterfly"],
+    "squirrel": ["squirrel"],
+
+    # 꽃
+    "rose": ["rose"],
+    "sunflower": ["sunflower"],
+    "daisy": ["daisy"],
+    "tulip": ["tulip"],
+    "dandelion": ["dandelion"],
+    "lily": ["lily"],
+    "lavender": ["lavender"],
+    "orchid": ["orchid"],
+    "iris": ["iris"],
+    "marigold": ["marigold"],
+    "aster": ["aster"],
+
+    # 과일
+    "apple": ["apple"],
+    "banana": ["banana"],
+    "strawberry": ["strawberry"],
+    "orange": ["orange"],
+    "carrot": ["carrot"],
+    "tomato": ["tomato"],
+    "cucumber": ["cucumber"],
+
+    # 탈것
+    "car": ["car"],
+    "bicycle": ["bicycle"],
+    "motorcycle": ["motorcycle"],
+    "airplane": ["airplane"],
+    "bus": ["bus"],
+
+    # 패션 및 잡화
+    "t-shirt": ["t shirt", "t-shirt"],
+    "sneakers": ["sneakers"],
+    "earrings": ["earring", "earrings"],
+    "glasses": ["glasses"],
+    "pants": ["pants"],
+    "bracelet": ["bracelet"],
+    "necklace": ["necklace"]
+}
+
+# ================================
+# 2. 경로
+# ================================
+HOME = os.path.expanduser("~")
+
+SRC_ROOT = os.path.join(
+    HOME,
+    "Desktop",
+    "raw_full_kg",
+    "extracted"
+)
+
+DST_ROOT = os.path.join(
+    HOME,
+    "Desktop",
+    "raw_kg"
+)
+
+os.makedirs(DST_ROOT, exist_ok=True)
+
+# ================================
+# 3. 해상도 필터
+# ================================
+def is_valid_image(path, min_size=MIN_RES):
+    try:
+        with Image.open(path) as img:
+            w, h = img.size
+            return w >= min_size and h >= min_size
+    except:
+        return False
+
+# ================================
+# 4. 중복 제거
+# ================================
+def get_hash(path):
+    try:
+        with open(path, "rb") as f:
+            return hashlib.md5(f.read()).hexdigest()
+    except:
+        return None
+
+seen_hashes = set()
+
+# ================================
+# 5. 클래스 매칭
+# ================================
+def match_class(folder_name):
+    name = folder_name.lower()
+    name = name.replace("-", " ").replace("_", " ")
+
+    words = name.split()
+
+    for target, keywords in CLASS_MAP.items():
+        for kw in keywords:
+            kw_words = kw.split()
+
+            if all(word in words for word in kw_words):
+                return target
+
+    return None
+
+# ================================
+# 6. 메인 로직
+# ================================
+class_counter = {
+    cls: 1 for cls in CLASS_MAP.keys()
+}
+
+copied = 0
+skipped = 0
+no_match = 0
+
+for root, dirs, files in os.walk(SRC_ROOT):
+
+    for d in dirs:
+        matched_class = match_class(d)
+
+        if matched_class is None:
+            no_match += 1
+            continue
+
+        src_path = os.path.join(root, d)
+        dst_path = os.path.join(DST_ROOT, matched_class)
+
+        for img in os.listdir(src_path):
+
+            src_file = os.path.join(src_path, img)
+
+            if not os.path.isfile(src_file):
+                continue
+
+            # 이미지 검증
+            try:
+                with Image.open(src_file) as im:
+                    im.verify()
+
+            except:
+                skipped += 1
+                continue
+
+            # 해상도 필터
+            if not is_valid_image(src_file):
+                skipped += 1
+                continue
+
+            # 중복 제거
+            img_hash = get_hash(src_file)
+
+            if img_hash is None or img_hash in seen_hashes:
+                skipped += 1
+                continue
+
+            seen_hashes.add(img_hash)
+
+            if not os.path.exists(dst_path):
+                os.makedirs(dst_path, exist_ok=True)
+
+            number = str(
+                class_counter[matched_class]
+            ).zfill(3)
+
+            class_name_for_file = matched_class.replace("_", "-")
+
+            new_name = (
+                f"{PREFIX}_{class_name_for_file}_{number}.jpg"
+            )
+
+            dst_file = os.path.join(dst_path, new_name)
+
+            try:
+                with Image.open(src_file) as im:
+                    im.convert("RGB").save(
+                        dst_file,
+                        "JPEG"
+                    )
+
+                class_counter[matched_class] += 1
+                copied += 1
+
+                if copied % 100 == 0:
+                    print(f"{copied}장 처리 중...")
+
+            except:
+                skipped += 1
+
+print("\n완료!")
+print(f"복사: {copied}")
+print(f"스킵: {skipped}")
+print(f"매칭 실패 폴더: {no_match}")

src/collection/count_label_hf.py

@@ -0,0 +1,115 @@
+from datasets import load_dataset
+from collections import Counter
+from dotenv import load_dotenv
+import os
+
+# ============================================================
+# [설정 부분]
+# ============================================================
+load_dotenv()
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+# 확인할 Hugging Face 데이터셋 이름
+DATASET_NAME = "jbarat/plant_species"  # 예: "uran66/animals"
+
+# 확인할 split 이름
+SPLIT_NAME = "train"
+
+# 라벨 필드명
+LABEL_FIELD_NAME = "label"
+
+# streaming 사용 여부
+# True  : 전체 데이터셋을 미리 다운로드하지 않고 하나씩 읽으면서 확인
+# False : 로컬 캐시에 데이터셋을 다운로드한 뒤 확인
+USE_STREAMING = True
+
+# 문자열 라벨 데이터셋일 경우 전체 데이터를 훑어야 정확한 개수를 알 수 있다.
+# None이면 전체 확인, 숫자를 넣으면 일부 샘플만 확인한다.
+MAX_SCAN_ITEMS = None
+
+# ============================================================
+
+
+def get_label_name(dataset, label_value):
+
+    label_feature = dataset.features[LABEL_FIELD_NAME]
+
+    # ClassLabel 타입이면 숫자 라벨을 문자열 라벨명으로 변환한다.
+    if hasattr(label_feature, "int2str") and isinstance(label_value, int):
+        return label_feature.int2str(label_value)
+
+    # 이미 문자열 라벨이면 그대로 문자열로 변환해서 사용한다.
+    return str(label_value)
+
+
+def get_unique_labels_with_counts():
+
+    print(f"[{DATASET_NAME}] 데이터셋 로드 중...")
+
+    dataset = load_dataset(
+        DATASET_NAME,
+        split=SPLIT_NAME,
+        streaming=USE_STREAMING,
+        token=HF_TOKEN
+    )
+
+    # 데이터셋의 feature 정보에서 라벨 필드를 가져온다.
+    label_feature = dataset.features[LABEL_FIELD_NAME]
+
+    # 클래스별 이미지 개수를 저장할 Counter
+    label_counter = Counter()
+
+    print("\n클래스별 이미지 개수 집계 중...")
+
+    # streaming=True인 경우에도 dataset을 순회하면서 개수를 셀 수 있다.
+    for idx, item in enumerate(dataset):
+        # MAX_SCAN_ITEMS가 설정되어 있으면 지정한 개수까지만 확인한다.
+        if MAX_SCAN_ITEMS is not None and idx >= MAX_SCAN_ITEMS:
+            break
+
+        label_value = item.get(LABEL_FIELD_NAME)
+
+        # 라벨 값이 없는 데이터는 건너뛴다.
+        if label_value is None:
+            continue
+
+        # 숫자 라벨이면 실제 라벨명으로 변환하고,
+        # 문자열 라벨이면 그대로 사용한다.
+        label_name = get_label_name(dataset, label_value)
+
+        # 해당 라벨의 이미지 개수를 1 증가시킨다.
+        label_counter[label_name] += 1
+
+    print("\n라벨 목록 및 클래스별 이미지 개수")
+    print("-" * 60)
+
+    # ------------------------------------------------------------
+    # 1. Food101처럼 label이 ClassLabel 타입인 경우
+    # ------------------------------------------------------------
+    # label_feature.names가 있으면 원래 데이터셋의 라벨 순서대로 출력한다.
+    if hasattr(label_feature, "names") and label_feature.names is not None:
+        label_names = label_feature.names
+
+        for idx, label_name in enumerate(label_names):
+            count = label_counter.get(label_name, 0)
+            print(f"{idx}: {label_name} - {count} 장")
+
+    # ------------------------------------------------------------
+    # 2. label이 문자열로 직접 들어있는 데이터셋인 경우
+    # ------------------------------------------------------------
+    # Counter에 모인 라벨명을 이름순으로 정렬해서 출력한다.
+    else:
+        label_names = sorted(label_counter.keys())
+
+        for idx, label_name in enumerate(label_names):
+            count = label_counter[label_name]
+            print(f"{idx}: {label_name} - {count} 장")
+
+    print("-" * 60)
+    print(f"총 라벨 개수: {len(label_counter)}")
+    print(f"총 이미지 개수: {sum(label_counter.values())}")
+
+    return label_counter
+
+if __name__ == "__main__":
+    get_unique_labels_with_counts()

src/collection/download_dataset_hf.py

@@ -0,0 +1,187 @@
+import os
+import re
+import glob
+from dotenv import load_dotenv
+from datasets import load_dataset
+from PIL import Image
+
+# =====================================================================
+# [설정 부분]
+# =====================================================================
+# 토큰
+load_dotenv()
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+print(f"이거 토큰 : {HF_TOKEN}")
+
+# 수집할 데이터셋
+DATASET_NAME = "KrushiJethe/fashion_data" 
+# 데이터셋 내의 이미지 데이터가 있는 필드명
+IMAGE_FIELD_NAME = "image"
+# 데이터셋 내의 라벨 데이터가 있는 필드명
+LABEL_FIELD_NAME = "articleType"
+
+# 여러 라벨을 하나의 대표 클래스로 묶는 매핑 딕셔너리
+CLASS_MAPPING = {
+    "t-shirt": ["Tshirts", "Tops"],
+    "sneakers":["Casual Shoes"],
+    #"umbrella":["Umbrellas"],
+    "glasses":["Sunglasses"],
+    "pants":["Jeans"],
+}
+
+# 클래스별로 수집할 이미지의 최대 개수
+NUM_IMAGES_PER_CLASS = 100
+# 저장할 이미지의 해상도 (width, height)
+TARGET_RESOLUTION = 256 
+# 이미지를 저장할 최상위 디렉토리명
+BASE_SAVE_DIR = "./dataset_output" 
+# 수집할 데이터셋의 split 이름 (예: "train", "validation", "test")
+SPLIT_NAME = "train"
+
+# 컨테이너를 실행한 상태에서는 컨테이너에 캐시 저장됨
+# 캐시 확인 -> ls -lah ~/.cache/huggingface
+# 캐시 삭제 -> rm -rf ~/.cache/huggingface
+USE_STREAMING = False
+# =====================================================================
+
+# 클래스 명명 규칙 적용
+def format_class_name(class_name: str) -> str:
+    """
+    클래스명은 소문자로 하고 띄어쓰기가 있을 경우 "-"로 대체
+    """
+    return str(class_name).lower().replace("_", "-").replace(" ", "-")
+
+# 마지막 이미지의 번호 + 1
+def get_next_image_index(save_dir: str, formatted_class_name: str) -> int:
+    """
+    이미지를 여러 차례 이어서 수집할 수 있도록 마지막 이미지 번호를 탐색
+    디렉토리를 스캔하여 가장 높은 번호를 찾은 뒤 +1을 반환
+    """
+    if not os.path.exists(save_dir):
+        return 1
+    
+    # jpg와 jpeg 확장자 모두 검색
+    search_pattern_jpg = os.path.join(save_dir, f"hf_{formatted_class_name}_*.jpg")
+    search_pattern_jpeg = os.path.join(save_dir, f"hf_{formatted_class_name}_*.jpeg")
+    
+    existing_files = glob.glob(search_pattern_jpg) + glob.glob(search_pattern_jpeg)
+    
+    max_idx = 0
+    # 파일명에서 정규표현식을 통해 번호 추출 (예: hf_fried-chicken_001.jpg -> 1)
+    regex = re.compile(rf"hf_{formatted_class_name}_(\d+)\.jpe?g$")
+    
+    for file_path in existing_files:
+        basename = os.path.basename(file_path)
+        match = regex.match(basename)
+        if match:
+            idx = int(match.group(1))
+            if idx > max_idx:
+                max_idx = idx
+                
+    return max_idx + 1
+
+def collect_hf_images():
+    """
+    메인 데이터 수집 함수.
+    Hugging Face 데이터셋에서 설정을 반영하여 이미지를 수집하고 저장
+    """
+
+    label_to_rep_class = {}
+    for rep_class, labels in CLASS_MAPPING.items():
+        for label in labels:
+            label_to_rep_class[label] = rep_class
+    
+
+    print(label_to_rep_class)
+    # 데이터셋별 낱개로 수집
+    # streaming=True 속성을 사용하면 전체 데이터셋을 메모리나 디스크에 한 번에 다운로드하지 않고
+    # generator 형태로 하나씩(낱개로) 가져오므로 메모리와 네트워크 효율성이 극대화
+    print(f"[{DATASET_NAME}] 데이터셋 스트리밍 로드 시작...")
+    dataset = load_dataset(DATASET_NAME, split=SPLIT_NAME, streaming=USE_STREAMING, token=HF_TOKEN)
+    
+    # 랜덤으로 가져오기
+    # random_seed = random.randint(0, 10000)
+    # dataset = load_dataset(DATASET_NAME, split=SPLIT_NAME, streaming=USE_STREAMING).shuffle(seed=random_seed, buffer_size=1000)
+
+    # 클래스별로 포맷팅된 폴더명과, 현재까지 수집된 개수, 그리고 저장될 시작 번호를 관리할 딕셔너리
+    class_info = {}
+    for label in CLASS_MAPPING.keys():
+        formatted_name = format_class_name(label)
+        save_path = os.path.join(BASE_SAVE_DIR, formatted_name)
+        
+        # [규칙 1, 4] 클래스를 폴더로 관리하며 폴더명은 변환된 클래스명을 따른다.
+        os.makedirs(save_path, exist_ok=True)
+        
+        # 이어서 수집하기 위한 시작 인덱스 탐색
+        start_idx = get_next_image_index(save_path, formatted_name)
+        
+        class_info[label] = {
+            "formatted_name": formatted_name,
+            "save_path": save_path,
+            "collected_count": 0,
+            "current_idx": start_idx
+        }
+
+    print("데이터 수집을 시작합니다...")
+    
+    # 스트리밍 데이터 순회
+    for item in dataset:
+
+        print("1. 데이터셋 로드 시작...")
+        # 모든 클래스가 목표 수집량을 채웠��지 확인
+        if all(info["collected_count"] >= NUM_IMAGES_PER_CLASS for info in class_info.values()):
+            print("모든 클래스의 이미지 수집이 완료되었습니다.")
+            break
+
+        print("2. 데이터셋 라벨 아이템 꺼내기...")    
+        current_label = item.get(LABEL_FIELD_NAME)
+        
+        print(current_label)
+        # 현재 뽑힌 라벨이 정의한 매핑 딕셔너리에 존재하는지 확인
+        if current_label in label_to_rep_class:
+            rep_class = label_to_rep_class[current_label]
+            target_info = class_info[rep_class]
+            
+
+            print("4. 이미지 유효성 검사...")
+            # 이미 목표 개수를 채운 클래스라면 스킵
+            if target_info["collected_count"] >= NUM_IMAGES_PER_CLASS:
+                continue
+                
+            # 이미지 유효성 체크
+            image = item.get(IMAGE_FIELD_NAME)
+            if image is None:
+                continue
+            
+            print("5. 이미지 변환...")
+            try:
+                # 이미지를 jpg/jpeg로만 취급하기 위해 RGB 모드로 변환 (알파 채널 등 제거)
+                if image.mode != "RGB":
+                    image = image.convert("RGB")
+                    
+                #이미지 해상도가 최소 256px만 수집
+                if image.width < TARGET_RESOLUTION or image.height < TARGET_RESOLUTION:
+                    continue
+
+                print("6. 클래스 명명 규칙에 따라...")
+                # [규칙 3, 4] 이미지 명명 규칙 (hf_[클래스명]_[3자리숫자].jpg)
+                # {:03d}를 통해 3자리 숫자로 맞추고 빈자리는 0으로 채움
+                file_name = f"hf_{target_info['formatted_name']}_{target_info['current_idx']:03d}.jpg"
+                file_path = os.path.join(target_info["save_path"], file_name)
+                
+                print("7. 이미지 저장...")
+                image.save(file_path, "JPEG", quality=95)
+                
+                # 카운트 및 인덱스 증가
+                target_info["collected_count"] += 1
+                target_info["current_idx"] += 1
+                
+                print(f"Saved: {file_path} ({target_info['collected_count']}/{NUM_IMAGES_PER_CLASS})")
+                
+            except Exception as e:
+                # 오류 발생 시 스크립트가 멈추지 않도록 예외 처리
+                print(f"이미지 저장 중 오류 발생 (Label: {current_label}): {e}")
+
+if __name__ == "__main__":
+    collect_hf_images()

src/collection/download_dataset_kg.py

@@ -0,0 +1,56 @@
+import os
+
+# ================================
+# 0. 설정
+# ================================
+TARGET_COUNT = 60
+MIN_RES = 128  # 해상도 128
+PREFIX = "kg"
+BASE_DIR = "./data/raw"
+
+# ================================
+# 1. 다운로드 경로
+# ================================
+DOWNLOAD_PATH = "data/raw_full_kg"
+
+os.makedirs(DOWNLOAD_PATH, exist_ok=True)
+
+# ================================
+# 2. 사용할 Kaggle 데이터셋 (slug 기준)
+# ================================
+DATASETS = [
+    # 음식 및 식재료
+    "kmader/food41",
+
+    # 동물
+    "alessiocorrado99/animals10",
+    "gpiosenka/100-bird-species",
+
+    # 꽃
+    "alxmamaev/flowers-recognition",
+
+    # 과일
+    "moltean/fruits",
+    "yihfeng/strawberry-maturity",
+
+    # 탈것
+    "sshikamaru/car-object-detection",
+    "jessicali9530/stanford-cars-dataset",
+    "dataclusterlabs/vehicle-detection-image-dataset",
+    "meowmeowmeowmeowmeow/vehicle-type-recognition",
+
+    # 패션 및 잡화
+    "promptcloudhq/jewelry-text-to-image-dataset",
+    "ashwingupta3012/glasses-dataset",
+    "agrigorev/clothing-dataset-full",
+    "paramaggarwal/fashion-product-images-small"
+]
+
+# ================================
+# 3. 다운로드 실행
+# ================================
+for ds in DATASETS:
+    print(f"\nDownloading {ds} ...")
+    os.system(f"kaggle datasets download -d {ds} -p {DOWNLOAD_PATH}")
+
+print("\n모든 데이터셋 다운로드 완료!")

src/collection/download_dataset_us.py

@@ -0,0 +1,233 @@
+import os
+import re
+import requests
+from PIL import Image
+from io import BytesIO
+import time
+from dotenv import load_dotenv
+
+# =========================
+# 1. 설정 (여기만 수정하면 됨)
+# =========================
+load_dotenv()
+US_TOKEN = os.environ.get("US_TOKEN")
+
+ACCESS_KEY = "US_TOKEN"
+
+TARGET_COUNT = 100
+MIN_WIDTH = 256
+MIN_HEIGHT = 256
+SLEEP_TIME = 2
+
+BASE_DIR = "un_images"
+
+MASTER_CLASSES = [
+    "pizza","hamburger","sushi","pasta","salad","steak","cake","sandwich","fried_chicken","bread",
+    "apple","banana","strawberry","orange","carrot",
+    "golden_retriever","bulldog","siamese_cat","persian_cat","eagle","owl","lion","elephant","zebra","giraffe",
+    "rose","sunflower","daisy","tulip","palm_tree","pine_tree","maple_tree","bamboo",
+    "laptop","watch","camera","chair","clock","microwave","refrigerator",
+    "car","bicycle","motorcycle","airplane","bus",
+    "backpack","sneakers","umbrella","glasses","hat"
+]
+
+SIMPLE_CLASSES = [
+    "pizza","burger","sushi","pasta","salad","steak","cake","sandwich","fried_chicken","bread",
+    "apple","banana","strawberry","orange","carrot",
+    "golden_retriever","bulldog","siamese_cat","persian_cat","eagle","owl","lion","elephant","zebra","giraffe",
+    "rose","sunflower","daisy","tulip","palm_tree","pine_tree","maple_tree","bamboo",
+    "laptop","wristwatch","camera","chair","wall_clock","microwave","refrigerator",
+    "car","bicycle","motorcycle","airplane","bus",
+    "backpack","sneakers","umbrella","glasses","hat"
+]
+
+os.makedirs(BASE_DIR, exist_ok=True)
+
+# =========================
+# 2. 유틸
+# =========================
+def format_name(name):
+    return name.replace("_", "-")
+
+def get_start_index(folder, simple_cls):
+    pattern = re.compile(rf"un_{simple_cls}_(\d+)\.jpg")
+    max_idx = 0
+
+    for f in os.listdir(folder):
+        match = pattern.match(f)
+        if match:
+            num = int(match.group(1))
+            max_idx = max(max_idx, num)
+
+    return max_idx + 1
+
+# =========================
+# 3. API
+# =========================
+def search_images(query, page):
+    url = "https://api.unsplash.com/search/photos"
+
+    headers = {
+        "Authorization": f"Client-ID {ACCESS_KEY}"
+    }
+
+    params = {
+        "query": query.replace("_", " "),
+        "per_page": 30,
+        "page": page
+    }
+
+    res = requests.get(url, headers=headers, params=params)
+
+    if res.status_code == 429:
+        print("⏳ Rate limit → 60초 대기")
+        time.sleep(60)
+        return []
+
+    if res.status_code != 200:
+        print("API ERROR:", res.text)
+        return []
+
+    return [item["urls"]["regular"] for item in res.json().get("results", [])]
+
+def download_image(url):
+    try:
+        res = requests.get(url, timeout=10)
+        if res.status_code != 200:
+            return None
+
+        img = Image.open(BytesIO(res.content))
+        w, h = img.size
+
+        if w < MIN_WIDTH or h < MIN_HEIGHT:
+            return None
+
+        return res.content
+    except:
+        return None
+
+# =========================
+# 4. 메인
+# =========================
+for idx in range(len(MASTER_CLASSES)):
+
+    master_cls = MASTER_CLASSES[idx]
+    simple_cls = format_name(SIMPLE_CLASSES[idx])
+
+    print(f"\n[START] {master_cls}")
+
+    class_dir = os.path.join(BASE_DIR, master_cls)
+    os.makedirs(class_dir, exist_ok=True)
+
+    start_idx = get_start_index(class_dir, simple_cls)
+    count = start_idx - 1
+
+    page = 1
+    seen = set()
+
+    while count < TARGET_COUNT:
+        urls = search_images(simple_cls, page)
+
+        if not urls:
+            print("이미지 없음")
+            break
+
+        for url in urls:
+            if count >= TARGET_COUNT:
+                break
+
+            if url in seen:
+                continue
+            seen.add(url)
+
+            img_data = download_image(url)
+            if img_data is None:
+                continue
+
+            count += 1
+
+            file_name = f"un_{simple_cls}_{count:03d}.jpg"
+            path = os.path.join(class_dir, file_name)
+
+            with open(path, "wb") as f:
+                f.write(img_data)
+
+            print(f"Saved: {path}")
+
+        page += 1
+        time.sleep(SLEEP_TIME)
+
+    print(f"[DONE] {master_cls} -> {count}/{TARGET_COUNT}")
+
+# =========================
+# 5. 검증 + 부족분 자동 보충
+# =========================
+print("\n[검증 시작]\n")
+
+for idx in range(len(MASTER_CLASSES)):
+
+    master_cls = MASTER_CLASSES[idx]
+    simple_cls = format_name(SIMPLE_CLASSES[idx])
+
+    class_dir = os.path.join(BASE_DIR, master_cls)
+    files = [f for f in os.listdir(class_dir) if f.endswith(".jpg")]
+
+    valid_files = []
+
+    for f in files:
+        path = os.path.join(class_dir, f)
+        try:
+            img = Image.open(path)
+            w, h = img.size
+
+            if w >= MIN_WIDTH and h >= MIN_HEIGHT:
+                valid_files.append(f)
+            else:
+                os.remove(path)
+        except:
+            os.remove(path)
+
+    count = len(valid_files)
+
+    print(f"{master_cls}: {count}/{TARGET_COUNT}")
+
+    if count < TARGET_COUNT:
+        print(f"→ 부족분 재수집 시작")
+
+        page = 1
+        seen = set()
+        start_idx = get_start_index(class_dir, simple_cls)
+
+        while count < TARGET_COUNT:
+            urls = search_images(simple_cls, page)
+
+            if not urls:
+                break
+
+            for url in urls:
+                if count >= TARGET_COUNT:
+                    break
+
+                if url in seen:
+                    continue
+                seen.add(url)
+
+                img_data = download_image(url)
+                if img_data is None:
+                    continue
+
+                file_name = f"un_{simple_cls}_{start_idx:03d}.jpg"
+                path = os.path.join(class_dir, file_name)
+
+                with open(path, "wb") as f:
+                    f.write(img_data)
+
+                print(f"ReSaved: {path}")
+
+                start_idx += 1
+                count += 1
+
+            page += 1
+            time.sleep(SLEEP_TIME)
+
+print("\n[완료]")

src/collection/get_label_list_hf.py

@@ -0,0 +1,98 @@
+from datasets import load_dataset
+
+
+# ============================================================
+# [설정 부분]
+# ============================================================
+
+# 확인할 Hugging Face 데이터셋 이름
+DATASET_NAME = "KrushiJethe/fashion_data" #uran66/animals
+
+# 확인할 split 이름
+SPLIT_NAME = "train"
+
+# 라벨 필드명
+LABEL_FIELD_NAME = "articleType"
+
+# streaming 사용 여부
+# 라벨 구조만 확인할 때는 streaming=True로 해도 된다.
+USE_STREAMING = True
+
+# 문자열 라벨 데이터셋일 경우 전체 데이터를 훑어야 할 수 있다.
+# None이면 전체 확인, 숫자를 넣으면 일부 샘플만 확인한다.
+MAX_SCAN_ITEMS = None
+
+# ============================================================
+
+
+def get_unique_labels():
+    """
+    Hugging Face 데이터셋에서 라벨 목록을 중복 없이 출력한다.
+    """
+
+    print(f"[{DATASET_NAME}] 데이터셋 로드 중...")
+
+    dataset = load_dataset(
+        DATASET_NAME,
+        split=SPLIT_NAME,
+        streaming=USE_STREAMING,
+    )
+
+    # 데이터셋의 feature 정보에서 라벨 필드를 가져온다.
+    label_feature = dataset.features[LABEL_FIELD_NAME]
+
+    # ------------------------------------------------------------
+    # 1. Food101처럼 label이 ClassLabel 타입인 경우
+    # ------------------------------------------------------------
+    # 이 경우 데이터 전체를 순회하지 않아도
+    # dataset.features["label"].names 에서 전체 라벨명을 바로 확인할 수 있다.
+    if hasattr(label_feature, "names") and label_feature.names is not None:
+        label_names = label_feature.names
+
+        print("\n라벨 목록")
+        print("-" * 50)
+
+        for idx, label_name in enumerate(label_names):
+            print(f"{idx}: {label_name}")
+
+        print("-" * 50)
+        print(f"총 라벨 개수: {len(label_names)}")
+
+        return label_names
+
+    # ------------------------------------------------------------
+    # 2. label이 문자열로 직접 들어있는 데이터셋인 경우
+    # ------------------------------------------------------------
+    # 이 경우에는 데이터를 직접 순회하면서 중복을 제거해야 한다.
+    unique_labels = set()
+
+    print("\n라벨 필드가 ClassLabel 타입이 아니므로 데이터를 순회합니다...")
+
+    for idx, item in enumerate(dataset):
+        if MAX_SCAN_ITEMS is not None and idx >= MAX_SCAN_ITEMS:
+            break
+
+        label_value = item.get(LABEL_FIELD_NAME)
+
+        if label_value is None:
+            continue
+
+        unique_labels.add(str(label_value))
+
+    label_names = sorted(unique_labels)
+
+    print("\n라벨 목록")
+    print("-" * 50)
+    
+    
+    for idx, label_name in enumerate(label_names):
+        print(f"{idx}: {label_name}")
+
+    print("-" * 50)
+    print(f"총 라벨 개수: {len(label_names)}")
+
+    return label_names
+
+
+if __name__ == "__main__":
+    get_unique_labels()

src/collection/select_60_images.py

@@ -0,0 +1,115 @@
+import os
+import random
+import shutil
+
+# ================================
+# 0. 설정
+# ================================
+TARGET_COUNT = 60
+MIN_RES = 128  # 해상도 128
+PREFIX = "kg"
+BASE_DIR = "./data/raw"
+
+# ================================
+# 1. 경로
+# ================================
+DATA_DIR = r"C:\Users\qud46\Desktop\raw_kg"
+
+# ================================
+# 2. 클래스 목록
+# ================================
+CLASS_LIST = [
+    # 음식 및 식재료
+    "pizza","hamburger","sushi","pasta","salad",
+    "steak","cup_cake","sandwich","waffle","dumpling",
+
+    # 동물
+    "golden-retriever","bulldog","siamese-cat",
+    "persian-cat","elephant","sheep","horse",
+    "penguin","butterfly","squirrel",
+
+    # 꽃
+    "rose","sunflower","daisy","tulip","dandelion",
+    "lily","lavender","orchid","iris","marigold","aster",
+
+    # 과일
+    "apple","banana","strawberry","orange",
+    "carrot","tomato","cucumber",
+
+    # 탈것
+    "car","bicycle","motorcycle","airplane","bus",
+
+    # 패션 및 잡화
+    "t-shirt","sneakers","earrings","glasses",
+    "pants","bracelet","necklace"
+]
+
+print("클래스별 이미지 60장 맞추기 시작\n")
+
+# ================================
+# 3. 메인 로직
+# ================================
+for cls in CLASS_LIST:
+
+    cls_path = os.path.join(DATA_DIR, cls)
+
+    if not os.path.exists(cls_path):
+        print(f"{cls}: 폴더 없음 (skip)")
+        continue
+
+    # 이미지 파일 목록
+    images = [
+        f for f in os.listdir(cls_path)
+        if os.path.isfile(os.path.join(cls_path, f))
+    ]
+
+    current_count = len(images)
+
+    print(
+        f"{cls}: 현재 {current_count}장 "
+        f"→ 목표 {TARGET_COUNT}장"
+    )
+
+    # ================================
+    # 1) 60장 초과 → 랜덤 삭제
+    # ================================
+    if current_count > TARGET_COUNT:
+
+        delete_count = current_count - TARGET_COUNT
+
+        to_delete = random.sample(
+            images,
+            delete_count
+        )
+
+        for file in to_delete:
+
+            file_path = os.path.join(cls_path, file)
+
+            try:
+                os.remove(file_path)
+
+            except:
+                continue
+
+        print(f"   → {delete_count}장 삭제 완료")
+
+    # ================================
+    # 2) 60장 미만 → 부족 개수 출력
+    # ================================
+    elif current_count < TARGET_COUNT:
+
+        need_count = TARGET_COUNT - current_count
+
+        print(
+            f"   → {need_count}장 부족 "
+            f"(추가 수집 필요)"
+        )
+
+    # ================================
+    # 3) 정확히 60장
+    # ================================
+    else:
+        print("   → 이미 60장 완료")
+
+print("\n전체 정리 완료!")

src/collection/unzip_data_kg.py

@@ -0,0 +1,45 @@
+import os
+import zipfile
+
+# ================================
+# 0. 설정
+# ================================
+TARGET_COUNT = 60
+MIN_RES = 128  # 해상도 128
+PREFIX = "kg"
+BASE_DIR = "./data/raw"
+
+# ================================
+# 1. 경로
+# ================================
+ZIP_DIR = "data/raw_full_kg"
+
+# 압축 해제 위치
+EXTRACT_DIR = os.path.join(ZIP_DIR, "extracted")
+
+os.makedirs(EXTRACT_DIR, exist_ok=True)
+
+# ================================
+# 2. zip 파일 목록
+# ================================
+zip_files = [
+    f for f in os.listdir(ZIP_DIR)
+    if f.endswith(".zip")
+]
+
+# ================================
+# 3. 압축 해제
+# ================================
+for zip_file in zip_files:
+    zip_path = os.path.join(ZIP_DIR, zip_file)
+
+    print(f"{zip_file} 압축 해제 중...")
+
+    try:
+        with zipfile.ZipFile(zip_path, "r") as zip_ref:
+            zip_ref.extractall(EXTRACT_DIR)
+
+    except Exception as e:
+        print(f"오류 발생: {zip_file} → {e}")
+
+print("모든 압축 해제 완료!")

src/dataset/.ipynb_checkpoints/captioning_dataset-checkpoint.py

@@ -0,0 +1,124 @@
+import os
+import json
+import random
+from PIL import Image
+from torch.utils.data import Dataset
+import torch
+
+
+class CaptionDataset(Dataset):
+
+    def __init__(
+        self,
+        json_path,
+        image_dir,
+        w2i,
+        tokenizer: callable,
+        split='train',
+        transform=None,
+        max_len=30,
+        train_num_caption=1,
+        debug=False,
+        use_subword=False,
+        sp_model_path="tokenizer.model"
+    ):
+
+        with open(json_path, 'r') as f:
+            self.data = json.load(f)
+
+        # 디버깅용
+        if debug:
+            self.data= self.data[:10]
+
+        if split == "val":
+            self.is_val = True
+        else:
+            self.is_val = False
+        
+        self.image_dir = image_dir
+        self.w2i = w2i
+        self.transform = transform
+        self.max_len = max_len
+        self.tokenizer = tokenizer
+        self.train_num_caption = train_num_caption
+        self.use_subword = use_subword
+        if self.use_subword:
+            import sentencepiece as spm
+
+            self.sp = spm.SentencePieceProcessor()
+            self.sp.load(sp_model_path)
+        
+
+    def __len__(self):
+        return len(self.data)
+    
+    def encode_caption(self, caption):
+
+        if self.use_subword:
+            words = self.sp.encode(caption, out_type=str)
+
+            tokens = (
+                [self.w2i["<sos>"]] +
+                [self.w2i.get(w, self.w2i["<unk>"]) for w in words] +
+                [self.w2i["<eos>"]]
+            )
+        else:
+            words = self.tokenizer(caption)
+
+            tokens = (
+                [self.w2i["<sos>"]] +
+                [self.w2i.get(w, self.w2i["<unk>"]) for w in words] +
+                [self.w2i["<eos>"]]
+                )
+        
+        # truncation
+        if len(tokens) > self.max_len:
+            tokens = (tokens[:self.max_len - 1])
+            tokens.append(self.w2i["<eos>"])
+        else:
+            tokens += ([self.w2i["<pad>"]] * (self.max_len - len(tokens)))
+
+        return torch.tensor(tokens, dtype=torch.long)
+
+    def __getitem__(self, index):
+
+        data = self.data[index]
+        file_name = data["file_name"]
+
+        image_path = os.path.join(self.image_dir, file_name)
+
+        image = Image.open(image_path).convert('RGB')
+
+        if self.transform:
+            image = self.transform(image)
+
+
+        captions = data["captions"]
+
+        captions = captions[:5] # 캡션 5개 초과시 5개까지만 씀
+
+        while len(captions) < 5: # 캡션 5개 보다 부족할 시 마지막 캡션 복제해서 씀
+            captions.append(captions[-1])
+
+        # validation
+        if self.is_val:
+            caption = random.choice(captions)
+
+            tokens = (self.encode_caption(caption))
+
+            return image, tokens, captions, file_name
+
+        # train
+        selected_captions = (random.sample(captions, k=self.train_num_caption))
+
+        images = []
+        token_list = []
+        for caption in selected_captions:
+            images.append(image)
+            token_list.append(self.encode_caption(caption))
+
+        images = torch.stack(images)
+        tokens = torch.stack(token_list)
+
+        return images, tokens
+

src/dataset/.ipynb_checkpoints/train_sub_tokenizer-checkpoint.py

@@ -0,0 +1,55 @@
+import json
+import sentencepiece as spm
+import yaml
+
+# params
+with open("/workspace/params.yaml", "r", encoding="utf-8") as f:
+    params = yaml.safe_load(f)
+
+
+def train_sentencepiece(
+        json_path,
+        model_prefix="sub_tokenizer",
+        vocab_size=500,
+        model_type="unigram"
+    ):
+
+    with open(json_path, 'r') as f:
+        data = json.load(f)
+
+    txt_path = "/workspace/src/dataset/sub_tokenizing_captions.txt"
+
+    with open(txt_path, "w", encoding="utf-8") as f:
+        for item in data:
+            captions = item["captions"]
+
+            for caption in captions:
+                f.write(caption.lower() + "\n")
+
+    spm.SentencePieceTrainer.train(
+        input=txt_path,
+        model_prefix=model_prefix,
+        vocab_size=vocab_size,
+        model_type=model_type,
+
+        pad_piece="<pad>",
+        bos_piece="<sos>",
+        eos_piece="<eos>",
+        unk_piece="<unk>",
+        
+        pad_id=0,
+        bos_id=1,
+        eos_id=2,
+        unk_id=3
+    )
+
+    print("tokenizer training done")
+
+
+if __name__ == "__main__":
+    train_sentencepiece(
+        json_path="/workspace/data/captioning/annotations/train.json",
+        model_prefix="/workspace/src/dataset/sub_tokenizer",
+        vocab_size=params["captioning"]["tokenizer"]["sp_vocab_size"],
+        model_type="unigram"
+    )

src/dataset/build_vocab.py

@@ -0,0 +1,62 @@
+import json
+from collections import Counter
+import re
+import sentencepiece as spm
+
+def tokenizer(captions):
+    text = captions.lower()
+    text = re.sub(r"([.,!?])", r" \1 ", text)  # 특수문자 제거
+    tokens = text.split()
+    
+    return tokens
+
+def sub_tokenizer(caption, sp):
+    tokens = sp.encode(caption, out_type=str)
+
+    return tokens
+
+
+def build_vocab(json_path, min_freq=3, max_size=10000, use_subword=False, sp_model_path="/workspace/src/dataset/sub_tokenizer.model"):
+    w2i = dict()
+    i2w = dict()
+
+    # ==================================================
+    # SentencePiece tokenizer 사용
+    # ==================================================
+    if use_subword:
+
+        sp = spm.SentencePieceProcessor()
+        sp.load(sp_model_path)
+
+        voca_size = sp.get_piece_size()
+
+        for i in range(voca_size):
+            token = sp.id_to_piece(i)
+
+            w2i[token] = i
+            i2w[i] = token
+    else:
+        with open(json_path, 'r') as f:
+            data = json.load(f)
+
+        counter = Counter()
+
+        for item in data:
+            captions = item["captions"]
+            for caption in captions:
+                tokens = tokenizer(caption)
+                counter.update(tokens)
+            
+        words = [w for w, freq in counter.most_common() if freq >= min_freq]
+
+        voca = ["<pad>", "<sos>", "<eos>", "<unk>"]
+        voca.extend(words[:max_size-4])
+        voca_size = len(voca)
+
+        for i, w in enumerate(voca):
+            w2i[w] = i
+            i2w[i] = w
+
+    print(voca_size)
+
+    return w2i, i2w, voca_size

src/dataset/captioning_dataset.py

@@ -0,0 +1,124 @@
+import os
+import json
+import random
+from PIL import Image
+from torch.utils.data import Dataset
+import torch
+
+
+class CaptionDataset(Dataset):
+
+    def __init__(
+        self,
+        json_path,
+        image_dir,
+        w2i,
+        tokenizer: callable,
+        split='train',
+        transform=None,
+        max_len=30,
+        train_num_caption=1,
+        debug=False,
+        use_subword=False,
+        sp_model_path="tokenizer.model"
+    ):
+
+        with open(json_path, 'r') as f:
+            self.data = json.load(f)
+
+        # 디버깅용
+        if debug:
+            self.data= self.data[:10]
+
+        if split == "val":
+            self.is_val = True
+        else:
+            self.is_val = False
+        
+        self.image_dir = image_dir
+        self.w2i = w2i
+        self.transform = transform
+        self.max_len = max_len
+        self.tokenizer = tokenizer
+        self.train_num_caption = train_num_caption
+        self.use_subword = use_subword
+        if self.use_subword:
+            import sentencepiece as spm
+
+            self.sp = spm.SentencePieceProcessor()
+            self.sp.load(sp_model_path)
+        
+
+    def __len__(self):
+        return len(self.data)
+    
+    def encode_caption(self, caption):
+
+        if self.use_subword:
+            words = self.sp.encode(caption.lower(), out_type=str)
+
+            tokens = (
+                [self.w2i["<sos>"]] +
+                [self.w2i.get(w, self.w2i["<unk>"]) for w in words] +
+                [self.w2i["<eos>"]]
+            )
+        else:
+            words = self.tokenizer(caption)
+
+            tokens = (
+                [self.w2i["<sos>"]] +
+                [self.w2i.get(w, self.w2i["<unk>"]) for w in words] +
+                [self.w2i["<eos>"]]
+                )
+        
+        # truncation
+        if len(tokens) > self.max_len:
+            tokens = (tokens[:self.max_len - 1])
+            tokens.append(self.w2i["<eos>"])
+        else:
+            tokens += ([self.w2i["<pad>"]] * (self.max_len - len(tokens)))
+
+        return torch.tensor(tokens, dtype=torch.long)
+
+    def __getitem__(self, index):
+
+        data = self.data[index]
+        file_name = data["file_name"]
+
+        image_path = os.path.join(self.image_dir, file_name)
+
+        image = Image.open(image_path).convert('RGB')
+
+        if self.transform:
+            image = self.transform(image)
+
+
+        captions = data["captions"]
+
+        captions = captions[:5] # 캡션 5개 초과시 5개까지만 씀
+
+        while len(captions) < 5: # 캡션 5개 보다 부족할 시 마지막 캡션 복제해서 씀
+            captions.append(captions[-1])
+
+        # validation
+        if self.is_val:
+            caption = random.choice(captions)
+
+            tokens = (self.encode_caption(caption))
+
+            return image, tokens, captions, file_name
+
+        # train
+        selected_captions = (random.sample(captions, k=self.train_num_caption))
+
+        images = []
+        token_list = []
+        for caption in selected_captions:
+            images.append(image)
+            token_list.append(self.encode_caption(caption))
+
+        images = torch.stack(images)
+        tokens = torch.stack(token_list)
+
+        return images, tokens
+

src/dataset/classification_dataset.py

@@ -0,0 +1,69 @@
+import os
+from PIL import Image
+from torch.utils.data import Dataset
+
+
+class ClassificationDataset(Dataset):
+
+    def __init__(
+        self,
+        root_dir,
+        class_to_idx,
+        split="train",
+        transform=None,
+        split_ratio=(0.7, 0.15, 0.15)
+    ):
+
+        self.transform = transform
+        self.samples = []
+
+        for class_name in sorted(os.listdir(root_dir)):
+            class_path = os.path.join(
+                root_dir,
+                class_name
+            )
+
+            if not os.path.isdir(class_path):
+                continue
+
+            images = sorted(os.listdir(class_path))
+
+            total = len(images)
+
+            train_end = int(total * split_ratio[0])
+            val_end = train_end + int(total * split_ratio[1])
+
+            if split == "train":
+                split_images = images[:train_end]
+
+            elif split == "val":
+                split_images = images[train_end:val_end]
+
+            else:
+                split_images = images[val_end:]
+
+            label = class_to_idx[class_name]
+
+            for image_name in split_images:
+                image_path = os.path.join(
+                    class_path,
+                    image_name
+                )
+
+                self.samples.append(
+                    (image_path, label)
+                )
+
+    def __len__(self):
+        return len(self.samples)
+
+    def __getitem__(self, index):
+        image_path, label = self.samples[index]
+        image = Image.open(
+            image_path
+        ).convert("RGB")
+
+        if self.transform:
+            image = self.transform(image)
+
+        return image, label, image_path

src/dataset/collate_caption.py

@@ -0,0 +1,14 @@
+import torch
+
+def collate_caption(batch):
+    images = []
+    tokens = []
+
+    for image, token in batch:
+        images.append(image)
+        tokens.append(token)
+
+    images = torch.cat(images, dim=0)
+    tokens = torch.cat(tokens, dim=0)
+
+    return images, tokens

src/dataset/sub_tokenizer1000.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f24d6fdba624cc28650fc27f6ef8bd2002d3a3b652205a2e0e8af2aa21ea7be
+size 254104

src/dataset/sub_tokenizer1000.vocab

@@ -0,0 +1,1000 @@
+<pad>	0
+<sos>	0
+<eos>	0
+<unk>	0
+▁a	-1.98322
+.	-2.9407
+s	-3.42779
+▁of	-3.61005
+▁on	-3.6172
+▁in	-3.86704
+▁the	-3.87608
+nd	-3.93335
+▁with	-3.9459
+ing	-4.4554
+▁is	-4.4607
+▁	-4.56396
+▁to	-4.59365
+e	-4.65398
+▁man	-4.85533
+▁sitting	-4.95263
+,	-5.03648
+ed	-5.04331
+y	-5.11805
+▁an	-5.13757
+▁next	-5.20493
+▁two	-5.22965
+d	-5.2952
+▁white	-5.36199
+n	-5.36722
+a	-5.37838
+i	-5.38911
+▁are	-5.39818
+▁holding	-5.40034
+▁standing	-5.40128
+o	-5.40533
+▁table	-5.46676
+p	-5.47577
+t	-5.48538
+▁it	-5.49508
+r	-5.51872
+le	-5.59946
+▁plate	-5.6039
+▁woman	-5.63334
+▁that	-5.6784
+▁up	-5.68785
+▁at	-5.71356
+er	-5.7243
+▁top	-5.76056
+▁people	-5.76236
+▁some	-5.79632
+▁person	-5.81136
+m	-5.8687
+▁black	-5.88285
+▁large	-5.88415
+▁street	-5.92398
+es	-5.93578
+▁dog	-5.93954
+▁red	-5.97105
+▁s	-6.00218
+▁f	-6.02117
+▁his	-6.02288
+▁small	-6.0281
+▁near	-6.04428
+l	-6.05039
+u	-6.10616
+▁front	-6.11685
+g	-6.13449
+b	-6.16149
+▁sign	-6.19078
+f	-6.22392
+▁cat	-6.23333
+▁bear	-6.2501
+▁by	-6.25248
+▁group	-6.27012
+▁field	-6.28847
+▁blue	-6.29729
+▁has	-6.30494
+▁food	-6.30817
+▁down	-6.34169
+▁green	-6.34463
+ar	-6.34581
+▁young	-6.36549
+in	-6.3685
+▁tennis	-6.38354
+▁snow	-6.3845
+▁other	-6.39216
+▁close	-6.39402
+c	-6.40273
+▁whi	-6.40908
+k	-6.41607
+▁water	-6.41957
+▁grass	-6.42378
+▁there	-6.42571
+st	-6.42804
+al	-6.42888
+w	-6.44482
+h	-6.46665
+▁side	-6.47196
+ll	-6.4794
+▁train	-6.51638
+▁computer	-6.52006
+▁lay	-6.53283
+ch	-6.54474
+▁stop	-6.56702
+▁board	-6.56877
+▁for	-6.56976
+▁bus	-6.57341
+▁baseball	-6.58612
+▁phone	-6.58992
+▁parked	-6.59899
+▁walking	-6.60393
+▁her	-6.614
+▁b	-6.6172
+▁riding	-6.62086
+▁sits	-6.62201
+▁bowl	-6.62631
+▁clock	-6.63455
+▁parking	-6.63607
+▁kitchen	-6.64843
+▁light	-6.65729
+or	-6.68298
+▁wearing	-6.69127
+▁picture	-6.69418
+▁boy	-6.69421
+▁looking	-6.69659
+▁building	-6.69858
+en	-6.71042
+▁out	-6.71159
+▁girl	-6.71339
+▁vase	-6.73144
+▁fire	-6.73145
+▁hand	-6.73375
+▁pizza	-6.7406
+an	-6.75261
+'	-6.75293
+▁thre	-6.76074
+▁t	-6.76468
+▁couple	-6.77331
+at	-6.7805
+▁be	-6.78194
+▁yellow	-6.78446
+▁bathroom	-6.78927
+▁glass	-6.79439
+▁woode	-6.79814
+▁toilet	-6.80878
+▁meter	-6.80878
+▁from	-6.82199
+▁m	-6.82553
+▁several	-6.82698
+▁bat	-6.82956
+▁skateboard	-6.83116
+▁boat	-6.84214
+▁brown	-6.84553
+on	-6.84954
+▁vegetables	-6.85327
+▁hydrant	-6.86097
+▁over	-6.87842
+▁beach	-6.88194
+▁scissors	-6.88724
+▁sink	-6.89616
+▁their	-6.89832
+▁room	-6.89974
+▁filled	-6.90334
+el	-6.9077
+▁umbrella	-6.93262
+re	-6.93408
+v	-6.93453
+▁desk	-6.93802
+▁wine	-6.94951
+ly	-6.95091
+▁elephant	-6.95316
+▁horse	-6.95723
+▁motorcycle	-6.95896
+▁hot	-6.96056
+▁road	-6.96278
+▁open	-6.96662
+▁st	-6.97948
+▁pair	-6.98018
+ro	-6.98102
+▁d	-6.98209
+it	-6.9932
+▁tie	-6.99493
+▁bench	-6.99587
+ur	-6.99703
+▁g	-7.0009
+▁piece	-7.0058
+▁cake	-7.01609
+▁pa	-7.02046
+▁keyboard	-7.02191
+▁each	-7.02191
+▁thi	-7.02415
+▁wall	-7.03008
+▁sheep	-7.03431
+▁orange	-7.03899
+▁ball	-7.03925
+▁different	-7.04036
+▁frisbee	-7.04036
+▁counter	-7.04036
+▁flying	-7.04119
+▁flowers	-7.04797
+▁traffic	-7.04866
+▁giraffe	-7.05405
+▁laptop	-7.05704
+▁car	-7.06113
+▁tree	-7.06714
+▁eating	-7.06818
+▁teddy	-7.07615
+▁its	-7.07656
+▁bunch	-7.07657
+▁around	-7.0783
+▁mouse	-7.08046
+li	-7.08325
+▁covered	-7.08612
+il	-7.09208
+▁bike	-7.10221
+▁broccoli	-7.10883
+▁cell	-7.11314
+▁through	-7.11997
+▁inside	-7.12178
+la	-7.12697
+▁fruit	-7.12824
+▁one	-7.13683
+▁remote	-7.14262
+▁c	-7.14674
+▁zebra	-7.14708
+▁outside	-7.15234
+very	-7.15784
+▁little	-7.15879
+▁microwave	-7.16345
+▁truck	-7.1658
+▁area	-7.16815
+▁hold	-7.16833
+ting	-7.16983
+et	-7.17782
+▁city	-7.18235
+▁old	-7.18386
+ter	-7.18462
+▁lot	-7.18947
+▁pe	-7.19085
+▁back	-7.19676
+▁bed	-7.19905
+▁surfboard	-7.19922
+▁sandwich	-7.20647
+▁w	-7.20741
+▁together	-7.21876
+▁air	-7.2197
+▁co	-7.22215
+▁playing	-7.22223
+▁sit	-7.24205
+▁player	-7.2433
+▁book	-7.25398
+▁under	-7.25907
+▁head	-7.25926
+▁racket	-7.25933
+▁behind	-7.26428
+▁fence	-7.26947
+▁bananas	-7.27123
+pped	-7.27309
+▁la	-7.27649
+▁bird	-7.27907
+▁ground	-7.27992
+▁men	-7.28063
+▁po	-7.28079
+▁background	-7.28255
+▁snowboard	-7.28561
+▁glasses	-7.2927
+▁child	-7.30655
+ers	-7.30667
+▁do	-7.31758
+▁carrots	-7.31859
+▁display	-7.31909
+▁stuff	-7.32014
+▁bag	-7.32028
+ic	-7.32392
+ous	-7.32735
+ol	-7.32786
+▁cup	-7.33429
+th	-7.33652
+ck	-7.33703
+▁baby	-7.34507
+▁full	-7.35071
+▁cut	-7.35128
+▁stands	-7.3625
+▁into	-7.37317
+▁luggage	-7.37353
+id	-7.37802
+ion	-7.38241
+▁being	-7.38631
+▁oven	-7.39099
+▁re	-7.39201
+▁sh	-7.39833
+▁beside	-7.40047
+▁photo	-7.40281
+▁sidewalk	-7.41477
+ve	-7.41626
+▁shirt	-7.42155
+▁paper	-7.42387
+▁di	-7.42522
+▁floor	-7.42687
+▁window	-7.42687
+▁dirt	-7.42994
+▁bottle	-7.43605
+▁knife	-7.43605
+▁cutt	-7.43621
+▁pink	-7.43913
+▁ski	-7.44157
+ra	-7.44446
+z	-7.44525
+▁cow	-7.44585
+▁trees	-7.45224
+▁game	-7.45465
+▁bicycle	-7.46093
+▁wood	-7.46571
+▁look	-7.47155
+▁pole	-7.47177
+▁big	-7.48325
+▁skis	-7.48476
+way	-7.48747
+▁sp	-7.49752
+▁chair	-7.50274
+ct	-7.50429
+ut	-7.50672
+▁view	-7.51264
+▁kite	-7.51347
+▁pile	-7.51817
+▁suitcase	-7.51929
+▁face	-7.52049
+▁airplane	-7.52599
+▁walk	-7.52786
+▁oranges	-7.53837
+▁mouth	-7.54637
+▁toothbrush	-7.54683
+▁k	-7.55424
+ce	-7.5561
+▁image	-7.55671
+▁apples	-7.55932
+▁zebras	-7.56039
+▁n	-7.56135
+▁sky	-7.56722
+▁plane	-7.57067
+ti	-7.574
+▁box	-7.5742
+▁stand	-7.57636
+▁kites	-7.5947
+▁sa	-7.59849
+▁ta	-7.60838
+▁meat	-7.62116
+▁apple	-7.62208
+▁ca	-7.62627
+j	-7.63227
+▁another	-7.63229
+▁park	-7.63289
+▁refrigerator	-7.65497
+▁rock	-7.65499
+ent	-7.66018
+age	-7.66522
+▁vari	-7.68129
+▁cr	-7.68348
+▁he	-7.68909
+▁grassy	-7.69146
+▁ma	-7.70065
+▁four	-7.70599
+▁slice	-7.7131
+▁colorful	-7.71405
+▁bears	-7.71567
+▁line	-7.72202
+▁hat	-7.72466
+▁lo	-7.72602
+▁someone	-7.72644
+▁donuts	-7.74847
+▁tall	-7.74884
+▁ha	-7.75025
+▁women	-7.75124
+▁fork	-7.75241
+▁banana	-7.75259
+us	-7.75503
+▁cows	-7.75741
+▁cheese	-7.7597
+▁stove	-7.75971
+▁control	-7.76392
+▁h	-7.76772
+▁them	-7.76867
+um	-7.76899
+▁coffee	-7.77252
+▁ra	-7.77705
+▁station	-7.77735
+▁mirror	-7.79426
+▁along	-7.79426
+▁ready	-7.79866
+▁can	-7.7987
+▁we	-7.80183
+▁no	-7.80262
+▁com	-7.80472
+▁pan	-7.80668
+te	-7.81235
+▁herd	-7.81529
+▁pot	-7.81626
+▁tak	-7.81771
+▁camera	-7.82099
+▁dish	-7.82169
+▁pi	-7.82775
+▁hanging	-7.83368
+▁items	-7.83463
+▁off	-7.83819
+▁mo	-7.84324
+▁tray	-7.84383
+▁cook	-7.84846
+to	-7.85812
+▁above	-7.86718
+▁giraffes	-7.87394
+▁spoon	-7.87669
+▁middle	-7.88628
+▁long	-7.88667
+▁wii	-7.88846
+▁suit	-7.89112
+▁half	-7.90085
+▁elephants	-7.90111
+▁posing	-7.90154
+▁metal	-7.90575
+▁silver	-7.90575
+▁get	-7.90787
+▁ne	-7.90954
+▁past	-7.91062
+▁grazing	-7.91067
+un	-7.91793
+mp	-7.92552
+▁bar	-7.92721
+▁stick	-7.93568
+▁ru	-7.94538
+▁day	-7.94592
+▁sand	-7.94793
+▁smiling	-7.95098
+▁like	-7.95098
+gain	-7.95197
+▁mountain	-7.95613
+▁carry	-7.95619
+▁brush	-7.96003
+qu	-7.96664
+▁fruits	-7.96821
+▁de	-7.96922
+ot	-7.97654
+▁plant	-7.97702
+▁living	-7.98231
+ate	-7.98572
+▁court	-7.98764
+▁surf	-7.9939
+▁basket	-7.99835
+▁drink	-8.00293
+ive	-8.0033
+▁ho	-8.00457
+▁door	-8.00632
+▁hill	-8.00921
+▁animals	-8.01801
+side	-8.02104
+ies	-8.02683
+▁signs	-8.03339
+▁crowd	-8.03682
+▁hair	-8.03684
+▁pull	-8.03822
+▁cars	-8.03923
+x	-8.05366
+▁row	-8.05382
+ping	-8.05595
+ng	-8.05924
+▁airport	-8.05948
+▁across	-8.05948
+▁bread	-8.05949
+▁animal	-8.06748
+how	-8.06865
+▁lean	-8.07115
+▁swing	-8.07118
+▁plastic	-8.07682
+▁cross	-8.07683
+▁who	-8.07684
+▁couch	-8.09448
+▁trick	-8.09464
+▁tracks	-8.11186
+▁horses	-8.1119
+led	-8.11552
+▁donut	-8.11644
+▁screen	-8.1185
+▁set	-8.11926
+▁prepar	-8.1246
+▁ride	-8.12466
+▁bu	-8.13179
+up	-8.13602
+▁skate	-8.13679
+▁birds	-8.1389
+▁type	-8.14312
+▁store	-8.14344
+ment	-8.14676
+▁driving	-8.14936
+▁empty	-8.14936
+ard	-8.15133
+▁un	-8.15811
+lying	-8.16103
+ew	-8.16374
+▁pu	-8.1642
+▁chocolate	-8.16835
+▁enclosure	-8.16835
+▁color	-8.16981
+▁something	-8.17484
+▁hands	-8.17591
+▁pen	-8.17919
+▁market	-8.1814
+▁kid	-8.18247
+▁seat	-8.19436
+▁purple	-8.20082
+▁television	-8.20082
+▁using	-8.20104
+▁displayed	-8.20337
+▁snowboarder	-8.20711
+▁house	-8.20748
+ake	-8.21209
+▁includ	-8.21411
+▁slope	-8.21411
+▁video	-8.21411
+▁hit	-8.21564
+▁con	-8.21607
+▁or	-8.21672
+▁skier	-8.21749
+▁controller	-8.22088
+▁shown	-8.22145
+own	-8.22178
+▁container	-8.22238
+▁pro	-8.23454
+▁multi	-8.24123
+▁case	-8.24262
+▁tooth	-8.24826
+▁gra	-8.25037
+▁flower	-8.25042
+▁outdoor	-8.25507
+uring	-8.2612
+▁number	-8.26206
+▁jet	-8.26208
+co	-8.26604
+▁tv	-8.26658
+ish	-8.27151
+▁double	-8.2762
+▁hotdog	-8.2762
+▁monitor	-8.2762
+▁salad	-8.27621
+▁lady	-8.27628
+▁doughnuts	-8.28027
+▁attached	-8.28334
+▁contain	-8.28887
+▁teeth	-8.29054
+▁she	-8.29109
+▁jump	-8.29778
+▁resting	-8.29782
+▁guy	-8.30513
+▁made	-8.30785
+▁ocean	-8.31244
+▁work	-8.31244
+▁tower	-8.31265
+▁bun	-8.31679
+▁corner	-8.31984
+▁meal	-8.31991
+ling	-8.32351
+▁mak	-8.32645
+▁passenger	-8.3273
+▁take	-8.33001
+▁racquet	-8.33482
+▁pose	-8.33792
+▁backpack	-8.35003
+▁high	-8.35773
+▁surround	-8.36548
+▁office	-8.36662
+▁zoo	-8.37329
+▁brick	-8.37353
+ation	-8.37577
+▁neck	-8.3822
+light	-8.38662
+▁restaurant	-8.3891
+▁dry	-8.3892
+▁cellphone	-8.38966
+▁photograph	-8.3971
+▁fresh	-8.3971
+▁surface	-8.3971
+▁shelf	-8.39711
+ween	-8.39718
+▁jacket	-8.40516
+▁sauce	-8.40517
+▁wave	-8.40517
+▁adult	-8.41329
+▁statue	-8.41329
+▁sc	-8.41573
+▁waiting	-8.41797
+▁branch	-8.42149
+▁cabinet	-8.42149
+▁kind	-8.42173
+▁watching	-8.42253
+▁painted	-8.42488
+▁play	-8.42802
+▁post	-8.42841
+▁polar	-8.42979
+▁track	-8.43055
+▁cloth	-8.43809
+colored	-8.44458
+▁runway	-8.4465
+▁glove	-8.45497
+▁time	-8.45497
+▁clear	-8.45499
+▁showing	-8.45502
+▁sun	-8.45814
+top	-8.45927
+▁toothbrushes	-8.46235
+▁helmet	-8.46351
+▁sleeping	-8.46351
+▁chicken	-8.46352
+▁rice	-8.46352
+▁dark	-8.46353
+▁children	-8.46355
+▁but	-8.46491
+▁arm	-8.4712
+▁electronic	-8.47214
+▁body	-8.47214
+▁eaten	-8.47334
+▁dressed	-8.4777
+▁skiing	-8.47879
+▁bet	-8.48197
+▁going	-8.48353
+▁tile	-8.48614
+ight	-8.48895
+ead	-8.503
+▁use	-8.50637
+▁talking	-8.50738
+▁smile	-8.51639
+▁pack	-8.51932
+▁doughnut	-8.52028
+fri	-8.52173
+ant	-8.52342
+▁go	-8.52519
+▁bright	-8.52549
+▁grey	-8.52551
+▁dock	-8.52768
+▁sliced	-8.52772
+▁eat	-8.53152
+▁wire	-8.53602
+▁rail	-8.53658
+▁cart	-8.54105
+▁place	-8.5424
+▁about	-8.54392
+▁ex	-8.54408
+▁gray	-8.552
+▁left	-8.55326
+▁fries	-8.55331
+▁strip	-8.5535
+▁toaster	-8.5552
+▁carrot	-8.55864
+▁star	-8.57154
+▁intersection	-8.57222
+▁toward	-8.58186
+▁rack	-8.5825
+▁pick	-8.58321
+▁have	-8.58323
+▁drinking	-8.59303
+▁device	-8.60135
+▁soup	-8.60137
+▁tub	-8.61128
+▁brushing	-8.61371
+▁beer	-8.61576
+▁blanket	-8.62125
+▁shower	-8.62126
+▁rain	-8.62579
+▁skateboarder	-8.62651
+▁leaves	-8.63135
+▁trunk	-8.63135
+eep	-8.63335
+▁variet	-8.63348
+▁stra	-8.63487
+▁dress	-8.63501
+▁watch	-8.65057
+▁male	-8.6518
+▁potatoes	-8.65186
+▁cover	-8.65653
+▁wi	-8.66218
+▁toppings	-8.6625
+▁night	-8.66297
+ver	-8.66416
+▁home	-8.67288
+▁signal	-8.68183
+▁eye	-8.68385
+▁vehicle	-8.69419
+▁shop	-8.69747
+▁batter	-8.69925
+▁decker	-8.70507
+▁was	-8.70587
+▁closeup	-8.70815
+▁dessert	-8.71605
+▁graffiti	-8.71605
+▁older	-8.72031
+line	-8.72433
+▁decorated	-8.72717
+▁traveling	-8.73226
+▁uniform	-8.7384
+▁not	-8.7476
+▁chi	-8.7524
+▁stacked	-8.75395
+▁hug	-8.76127
+▁shaped	-8.76411
+phone	-8.7734
+▁rest	-8.7802
+▁river	-8.78465
+▁motor	-8.78468
+▁roll	-8.78475
+▁lit	-8.7856
+able	-8.78876
+▁egg	-8.79656
+ical	-8.79692
+placed	-8.79851
+▁shot	-8.81013
+▁setting	-8.81027
+▁cloud	-8.8208
+▁assortment	-8.82108
+lic	-8.82399
+▁served	-8.83315
+▁onions	-8.83693
+▁ice	-8.85444
+▁appliances	-8.8583
+▁giant	-8.85831
+lush	-8.85832
+▁underneath	-8.85857
+▁vegetable	-8.86435
+▁assort	-8.87084
+▁machine	-8.87112
+▁produc	-8.87113
+▁gear	-8.87123
+▁forest	-8.8714
+▁smart	-8.88411
+▁electric	-8.88411
+▁reading	-8.88411
+▁purse	-8.88412
+▁mitt	-8.88414
+▁moving	-8.88414
+▁seen	-8.88435
+▁van	-8.88446
+▁fish	-8.88869
+unny	-8.89689
+▁edge	-8.8973
+▁feet	-8.89741
+▁part	-8.89746
+▁surfer	-8.90804
+▁tea	-8.91376
+▁fridge	-8.93781
+▁perched	-8.93781
+▁running	-8.93782
+▁size	-8.93785
+▁yard	-8.93796
+▁mount	-8.9517
+▁female	-8.9517
+▁mustard	-8.9517
+▁curb	-8.95172
+▁turn	-8.95174
+▁him	-8.95175
+▁scene	-8.95183
+▁ramp	-8.95295
+▁log	-8.95909
+▁object	-8.96579
+▁draw	-8.9658
+▁drive	-8.9658
+▁cattle	-8.96602
+▁cla	-8.9711
+▁wheel	-8.98007
+▁towel	-8.98022
+▁point	-8.9804
+▁putt	-8.98076
+▁tomatoes	-8.98212
+▁beautiful	-8.99457
+▁mother	-8.99457
+▁among	-8.99458
+▁gold	-8.99461
+▁hang	-8.99878
+intend	-9.00038
+▁pasture	-9.00931
+▁cage	-9.01343
+▁bridge	-9.0242
+▁clean	-9.02423
+▁frost	-9.02425
+▁leg	-9.0249
+▁reflect	-9.03935
+▁sausage	-9.03935
+▁right	-9.03942
+▁stone	-9.03967
+▁candle	-9.03972
+▁wrappe	-9.03975
+▁single	-9.04037
+▁goat	-9.04131
+▁arranged	-9.05473
+▁writ	-9.05476
+▁blow	-9.05488
+▁desktop	-9.05653
+board	-9.05754
+▁equipment	-9.07036
+▁cream	-9.07036
+▁trail	-9.07036
+▁mug	-9.0704
+▁wild	-9.0704
+▁sub	-9.07047
+▁beans	-9.07074
+▁paint	-9.08035
+▁below	-9.08623
+▁featur	-9.08623
+▁pitch	-9.08625
+▁perform	-9.10236
+▁toothpaste	-9.10236
+▁held	-9.10243
+▁float	-9.11875
+▁broken	-9.11875
+▁glazed	-9.11875
+▁ketchup	-9.11875
+▁check	-9.11876
+▁dinner	-9.11876
+▁grill	-9.11876
+▁police	-9.11877
+▁pre	-9.11932
+▁design	-9.13542
+▁flock	-9.13542
+▁gather	-9.13542
+▁platform	-9.13542
+▁trash	-9.13542
+▁veggie	-9.13542
+▁short	-9.13545
+▁foot	-9.13546
+▁flip	-9.14075
+▁birthday	-9.15237
+▁pretty	-9.15237
+▁soda	-9.15237
+▁reach	-9.15241
+▁nice	-9.15242
+▁public	-9.16961
+▁round	-9.16961
+having	-9.17397
+▁alone	-9.18719
+▁bushes	-9.18722
+▁lie	-9.19505
+▁stack	-9.19848
+just	-9.20508
+▁lake	-9.20519
+▁stopped	-9.20543
+▁fac	-9.23775
+▁engine	-9.24171
+▁tarmac	-9.24171
+▁garden	-9.24172
+▁vin	-9.2419
+tuck	-9.24618
+▁toast	-9.25758
+▁flat	-9.26074
+▁try	-9.26088
+▁french	-9.27981
+▁beverage	-9.27981
+▁shore	-9.27993
+▁grow	-9.28019
+▁fall	-9.28069
+▁supplies	-9.29942
+▁farm	-9.29955
+▁mushroom	-9.31942
+▁mix	-9.31942
+▁chain	-9.31942
+▁throw	-9.31942
+▁good	-9.31947
+loaded	-9.31949
+▁steel	-9.31954
+▁wait	-9.32815
+▁shape	-9.33477
+▁advertis	-9.33983
+▁lunch	-9.33983
+▁modern	-9.33983
+▁square	-9.33983
+▁smo	-9.34026
+▁tomato	-9.35768
+▁enjoy	-9.36066
+▁shoulder	-9.36066
+▁lemon	-9.36066
+▁pastries	-9.36067
+▁milk	-9.36068
+▁lamb	-9.36071
+▁match	-9.36071
+▁chew	-9.36072
+▁rose	-9.36084
+▁style	-9.36108
+▁appear	-9.38194
+▁breakfast	-9.38194
+▁foreground	-9.38194
+▁napkin	-9.38194
+▁platter	-9.38194
+▁ripe	-9.38194
+▁strawberries	-9.38194
+▁sunglasses	-9.38194
+▁chopped	-9.38194
+▁desert	-9.38194
+▁blender	-9.38194
+▁leaf	-9.38194
+▁shade	-9.38197
+▁lawn	-9.382
+▁bri	-9.39481
+▁decoration	-9.40368
+▁pedestrian	-9.40368
+▁baked	-9.4037
+▁kneel	-9.4259
+▁bottom	-9.4259
+▁christmas	-9.4259
+▁country	-9.4259
+▁decorative	-9.4259
+▁scooter	-9.4259
+▁sculpture	-9.4259
+▁sprinkles	-9.4259
+▁things	-9.42606
+▁chips	-9.42617
+▁catcher	-9.42642
+ball	-9.43836
+▁serving	-9.44863
+▁cub	-9.44867
+▁horn	-9.44874
+▁bathtub	-9.47188
+▁concrete	-9.47188
+▁distance	-9.47188
+where	-9.47193
+▁balloon	-9.47194
+▁tri	-9.47568
+bow	-9.50266
+▁bacon	-9.52008
+▁bedroom	-9.52008
+▁carriage	-9.52008
+▁kitten	-9.52008
+▁stainless	-9.52008
+▁reads	-9.52009
+▁bite	-9.52042
+▁graze	-9.52045
+▁carrie	-9.52051
+▁juice	-9.54508
+▁lettuce	-9.54508
+▁partially	-9.54508
+▁swimming	-9.54508
+▁position	-9.54508
+▁carpet	-9.54512
+▁sort	-9.54545
+▁travel	-9.55899
+▁blurr	-9.57072
+▁pigeons	-9.57072
+▁what	-9.57074
+▁cement	-9.57078
+▁word	-9.57079
+▁same	-9.57092
+▁ship	-9.58422
+▁climb	-9.59704
+▁arrangement	-9.59704
+▁collecti	-9.59704
+▁shadow	-9.59704
+▁parade	-9.59704
+▁bucket	-9.59704
+▁lift	-9.59704
+▁center	-9.59707
+berry	-9.5971
+▁flag	-9.5971
+▁lead	-9.59712
+▁giv	-9.61391
+▁you	-9.62387
+▁family	-9.62407
+▁military	-9.62407
+▁picnic	-9.62407
+▁soccer	-9.62407
+▁pavement	-9.62407
+▁peanut	-9.62407
+▁space	-9.62409
+▁mark	-9.65112
+▁curtain	-9.65184
+▁himself	-9.65184
+▁railroad	-9.65184
+▁ledge	-9.65184
+ddler	-9.65185
+▁duck	-9.65187
+▁model	-9.65188
+rcial	-9.65191
+▁base	-9.65213
+arrow	-9.65254
+made	-9.6699
+▁propeller	-9.68041
+▁school	-9.68041
+▁puppy	-9.68042
+▁cupcake	-9.68042
+▁built	-9.68043
+▁block	-9.70983
+▁event	-9.70983
+▁spread	-9.70983
+▁winter	-9.70985
+▁sport	-9.70997
+▁antique	-9.74013
+▁pattern	-9.74013
+▁professional	-9.74013
+▁balanc	-9.74013
+▁consist	-9.74013
+▁spray	-9.74016
+ough	-9.74019
+▁figure	-9.77138
+▁furniture	-9.77138
+▁notebook	-9.77138
+▁parrot	-9.77138
+▁sofa	-9.77138
+q	-10.8674

src/dataset/sub_tokenizer1500.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33febb00379c559196d197de44fa529e8e3cfe5853e9a778dee369f6f46aa4fe
+size 262628

src/dataset/sub_tokenizer1500.vocab

@@ -0,0 +1,1500 @@
+<pad>	0
+<sos>	0
+<eos>	0
+<unk>	0
+▁a	-1.93118
+.	-2.88169
+▁of	-3.55221
+▁on	-3.56188
+s	-3.62067
+▁in	-3.81444
+▁the	-3.81817
+▁with	-3.88689
+nd	-3.89226
+▁is	-4.40178
+ing	-4.52367
+▁to	-4.57599
+▁	-4.76533
+▁man	-4.79839
+▁sitting	-4.89362
+,	-4.97747
+▁an	-5.06532
+ed	-5.14206
+▁next	-5.14591
+e	-5.16728
+▁two	-5.17064
+▁white	-5.30298
+y	-5.31383
+▁are	-5.32805
+▁holding	-5.34089
+▁standing	-5.34217
+▁table	-5.41202
+▁it	-5.45231
+d	-5.5279
+▁plate	-5.54488
+▁woman	-5.57433
+▁at	-5.59347
+n	-5.59711
+▁that	-5.61939
+▁up	-5.62864
+▁top	-5.69587
+▁people	-5.70335
+▁some	-5.73733
+▁person	-5.75235
+▁black	-5.82383
+▁large	-5.82513
+▁street	-5.86497
+le	-5.9101
+▁red	-5.91133
+a	-5.94493
+▁his	-5.96388
+▁small	-5.96909
+t	-5.98331
+r	-5.98939
+▁dog	-6.01085
+▁near	-6.01845
+o	-6.04639
+▁front	-6.05784
+p	-6.09572
+▁sign	-6.13869
+er	-6.16551
+▁cat	-6.17326
+m	-6.1807
+▁by	-6.19106
+▁bear	-6.19634
+▁group	-6.2111
+▁field	-6.22948
+▁has	-6.24171
+▁food	-6.24916
+▁blue	-6.25015
+▁green	-6.28562
+▁down	-6.295
+▁young	-6.30647
+▁tennis	-6.32453
+▁snow	-6.32548
+▁other	-6.33314
+▁close	-6.335
+▁whi	-6.35017
+▁water	-6.36056
+▁there	-6.36118
+▁grass	-6.36577
+▁side	-6.41189
+▁train	-6.45737
+▁computer	-6.46105
+▁lay	-6.47228
+▁board	-6.50962
+▁baseball	-6.52711
+▁phone	-6.53028
+▁parked	-6.53973
+▁for	-6.54057
+▁walking	-6.54314
+b	-6.54853
+▁her	-6.5544
+g	-6.55726
+▁sits	-6.55971
+▁riding	-6.56185
+▁bowl	-6.5673
+i	-6.56754
+▁clock	-6.57554
+▁stop	-6.57563
+▁parking	-6.57676
+▁kitchen	-6.58942
+▁wearing	-6.63272
+▁picture	-6.63517
+▁boy	-6.63583
+▁looking	-6.63721
+▁bus	-6.64201
+▁girl	-6.65437
+▁vase	-6.67243
+▁pizza	-6.68159
+u	-6.68776
+es	-6.68791
+▁fire	-6.68928
+▁f	-6.6895
+'	-6.69392
+▁thre	-6.70171
+▁out	-6.70226
+▁woode	-6.70691
+▁couple	-6.7143
+▁yellow	-6.72545
+▁bathroom	-6.73026
+▁glass	-6.73696
+▁toilet	-6.74976
+al	-6.75068
+▁from	-6.76297
+▁several	-6.76797
+▁skateboard	-6.77339
+▁building	-6.78257
+▁brown	-6.78652
+▁vegetables	-6.79345
+▁light	-6.79579
+▁hydrant	-6.80196
+▁hand	-6.80907
+l	-6.82187
+▁beach	-6.82292
+▁scissors	-6.82898
+▁sink	-6.83714
+st	-6.83938
+▁room	-6.84073
+▁be	-6.84238
+▁their	-6.84242
+▁filled	-6.84433
+k	-6.85956
+▁over	-6.86437
+w	-6.87453
+▁desk	-6.87899
+in	-6.88124
+▁wine	-6.89047
+▁bat	-6.89335
+▁elephant	-6.90038
+▁hot	-6.90048
+▁road	-6.90377
+▁open	-6.90761
+▁meter	-6.92037
+▁pair	-6.92121
+▁tie	-6.93415
+▁bench	-6.93685
+▁horse	-6.94242
+▁one	-6.95584
+▁s	-6.95612
+▁cake	-6.95696
+▁thi	-6.96189
+▁each	-6.9629
+▁keyboard	-6.96291
+▁wall	-6.97106
+▁sheep	-6.9752
+▁ball	-6.98006
+▁counter	-6.98135
+▁different	-6.98135
+▁frisbee	-6.98135
+ch	-6.98225
+▁flying	-6.98315
+▁orange	-6.98616
+▁flowers	-6.98714
+▁traffic	-6.98965
+▁d	-6.99022
+c	-6.99385
+▁laptop	-6.99806
+▁giraffe	-7.00126
+▁eating	-7.00898
+en	-7.0135
+▁tree	-7.01429
+▁car	-7.01644
+▁teddy	-7.01714
+▁bunch	-7.01741
+▁around	-7.01929
+▁mouse	-7.02144
+▁covered	-7.02696
+▁its	-7.04127
+▁broccoli	-7.04982
+▁boat	-7.05254
+▁cell	-7.05417
+▁inside	-7.05999
+▁through	-7.06096
+▁fruit	-7.07545
+▁remote	-7.08361
+▁motorcycle	-7.08501
+▁outside	-7.09315
+on	-7.09401
+▁zebra	-7.09424
+▁area	-7.0966
+very	-7.09841
+▁little	-7.09978
+▁microwave	-7.10444
+▁truck	-7.10678
+ar	-7.11999
+▁city	-7.12334
+ly	-7.12514
+▁old	-7.12601
+▁lot	-7.12874
+v	-7.13036
+▁bed	-7.13144
+▁surfboard	-7.14022
+▁umbrella	-7.15312
+▁together	-7.15975
+▁playing	-7.16295
+or	-7.17949
+▁back	-7.17972
+▁player	-7.18388
+▁sit	-7.18649
+▁under	-7.20005
+▁racket	-7.20024
+h	-7.20133
+el	-7.20277
+▁behind	-7.20527
+▁bananas	-7.20846
+ting	-7.21381
+▁men	-7.22024
+▁ground	-7.22091
+▁background	-7.22354
+▁b	-7.22428
+▁bird	-7.22629
+▁snowboard	-7.22802
+▁bike	-7.22918
+▁glasses	-7.23108
+▁piece	-7.23459
+f	-7.24452
+▁child	-7.24754
+▁carrots	-7.25782
+▁air	-7.25869
+▁display	-7.26076
+▁stuff	-7.26113
+▁head	-7.2612
+▁sandwich	-7.26248
+▁cup	-7.27509
+pped	-7.27733
+▁baby	-7.28606
+▁full	-7.2917
+▁hold	-7.2918
+an	-7.29489
+▁stands	-7.29853
+ous	-7.30242
+▁luggage	-7.31451
+▁into	-7.31494
+▁being	-7.32332
+▁oven	-7.33198
+▁beside	-7.3412
+ll	-7.3429
+▁photo	-7.3438
+it	-7.34923
+▁sidewalk	-7.35576
+▁cutt	-7.35913
+▁shirt	-7.36241
+▁paper	-7.36492
+▁floor	-7.36792
+▁dirt	-7.37091
+▁knife	-7.37703
+▁g	-7.37863
+▁pink	-7.38024
+▁trees	-7.38423
+▁fence	-7.38543
+▁cow	-7.3919
+▁game	-7.39564
+▁bicycle	-7.40192
+▁window	-7.40507
+▁pole	-7.41001
+▁look	-7.41334
+▁skis	-7.41535
+re	-7.42092
+▁big	-7.42447
+▁t	-7.42848
+▁m	-7.4504
+▁pile	-7.45319
+▁view	-7.45363
+▁face	-7.4605
+▁kite	-7.4606
+il	-7.46525
+▁airplane	-7.46698
+▁oranges	-7.46926
+▁wood	-7.47265
+ro	-7.47918
+▁mouth	-7.48736
+▁do	-7.48758
+▁toothbrush	-7.48941
+▁zebras	-7.49212
+▁apples	-7.49458
+▁image	-7.4977
+▁plane	-7.51166
+ur	-7.51727
+▁stand	-7.52431
+ers	-7.52482
+▁kites	-7.52906
+▁sky	-7.54382
+▁ski	-7.54406
+▁meat	-7.56227
+▁cut	-7.56842
+▁apple	-7.56932
+▁another	-7.57328
+▁park	-7.57555
+▁refrigerator	-7.59596
+la	-7.60542
+▁vari	-7.6226
+▁st	-7.62544
+▁co	-7.6269
+▁grassy	-7.62858
+▁bears	-7.62972
+▁bag	-7.63261
+et	-7.63472
+▁four	-7.64696
+▁colorful	-7.65507
+▁hat	-7.66174
+▁someone	-7.66735
+▁them	-7.66998
+▁book	-7.67032
+▁donuts	-7.68524
+▁tall	-7.68824
+▁cows	-7.68997
+▁bottle	-7.69113
+▁fork	-7.69134
+▁women	-7.69223
+▁banana	-7.69969
+▁stove	-7.70069
+▁off	-7.70246
+▁n	-7.70318
+▁box	-7.70362
+▁control	-7.70493
+▁coffee	-7.71351
+▁station	-7.71806
+▁chair	-7.72184
+at	-7.72473
+id	-7.73245
+ter	-7.73264
+▁mirror	-7.73525
+▁along	-7.73525
+▁ready	-7.73965
+▁herd	-7.75265
+▁cr	-7.75359
+▁camera	-7.76197
+li	-7.76254
+▁suitcase	-7.76274
+▁w	-7.7628
+▁c	-7.76326
+ck	-7.76521
+to	-7.76797
+▁cheese	-7.77104
+▁hanging	-7.77442
+ic	-7.77562
+▁items	-7.77855
+▁la	-7.78103
+▁line	-7.78226
+▁tray	-7.78482
+▁giraffes	-7.80095
+▁above	-7.80817
+ent	-7.80837
+▁he	-7.81318
+▁spoon	-7.81767
+▁can	-7.82528
+▁elephants	-7.82618
+▁middle	-7.82727
+▁long	-7.82739
+▁wii	-7.82816
+ion	-7.83194
+▁suit	-7.8321
+▁re	-7.83355
+▁po	-7.83624
+▁half	-7.84183
+▁posing	-7.84212
+ra	-7.84547
+▁metal	-7.84673
+▁pa	-7.84758
+▁tak	-7.84771
+▁grazing	-7.85166
+▁get	-7.85233
+ve	-7.85276
+▁bar	-7.86318
+th	-7.86829
+▁sand	-7.87105
+way	-7.87305
+▁stick	-7.87666
+▁walk	-7.8805
+ut	-7.88539
+▁day	-7.88687
+▁smiling	-7.89196
+▁like	-7.89197
+gain	-7.89252
+▁fruits	-7.89493
+ce	-7.89643
+▁mountain	-7.89712
+▁carry	-7.89715
+▁sh	-7.90562
+▁living	-7.9233
+▁pe	-7.92486
+▁court	-7.92862
+▁all	-7.92962
+▁signs	-7.93469
+▁surf	-7.93646
+▁basket	-7.93934
+ol	-7.94429
+▁silver	-7.94475
+▁drink	-7.94482
+ct	-7.94505
+▁door	-7.94597
+▁animals	-7.95318
+▁ma	-7.95935
+▁cars	-7.96219
+▁hair	-7.97801
+▁pull	-7.9781
+▁pan	-7.98197
+▁dogs	-7.98348
+▁row	-7.99477
+▁h	-7.99801
+▁across	-8.00047
+▁airport	-8.00047
+▁bread	-8.00047
+▁lean	-8.01205
+▁animal	-8.01461
+▁plastic	-8.01781
+▁who	-8.01782
+te	-8.02911
+▁horses	-8.03068
+▁trick	-8.03557
+▁couch	-8.03559
+▁no	-8.03695
+▁com	-8.03871
+▁dish	-8.04053
+z	-8.04088
+▁tracks	-8.04841
+▁mo	-8.05569
+▁set	-8.05572
+▁ca	-8.05817
+▁screen	-8.05949
+▁slice	-8.05954
+▁donut	-8.06354
+▁rock	-8.06384
+▁birds	-8.06533
+▁hands	-8.06726
+▁ra	-8.06906
+▁skate	-8.07777
+▁store	-8.0842
+▁driving	-8.09035
+▁empty	-8.09035
+▁un	-8.09227
+lying	-8.09266
+▁pot	-8.09951
+up	-8.1032
+▁chocolate	-8.10934
+▁enclosure	-8.10934
+▁something	-8.1158
+▁market	-8.12232
+▁seat	-8.14076
+▁purple	-8.14181
+▁television	-8.14181
+▁using	-8.14191
+▁displayed	-8.1427
+▁snowboarder	-8.1446
+▁house	-8.14844
+▁skier	-8.15227
+▁slope	-8.1551
+▁video	-8.1551
+▁color	-8.15566
+▁hit	-8.15587
+▁controller	-8.16183
+▁container	-8.16206
+▁shown	-8.16321
+▁sp	-8.16484
+ies	-8.16508
+us	-8.1671
+▁hill	-8.16834
+▁di	-8.18087
+▁multi	-8.18222
+uring	-8.1826
+▁case	-8.18262
+▁tooth	-8.18916
+▁outdoor	-8.19606
+▁flower	-8.1975
+▁number	-8.20305
+▁tv	-8.20468
+▁brush	-8.20488
+ping	-8.21358
+▁doughnuts	-8.21646
+▁double	-8.21719
+▁hotdog	-8.21719
+▁monitor	-8.21719
+▁salad	-8.21719
+▁lady	-8.21722
+▁pen	-8.22209
+▁attached	-8.22433
+is	-8.22731
+▁boats	-8.23018
+um	-8.23022
+▁contain	-8.23126
+▁teeth	-8.23161
+▁k	-8.23797
+▁resting	-8.23827
+▁ha	-8.24136
+▁bun	-8.24273
+▁guy	-8.24609
+▁made	-8.24831
+ot	-8.25082
+▁de	-8.25114
+▁ocean	-8.25342
+▁tower	-8.25351
+ew	-8.25769
+▁meal	-8.26085
+▁corner	-8.26138
+▁passenger	-8.26829
+▁racquet	-8.27581
+▁crowd	-8.27749
+un	-8.28344
+▁umbrellas	-8.28864
+▁backpack	-8.29102
+▁high	-8.29871
+▁toy	-8.30503
+▁surround	-8.30647
+▁zoo	-8.31428
+▁brick	-8.3144
+▁lights	-8.31599
+colored	-8.32935
+▁restaurant	-8.33009
+▁cellphone	-8.33042
+ng	-8.3362
+▁pieces	-8.33696
+led	-8.33796
+▁fresh	-8.33809
+▁photograph	-8.33809
+▁shelf	-8.33809
+▁surface	-8.33809
+ween	-8.33811
+▁jacket	-8.34615
+▁sauce	-8.34615
+▁wave	-8.34616
+▁ride	-8.35186
+▁statue	-8.35428
+▁adult	-8.35428
+▁waiting	-8.3586
+▁watching	-8.35954
+▁plant	-8.36065
+▁branch	-8.36248
+▁jet	-8.36291
+▁post	-8.36497
+▁painted	-8.3655
+▁books	-8.36582
+ie	-8.36742
+▁past	-8.37075
+▁polar	-8.37096
+▁play	-8.37126
+▁so	-8.37503
+▁track	-8.37767
+▁runway	-8.38762
+side	-8.38829
+▁glove	-8.39596
+▁clear	-8.39596
+▁time	-8.39597
+▁toothbrushes	-8.39938
+▁showing	-8.40062
+▁chicken	-8.4045
+▁helmet	-8.4045
+▁dark	-8.40451
+▁children	-8.40452
+▁rice	-8.40464
+▁sleeping	-8.4071
+▁arm	-8.40782
+▁bet	-8.40891
+▁skiing	-8.41273
+▁electronic	-8.41312
+▁body	-8.41314
+▁eaten	-8.41397
+ation	-8.41581
+top	-8.41755
+▁dressed	-8.4181
+▁tile	-8.42344
+▁going	-8.42404
+▁prepar	-8.43043
+co	-8.43223
+ple	-8.43542
+if	-8.43911
+▁go	-8.43936
+▁talking	-8.45039
+ling	-8.45079
+ive	-8.45333
+▁or	-8.46099
+▁sliced	-8.46327
+▁bright	-8.46647
+▁grey	-8.46648
+▁doughnut	-8.46741
+▁dock	-8.46803
+▁lo	-8.47278
+▁wire	-8.47628
+▁rail	-8.47653
+▁eat	-8.47835
+▁sun	-8.47862
+▁types	-8.47913
+ig	-8.48178
+▁ex	-8.48495
+▁about	-8.48501
+▁gray	-8.4893
+▁left	-8.49425
+▁fries	-8.49426
+▁strip	-8.49429
+▁toaster	-8.49558
+age	-8.49645
+▁includ	-8.50229
+▁neck	-8.5041
+▁carrot	-8.5057
+▁star	-8.50571
+▁intersection	-8.51321
+▁ne	-8.51381
+mp	-8.51572
+▁toward	-8.52283
+▁rack	-8.523
+▁have	-8.52335
+ti	-8.52431
+▁drinking	-8.53239
+▁device	-8.54233
+▁soup	-8.5427
+▁beer	-8.54704
+▁brushing	-8.54769
+▁tub	-8.55224
+▁skateboarder	-8.5601
+▁blanket	-8.56224
+▁shower	-8.56331
+▁ho	-8.56386
+▁rain	-8.56402
+▁leaves	-8.57234
+▁trunk	-8.57234
+▁variet	-8.57362
+▁walks	-8.57371
+▁stra	-8.57436
+▁dress	-8.57668
+▁bags	-8.57846
+▁rocks	-8.5855
+▁male	-8.58767
+▁potatoes	-8.59314
+▁place	-8.59391
+▁take	-8.597
+▁cover	-8.59788
+▁gra	-8.59976
+▁swing	-8.60294
+x	-8.60304
+▁toppings	-8.60342
+▁night	-8.60375
+ight	-8.60386
+ian	-8.60506
+▁meters	-8.60745
+placed	-8.61271
+▁cross	-8.6135
+▁home	-8.61381
+ment	-8.61483
+▁cart	-8.61607
+▁signal	-8.61711
+▁bo	-8.6237
+▁shows	-8.62811
+▁kid	-8.63083
+▁vehicle	-8.63518
+▁shop	-8.63771
+▁batter	-8.63776
+▁office	-8.64358
+▁decker	-8.64688
+▁closeup	-8.64917
+▁sa	-8.6532
+▁she	-8.6543
+▁older	-8.65575
+ten	-8.6566
+▁dessert	-8.65704
+▁graffiti	-8.65704
+▁buildings	-8.66063
+▁decorated	-8.66817
+▁traveling	-8.67289
+▁uniform	-8.67939
+▁motorcycles	-8.68378
+▁not	-8.68576
+ke	-8.68938
+ca	-8.68955
+ish	-8.69322
+▁bottles	-8.69376
+▁stacked	-8.6945
+▁hug	-8.70225
+▁shaped	-8.70448
+ir	-8.70499
+ake	-8.70533
+▁work	-8.71475
+▁ru	-8.71526
+ate	-8.71737
+▁rest	-8.71757
+im	-8.72285
+▁se	-8.72481
+▁motor	-8.72565
+▁roll	-8.72566
+able	-8.72568
+▁lit	-8.72596
+▁river	-8.72599
+▁egg	-8.73755
+op	-8.74301
+ta	-8.74331
+▁cabinets	-8.74989
+▁shot	-8.75016
+▁setting	-8.75117
+▁dry	-8.75973
+▁cloud	-8.76179
+▁assortment	-8.76196
+ver	-8.76214
+▁slices	-8.76601
+▁cooking	-8.77002
+▁served	-8.77589
+▁onions	-8.77609
+▁way	-8.78289
+qu	-8.78524
+▁cute	-8.78994
+▁ice	-8.7915
+▁we	-8.79685
+▁cloth	-8.7979
+▁bu	-8.79794
+▁appliances	-8.79929
+lush	-8.79929
+▁giant	-8.79929
+▁underneath	-8.79957
+▁pu	-8.80217
+▁suitcases	-8.80993
+▁vegetable	-8.81141
+▁assort	-8.81193
+▁machine	-8.81211
+▁gear	-8.81213
+▁forest	-8.81221
+▁pack	-8.813
+ard	-8.81466
+▁mak	-8.81679
+▁jumping	-8.82059
+▁bikes	-8.82337
+▁electric	-8.8251
+▁purse	-8.8251
+▁reading	-8.8251
+▁smart	-8.8251
+▁mitt	-8.82511
+▁moving	-8.82512
+▁swinging	-8.8256
+▁fish	-8.82654
+▁seen	-8.82973
+▁edge	-8.83827
+▁feet	-8.83828
+unny	-8.84257
+▁surfer	-8.84536
+▁part	-8.84568
+ha	-8.85162
+▁tea	-8.85311
+▁ta	-8.85313
+ul	-8.85659
+▁pi	-8.8584
+▁con	-8.86021
+▁see	-8.86344
+▁chairs	-8.86605
+tra	-8.86699
+light	-8.86717
+ant	-8.87635
+ical	-8.87771
+▁fridge	-8.8788
+▁perched	-8.8788
+▁running	-8.87883
+▁yard	-8.87885
+▁din	-8.87895
+▁pro	-8.88325
+▁cooked	-8.88379
+▁but	-8.88445
+▁cap	-8.88554
+▁end	-8.8862
+▁hay	-8.88776
+▁mount	-8.89269
+▁female	-8.89269
+▁mustard	-8.89269
+▁curb	-8.89269
+▁turn	-8.8927
+▁scene	-8.89274
+▁him	-8.89311
+▁ramp	-8.89321
+▁log	-8.89489
+▁wi	-8.90234
+▁pie	-8.90312
+▁object	-8.90678
+▁draw	-8.90678
+▁drive	-8.90678
+▁cattle	-8.90688
+▁cla	-8.91037
+lic	-8.91151
+▁sale	-8.91173
+ead	-8.91193
+▁wheel	-8.92106
+▁towel	-8.92112
+▁point	-8.92116
+▁tomatoes	-8.92201
+▁size	-8.92222
+▁holder	-8.92301
+fri	-8.92629
+▁bath	-8.92676
+tage	-8.92972
+▁poses	-8.93077
+▁shoes	-8.93101
+per	-8.93464
+▁buses	-8.93471
+▁beautiful	-8.93555
+▁mother	-8.93556
+▁among	-8.93556
+▁gold	-8.93557
+▁hang	-8.93983
+intend	-8.9417
+▁pasture	-8.95027
+▁cage	-8.9514
+▁plants	-8.95357
+▁bridge	-8.96519
+▁clean	-8.96519
+▁frost	-8.9652
+▁leg	-8.96545
+▁smiles	-8.96762
+ff	-8.96829
+▁sc	-8.96887
+▁pick	-8.97804
+▁putt	-8.97847
+▁sausage	-8.98034
+▁right	-8.98037
+▁stone	-8.98045
+▁candle	-8.98051
+▁wrappe	-8.98052
+▁single	-8.98062
+▁crossing	-8.98077
+▁goat	-8.98108
+▁lap	-8.98755
+▁arranged	-8.99572
+▁writ	-8.99573
+▁blow	-8.99574
+▁van	-8.99579
+▁desktop	-8.99754
+▁war	-8.99867
+board	-9.00012
+▁handle	-9.00077
+ated	-9.00415
+▁cream	-9.01135
+▁equipment	-9.01135
+▁mug	-9.01136
+▁beans	-9.01151
+▁pose	-9.01615
+▁kinds	-9.01994
+▁paint	-9.02154
+▁below	-9.02722
+▁pitch	-9.02722
+▁fly	-9.04102
+▁cook	-9.04289
+▁toothpaste	-9.04335
+▁perform	-9.04335
+▁held	-9.04337
+eat	-9.04504
+▁fenced	-9.0479
+▁kids	-9.05097
+▁peppers	-9.05207
+▁make	-9.05829
+▁broken	-9.05974
+▁check	-9.05974
+▁glazed	-9.05974
+▁grill	-9.05974
+▁ketchup	-9.05974
+▁police	-9.05975
+▁dinner	-9.05975
+▁pre	-9.05987
+▁float	-9.05998
+lar	-9.06452
+▁jump	-9.06548
+▁coat	-9.06712
+▁flip	-9.07263
+▁eyes	-9.07455
+▁design	-9.07641
+▁platform	-9.07641
+▁trash	-9.07641
+▁veggie	-9.07641
+▁short	-9.07642
+▁foot	-9.07642
+▁gather	-9.07647
+▁flock	-9.0766
+less	-9.07765
+▁watch	-9.08105
+▁use	-9.09076
+▁birthday	-9.09336
+▁pretty	-9.09336
+▁reach	-9.09337
+▁nice	-9.09337
+▁soda	-9.09338
+▁five	-9.09576
+▁kind	-9.10127
+ial	-9.10589
+▁public	-9.1106
+▁round	-9.1106
+having	-9.11272
+▁alone	-9.12815
+▁bushes	-9.12816
+▁lie	-9.13024
+▁taken	-9.13789
+▁stack	-9.14013
+just	-9.14601
+▁stopped	-9.14618
+▁lake	-9.14671
+▁new	-9.14915
+▁sea	-9.15017
+out	-9.15305
+▁pet	-9.15388
+eep	-9.16366
+▁produce	-9.16723
+▁used	-9.16863
+tic	-9.17655
+▁fac	-9.18216
+▁engine	-9.1827
+▁tarmac	-9.1827
+▁garden	-9.1827
+▁wild	-9.18272
+▁vin	-9.18277
+tuck	-9.18447
+tro	-9.18453
+▁mid	-9.18527
+▁dishes	-9.18707
+▁toast	-9.1992
+▁flat	-9.20159
+▁try	-9.20164
+▁both	-9.20226
+▁jar	-9.20273
+own	-9.20329
+▁hole	-9.20671
+▁wet	-9.20733
+▁feeding	-9.20948
+▁type	-9.2135
+j	-9.21763
+▁french	-9.2208
+▁beverage	-9.2208
+▁shore	-9.22082
+▁grow	-9.22087
+▁fall	-9.22102
+▁put	-9.22688
+▁show	-9.23742
+▁supplies	-9.24041
+▁farm	-9.24043
+▁pool	-9.24189
+▁stoplight	-9.24247
+▁working	-9.2589
+▁mix	-9.2604
+▁mushroom	-9.26041
+▁chain	-9.26041
+▁throw	-9.26041
+loaded	-9.26042
+▁good	-9.26043
+▁steel	-9.26043
+▁prepare	-9.26078
+▁brushes	-9.2644
+▁wait	-9.27001
+▁lin	-9.27581
+▁shape	-9.27684
+▁lunch	-9.28082
+▁modern	-9.28082
+▁square	-9.28082
+▁smo	-9.28088
+▁ear	-9.28125
+▁wash	-9.28147
+▁boxes	-9.28756
+phone	-9.29231
+▁pad	-9.29713
+ock	-9.29772
+▁tomato	-9.30026
+▁enjoy	-9.30165
+▁shoulder	-9.30165
+▁lemon	-9.30165
+▁pastries	-9.30165
+▁milk	-9.30165
+▁match	-9.30166
+▁lamb	-9.30166
+▁chew	-9.30167
+▁rose	-9.30169
+▁style	-9.30179
+▁well	-9.30274
+▁appear	-9.32293
+▁breakfast	-9.32293
+▁desert	-9.32293
+▁foreground	-9.32293
+▁napkin	-9.32293
+▁platter	-9.32293
+▁strawberries	-9.32293
+▁sunglasses	-9.32293
+▁blender	-9.32293
+▁leaf	-9.32293
+▁chopped	-9.32293
+▁shade	-9.32294
+▁lawn	-9.32294
+▁ripe	-9.32297
+▁sail	-9.32409
+▁doll	-9.32734
+▁gate	-9.33599
+▁crowded	-9.3398
+▁pedestrian	-9.34466
+▁baked	-9.34467
+▁decoration	-9.34474
+▁mess	-9.34492
+▁pasta	-9.3516
+line	-9.3603
+▁bottom	-9.36689
+▁christmas	-9.36689
+▁country	-9.36689
+▁decorative	-9.36689
+▁kneel	-9.36689
+▁scooter	-9.36689
+▁sculpture	-9.36689
+▁sprinkles	-9.36689
+▁chips	-9.36694
+▁things	-9.36697
+▁catcher	-9.36718
+▁butter	-9.36763
+▁sandwiches	-9.37834
+▁potte	-9.3786
+uch	-9.37978
+ball	-9.38093
+▁smile	-9.38591
+▁serving	-9.38961
+▁horn	-9.38965
+▁says	-9.3899
+▁cub	-9.39066
+▁includes	-9.393
+▁dryer	-9.39362
+▁skies	-9.39838
+▁was	-9.41263
+▁bathtub	-9.41287
+▁concrete	-9.41287
+▁distance	-9.41287
+where	-9.41288
+▁balloon	-9.41289
+▁nearby	-9.41308
+▁spot	-9.41374
+▁lamp	-9.41404
+▁tri	-9.41457
+▁path	-9.41472
+▁chi	-9.42199
+▁par	-9.42369
+▁ri	-9.42577
+air	-9.43144
+▁nose	-9.43838
+ular	-9.44005
+▁em	-9.44072
+▁run	-9.44797
+▁gl	-9.45232
+▁bacon	-9.46107
+▁bedroom	-9.46107
+▁carriage	-9.46107
+▁kitten	-9.46107
+▁stainless	-9.46107
+▁reads	-9.46107
+▁graze	-9.46125
+▁carrie	-9.46131
+▁bull	-9.46233
+▁race	-9.46325
+▁clothes	-9.46378
+▁low	-9.46384
+▁rider	-9.46863
+▁bite	-9.4693
+▁juice	-9.48607
+▁lettuce	-9.48607
+▁partially	-9.48607
+▁position	-9.48607
+▁swimming	-9.48607
+▁carpet	-9.48608
+▁sort	-9.48618
+▁plain	-9.48756
+▁paw	-9.49126
+▁travel	-9.50079
+▁cabinet	-9.51108
+▁blurr	-9.51171
+▁fashion	-9.51171
+▁pigeons	-9.51171
+▁what	-9.51171
+▁cement	-9.51173
+▁word	-9.51173
+▁same	-9.51174
+▁reflection	-9.51199
+▁after	-9.51234
+▁tiny	-9.51597
+▁pin	-9.51823
+▁ship	-9.51836
+▁feed	-9.52763
+▁arrangement	-9.53803
+▁bucket	-9.53803
+▁climb	-9.53803
+▁collecti	-9.53803
+▁shadow	-9.53803
+▁lift	-9.53803
+▁parade	-9.53803
+▁center	-9.53804
+▁flag	-9.53804
+berry	-9.53804
+▁lead	-9.53805
+▁pears	-9.53813
+▁sheet	-9.5393
+▁tape	-9.54224
+▁pickle	-9.5429
+▁giv	-9.54341
+▁bri	-9.55714
+▁you	-9.56488
+▁cluttered	-9.56505
+▁family	-9.56505
+▁military	-9.56505
+▁pavement	-9.56505
+▁picnic	-9.56505
+▁soccer	-9.56505
+▁peanut	-9.56505
+▁space	-9.56506
+▁pastry	-9.56509
+▁lone	-9.56512
+▁finger	-9.56707
+▁watches	-9.56732
+most	-9.56742
+bow	-9.57068
+▁officer	-9.57132
+ful	-9.57246
+▁serve	-9.58911
+▁mark	-9.59236
+▁deck	-9.59241
+▁curtain	-9.59283
+▁himself	-9.59283
+▁ledge	-9.59283
+▁railroad	-9.59283
+���duck	-9.59284
+▁base	-9.59284
+▁model	-9.59285
+rcial	-9.59286
+▁comme	-9.59289
+arrow	-9.59302
+▁hillside	-9.59378
+▁tools	-9.5943
+▁flor	-9.59538
+ci	-9.6108
+made	-9.6127
+ddler	-9.61507
+fore	-9.61889
+▁puppy	-9.6214
+▁school	-9.6214
+▁propeller	-9.6214
+▁cupcake	-9.6214
+▁built	-9.62141
+▁mini	-9.62144
+▁step	-9.62151
+▁string	-9.6216
+▁panda	-9.62177
+▁port	-9.62248
+▁pipe	-9.62495
+▁qui	-9.62497
+▁spi	-9.64844
+▁event	-9.65082
+▁block	-9.65082
+▁spread	-9.65082
+▁winter	-9.65082
+▁flies	-9.65083
+▁still	-9.65083
+▁sport	-9.65085
+▁fried	-9.65092
+▁direction	-9.65107
+cal	-9.65183
+▁landing	-9.65229
+▁trailer	-9.65376
+▁eye	-9.65481
+▁bit	-9.6791
+▁sub	-9.68077
+▁balanc	-9.68112
+▁pattern	-9.68112
+▁professional	-9.68112
+▁consist	-9.68112
+▁grapes	-9.68113
+▁spray	-9.68113
+▁antique	-9.68113
+ough	-9.68114
+▁stall	-9.68115
+▁package	-9.68161
+▁corn	-9.68243
+▁town	-9.6827
+▁tag	-9.68465
+▁tin	-9.68496
+▁figure	-9.71237
+▁furniture	-9.71237
+▁notebook	-9.71237
+▁lime	-9.71237
+▁parrot	-9.71237
+▁sofa	-9.71237
+▁outfit	-9.71238
+▁power	-9.71239
+▁disc	-9.71248
+▁fry	-9.71262
+▁wide	-9.71277
+▁chili	-9.71294
+▁hard	-9.71328
+▁features	-9.71373
+▁tail	-9.71384
+▁featur	-9.74332
+▁contents	-9.74463
+▁delicious	-9.74463
+▁expired	-9.74463
+▁guitar	-9.74463
+▁leash	-9.74463
+▁snack	-9.74463
+▁steak	-9.74463
+▁name	-9.74463
+▁push	-9.74468
+▁touching	-9.74479
+▁subway	-9.74504
+▁stir	-9.74607
+▁wear	-9.76934
+▁trail	-9.77464
+neath	-9.7773
+▁berries	-9.77796
+▁ceramic	-9.77796
+▁condiments	-9.77796
+▁fabric	-9.77796
+▁fancy	-9.77796
+▁blond	-9.77796
+▁stairs	-9.77798
+▁pants	-9.77835
+uzz	-9.77851
+▁tee	-9.79758
+▁craft	-9.81244
+▁visible	-9.81244
+▁chop	-9.81246
+▁rope	-9.81249
+▁beef	-9.81252
+▁key	-9.81259
+▁rais	-9.8143
+▁sleep	-9.83725
+bby	-9.84671
+▁bouquet	-9.84816
+▁museum	-9.84816
+▁restroom	-9.84816
+▁shelves	-9.84816
+▁advertisement	-9.84816
+▁flown	-9.84816
+▁tank	-9.84818
+▁vest	-9.84824
+ature	-9.84836
+soft	-9.84883
+▁icing	-9.84891
+▁flo	-9.88084
+band	-9.8828
+▁reflect	-9.88479
+▁amount	-9.88519
+▁owl	-9.88519
+▁steam	-9.88519
+▁tongue	-9.88519
+▁business	-9.88519
+▁costume	-9.88519
+▁heart	-9.88521
+▁calf	-9.88521
+▁worn	-9.88522
+▁sill	-9.88598
+▁propp	-9.91136
+ized	-9.92239
+how	-9.92279
+▁happy	-9.92365
+▁harbor	-9.92365
+▁pillow	-9.92365
+▁roof	-9.92365
+▁sugar	-9.92365
+▁airliner	-9.92366
+▁ornate	-9.92366
+▁indoor	-9.92366
+▁frame	-9.92366
+▁itself	-9.92367
+▁residen	-9.92367
+▁rusted	-9.92368
+▁selling	-9.92368
+▁ja	-9.96136
+▁figurine	-9.96365
+▁freezer	-9.96365
+▁garbage	-9.96365
+▁goggles	-9.96365
+▁waffle	-9.96365
+▁overhead	-9.96366
+▁section	-9.96366
+▁patio	-9.96366
+▁tasty	-9.96367
+▁frog	-9.96368
+▁mud	-9.96372
+▁belt	-9.96374
+▁fast	-9.96384
+▁curl	-9.96391
+▁item	-9.97951
+shirt	-9.99666
+▁celery	-10.0053
+▁faucet	-10.0053
+▁kept	-10.0053
+▁leather	-10.0053
+▁structure	-10.0053
+▁loading	-10.0053
+▁relax	-10.0053
+▁scatter	-10.0053
+▁numer	-10.0053
+▁six	-10.0053
+▁asleep	-10.0053
+ixture	-10.0053
+▁iron	-10.0053
+▁hood	-10.0054
+▁owner	-10.0055
+▁also	-10.0056
+rban	-10.0135
+▁produc	-10.0415
+thered	-10.0476
+▁accessories	-10.0488
+▁bakery	-10.0488
+▁cereal	-10.0488
+▁champagne	-10.0488
+▁commuter	-10.0488
+▁individual	-10.0488
+▁practic	-10.0488
+▁sweater	-10.0488
+▁ceiling	-10.0488
+▁text	-10.0488
+▁neatly	-10.0488
+▁rusty	-10.0488
+▁headphones	-10.0488
+▁cord	-10.0489
+▁lid	-10.0492
+▁je	-10.0523
+▁direct	-10.0939
+▁circle	-10.0943
+▁crochet	-10.0943
+▁magazine	-10.0943
+▁marble	-10.0943
+▁marina	-10.0943
+▁measur	-10.0943
+▁monkey	-10.0943
+▁roman	-10.0943
+▁urinal	-10.0943
+▁garage	-10.0943
+▁speak	-10.0943
+▁shoe	-10.109
+form	-10.1219
+▁talk	-10.131
+▁potato	-10.1405
+▁advertis	-10.1419
+▁approach	-10.1419
+▁burger	-10.1419
+▁character	-10.1419
+▁depict	-10.1419
+▁jockey	-10.1419
+▁kiwi	-10.1419
+▁stream	-10.1419
+▁terminal	-10.1419
+▁attempt	-10.1419
+▁jetliner	-10.1419
+▁vendor	-10.1419
+▁stunt	-10.1419
+▁collar	-10.1419
+foam	-10.1419
+▁palm	-10.1419
+▁necktie	-10.1419
+▁indicat	-10.1419
+▁foil	-10.1419
+▁than	-10.1419
+▁burn	-10.1419
+▁help	-10.1421
+▁tire	-10.1427
+▁jo	-10.1432
+▁pepper	-10.1649
+▁shak	-10.1781
+▁laughing	-10.1919
+▁ribbon	-10.1919
+▁shrimp	-10.1919
+▁sniff	-10.1919
+▁custom	-10.1919
+▁pepperoni	-10.1919
+▁missing	-10.1919
+▁rubb	-10.1919
+▁wool	-10.1919
+▁cartoon	-10.1919
+ique	-10.1919
+▁slid	-10.1919
+▁needle	-10.192
+▁canoe	-10.1922
+▁paddle	-10.2032
+▁bikini	-10.2445
+▁connect	-10.2445
+▁focus	-10.2445
+▁furry	-10.2445
+▁garnish	-10.2445
+▁grizzl	-10.2445
+▁horseback	-10.2445
+▁jersey	-10.2445
+▁liquid	-10.2445
+▁motorbike	-10.2445
+▁newspaper	-10.2445
+▁opposite	-10.2445
+▁ostrich	-10.2445
+▁powder	-10.2445
+▁selection	-10.2445
+▁silverware	-10.2445
+▁america	-10.2445
+▁money	-10.2445
+▁process	-10.2445
+▁pocket	-10.2445
+▁relish	-10.2445
+▁jelly	-10.2445
+▁odd	-10.2445
+▁santa	-10.2445
+▁fighter	-10.2445
+▁patch	-10.2445
+▁typing	-10.2445
+▁cold	-10.2446
+▁tasting	-10.2446
+lumin	-10.2456
+ador	-10.2834
+▁aircraft	-10.3001
+▁bookshelf	-10.3001
+▁cigarette	-10.3001
+▁digital	-10.3001
+▁exhibit	-10.3001
+▁interesting	-10.3001
+▁meadow	-10.3001
+▁muffin	-10.3001
+▁natural	-10.3001
+▁organiz	-10.3001
+▁overlook	-10.3001
+▁sweet	-10.3001
+plug	-10.3001
+▁baking	-10.3001
+▁celebrat	-10.3001
+▁remov	-10.3001
+▁wedding	-10.3001
+▁dozen	-10.3001
+▁forward	-10.3001
+▁jeans	-10.3001
+▁dust	-10.3002
+guard	-10.3004
+▁beak	-10.3005
+ield	-10.3036
+book	-10.3502
+▁avocado	-10.3589
+▁construction	-10.3589
+▁grapefruit	-10.3589
+▁ingredients	-10.3589
+▁instruction	-10.3589
+▁ipod	-10.3589
+▁knives	-10.3589
+▁learning	-10.3589
+▁liquor	-10.3589
+▁ornament	-10.3589
+▁pencils	-10.3589
+▁pineapple	-10.3589
+▁shallow	-10.3589
+▁tourist	-10.3589
+▁transport	-10.3589
+▁trolley	-10.3589
+▁tulips	-10.3589
+▁knitt	-10.3589
+▁magnet	-10.3589
+▁clown	-10.3589
+▁loung	-10.3589
+▁finish	-10.3589
+▁first	-10.3589
+▁operat	-10.3589
+▁backyard	-10.3589
+▁semi	-10.359
+▁medi	-10.372
+▁scissor	-10.3945
+▁plai	-10.4176
+▁calculator	-10.4214
+▁chrome	-10.4214
+▁church	-10.4214
+▁extreme	-10.4214
+▁fixing	-10.4214
+▁flavor	-10.4214
+▁fluffy	-10.4214
+▁motorcyclist	-10.4214
+▁nokia	-10.4214
+▁project	-10.4214
+▁selfie	-10.4214
+▁skating	-10.4214
+▁sneakers	-10.4214
+▁spinach	-10.4214
+▁stretch	-10.4214
+▁transit	-10.4214
+▁vegetation	-10.4214
+▁console	-10.4214
+▁material	-10.4214
+▁roast	-10.4214
+▁cheesecake	-10.4214
+▁crouch	-10.4214
+▁hung	-10.4214
+▁taxi	-10.4214
+▁weather	-10.4214
+▁swan	-10.4214
+▁crib	-10.4214
+▁safe	-10.4221
+▁decor	-10.4877
+▁antelope	-10.4881
+▁bamboo	-10.4881
+▁blood	-10.4881
+▁circu	-10.4881
+q	-10.8094

src/dataset/sub_tokenizer2000.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c60fef05dfb996f1a074bc5a252a0bd87eba570e346493a778d47dd3ee929f95
+size 271282

src/dataset/sub_tokenizer2000.vocab

@@ -0,0 +1,2000 @@
+<pad>	0
+<sos>	0
+<eos>	0
+<unk>	0
+▁a	-1.90379
+.	-2.84988
+▁of	-3.52043
+▁on	-3.53193
+▁in	-3.7858
+▁the	-3.78741
+s	-3.81805
+▁with	-3.85508
+nd	-3.87567
+▁is	-4.37004
+▁to	-4.55105
+ing	-4.63649
+▁man	-4.76743
+▁	-4.8183
+▁sitting	-4.86181
+,	-4.94566
+▁an	-5.02722
+▁next	-5.11411
+▁two	-5.13883
+ed	-5.23179
+▁white	-5.27115
+▁are	-5.29061
+▁holding	-5.3089
+▁standing	-5.31022
+▁table	-5.3802
+▁it	-5.42532
+y	-5.47436
+e	-5.47748
+▁at	-5.5327
+▁woman	-5.54252
+▁plate	-5.56616
+▁that	-5.58758
+▁up	-5.61664
+▁top	-5.64691
+▁people	-5.67154
+d	-5.70035
+▁some	-5.70552
+n	-5.71926
+▁person	-5.72054
+▁black	-5.79203
+▁large	-5.79333
+▁street	-5.83316
+▁red	-5.87919
+▁his	-5.93696
+▁small	-5.93728
+▁dog	-5.98374
+▁near	-5.98661
+▁front	-6.02607
+le	-6.03631
+▁sign	-6.112
+▁by	-6.15826
+▁bear	-6.16926
+▁cat	-6.16953
+▁group	-6.1793
+▁field	-6.19767
+▁has	-6.20777
+▁food	-6.21735
+▁blue	-6.21834
+▁green	-6.25381
+▁down	-6.26316
+▁young	-6.27466
+▁tennis	-6.29272
+▁snow	-6.29366
+▁other	-6.30134
+▁close	-6.3032
+▁whi	-6.31842
+▁there	-6.32654
+▁grass	-6.33462
+▁water	-6.33544
+▁side	-6.39473
+▁train	-6.42556
+▁computer	-6.42924
+t	-6.44999
+a	-6.46645
+▁baseball	-6.4953
+▁parked	-6.50776
+▁walking	-6.51064
+▁board	-6.51169
+▁sits	-6.52527
+▁riding	-6.53004
+o	-6.53714
+▁stop	-6.54352
+▁parking	-6.5445
+▁for	-6.55358
+▁kitchen	-6.55761
+r	-6.58109
+▁phone	-6.58369
+▁her	-6.59094
+▁wearing	-6.60086
+▁boy	-6.60371
+▁looking	-6.60483
+er	-6.60786
+▁clock	-6.61485
+▁bowl	-6.62845
+▁bus	-6.64378
+▁pizza	-6.64978
+▁fire	-6.6575
+▁woode	-6.65812
+'	-6.66211
+▁out	-6.6706
+m	-6.67177
+▁couple	-6.68249
+p	-6.68396
+▁picture	-6.68538
+▁thre	-6.69363
+▁yellow	-6.69364
+▁bathroom	-6.69845
+▁glass	-6.70559
+▁girl	-6.70816
+▁from	-6.73116
+▁toilet	-6.73449
+▁several	-6.73616
+▁skateboard	-6.74251
+▁vase	-6.74527
+▁brown	-6.75471
+▁building	-6.75554
+▁vegetables	-6.76098
+▁light	-6.76845
+▁hydrant	-6.77015
+▁hand	-6.79109
+▁beach	-6.79111
+▁scissors	-6.79703
+▁sink	-6.80533
+▁room	-6.80892
+▁their	-6.80978
+▁filled	-6.81271
+g	-6.8148
+▁over	-6.83258
+▁desk	-6.84733
+▁wine	-6.85868
+▁bat	-6.86169
+▁road	-6.87196
+▁elephant	-6.87345
+▁open	-6.8758
+▁one	-6.88385
+▁pair	-6.88936
+▁meter	-6.89336
+▁be	-6.91028
+▁horse	-6.91547
+▁hot	-6.92288
+▁cake	-6.92509
+▁thi	-6.92907
+▁each	-6.93109
+▁keyboard	-6.9311
+▁wall	-6.93925
+▁sheep	-6.94337
+▁different	-6.94954
+▁frisbee	-6.9496
+▁flying	-6.95108
+▁flowers	-6.95386
+b	-6.95499
+▁traffic	-6.95785
+▁orange	-6.95925
+▁laptop	-6.96623
+▁giraffe	-6.97434
+▁eating	-6.97688
+▁bench	-6.9801
+▁counter	-6.98527
+▁teddy	-6.98533
+▁its	-6.98622
+▁f	-6.98731
+▁tree	-6.98738
+▁around	-6.98749
+▁mouse	-6.98963
+▁covered	-6.99463
+▁tie	-7.00138
+st	-7.00458
+▁lay	-7.00734
+u	-7.01767
+▁broccoli	-7.01801
+▁cell	-7.02242
+▁inside	-7.0251
+▁boat	-7.02576
+▁through	-7.02915
+▁fruit	-7.04853
+▁remote	-7.05181
+▁ball	-7.0541
+▁bunch	-7.05562
+▁motorcycle	-7.05802
+▁area	-7.05888
+▁outside	-7.06116
+very	-7.06637
+al	-7.06708
+▁zebra	-7.06733
+▁little	-7.06797
+▁microwave	-7.07264
+▁truck	-7.07512
+es	-7.08244
+▁city	-7.09155
+▁old	-7.09496
+▁lot	-7.09663
+▁bed	-7.11485
+▁car	-7.11886
+▁umbrella	-7.12617
+▁together	-7.12794
+▁playing	-7.12974
+▁back	-7.14791
+ly	-7.14982
+▁sit	-7.15914
+▁under	-7.16824
+▁behind	-7.17346
+▁bananas	-7.17365
+ting	-7.18579
+▁men	-7.1879
+en	-7.18889
+▁ground	-7.18911
+▁s	-7.19022
+▁background	-7.19173
+▁glasses	-7.19856
+▁bird	-7.19888
+▁bike	-7.20221
+l	-7.20434
+▁piece	-7.20762
+▁child	-7.21573
+▁carrots	-7.2246
+▁stuff	-7.22932
+▁head	-7.22934
+▁display	-7.22943
+▁sandwich	-7.23111
+▁air	-7.23995
+▁surfboard	-7.24968
+▁baby	-7.25425
+▁full	-7.25989
+▁stands	-7.2628
+▁hold	-7.2628
+▁into	-7.2794
+▁luggage	-7.28271
+▁being	-7.28796
+▁oven	-7.30017
+w	-7.3046
+▁player	-7.30562
+▁beside	-7.30918
+▁photo	-7.31199
+ous	-7.31987
+▁sidewalk	-7.32395
+▁shirt	-7.33037
+laying	-7.33119
+▁paper	-7.33304
+▁cutt	-7.33384
+▁racket	-7.33603
+▁floor	-7.33609
+▁dirt	-7.3391
+▁knife	-7.34523
+▁trees	-7.34536
+▁snowboard	-7.34713
+▁pink	-7.34836
+▁fence	-7.35427
+▁game	-7.36383
+▁cup	-7.36772
+pped	-7.37284
+▁window	-7.37327
+i	-7.37392
+▁skis	-7.37715
+▁look	-7.38279
+▁cow	-7.3872
+▁big	-7.3926
+▁pile	-7.42014
+▁view	-7.42185
+▁d	-7.42569
+▁face	-7.4275
+▁oranges	-7.42956
+on	-7.43198
+▁kite	-7.4337
+ch	-7.44104
+▁zebras	-7.45306
+▁mouth	-7.45555
+▁toothbrush	-7.45804
+▁apples	-7.45822
+▁image	-7.46589
+▁plane	-7.47986
+c	-7.48139
+▁kites	-7.49201
+▁stand	-7.49869
+▁sky	-7.51201
+▁cut	-7.52986
+▁meat	-7.53038
+h	-7.53967
+▁another	-7.54147
+▁apple	-7.54237
+▁park	-7.54538
+▁refrigerator	-7.56415
+▁do	-7.57386
+���bears	-7.57774
+▁airplane	-7.58145
+▁pole	-7.58642
+▁vari	-7.59082
+▁grassy	-7.59426
+k	-7.6016
+▁bicycle	-7.6096
+▁four	-7.61515
+▁them	-7.62206
+▁colorful	-7.62325
+f	-7.62581
+▁hat	-7.62834
+▁someone	-7.63551
+▁ski	-7.63961
+▁book	-7.64343
+ar	-7.64588
+▁donuts	-7.65007
+▁cows	-7.65097
+▁tall	-7.65632
+▁fork	-7.65773
+▁women	-7.66042
+▁bottle	-7.6642
+▁stove	-7.66888
+▁off	-7.66944
+▁box	-7.67222
+▁banana	-7.67282
+▁control	-7.67313
+or	-7.67649
+▁coffee	-7.6817
+▁station	-7.68613
+▁bag	-7.6901
+▁chair	-7.69491
+it	-7.69954
+▁wood	-7.70204
+▁mirror	-7.70344
+▁ready	-7.70803
+▁herd	-7.71799
+an	-7.72101
+ll	-7.72672
+▁camera	-7.73017
+▁suitcase	-7.73581
+▁cheese	-7.73945
+▁hanging	-7.74151
+▁items	-7.74619
+▁tray	-7.75301
+▁line	-7.75759
+▁giraffes	-7.75836
+▁above	-7.77636
+▁elephants	-7.78217
+▁middle	-7.79546
+▁long	-7.79554
+▁wii	-7.79565
+v	-7.79936
+▁suit	-7.80029
+▁half	-7.81002
+▁posing	-7.81013
+▁metal	-7.81492
+in	-7.8172
+▁grazing	-7.81985
+▁get	-7.82
+▁tak	-7.82337
+▁can	-7.82451
+▁along	-7.82967
+▁sand	-7.83648
+▁t	-7.85072
+▁fruits	-7.85213
+▁day	-7.85504
+▁smiling	-7.86016
+▁like	-7.86016
+gain	-7.86042
+▁carry	-7.86556
+▁b	-7.8661
+▁signs	-7.87456
+▁all	-7.88738
+▁living	-7.89149
+▁co	-7.89325
+▁court	-7.89681
+ur	-7.90155
+▁surf	-7.90671
+▁basket	-7.90753
+▁cars	-7.91145
+▁silver	-7.91294
+▁animals	-7.91676
+▁dogs	-7.91824
+▁n	-7.92221
+▁m	-7.93004
+▁cr	-7.93076
+▁la	-7.93381
+re	-7.93666
+▁he	-7.9454
+la	-7.94666
+▁pan	-7.94721
+▁row	-7.96295
+il	-7.96838
+▁across	-7.96866
+▁airport	-7.96866
+▁bread	-7.96866
+el	-7.97626
+et	-7.97645
+▁lean	-7.98022
+ers	-7.9831
+▁horses	-7.98457
+▁plastic	-7.986
+▁hair	-7.98608
+▁animal	-7.98773
+ck	-8.00089
+▁trick	-8.00374
+▁couch	-8.00375
+▁dish	-8.00894
+▁tracks	-8.01307
+▁walk	-8.0142
+▁hands	-8.01734
+▁spoon	-8.01806
+▁set	-8.02221
+▁birds	-8.02343
+to	-8.02705
+▁screen	-8.02768
+ve	-8.02892
+▁slice	-8.03331
+▁donut	-8.03666
+▁door	-8.04045
+▁skate	-8.04596
+▁store	-8.05233
+▁g	-8.05489
+lying	-8.05641
+▁driving	-8.05854
+▁empty	-8.05854
+id	-8.06796
+▁st	-8.06843
+▁chocolate	-8.07753
+▁enclosure	-8.07754
+▁something	-8.08397
+▁pot	-8.08501
+ion	-8.08685
+▁market	-8.09047
+ol	-8.09804
+▁snowboarder	-8.10701
+▁displayed	-8.10975
+▁purple	-8.11
+▁television	-8.11
+▁using	-8.11007
+up	-8.11132
+▁house	-8.11663
+te	-8.12195
+▁slope	-8.12329
+▁video	-8.12329
+▁hit	-8.12384
+▁controller	-8.13001
+▁shown	-8.13113
+ct	-8.13323
+▁hill	-8.13673
+ter	-8.13878
+uring	-8.14849
+▁mountain	-8.14942
+▁case	-8.15073
+▁tooth	-8.15733
+ce	-8.16723
+▁flower	-8.17063
+▁number	-8.17125
+▁tv	-8.1718
+▁brush	-8.17476
+th	-8.17897
+▁doughnuts	-8.18083
+▁boats	-8.18276
+▁who	-8.18294
+▁double	-8.18538
+▁monitor	-8.18538
+▁salad	-8.18538
+▁lady	-8.1854
+▁un	-8.19178
+▁attached	-8.19252
+▁w	-8.19366
+▁teeth	-8.19978
+▁contain	-8.20528
+▁resting	-8.20567
+▁c	-8.21269
+ic	-8.21423
+▁guy	-8.21427
+▁made	-8.21638
+▁ocean	-8.22162
+▁tower	-8.22165
+▁pen	-8.22303
+▁pull	-8.22371
+▁rock	-8.22534
+▁meal	-8.22903
+▁corner	-8.22938
+ut	-8.23325
+▁drink	-8.23517
+▁passenger	-8.23649
+▁bun	-8.23926
+▁umbrellas	-8.24185
+▁racquet	-8.244
+▁crowd	-8.24616
+▁bar	-8.249
+▁toy	-8.24938
+▁backpack	-8.25921
+▁lights	-8.26244
+us	-8.27446
+▁surround	-8.27466
+li	-8.27745
+▁zoo	-8.28247
+▁brick	-8.28255
+▁pieces	-8.29064
+▁multi	-8.29803
+▁restaurant	-8.29828
+▁cellphone	-8.29839
+▁fresh	-8.30628
+▁photograph	-8.30628
+▁shelf	-8.30628
+▁surface	-8.30628
+▁jacket	-8.31435
+▁sauce	-8.31435
+ra	-8.31944
+▁ride	-8.32111
+▁adult	-8.32248
+▁statue	-8.32248
+▁books	-8.3242
+▁waiting	-8.32447
+▁watching	-8.32712
+▁jet	-8.33094
+▁painted	-8.33164
+▁plant	-8.33394
+▁post	-8.33454
+ro	-8.33538
+▁seat	-8.33698
+▁polar	-8.33905
+▁re	-8.34023
+▁de	-8.34346
+ween	-8.34728
+▁track	-8.3508
+▁runway	-8.35575
+▁glove	-8.36415
+▁clear	-8.36415
+▁time	-8.36415
+▁toothbrushes	-8.36648
+colored	-8.36734
+▁showing	-8.36824
+▁bet	-8.37115
+▁chicken	-8.37269
+▁helmet	-8.37269
+▁dark	-8.3727
+▁children	-8.3727
+▁rice	-8.3728
+▁arm	-8.37381
+▁sleeping	-8.37499
+▁skiing	-8.37539
+▁electronic	-8.38132
+▁body	-8.38132
+▁eaten	-8.38205
+▁dressed	-8.38589
+▁po	-8.38869
+z	-8.38961
+▁going	-8.39045
+▁tile	-8.39077
+um	-8.39819
+▁high	-8.40741
+▁no	-8.41272
+▁ca	-8.41279
+▁container	-8.41439
+▁talking	-8.41835
+ping	-8.4213
+▁sliced	-8.43047
+▁grey	-8.43467
+▁doughnut	-8.44054
+▁rail	-8.44409
+▁wire	-8.44417
+▁types	-8.44495
+▁eat	-8.44589
+▁go	-8.4494
+▁outdoor	-8.44991
+way	-8.45313
+▁about	-8.45317
+▁gray	-8.45481
+▁stick	-8.45768
+▁left	-8.46244
+▁fries	-8.46244
+▁strip	-8.46332
+▁toaster	-8.46334
+ple	-8.46803
+▁includ	-8.47093
+▁neck	-8.47215
+ew	-8.47727
+▁carrot	-8.47883
+���plates	-8.48014
+▁intersection	-8.4814
+ies	-8.48551
+▁toward	-8.49102
+▁rack	-8.49106
+▁have	-8.49121
+▁drinking	-8.49373
+▁mo	-8.49608
+led	-8.49823
+at	-8.49873
+is	-8.4997
+▁h	-8.50948
+▁beer	-8.50953
+▁ma	-8.51031
+▁device	-8.51053
+▁soup	-8.51077
+▁brushing	-8.51413
+▁tub	-8.52043
+▁skateboarder	-8.52274
+▁walks	-8.52647
+▁bags	-8.53018
+▁blanket	-8.53043
+▁rain	-8.53077
+placed	-8.53093
+▁shower	-8.53111
+▁com	-8.53531
+▁rocks	-8.53974
+▁leaves	-8.54053
+▁trunk	-8.54053
+▁variet	-8.54173
+▁skier	-8.54242
+▁dress	-8.54535
+▁meters	-8.55009
+▁male	-8.55324
+se	-8.55394
+▁take	-8.55728
+▁potatoes	-8.56124
+▁branch	-8.5705
+▁toppings	-8.57155
+▁night	-8.57165
+▁swing	-8.57239
+▁color	-8.57781
+▁signal	-8.58144
+▁home	-8.58199
+▁cross	-8.58294
+▁she	-8.58596
+▁hotdog	-8.59239
+▁shows	-8.59355
+▁or	-8.59559
+ive	-8.59638
+▁buildings	-8.59816
+▁kid	-8.60352
+co	-8.60403
+▁batter	-8.60504
+▁sun	-8.60763
+▁office	-8.61285
+▁lo	-8.61329
+▁decker	-8.61477
+un	-8.6149
+▁closeup	-8.61732
+▁past	-8.62
+▁older	-8.62047
+▁dessert	-8.62523
+▁graffiti	-8.62523
+▁motorcycles	-8.62848
+▁decorated	-8.63636
+▁cart	-8.63985
+▁traveling	-8.64054
+▁bright	-8.64687
+▁uniform	-8.64758
+▁wave	-8.64843
+▁bottles	-8.64876
+▁sp	-8.65093
+▁stacked	-8.66238
+ation	-8.66312
+▁prepar	-8.66412
+▁hug	-8.67044
+▁shaped	-8.67231
+ie	-8.67365
+▁ne	-8.68267
+as	-8.68526
+▁rest	-8.68555
+ake	-8.69332
+▁motor	-8.69384
+▁roll	-8.69384
+▁lit	-8.694
+▁river	-8.69433
+ng	-8.69758
+▁di	-8.70117
+mp	-8.7023
+▁cabinets	-8.71577
+▁shot	-8.71788
+ian	-8.71819
+ped	-8.71821
+▁setting	-8.71935
+▁so	-8.7209
+▁slices	-8.72431
+▁cloud	-8.72998
+▁assortment	-8.73008
+▁dry	-8.73142
+▁cooking	-8.7332
+ca	-8.73336
+▁way	-8.74097
+▁onions	-8.74355
+▁served	-8.7438
+op	-8.74553
+ton	-8.74844
+▁cute	-8.75342
+▁place	-8.75415
+▁ice	-8.75825
+▁suitcases	-8.76433
+▁cloth	-8.76653
+▁giant	-8.76748
+▁ex	-8.76751
+lush	-8.76765
+▁underneath	-8.76778
+▁bikes	-8.76815
+▁appliances	-8.77018
+x	-8.78016
+▁assort	-8.7802
+▁machine	-8.7803
+▁gear	-8.78031
+▁forest	-8.78036
+ard	-8.78049
+▁ra	-8.7816
+if	-8.78346
+▁vegetable	-8.78455
+▁jumping	-8.7878
+ti	-8.78849
+▁swinging	-8.79217
+▁electric	-8.79329
+▁purse	-8.79329
+▁smart	-8.79329
+▁mitt	-8.79329
+▁moving	-8.7933
+▁fish	-8.79417
+▁reading	-8.79453
+▁mak	-8.7957
+▁seen	-8.79576
+▁play	-8.79971
+ent	-8.80641
+▁edge	-8.80646
+▁feet	-8.80646
+▁surfer	-8.8085
+unny	-8.81131
+▁ho	-8.81569
+able	-8.81641
+▁chairs	-8.81908
+ling	-8.82076
+▁part	-8.82764
+▁din	-8.83845
+▁fridge	-8.847
+▁running	-8.84701
+▁yard	-8.84703
+▁perched	-8.84816
+▁cooked	-8.84902
+▁but	-8.85047
+▁hay	-8.8505
+ish	-8.85124
+▁end	-8.85212
+▁female	-8.86088
+▁mount	-8.86088
+▁mustard	-8.86088
+▁curb	-8.86088
+▁scene	-8.86089
+▁him	-8.86148
+▁ramp	-8.86157
+▁log	-8.86229
+ir	-8.86636
+tage	-8.87072
+ant	-8.87295
+ha	-8.87336
+▁object	-8.87497
+▁draw	-8.87497
+▁cattle	-8.87502
+▁sale	-8.87809
+ver	-8.87839
+▁holder	-8.88327
+▁dock	-8.8855
+▁poses	-8.88726
+▁we	-8.88797
+▁wheel	-8.88925
+▁size	-8.88927
+▁towel	-8.88927
+▁tomatoes	-8.8899
+▁bath	-8.89331
+ten	-8.89694
+▁shoes	-8.89819
+▁buses	-8.89825
+▁atop	-8.90093
+▁beautiful	-8.90375
+▁among	-8.90375
+▁gold	-8.90375
+▁mother	-8.90388
+▁bu	-8.90421
+intend	-8.90907
+▁plants	-8.9126
+▁pasture	-8.91856
+▁cage	-8.91891
+▁pa	-8.9214
+▁bicycles	-8.92412
+▁k	-8.92686
+▁con	-8.93124
+▁smiles	-8.93241
+▁shop	-8.93336
+▁bridge	-8.93338
+▁clean	-8.93338
+▁o	-8.94138
+▁sa	-8.94237
+king	-8.94412
+ty	-8.94564
+▁pick	-8.94631
+▁crossing	-8.94716
+tra	-8.94722
+▁sausage	-8.94853
+▁wrappe	-8.94861
+▁stone	-8.94862
+▁candle	-8.94864
+▁single	-8.9487
+▁right	-8.95142
+▁lap	-8.95192
+ig	-8.95197
+lic	-8.95305
+▁putt	-8.95306
+▁vases	-8.95733
+▁handle	-8.96353
+▁arranged	-8.96391
+▁blow	-8.96392
+▁desktop	-8.96441
+▁bowls	-8.96528
+board	-8.96861
+▁work	-8.97108
+ta	-8.97472
+▁cream	-8.97954
+▁mug	-8.97954
+▁equipment	-8.97955
+▁beans	-8.97962
+▁sh	-8.97982
+▁kinds	-8.98358
+ate	-8.98666
+▁wi	-8.98951
+per	-8.99501
+▁below	-8.99541
+▁phones	-9.00677
+▁fly	-9.00787
+▁toothpaste	-9.01154
+▁perform	-9.01154
+▁held	-9.01155
+▁kids	-9.01168
+▁fenced	-9.01268
+▁ha	-9.01676
+▁peppers	-9.01863
+▁broken	-9.02794
+▁ketchup	-9.02794
+▁police	-9.02794
+▁grill	-9.02794
+▁dinner	-9.02794
+▁flip	-9.02803
+▁float	-9.02807
+▁glazed	-9.0284
+lar	-9.0287
+▁make	-9.02938
+▁coat	-9.03253
+▁jump	-9.03492
+less	-9.03789
+▁eyes	-9.03995
+▁pe	-9.04206
+▁platform	-9.0446
+▁trash	-9.0446
+▁foot	-9.0446
+▁design	-9.0446
+▁gather	-9.04462
+▁flock	-9.04476
+▁rackets	-9.04532
+▁cover	-9.04694
+ot	-9.04859
+▁watch	-9.05093
+▁pro	-9.05256
+▁poles	-9.05323
+▁not	-9.05548
+▁use	-9.05927
+▁birthday	-9.06155
+▁pretty	-9.06155
+▁reach	-9.06155
+▁nice	-9.06156
+▁soda	-9.06156
+▁five	-9.06325
+ry	-9.06769
+▁kind	-9.07436
+out	-9.07513
+im	-9.07559
+▁cap	-9.07768
+▁public	-9.07879
+▁round	-9.0788
+having	-9.07927
+ke	-9.09157
+▁woods	-9.09593
+▁alone	-9.09634
+▁bushes	-9.09672
+▁sea	-9.09702
+▁lie	-9.0974
+▁pie	-9.09898
+▁players	-9.10129
+▁taken	-9.1048
+ated	-9.10616
+▁ta	-9.10644
+▁stack	-9.1088
+▁lake	-9.11458
+▁tea	-9.11498
+▁new	-9.11598
+▁stopped	-9.11657
+▁bo	-9.12561
+▁girls	-9.13226
+▁produce	-9.1347
+▁used	-9.13607
+eep	-9.13774
+▁surfboards	-9.14452
+ul	-9.14582
+▁engine	-9.15089
+▁garden	-9.15089
+▁tarmac	-9.15089
+▁wild	-9.1509
+tuck	-9.1517
+▁mid	-9.15203
+▁vin	-9.15343
+▁dishes	-9.15387
+▁pictures	-9.15426
+▁skiers	-9.15512
+light	-9.1598
+side	-9.16681
+▁lin	-9.16698
+▁toast	-9.16817
+▁flat	-9.16977
+▁both	-9.17007
+▁try	-9.17021
+▁jar	-9.17041
+▁wet	-9.17266
+▁hole	-9.17314
+▁pet	-9.17527
+▁feeding	-9.17672
+▁hang	-9.17911
+▁put	-9.18438
+▁type	-9.18665
+▁snowboards	-9.18809
+▁beverage	-9.18899
+▁french	-9.18899
+▁shore	-9.189
+▁grow	-9.19133
+ment	-9.1917
+tro	-9.20361
+▁van	-9.20857
+▁supplies	-9.2086
+▁pre	-9.20865
+▁pool	-9.20927
+ical	-9.20931
+▁stoplight	-9.21012
+▁show	-9.21323
+▁working	-9.22095
+▁pack	-9.22349
+j	-9.22633
+▁drinks	-9.22648
+▁drive	-9.22736
+▁mix	-9.22837
+▁chain	-9.2286
+▁throw	-9.2286
+loaded	-9.22861
+▁steel	-9.22861
+▁good	-9.22861
+▁brushes	-9.23132
+▁clocks	-9.2325
+▁star	-9.24495
+▁shape	-9.24568
+qu	-9.24648
+▁egg	-9.24781
+tic	-9.2479
+age	-9.24878
+▁lunch	-9.24901
+▁modern	-9.24901
+▁square	-9.24901
+▁ear	-9.24918
+▁wash	-9.24948
+ther	-9.2503
+▁pad	-9.25358
+▁boxes	-9.25361
+▁balls	-9.25453
+▁ties	-9.2653
+▁sc	-9.26645
+ight	-9.26724
+▁tomato	-9.2689
+▁shoulder	-9.26984
+▁enjoy	-9.26984
+▁pastries	-9.26984
+▁milk	-9.26984
+▁lamb	-9.26985
+▁chew	-9.26985
+▁rose	-9.26986
+▁style	-9.26992
+▁match	-9.27015
+▁well	-9.27046
+▁mountains	-9.27287
+▁veggies	-9.27527
+ma	-9.28982
+▁vehicles	-9.29095
+▁appear	-9.29112
+▁blender	-9.29112
+▁breakfast	-9.29112
+▁desert	-9.29112
+▁foreground	-9.29112
+▁leaf	-9.29112
+▁napkin	-9.29112
+▁platter	-9.29112
+▁strawberries	-9.29112
+▁sunglasses	-9.29112
+▁shade	-9.29112
+▁lawn	-9.29113
+▁chopped	-9.2912
+▁hotdogs	-9.29181
+▁stra	-9.2927
+▁ripe	-9.29287
+▁doll	-9.29379
+▁cla	-9.29632
+▁dr	-9.29828
+▁potte	-9.29836
+▁gate	-9.29845
+▁see	-9.29895
+ff	-9.30004
+▁crowded	-9.30661
+▁ru	-9.30863
+▁pedestrian	-9.31286
+▁baked	-9.31286
+▁decoration	-9.31291
+▁mess	-9.31294
+▁vehicle	-9.31303
+ac	-9.31455
+▁par	-9.31519
+▁pasta	-9.31593
+▁noodles	-9.31606
+ock	-9.32467
+▁bottom	-9.33508
+▁christmas	-9.33508
+▁country	-9.33508
+▁decorative	-9.33508
+▁scooter	-9.33508
+▁sculpture	-9.33508
+▁sprinkles	-9.33508
+▁chips	-9.3351
+▁things	-9.33513
+▁butter	-9.33556
+▁war	-9.33657
+▁kneel	-9.33963
+uch	-9.3405
+▁catcher	-9.34198
+▁sandwiches	-9.34288
+▁pi	-9.34733
+▁gra	-9.35321
+▁se	-9.35518
+▁serving	-9.35781
+▁horn	-9.35782
+▁says	-9.35794
+▁smile	-9.35928
+▁includes	-9.36011
+▁dryer	-9.36038
+▁skies	-9.36308
+▁i	-9.36614
+line	-9.36833
+▁spoons	-9.37146
+▁pulled	-9.37477
+▁prepare	-9.37951
+own	-9.37978
+▁colors	-9.38069
+▁bathtub	-9.38106
+▁concrete	-9.38106
+▁distance	-9.38106
+where	-9.38106
+▁balloon	-9.38107
+▁nearby	-9.38118
+▁was	-9.38129
+▁path	-9.38151
+▁tri	-9.38157
+▁lamp	-9.38182
+▁spot	-9.38335
+▁chi	-9.38405
+▁fac	-9.38913
+▁cups	-9.40016
+▁check	-9.40265
+▁nose	-9.40615
+ular	-9.40686
+▁run	-9.40878
+ger	-9.41062
+▁passing	-9.41115
+▁card	-9.41367
+▁ri	-9.42765
+▁bacon	-9.42926
+▁bedroom	-9.42926
+▁carriage	-9.42926
+▁kitten	-9.42926
+▁stainless	-9.42926
+▁graze	-9.42934
+▁carrie	-9.42936
+▁bull	-9.43
+▁low	-9.43078
+▁clothes	-9.43111
+▁fi	-9.43248
+▁vi	-9.4333
+▁rider	-9.43351
+▁painting	-9.4345
+▁reads	-9.4346
+▁bite	-9.43579
+▁pond	-9.43614
+▁airplanes	-9.44209
+▁legs	-9.44344
+cal	-9.44764
+▁turn	-9.44882
+▁juice	-9.45426
+▁lettuce	-9.45426
+▁partially	-9.45426
+▁position	-9.45426
+▁carpet	-9.45427
+▁swimming	-9.45427
+▁sort	-9.45431
+▁pose	-9.45529
+▁paw	-9.4554
+▁plain	-9.45556
+▁docked	-9.4639
+ine	-9.46494
+eat	-9.46557
+▁staring	-9.46617
+▁travel	-9.47023
+▁here	-9.47823
+▁blurr	-9.4799
+▁pigeons	-9.4799
+▁what	-9.4799
+▁cement	-9.47991
+▁word	-9.47991
+▁same	-9.47991
+▁ship	-9.47999
+▁reflection	-9.48012
+▁fashion	-9.48133
+▁tiny	-9.48351
+▁pin	-9.48409
+▁cabinet	-9.48423
+▁spi	-9.48493
+▁mushrooms	-9.48603
+ci	-9.49338
+▁point	-9.49645
+▁feed	-9.49684
+▁giv	-9.49926
+just	-9.50571
+▁arrangement	-9.50622
+▁bucket	-9.50622
+▁collecti	-9.50622
+▁lift	-9.50622
+▁parade	-9.50622
+▁shadow	-9.50622
+▁climb	-9.50622
+▁flag	-9.50622
+▁lead	-9.50623
+▁center	-9.50623
+▁pears	-9.50624
+▁sheet	-9.5071
+▁life	-9.50858
+▁utensils	-9.50952
+▁gl	-9.50966
+▁pickle	-9.51051
+▁writing	-9.51316
+▁pu	-9.51557
+▁containers	-9.51904
+▁tape	-9.52638
+▁race	-9.53101
+▁pitch	-9.53126
+▁you	-9.53311
+▁family	-9.53325
+▁military	-9.53325
+▁pavement	-9.53325
+▁peanut	-9.53325
+▁picnic	-9.53325
+▁soccer	-9.53325
+▁space	-9.53325
+▁lone	-9.53328
+▁pastry	-9.53355
+▁finger	-9.53421
+▁cluttered	-9.53437
+▁watches	-9.53479
+▁officer	-9.53677
+ful	-9.53806
+▁upside	-9.53856
+▁bell	-9.54519
+▁benches	-9.54552
+man	-9.54553
+▁covering	-9.549
+ure	-9.54966
+▁goats	-9.55088
+▁plays	-9.55139
+▁serve	-9.55787
+▁deck	-9.56079
+▁curtain	-9.56102
+▁himself	-9.56102
+▁ledge	-9.56102
+▁railroad	-9.56102
+▁duck	-9.56103
+▁model	-9.56103
+▁comme	-9.56104
+rcial	-9.56105
+arrow	-9.56106
+▁base	-9.56112
+▁hillside	-9.56146
+▁tools	-9.56201
+▁mark	-9.56362
+ial	-9.56538
+▁frosting	-9.56948
+▁sticks	-9.57019
+▁outdoors	-9.57072
+ral	-9.57163
+cer	-9.57499
+made	-9.5813
+ddler	-9.58536
+▁waves	-9.58741
+▁propeller	-9.58959
+▁puppy	-9.58959
+▁school	-9.58959
+▁built	-9.5896
+▁cupcake	-9.5896
+▁mini	-9.58962
+▁step	-9.58962
+▁string	-9.58966
+▁panda	-9.58968
+▁port	-9.59003
+▁eggs	-9.59129
+fore	-9.59259
+▁pipe	-9.59269
+ster	-9.59617
+ead	-9.5992
+▁flor	-9.6181
+▁block	-9.61901
+▁friend	-9.61901
+▁spread	-9.61901
+▁winter	-9.61901
+▁still	-9.61901
+▁flies	-9.61901
+▁sport	-9.61902
+▁fried	-9.61904
+▁direction	-9.61921
+▁event	-9.61939
+▁landing	-9.62011
+▁cub	-9.62039
+▁trailer	-9.62104
+▁cookie	-9.62399
+▁bi	-9.62462
+ary	-9.62694
+▁eye	-9.62776
+▁barn	-9.63204
+▁prepared	-9.63753
+ball	-9.64097
+▁comp	-9.64222
+▁tin	-9.64718
+▁corn	-9.64787
+▁dip	-9.64922
+▁balanc	-9.64931
+▁pattern	-9.64931
+▁professional	-9.64931
+▁consist	-9.64931
+▁spray	-9.64931
+▁antique	-9.64932
+▁stall	-9.64932
+▁package	-9.64945
+▁town	-9.65018
+▁tag	-9.6513
+▁grapes	-9.6552
+▁does	-9.65696
+▁lemons	-9.65937
+▁note	-9.66422
+▁stickers	-9.66544
+▁figure	-9.68056
+▁furniture	-9.68056
+▁lime	-9.68056
+▁parrot	-9.68056
+▁sofa	-9.68056
+▁outfit	-9.68056
+▁power	-9.68058
+▁disc	-9.68059
+▁notebook	-9.68059
+▁wide	-9.68085
+▁after	-9.6809
+▁chili	-9.68091
+▁hard	-9.68108
+▁tail	-9.68122
+▁features	-9.68149
+▁fry	-9.68236
+▁coin	-9.6829
+▁bit	-9.6863
+▁turned	-9.68744
+▁pointing	-9.69232
+▁bunches	-9.69379
+▁seated	-9.70181
+▁frost	-9.70307
+▁featur	-9.71192
+▁contents	-9.71282
+▁delicious	-9.71282
+▁expired	-9.71282
+▁guitar	-9.71282
+▁leash	-9.71282
+▁snack	-9.71282
+▁steak	-9.71282
+▁name	-9.71282
+▁push	-9.71283
+▁touching	-9.71297
+▁subway	-9.71301
+▁stir	-9.71352
+▁catch	-9.72328
+▁short	-9.72397
+▁whole	-9.72414
+▁waits	-9.72591
+▁wear	-9.73765
+▁trail	-9.74385
+neath	-9.74541
+▁berries	-9.74615
+▁blond	-9.74615
+▁ceramic	-9.74615
+▁condiments	-9.74615
+▁fabric	-9.74615
+▁fancy	-9.74615
+▁stairs	-9.74616
+▁pants	-9.74635
+uzz	-9.74642
+▁ham	-9.7487
+▁peeled	-9.7509
+▁tee	-9.75839
+▁paste	-9.75922
+▁goat	-9.75928
+how	-9.75951
+▁bri	-9.76824
+▁shorts	-9.76884
+top	-9.77115
+ity	-9.77287
+▁kn	-9.77848
+cent	-9.77932
+▁smo	-9.78057
+▁key	-9.78062
+▁craft	-9.78063
+▁visible	-9.78063
+▁beef	-9.78067
+▁rope	-9.78084
+▁flo	-9.78113
+▁rais	-9.78124
+▁sail	-9.78141
+▁net	-9.78173
+▁bro	-9.78227
+▁ram	-9.78314
+▁most	-9.78953
+▁propp	-9.79941
+▁rac	-9.80259
+▁farm	-9.80443
+▁sleep	-9.80669
+▁fall	-9.81572
+▁advertisement	-9.81635
+▁bouquet	-9.81635
+▁museum	-9.81635
+▁restroom	-9.81635
+▁shelves	-9.81635
+▁flown	-9.81635
+▁vest	-9.8164
+▁icing	-9.81676
+▁tank	-9.81686
+▁walkway	-9.81766
+▁seem	-9.81848
+▁shiny	-9.82033
+▁hid	-9.8206
+▁pier	-9.82143
+▁hotel	-9.82262
+▁packed	-9.82626
+band	-9.82788
+vi	-9.82908
+ature	-9.8321
+▁cook	-9.83224
+▁tan	-9.83303
+▁rocky	-9.83744
+▁writ	-9.84364
+▁pass	-9.84387
+bby	-9.85172
+▁bin	-9.8518
+▁reflect	-9.85308
+▁sub	-9.85318
+▁amount	-9.85338
+▁steam	-9.85338
+▁tongue	-9.85338
+▁business	-9.85339
+▁costume	-9.85339
+▁owl	-9.85339
+▁worn	-9.85339
+▁heart	-9.85339
+▁calf	-9.8534
+▁sill	-9.85379
+▁alongside	-9.85419
+▁tent	-9.85452
+▁gro	-9.85528
+▁print	-9.85611
+air	-9.85803
+min	-9.86074
+▁leg	-9.87016
+▁peel	-9.88722
+▁shin	-9.89114
+▁happy	-9.89185
+▁harbor	-9.89185
+▁pillow	-9.89185
+▁roof	-9.89185
+▁sugar	-9.89185
+▁airliner	-9.89185
+▁indoor	-9.89185
+▁ornate	-9.89185
+▁frame	-9.89185
+▁itself	-9.89185
+▁residen	-9.89185
+▁rusted	-9.89186
+▁shopping	-9.89246
+ized	-9.89285
+▁gas	-9.8932
+ator	-9.89406
+▁pitcher	-9.8947
+▁branches	-9.89544
+▁selling	-9.89747
+▁posed	-9.89757
+▁boarder	-9.8982
+▁em	-9.90128
+▁gr	-9.9121
+▁son	-9.9131
+▁hi	-9.91573
+▁ja	-9.93084
+▁fun	-9.93085
+▁figurine	-9.93185
+▁freezer	-9.93185
+▁garbage	-9.93185
+▁goggles	-9.93185
+▁overhead	-9.93185
+▁patio	-9.93185
+▁section	-9.93185
+▁waffle	-9.93185
+▁frog	-9.93185
+▁tasty	-9.93186
+▁mud	-9.93187
+▁belt	-9.93188
+ough	-9.93191
+▁curl	-9.93195
+▁fast	-9.93216
+▁rug	-9.93392
+▁ring	-9.93443
+▁bra	-9.93457
+▁mat	-9.94022
+tial	-9.94446
+▁item	-9.95206
+fri	-9.95832
+▁scatter	-9.97351
+▁celery	-9.97351
+▁faucet	-9.97351
+▁kept	-9.97351
+▁leather	-9.97351
+▁loading	-9.97351
+▁relax	-9.97351
+▁structure	-9.97351
+▁numer	-9.97352
+▁six	-9.97352
+▁asleep	-9.97352
+▁hood	-9.97353
+▁iron	-9.97353
+▁owner	-9.97359
+▁also	-9.9736
+▁qui	-9.97365
+▁let	-9.97421
+▁fallen	-9.97444
+▁class	-9.97447
+▁straw	-9.97482
+rban	-9.97534
+▁bow	-9.98064
+right	-10.0112
+▁produc	-10.0114
+▁accessories	-10.017
+▁bakery	-10.017
+▁ceiling	-10.017
+▁cereal	-10.017
+▁champagne	-10.017
+▁commuter	-10.017
+▁individual	-10.017
+▁sweater	-10.017
+▁neatly	-10.017
+▁practic	-10.017
+▁text	-10.017
+▁headphones	-10.017
+▁rusty	-10.017
+▁cord	-10.017
+▁shar	-10.0172
+▁lid	-10.0172
+▁passes	-10.0181
+thered	-10.0262
+ign	-10.0263
+▁pay	-10.0297
+▁farmer	-10.0318
+▁bush	-10.0319
+▁move	-10.033
+ok	-10.0372
+▁sticker	-10.047
+▁paint	-10.0473
+ador	-10.0608
+▁direct	-10.0621
+▁circle	-10.0624
+▁crochet	-10.0624
+▁garage	-10.0624
+▁magazine	-10.0624
+▁measur	-10.0624
+▁monkey	-10.0624
+▁urinal	-10.0624
+▁marina	-10.0624
+▁speak	-10.0624
+▁roman	-10.0625
+▁marble	-10.0625
+▁crate	-10.0636
+▁doorway	-10.0641
+▁je	-10.0645
+▁real	-10.0673
+▁shoe	-10.0804
+▁talk	-10.1004
+▁potato	-10.1091
+▁chop	-10.1099
+▁advertis	-10.1101
+▁approach	-10.1101
+▁attempt	-10.1101
+▁character	-10.1101
+▁depict	-10.1101
+▁jockey	-10.1101
+▁stream	-10.1101
+▁terminal	-10.1101
+▁burger	-10.1101
+▁jetliner	-10.1101
+▁kiwi	-10.1101
+▁vendor	-10.1101
+▁stunt	-10.1101
+▁collar	-10.1101
+▁palm	-10.1101
+foam	-10.1101
+▁foil	-10.1101
+▁necktie	-10.1101
+▁indicat	-10.1101
+▁burn	-10.1101
+▁than	-10.1101
+▁help	-10.1102
+▁tire	-10.1105
+▁workers	-10.1113
+▁jo	-10.1113
+▁grab	-10.1122
+▁driver	-10.1135
+▁brightly	-10.1136
+▁mar	-10.1161
+▁lab	-10.1183
+▁shak	-10.1276
+ound	-10.1287
+▁pepper	-10.138
+▁dea	-10.1406
+▁fill	-10.1557
+▁laughing	-10.1601
+▁missing	-10.1601
+▁pepperoni	-10.1601
+▁ribbon	-10.1601
+▁shrimp	-10.1601
+▁custom	-10.1601
+▁sniff	-10.1601
+▁rubb	-10.1601
+ique	-10.1601
+▁wool	-10.1601
+▁cartoon	-10.1601
+▁slid	-10.1601
+▁needle	-10.1601
+▁iphone	-10.1601
+▁speed	-10.1602
+▁canoe	-10.1602
+berry	-10.1603
+▁multicolored	-10.1616
+▁bagel	-10.1642
+▁bent	-10.1671
+▁paddle	-10.1676
+▁sla	-10.2031
+most	-10.2043
+▁garnish	-10.2127
+▁america	-10.2127
+▁bikini	-10.2127
+▁connect	-10.2127
+▁focus	-10.2127
+▁grizzl	-10.2127
+▁horseback	-10.2127
+▁jersey	-10.2127
+▁liquid	-10.2127
+▁money	-10.2127
+▁motorbike	-10.2127
+▁newspaper	-10.2127
+▁opposite	-10.2127
+▁ostrich	-10.2127
+▁powder	-10.2127
+▁process	-10.2127
+▁selection	-10.2127
+▁silverware	-10.2127
+▁furry	-10.2127
+▁pocket	-10.2127
+▁relish	-10.2127
+▁jelly	-10.2127
+▁patch	-10.2127
+▁fighter	-10.2127
+▁santa	-10.2127
+▁typing	-10.2127
+▁cold	-10.2127
+▁odd	-10.2127
+▁tasting	-10.2128
+lumin	-10.2138
+▁sharp	-10.2142
+▁wings	-10.2143
+▁stew	-10.2146
+▁team	-10.2165
+▁checker	-10.2177
+▁wait	-10.2362
+form	-10.2464
+bow	-10.2598
+plug	-10.2683
+▁aircraft	-10.2683
+▁bookshelf	-10.2683
+▁cigarette	-10.2683
+▁digital	-10.2683
+▁exhibit	-10.2683
+▁interesting	-10.2683
+▁meadow	-10.2683
+▁muffin	-10.2683
+▁natural	-10.2683
+▁organiz	-10.2683
+▁remov	-10.2683
+▁sweet	-10.2683
+▁celebrat	-10.2683
+▁overlook	-10.2683
+▁baking	-10.2683
+▁wedding	-10.2683
+▁dozen	-10.2683
+▁forward	-10.2683
+▁jeans	-10.2683
+▁sailboat	-10.2683
+guard	-10.2684
+▁dust	-10.2684
+▁hook	-10.2685
+▁beak	-10.2686
+▁action	-10.2686
+▁adjust	-10.2693
+ield	-10.2699
+▁more	-10.2703
+tain	-10.275
+tted	-10.2755
+▁layer	-10.2867
+▁medi	-10.2971
+▁start	-10.3157
+book	-10.3187
+▁avocado	-10.3271
+▁construction	-10.3271
+▁grapefruit	-10.3271
+▁ingredients	-10.3271
+▁instruction	-10.3271
+▁ipod	-10.3271
+▁learning	-10.3271
+▁liquor	-10.3271
+▁ornament	-10.3271
+▁pencils	-10.3271
+▁shallow	-10.3271
+▁tourist	-10.3271
+▁transport	-10.3271
+▁trolley	-10.3271
+▁tulips	-10.3271
+▁clown	-10.3271
+▁magnet	-10.3271
+▁pineapple	-10.3271
+▁loung	-10.3271
+▁finish	-10.3271
+▁first	-10.3271
+▁knives	-10.3271
+▁backyard	-10.3271
+▁operat	-10.3271
+▁knitt	-10.3271
+▁skirt	-10.3271
+▁semi	-10.3271
+▁formal	-10.3271
+▁apartment	-10.3272
+▁mask	-10.3274
+soft	-10.3274
+▁peas	-10.3278
+▁highway	-10.3286
+▁countertop	-10.3288
+▁flight	-10.3301
+▁scissor	-10.3677
+shirt	-10.3813
+phone	-10.3817
+▁perch	-10.3842
+▁plai	-10.3863
+▁calculator	-10.3896
+▁chrome	-10.3896
+▁church	-10.3896
+▁console	-10.3896
+▁flavor	-10.3896
+▁fluffy	-10.3896
+▁motorcyclist	-10.3896
+▁nokia	-10.3896
+▁project	-10.3896
+▁selfie	-10.3896
+▁skating	-10.3896
+▁sneakers	-10.3896
+▁spinach	-10.3896
+▁stretch	-10.3896
+▁transit	-10.3896
+▁vegetation	-10.3896
+▁cheesecake	-10.3896
+▁extreme	-10.3896
+▁material	-10.3896
+▁roast	-10.3896
+▁crouch	-10.3896
+▁hung	-10.3896
+▁fixing	-10.3896
+▁taxi	-10.3896
+▁crib	-10.3896
+▁swan	-10.3896
+▁boots	-10.3896
+▁weather	-10.3896
+▁safe	-10.3896
+▁sunset	-10.3896
+▁strawberry	-10.3896
+▁disk	-10.3896
+▁great	-10.3897
+▁competition	-10.3898
+▁alley	-10.3899
+▁label	-10.3901
+▁hangar	-10.3912
+▁bundle	-10.3928
+ither	-10.3942
+ould	-10.3944
+▁knee	-10.395
+▁page	-10.4205
+▁lemon	-10.4341
+▁sell	-10.4478
+amon	-10.4558
+▁decor	-10.456
+▁antelope	-10.4562
+▁bamboo	-10.4562
+▁blood	-10.4562
+▁circu	-10.4562
+▁dishwasher	-10.4562
+▁habitat	-10.4562
+▁library	-10.4562
+▁located	-10.4562
+▁pancakes	-10.4562
+▁separate	-10.4562
+▁special	-10.4562
+▁tissue	-10.4562
+▁wagon	-10.4562
+▁basin	-10.4562
+▁elaborate	-10.4562
+▁follow	-10.4562
+▁garlic	-10.4562
+▁gravel	-10.4562
+▁mobile	-10.4562
+▁roam	-10.4562
+▁baggage	-10.4562
+▁bank	-10.4562
+▁complete	-10.4562
+▁gravy	-10.4562
+▁attire	-10.4563
+▁dried	-10.4563
+▁ruler	-10.4563
+▁cafe	-10.4563
+▁nap	-10.4563
+▁sold	-10.4563
+▁saying	-10.4564
+▁thread	-10.4565
+▁smoke	-10.4565
+▁candy	-10.4571
+shells	-10.4622
+▁enclos	-10.5276
+▁artistic	-10.5277
+▁cauliflower	-10.5277
+▁charger	-10.5277
+▁chinese	-10.5277
+▁cushion	-10.5277
+▁dispenser	-10.5277
+▁grip	-10.5277
+▁illuminated	-10.5277
+▁industr	-10.5277
+▁kissing	-10.5277
+▁limb	-10.5277
+▁locomotive	-10.5277
+▁monument	-10.5277
+▁price	-10.5277
+▁splash	-10.5277
+▁system	-10.5277
+▁telephone	-10.5277
+▁turkey	-10.5277
+▁tusks	-10.5277
+▁wetsuit	-10.5277
+▁elder	-10.5277
+▁repair	-10.5277
+▁scarf	-10.5277
+terior	-10.5277
+▁wax	-10.5277
+▁groom	-10.5277
+▁punch	-10.5277
+▁storm	-10.5277
+▁barrel	-10.5277
+▁pudd	-10.5277
+▁extend	-10.5277
+▁dump	-10.5277
+▁trop	-10.5277
+▁chas	-10.5278
+▁enter	-10.5278
+▁heav	-10.5279
+▁vanity	-10.5279
+▁lick	-10.528
+▁wok	-10.5282
+ixture	-10.5285
+boat	-10.5962
+work	-10.5963
+▁apart	-10.6045
+▁micro	-10.6046
+▁donkey	-10.6046
+▁honey	-10.6046
+▁huddle	-10.6046
+▁jumbo	-10.6046
+▁pedestal	-10.6046
+▁protect	-10.6046
+▁symbol	-10.6046
+▁wrapping	-10.6046
+▁bloom	-10.6046
+▁break	-10.6046
+▁chest	-10.6046
+▁cruise	-10.6046
+▁gaze	-10.6046
+▁polish	-10.6046
+▁serious	-10.6046
+▁wrist	-10.6046
+▁booth	-10.6046
+▁soldier	-10.6046
+usual	-10.6046
+▁nail	-10.6046
+▁coach	-10.6046
+▁formation	-10.6046
+▁fighting	-10.6046
+▁dvd	-10.6046
+▁bald	-10.6046
+▁sewing	-10.6047
+▁warning	-10.6047
+▁sideways	-10.6049
+▁range	-10.6062
+▁clip	-10.6085
+▁veggie	-10.6658
+▁swim	-10.6879
+▁cluster	-10.6879
+▁mattress	-10.6879
+▁suspende	-10.6879
+▁blueberries	-10.6879
+▁brocolli	-10.6879
+▁conveyor	-10.6879
+▁cucumber	-10.6879
+▁downhill	-10.6879
+▁festival	-10.6879
+▁fountain	-10.6879
+▁frozen	-10.6879
+▁handbag	-10.6879
+▁kayak	-10.6879
+▁language	-10.6879
+▁league	-10.6879
+▁mustache	-10.6879
+▁sleek	-10.6879
+▁vandalized	-10.6879
+▁windowsill	-10.6879
+▁yogurt	-10.6879
+▁twig	-10.6879
+▁strange	-10.6879
+▁peach	-10.6879
+▁drain	-10.6879
+▁root	-10.6879
+▁soap	-10.6879
+▁grind	-10.688
+▁microphone	-10.688
+▁melt	-10.6881
+▁peek	-10.6882
+▁tip	-10.7229
+mote	-10.7542
+▁mushroom	-10.7567
+dog	-10.7698
+white	-10.7706
+life	-10.7736
+q	-10.7779
+▁navy	-10.7788
+tangerines	-10.7788
+▁alcohol	-10.7788
+▁artwork	-10.7788
+▁asphalt	-10.7788
+▁belong	-10.7788
+▁bronze	-10.7788
+▁casserole	-10.7788
+▁charging	-10.7788
+▁chef	-10.7788
+▁cherries	-10.7788
+▁citrus	-10.7788
+▁crosswalk	-10.7788
+▁factory	-10.7788
+▁fireplace	-10.7788
+▁gentlem	-10.7788
+▁herself	-10.7788
+▁junk	-10.7788
+▁kettle	-10.7788
+▁krisp	-10.7788
+▁landscape	-10.7788
+▁leaving	-10.7788
+▁macaroni	-10.7788
+▁squash	-10.7788
+▁thumb	-10.7788
+▁interact	-10.7788
+▁softball	-10.7788
+▁storage	-10.7788
+▁depot	-10.7788
+▁human	-10.7788
+▁quilt	-10.7788
+▁wicker	-10.7788
+▁second	-10.7789
+▁stalk	-10.7789
+▁twin	-10.7789
+▁order	-10.7789
+▁steer	-10.7789
+▁music	-10.7789
+▁mash	-10.779
+▁crane	-10.779
+▁heels	-10.779
+eagle	-10.779
+▁carving	-10.779
+▁appli	-10.7793
+▁swis	-10.7796
+▁ferr	-10.78
+dual	-10.7817
+pokes	-10.7818
+▁lu	-10.7963
+ello	-10.7972
+▁read	-10.8426
+▁appliance	-10.8568
+ripe	-10.8612
+wood	-10.8706
+▁barrier	-10.8788
+▁bicyclist	-10.8788
+▁briefcase	-10.8788
+▁change	-10.8788
+▁convention	-10.8788
+▁conversation	-10.8788
+▁expos	-10.8788
+▁kreme	-10.8788
+▁mannequin	-10.8788
+▁plunger	-10.8788
+▁polka	-10.8788
+▁resembl	-10.8788
+▁savannah	-10.8788
+▁seafood	-10.8788
+▁service	-10.8788
+▁shaggy	-10.8788
+▁skyscraper	-10.8788
+▁spectators	-10.8788
+▁sunflower	-10.8788
+▁tablecloth	-10.8788
+▁wilderness	-10.8788
+▁zone	-10.8788
+▁elev	-10.8788
+▁hallway	-10.8788
+▁pepsi	-10.8788
+▁retriev	-10.8788
+▁sparse	-10.8788
+▁squat	-10.8788
+▁student	-10.8788
+▁cubicle	-10.8788
+▁freight	-10.8788
+▁bubbl	-10.8788
+▁duff	-10.8788
+▁skillet	-10.8788
+▁curve	-10.8788
+▁german	-10.8788
+▁frown	-10.8788
+▁mural	-10.8788
+▁mesh	-10.8788
+▁porch	-10.8788
+▁broad	-10.8789
+▁present	-10.8789
+▁toiletries	-10.8789
+▁mice	-10.8789
+▁fixtures	-10.8789
+▁stem	-10.8789
+were	-10.9497
+▁grape	-10.9678
+print	-10.9815
+waiting	-10.9816
+▁chees	-10.9844
+▁glaze	-10.9866
+▁even	-10.9884
+ploye	-10.9899
+▁audience	-10.9899
+▁bristles	-10.9899
+▁coleslaw	-10.9899
+▁commode	-10.9899
+▁consumption	-10.9899
+▁damage	-10.9899
+▁elegant	-10.9899
+▁entree	-10.9899
+▁environment	-10.9899
+▁inspect	-10.9899
+▁jungle	-10.9899
+▁radio	-10.9899
+▁receipt	-10.9899
+▁return	-10.9899
+▁scratch	-10.9899
+▁slow	-10.9899
+▁spaghetti	-10.9899
+▁surprised	-10.9899
+▁world	-10.9899
+▁describ	-10.9899
+▁pilot	-10.9899
+▁roost	-10.9899
+▁shoot	-10.9899
+▁smell	-10.9899
+▁wedge	-10.9899
+▁hawk	-10.99
+▁placemat	-10.99
+▁sauerk	-10.99
+▁peace	-10.99
+▁loaf	-10.99
+▁dairy	-10.99
+▁worm	-10.99
+▁carousel	-10.99
+▁apron	-10.99
+▁kick	-10.99
+▁begin	-10.99

src/dataset/train_sub_tokenizer.py

@@ -0,0 +1,55 @@
+import json
+import os
+import sentencepiece as spm
+import yaml
+
+# params
+with open("/workspace/params.yaml", "r", encoding="utf-8") as f:
+    params = yaml.safe_load(f)
+
+
+def train_sentencepiece(
+        json_path,
+        model_prefix="sub_tokenizer",
+        vocab_size=500,
+        model_type="unigram"
+    ):
+
+    with open(json_path, 'r') as f:
+        data = json.load(f)
+
+    txt_path = "/workspace/src/dataset/sub_tokenizing_captions.txt"
+
+    with open(txt_path, "w", encoding="utf-8") as f:
+        for item in data:
+            captions = item["captions"]
+
+            for caption in captions:
+                f.write(caption.lower() + "\n")
+                
+    spm.SentencePieceTrainer.train(
+        input=txt_path,
+        model_prefix=model_prefix,
+        vocab_size=vocab_size,
+        model_type=model_type,
+
+        pad_piece="<pad>",
+        bos_piece="<sos>",
+        eos_piece="<eos>",
+        unk_piece="<unk>",
+        
+        pad_id=0,
+        bos_id=1,
+        eos_id=2,
+        unk_id=3
+    )
+
+    print("tokenizer training done")
+
+if __name__ == "__main__":
+    train_sentencepiece(
+        json_path="/workspace/data/captioning/annotations/train.json",
+        model_prefix="/workspace/src/dataset/sub_tokenizer2000",
+        vocab_size=params["captioning"]["tokenizer"]["sp_vocab_size"],
+        model_type="unigram"
+    )

src/debug/test_forward.py

@@ -0,0 +1,208 @@
+import sys
+sys.path.append("/workspace/src/models")
+import torch
+
+
+# model imports
+from lstm import DecoderLSTM
+from gru import DecoderGRU
+from transformer import DecoderTransformer
+# from transformer_scratch import DecoderTransformer
+
+from resnet18 import EncoderResnet18
+from efficientnet import EncoderEfficientNetB0
+from convnext import EncoderConvNextTiny
+from mobilenet import EncoderMobileNetV3Small
+from vit import EncoderViTB16
+from swin import EncoderSwinTiny
+from deit import EncoderDeiTTiny
+
+
+
+# device
+device = torch.device(
+    "cuda" if torch.cuda.is_available() else "cpu"
+)
+
+print(f"device: {device}")
+
+
+
+# caption model dummy input
+feature = torch.randn(1, 512).to(device)
+# feature = torch.randn(1, 49, 512).to(device)
+
+caption = torch.tensor(
+    [[0, 1, 2, 3, 4]]
+).to(device)
+
+
+
+### LSTM Forward ###
+lstm_model = DecoderLSTM().to(device)
+
+lstm_out = lstm_model(
+    feature,
+    caption
+)
+
+print(f"LSTM: {lstm_out.shape}")
+
+
+
+### GRU Forward ###
+gru_model = DecoderGRU().to(device)
+
+gru_out = gru_model(
+    feature,
+    caption
+)
+
+print(f"GRU: {gru_out.shape}")
+
+
+
+### Transformer Forward ###
+transformer_model = DecoderTransformer().to(device)
+
+transformer_out, map, map = transformer_model(
+    caption,
+    feature,
+    0
+)
+
+print(f"Transformer: {transformer_out.shape}")
+
+
+
+### ResNet18 Forward ###
+NUM_CLASSES = 50
+
+resnet18_model = EncoderResnet18(
+    num_classes=NUM_CLASSES
+).to(device)
+
+dummy_images = torch.randn(
+    8, 3, 224, 224
+).to(device)
+
+logits, features = resnet18_model(
+    dummy_images
+)
+
+print(f"ResNet18 logits: {logits.shape}")
+print(f"ResNet18 features: {features.shape}")
+
+
+
+### EfficientNet-B0 Forward ###
+efficientnet_model = EncoderEfficientNetB0(
+    num_classes=NUM_CLASSES
+).to(device)
+
+efficientnet_out = efficientnet_model(
+    dummy_images
+)
+
+print(
+    f"EfficientNet-B0: "
+    f"{efficientnet_out.shape}"
+)
+
+# expected:
+# torch.Size([8, 50])
+
+
+
+### ConvNeXt-Tiny Forward ###
+convnext_model = EncoderConvNextTiny(
+    num_classes=NUM_CLASSES
+).to(device)
+
+convnext_out = convnext_model(
+    dummy_images
+)
+
+print(
+    f"ConvNeXt-Tiny: "
+    f"{convnext_out.shape}"
+)
+
+# expected:
+# torch.Size([8, 50])
+
+
+
+### MobileNetV3 Small Forward ###
+mobilenet_model = EncoderMobileNetV3Small(
+    num_classes=NUM_CLASSES
+).to(device)
+
+mobilenet_out = mobilenet_model(
+    dummy_images
+)
+
+print(
+    f"MobileNetV3 Small: "
+    f"{mobilenet_out.shape}"
+)
+
+# expected:
+# torch.Size([8, 50])
+
+
+
+### ViT-B/16 Forward ###
+vit_model = EncoderViTB16(
+    num_classes=NUM_CLASSES
+).to(device)
+
+vit_out = vit_model(
+    dummy_images
+)
+
+print(
+    f"ViT-B/16: "
+    f"{vit_out.shape}"
+)
+
+# expected:
+# torch.Size([8, 50])
+
+
+
+### Swin-T Forward ###
+swin_model = EncoderSwinTiny(
+    num_classes=NUM_CLASSES
+).to(device)
+
+swin_out = swin_model(
+    dummy_images
+)
+
+print(
+    f"Swin-T: "
+    f"{swin_out.shape}"
+)
+
+# expected:
+# torch.Size([8, 50])
+
+
+
+### DeiT-Tiny Forward ###
+deit_model = EncoderDeiTTiny(
+    num_classes=NUM_CLASSES
+).to(device)
+
+deit_out = deit_model(
+    dummy_images
+)
+
+print(
+    f"DeiT-Tiny: "
+    f"{deit_out.shape}"
+)
+
+# expected:
+# torch.Size([8, 50])

src/engines/captioning_trainer.py

@@ -0,0 +1,43 @@
+import torch
+
+
+def train_one_epoch(
+    encoder,
+    decoder,
+    loader,
+    criterion,
+    optimizer,
+    device,
+    scheduler=None
+    ):
+
+    encoder.train()
+    decoder.train()
+
+    total_loss = 0
+    for images, captions in loader:
+        images = images.to(device)
+        captions = captions.to(device)
+
+        feature = encoder(images, return_features=True)
+
+        input_caption = captions[:, :-1]
+        target_caption = captions[:, 1:]
+
+        outputs = decoder(feature, input_caption)
+
+        loss = criterion(
+            outputs.reshape(-1, outputs.shape[-1]),
+            target_caption.reshape(-1)
+        )
+
+        if scheduler is not None:
+            scheduler.step()
+            
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+
+        total_loss += loss.item()
+
+    return total_loss / len(loader)

src/engines/captioning_validator.py

@@ -0,0 +1,36 @@
+import torch
+
+
+def validation_one_epoch(
+    encoder,
+    decoder,
+    loader,
+    criterion,
+    device,
+):
+
+    encoder.eval()
+    decoder.eval()
+
+    with torch.no_grad():
+        total_loss = 0
+        for images, captions, _, __ in loader:
+
+            images = images.to(device) # B, 3, 224, 224
+            captions = captions.to(device) # B, seq_len
+
+            feature = encoder(images, return_features=True) # B, 49, 512
+            
+            input_caption = captions[:, :-1] # B, seq_len-1 
+            target_caption = captions[:, 1:] # B, seq_len-1
+
+            outputs = decoder(feature, input_caption) # B, seq_len-1, voca_size
+
+            loss = criterion(
+                outputs.reshape(-1, outputs.shape[-1]), # B*(seq_len-1), voca_size
+                target_caption.reshape(-1) # B*seq_len-1
+            )
+
+            total_loss += loss.item()
+
+    return total_loss / len(loader)

src/engines/classification_trainer.py

@@ -0,0 +1,70 @@
+from torchmetrics.classification import (
+    MulticlassAccuracy
+)
+
+from transforms.mixup import mixup_data
+from transforms.cutmix import cutmix_data
+
+
+def train_one_epoch(
+    model,
+    loader,
+    criterion,
+    optimizer,
+    device,
+    num_classes,
+    augmentation=None
+):
+
+    model.train()
+
+    metric = MulticlassAccuracy(
+        num_classes=num_classes
+    ).to(device)
+
+    total_loss = 0
+
+    for images, labels in loader:
+        images = images.to(device)
+        labels = labels.to(device)
+
+        if augmentation == "mixup":
+            images, labels_a, labels_b, lam = mixup_data(
+                images,
+                labels
+            )
+
+        elif augmentation == "cutmix":
+            images, labels_a, labels_b, lam = cutmix_data(
+                images,
+                labels
+            )
+
+        outputs = model(images)
+
+        if augmentation in ["mixup", "cutmix"]:
+            loss = (
+                lam * criterion(outputs, labels_a)
+                + (1 - lam) * criterion(outputs, labels_b)
+            )
+
+        else:
+            loss = criterion(
+                outputs,
+                labels
+            )
+
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        total_loss += loss.item()
+        preds = outputs.argmax(dim=1)
+
+        metric.update(
+            preds,
+            labels
+        )
+
+    acc = metric.compute().item()
+
+    return total_loss / len(loader), acc

src/engines/classification_validator.py

@@ -0,0 +1,90 @@
+import torch
+
+from torchmetrics.classification import (
+    MulticlassAccuracy,
+    MulticlassF1Score,
+
+    # precision / recall
+    MulticlassPrecision,
+    MulticlassRecall
+)
+
+
+def validation_one_epoch(
+    model,
+    loader,
+    criterion,
+    device,
+    num_classes
+):
+
+    model.eval()
+
+    acc_metric = MulticlassAccuracy(
+        num_classes=num_classes
+    ).to(device)
+
+    f1_metric = MulticlassF1Score(
+        num_classes=num_classes,
+        average="macro"
+    ).to(device)
+
+    # precision metric
+    # precision_metric = MulticlassPrecision(
+    #     num_classes=num_classes,
+    #     average="macro"
+    # ).to(device)
+
+    # recall metric
+    # recall_metric = MulticlassRecall(
+    #     num_classes=num_classes,
+    #     average="macro"
+    # ).to(device)
+
+    total_loss = 0
+
+    with torch.no_grad():
+
+        for images, labels in loader:
+            images = images.to(device)
+            labels = labels.to(device)
+            outputs = model(images)
+            loss = criterion(
+                outputs,
+                labels
+            )
+            total_loss += loss.item()
+            preds = outputs.argmax(dim=1)
+
+            acc_metric.update(
+                preds,
+                labels
+            )
+
+            f1_metric.update(
+                preds,
+                labels
+            )
+
+            # precision_metric.update(
+            #     preds,
+            #     labels
+            # )
+
+            # recall_metric.update(
+            #     preds,
+            #     labels
+            # )
+
+    acc = acc_metric.compute().item()
+    f1 = f1_metric.compute().item()
+    # precision = precision_metric.compute().item()
+    # recall = recall_metric.compute().item()
+
+    return (
+        total_loss / len(loader),
+        acc,
+        f1,
+        # precision,
+        # recall
+    )

src/metrics/.ipynb_checkpoints/evaluate_caption-checkpoint.py

@@ -0,0 +1,35 @@
+import torch
+from pycocoevalcap.bleu.bleu import Bleu
+from pycocoevalcap.cider.cider import Cider
+
+def evaluate_caption(
+    all_generated_sentence,
+    all_references    
+    ):
+
+    references_dict = {i:list(sentences) for i, sentences in enumerate(all_references)}
+    generated_dict = {i:[sentence] for i, sentence in enumerate(all_generated_sentence)}
+
+    bleu_scorer = Bleu(4)
+    bleu_score, _ = bleu_scorer.compute_score(
+        references_dict,
+        generated_dict
+    )
+
+    cider_scorer = Cider()
+    cider_score, _ = cider_scorer.compute_score(
+        references_dict,
+        generated_dict
+    )
+
+    metric_result = {
+        "bleu1": bleu_score[0],
+        "bleu2": bleu_score[1],
+        "bleu3": bleu_score[2],
+        "bleu4": bleu_score[3],
+        "cider": cider_score,
+        "generated": generated_dict,
+        "references": references_dict
+    }
+
+    return metric_result

src/metrics/.ipynb_checkpoints/make_show_all_caption-checkpoint.py

@@ -0,0 +1,105 @@
+###### best val loss 지점에서 모든 생성 캡션 출력 및 반환, heatmap 저장 #####
+import torch
+import random
+from utils.checkpoint_manager import load_checkpoint
+
+def make_show_all_caption(
+        loader,
+        encoder,
+        decoder,
+        optimizer,
+        w2i,
+        i2w,
+        best_path,
+        dec_atten_dir,
+        enc_dec_atten_dir,
+        SEED,
+        heatmap_sample,
+        layer,
+        device,
+        use_subword,
+        sp_model_path
+):
+    
+    _, best_val_loss = load_checkpoint(
+        best_path,
+        encoder,
+        decoder,
+        optimizer,
+        device
+    )
+
+    all_references = []
+    all_generated_token = []
+    all_dec_atten = []
+    all_enc_dec_atten = []
+    all_images = []
+    all_file_name = []
+    for images, _, batch_references, file_name in loader:
+        images = images.to(device)
+
+        features = encoder(images, return_features=True)
+
+        generated_token, dec_atten, enc_dec_atten = decoder.generate(
+                features, # B, 49, 512
+                torch.full((features.size(0),), w2i["<sos>"], device=device), # B,
+                w2i["<eos>"],
+            )
+        all_dec_atten.extend(dec_atten) # all_B, layers, nhead, seq_len, seq_len
+        all_enc_dec_atten.extend(enc_dec_atten) # all_B, layers, nhead, seq_len, 49
+        all_images.extend(images.cpu())
+        all_references.extend(list(zip(*batch_references)))
+        all_generated_token.extend(generated_token) # all_B, seq_len-1
+        all_file_name.extend(file_name)
+
+    
+    all_generated_sentence = []
+    for sentence_token in all_generated_token:
+        if w2i["<eos>"] in sentence_token:
+            end_inx = sentence_token.index(w2i["<eos>"])
+            sentence_token = sentence_token[:end_inx]
+        
+        # ==================================
+        # SentencePiece tokenizer
+        # ==================================
+        if use_subword:
+            import sentencepiece as spm
+
+            sp = spm.SentencePieceProcessor()
+            sp.load(sp_model_path)
+            # special token 제거
+            sentence_token = [token for token in sentence_token
+                if token not in [
+                    w2i["<pad>"],
+                    w2i["<sos>"],
+                    w2i["<eos>"]
+                ]
+            ]
+            sentence = sp.decode(sentence_token)
+
+        else:
+            words = [i2w[i] for i in sentence_token]
+            sentence = ' '.join(words)
+        
+        all_generated_sentence.append(sentence) # all_B, 1(문장)
+
+
+    decoder.show_dec_atten(all_dec_atten[heatmap_sample], all_generated_sentence[heatmap_sample].split(), layer, dec_atten_dir)
+    decoder.show_cross_atten(all_enc_dec_atten[heatmap_sample], all_generated_sentence[heatmap_sample].split(), layer, all_images[heatmap_sample], enc_dec_atten_dir)
+
+    random.seed(SEED)
+    all_B = len(all_generated_sentence)
+    sample = random.sample(range(0, all_B), 5)
+    for i in sample:
+        print("-" * 60)
+        print(f' {all_file_name[i]}: {all_generated_sentence[i]}')
+        print("-" * 60)
+
+        for inx, reference in enumerate(all_references[i], start=1):
+            print(f'Reference {inx}: {reference}')
+        print("=" * 60)
+
+    print(f'Best Val Loss: {best_val_loss}')
+
+    return all_generated_sentence, all_references
+