Add files using upload-large-folder tool

download_data.py

@@ -0,0 +1,113 @@
+"""Download multiple datasets from Hugging Face and COCO."""
+
+import argparse
+import os
+import sys
+from huggingface_hub import snapshot_download
+
+
+COCO_VAL_URL = "http://images.cocodataset.org/zips/val2014.zip"
+
+
+def download_huggingface_dataset(repo_id: str, output_dir: str, use_auth_token: bool) -> str:
+    os.makedirs(output_dir, exist_ok=True)
+    print(f"Downloading Hugging Face dataset {repo_id} to '{output_dir}'...")
+    local_dir = snapshot_download(
+        repo_id=repo_id,
+        repo_type="dataset",
+        local_dir=output_dir,
+        use_auth_token=use_auth_token,
+    )
+    print(f"Finished: {repo_id} -> {local_dir}\n")
+    return local_dir
+
+
+def download_file(url: str, output_path: str) -> None:
+    try:
+        import requests
+    except ImportError:
+        import urllib.request
+
+        print("requests not found, falling back to urllib.request")
+        urllib.request.urlretrieve(url, output_path)
+        return
+
+    print(f"Downloading file from {url} to '{output_path}'...")
+    with requests.get(url, stream=True) as response:
+        response.raise_for_status()
+        total_size = int(response.headers.get("content-length", 0))
+        downloaded = 0
+        chunk_size = 8192
+        with open(output_path, "wb") as f:
+            for chunk in response.iter_content(chunk_size=chunk_size):
+                if chunk:
+                    f.write(chunk)
+                    downloaded += len(chunk)
+                    if total_size:
+                        percent = downloaded * 100 / total_size
+                        print(f"\rDownloaded {downloaded}/{total_size} bytes ({percent:.1f}%)", end="")
+    if total_size:
+        print()
+    print(f"Finished: {output_path}\n")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Download VGMED, Medical Multimodal Evaluation Data, and COCO 2014 val images.")
+    parser.add_argument(
+        "--output-dir",
+        default="/home/jinhua/ExplainableVQA/benchmarks",
+        help="Base output directory for all downloads. Default: downloads",
+    )
+    parser.add_argument(
+        "--use-auth-token",
+        action="store_true",
+        help="Use HF authentication token from HUGGINGFACE_HUB_TOKEN environment variable if required.",
+    )
+    parser.add_argument(
+        "--skip-vgmed",
+        action="store_true",
+        help="Skip downloading guimeng-liu/VGMED.",
+    )
+    parser.add_argument(
+        "--skip-med-eval",
+        action="store_true",
+        help="Skip downloading FreedomIntelligence/Medical_Multimodal_Evaluation_Data.",
+    )
+    parser.add_argument(
+        "--skip-coco",
+        action="store_true",
+        help="Skip downloading COCO 2014 validation images.",
+    )
+    parser.add_argument(
+        "--coco-url",
+        default=COCO_VAL_URL,
+        help="COCO 2014 val images download URL. Default is the official val2014 zip URL.",
+    )
+    args = parser.parse_args()
+
+    base_dir = os.path.abspath(args.output_dir)
+    os.makedirs(base_dir, exist_ok=True)
+
+    if not args.skip_vgmed:
+        download_huggingface_dataset(
+            repo_id="guimeng-liu/VGMED",
+            output_dir=os.path.join(base_dir, "VGMED"),
+            use_auth_token=args.use_auth_token,
+        )
+
+    if not args.skip_med_eval:
+        download_huggingface_dataset(
+            repo_id="FreedomIntelligence/Medical_Multimodal_Evaluation_Data",
+            output_dir=os.path.join(base_dir, "Medical_Multimodal_Evaluation_Data"),
+            use_auth_token=args.use_auth_token,
+        )
+
+    if not args.skip_coco:
+        coco_target = os.path.join(base_dir, "val2014.zip")
+        download_file(args.coco_url, coco_target)
+
+    print("All requested downloads completed.")
+
+
+if __name__ == "__main__":
+    main()

download_sam3_assets.py

@@ -0,0 +1,89 @@
+"""
+Download the external assets needed by MedSAM3 that are not part of the
+git repo itself (sam3.pt is too large to commit; bpe_simple_vocab_16e6.txt.gz
+ships with OpenAI's CLIP repo, not this one; the LoRA weights are the
+project's own released checkpoint on HuggingFace):
+
+  - facebook/sam3 model weights -> models/sam3.pt              (via ModelScope, ~3.45GB)
+  - CLIP BPE vocab               -> models/bpe_simple_vocab_16e6.txt.gz
+  - MedSAM3-v1 LoRA weights      -> models/best_lora_weights.pt (huggingface.co/lal-Joey/MedSAM3_v1)
+
+Usage:
+    python download_sam3_assets.py
+"""
+
+import os
+import shutil
+import urllib.request
+
+MODELS_DIR = "/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/models"
+
+BPE_VOCAB_URL = "https://raw.githubusercontent.com/facebookresearch/sam3/main/sam3/assets/bpe_simple_vocab_16e6.txt.gz"
+BPE_VOCAB_PATH = os.path.join(MODELS_DIR, "bpe_simple_vocab_16e6.txt.gz")
+
+SAM3_WEIGHTS_PATH = os.path.join(MODELS_DIR, "sam3.pt")
+
+LORA_REPO_ID = "lal-Joey/MedSAM3_v1"
+LORA_WEIGHTS_PATH = os.path.join(MODELS_DIR, "best_lora_weights.pt")
+
+
+def download_bpe_vocab():
+    if os.path.exists(BPE_VOCAB_PATH):
+        print(f"[Skip] {BPE_VOCAB_PATH} already exists.")
+        return
+    print(f"==> Downloading CLIP BPE vocab from {BPE_VOCAB_URL}")
+    os.makedirs(MODELS_DIR, exist_ok=True)
+    urllib.request.urlretrieve(BPE_VOCAB_URL, BPE_VOCAB_PATH)
+    print(f"[Done] Saved to {BPE_VOCAB_PATH}")
+
+
+def download_sam3_weights():
+    if os.path.exists(SAM3_WEIGHTS_PATH):
+        print(f"[Skip] {SAM3_WEIGHTS_PATH} already exists.")
+        return
+    from modelscope import snapshot_download
+
+    print("==> Downloading facebook/sam3 weights via ModelScope (~3.45GB)")
+    model_dir = snapshot_download("facebook/sam3")
+    print(f"    Downloaded to {model_dir}")
+
+    src = os.path.join(model_dir, "sam3.pt")
+    if not os.path.exists(src):
+        candidates = [f for f in os.listdir(model_dir) if f.endswith(".pt")]
+        if not candidates:
+            raise FileNotFoundError(f"No .pt file found in {model_dir}")
+        src = os.path.join(model_dir, candidates[0])
+
+    os.makedirs(MODELS_DIR, exist_ok=True)
+    shutil.copy(src, SAM3_WEIGHTS_PATH)
+    print(f"[Done] Copied to {SAM3_WEIGHTS_PATH}")
+
+
+def download_lora_weights():
+    if os.path.exists(LORA_WEIGHTS_PATH):
+        print(f"[Skip] {LORA_WEIGHTS_PATH} already exists.")
+        return
+    from huggingface_hub import snapshot_download
+
+    print(f"==> Downloading LoRA weights from huggingface.co/{LORA_REPO_ID}")
+    repo_dir = snapshot_download(repo_id=LORA_REPO_ID)
+    print(f"    Downloaded to {repo_dir}")
+
+    candidates = [f for f in os.listdir(repo_dir) if f.endswith(".pt")]
+    if not candidates:
+        raise FileNotFoundError(f"No .pt file found in {repo_dir}")
+    src = os.path.join(repo_dir, candidates[0])
+
+    os.makedirs(MODELS_DIR, exist_ok=True)
+    shutil.copy(src, LORA_WEIGHTS_PATH)
+    print(f"[Done] Copied to {LORA_WEIGHTS_PATH}")
+
+
+def main():
+    download_bpe_vocab()
+    download_sam3_weights()
+    download_lora_weights()
+
+
+if __name__ == "__main__":
+    main()

gpu1_hold_30g.log

@@ -0,0 +1,2 @@
+nohup: ignoring input
+GPU 1: holding 30 GiB for 1 hour

predownload_models-ori.py

@@ -0,0 +1,99 @@
+"""
+Pre-download all HuggingFace model repos used by:
+  - MedEvalKit/eval_scripts/grounding/diff_region/*.sh (grounding diff-region analysis)
+
+Downloads go to a custom HuggingFace cache directory
+(/home/jinhua/ExplainableVQA/checkpoint/models) instead of the default
+~/.cache/huggingface. As long as HF_HOME points to this same directory at
+load time, `from_pretrained(repo_id, ...)` will transparently hit this cache
+instead of re-downloading - it behaves exactly like the normal HF cache,
+just relocated.
+
+Each repo is retried several times with resume (snapshot_download skips files
+already fully downloaded and resumes partial ones), since mirrors like
+hf-mirror.com often time out near the end of large files.
+
+Usage:
+    python predownload_models_to_checkpoint.py
+"""
+
+import os
+import time
+
+# Custom cache location - from_pretrained(repo_id) will read from here as
+# long as HF_HOME is set to the same path when the model is actually loaded.
+#os.environ.setdefault("HF_HOME", "/home/jinhua/ExplainableVQA/checkpoint/models")
+
+# Use the hf-mirror.com mirror (much faster from China than huggingface.co directly).
+#os.environ.setdefault("HF_ENDPOINT", "https://hf-mirror.com")
+
+# Raise the default 10s timeout - hf-mirror.com often stalls on the final
+# etag/HEAD check for large files.
+os.environ.setdefault("HF_HUB_DOWNLOAD_TIMEOUT", "60")
+
+# Disable the Xet backend: it generates a fresh random temp file on every
+# resume instead of continuing the previous .incomplete file, so interrupted
+# downloads restart from 0 instead of resuming.
+os.environ.setdefault("HF_HUB_DISABLE_XET", "1")
+
+from huggingface_hub import snapshot_download
+
+REPO_IDS = [
+    # ---- MedEvalKit/eval_scripts/grounding/diff_region/*.sh ----
+    # "Qwen/Qwen2.5-VL-3B-Instruct",                 # Qwen2.5VL-3B
+    # "Qwen/Qwen2.5-VL-7B-Instruct",                 # Qwen2.5VL-7B
+    # "Qwen/Qwen2.5-VL-32B-Instruct",                # Qwen2.5VL-32B
+    "llava-hf/llava-1.5-7b-hf",                    # LLaVA-7B
+    "llava-hf/llava-1.5-13b-hf",                   # LLaVA-13B
+    # "microsoft/llava-med-v1.5-mistral-7b",         # LLaVA-Med-7B
+    # "FreedomIntelligence/HuatuoGPT-Vision-7B",     # HuatuoGPT-V-7B
+    # "FreedomIntelligence/HuatuoGPT-Vision-34B",    # HuatuoGPT-V-34B
+    # "google/medgemma-4b-it",                       # MedGemma-4B-IT
+    # "google/medgemma-27b-it",                      # MedGemma-27B-IT (gated repo, multimodal variant)
+]
+
+DATASET_REPO_IDS = []
+
+MAX_RETRIES = 10
+RETRY_DELAY_SECONDS = 5
+
+
+def download_with_retries(repo_id, repo_type="model"):
+    for attempt in range(1, MAX_RETRIES + 1):
+        try:
+            snapshot_download(repo_id=repo_id, repo_type=repo_type, resume_download=True)
+            return True
+        except Exception as e:
+            print(f"    attempt {attempt}/{MAX_RETRIES} failed: {e}")
+            if attempt < MAX_RETRIES:
+                time.sleep(RETRY_DELAY_SECONDS)
+    return False
+
+
+def main():
+    #print(f"HF_HOME = {os.environ['HF_HOME']}")
+
+    failed = []
+    for repo_id in REPO_IDS:
+        print(f"==> Downloading {repo_id}")
+        if not download_with_retries(repo_id):
+            failed.append(repo_id)
+            print(f"    GAVE UP: {repo_id}")
+
+    for repo_id in DATASET_REPO_IDS:
+        print(f"==> Downloading dataset {repo_id}")
+        if not download_with_retries(repo_id, repo_type="dataset"):
+            failed.append(repo_id)
+            print(f"    GAVE UP: {repo_id}")
+
+    print("\n==== Summary ====")
+    if failed:
+        print("Failed repos (re-run this script to resume them):")
+        for repo_id in failed:
+            print(f"  - {repo_id}")
+    else:
+        print("All repos downloaded successfully.")
+
+
+if __name__ == "__main__":
+    main()

predownload_models.py

@@ -0,0 +1,102 @@
+"""
+Pre-download all HuggingFace model repos used by:
+  - MedEvalKit/eval_scripts/grounding/diff_region/*.sh (grounding diff-region analysis)
+
+Downloads go to a custom HuggingFace cache directory
+(/home/jinhua/ExplainableVQA/checkpoint/models) instead of the default
+~/.cache/huggingface. As long as HF_HOME points to this same directory at
+load time, `from_pretrained(repo_id, ...)` will transparently hit this cache
+instead of re-downloading - it behaves exactly like the normal HF cache,
+just relocated.
+
+Each repo is retried several times with resume (snapshot_download skips files
+already fully downloaded and resumes partial ones), since mirrors like
+hf-mirror.com often time out near the end of large files.
+
+Usage:
+    python predownload_models_to_checkpoint.py
+"""
+
+import os
+import time
+
+# Custom cache location - from_pretrained(repo_id) will read from here as
+# long as HF_HOME is set to the same path when the model is actually loaded.
+#os.environ.setdefault("HF_HOME", "/home/jinhua/ExplainableVQA/checkpoint/models")
+
+# Use the hf-mirror.com mirror (much faster from China than huggingface.co directly).
+#os.environ.setdefault("HF_ENDPOINT", "https://hf-mirror.com")
+
+# Raise the default 10s timeout - hf-mirror.com often stalls on the final
+# etag/HEAD check for large files.
+os.environ.setdefault("HF_HUB_DOWNLOAD_TIMEOUT", "60")
+
+# Disable the Xet backend: it generates a fresh random temp file on every
+# resume instead of continuing the previous .incomplete file, so interrupted
+# downloads restart from 0 instead of resuming.
+os.environ.setdefault("HF_HUB_DISABLE_XET", "1")
+
+from huggingface_hub import snapshot_download
+
+REPO_IDS = [
+    # ---- MedEvalKit/eval_scripts/grounding/diff_region/*.sh ----
+    # "Qwen/Qwen2.5-VL-3B-Instruct",                 # Qwen2.5VL-3B
+    # "Qwen/Qwen2.5-VL-7B-Instruct",                 # Qwen2.5VL-7B
+    # "Qwen/Qwen2.5-VL-32B-Instruct",                # Qwen2.5VL-32B
+    # "llava-hf/llava-1.5-7b-hf",                    # LLaVA-7B
+    # "llava-hf/llava-1.5-13b-hf",                   # LLaVA-13B
+    # "microsoft/llava-med-v1.5-mistral-7b",         # LLaVA-Med-7B
+    # "FreedomIntelligence/HuatuoGPT-Vision-7B",     # HuatuoGPT-V-7B
+    # "FreedomIntelligence/HuatuoGPT-Vision-34B",    # HuatuoGPT-V-34B
+    # "google/medgemma-4b-it",                       # MedGemma-4B-IT
+    # "google/medgemma-27b-it",                      # MedGemma-27B-IT (gated repo, multimodal variant)
+    "Qwen/Qwen3-VL-2B-Instruct",                     # Qwen3VL-2B
+    "Qwen/Qwen3-VL-4B-Instruct",                     # Qwen3VL-4B
+    "Qwen/Qwen3-VL-8B-Instruct",                     # Qwen3VL-8B
+]
+
+DATASET_REPO_IDS = []
+
+MAX_RETRIES = 10
+RETRY_DELAY_SECONDS = 5
+
+
+def download_with_retries(repo_id, repo_type="model"):
+    for attempt in range(1, MAX_RETRIES + 1):
+        try:
+            snapshot_download(repo_id=repo_id, repo_type=repo_type, resume_download=True)
+            return True
+        except Exception as e:
+            print(f"    attempt {attempt}/{MAX_RETRIES} failed: {e}")
+            if attempt < MAX_RETRIES:
+                time.sleep(RETRY_DELAY_SECONDS)
+    return False
+
+
+def main():
+    #print(f"HF_HOME = {os.environ['HF_HOME']}")
+
+    failed = []
+    for repo_id in REPO_IDS:
+        print(f"==> Downloading {repo_id}")
+        if not download_with_retries(repo_id):
+            failed.append(repo_id)
+            print(f"    GAVE UP: {repo_id}")
+
+    for repo_id in DATASET_REPO_IDS:
+        print(f"==> Downloading dataset {repo_id}")
+        if not download_with_retries(repo_id, repo_type="dataset"):
+            failed.append(repo_id)
+            print(f"    GAVE UP: {repo_id}")
+
+    print("\n==== Summary ====")
+    if failed:
+        print("Failed repos (re-run this script to resume them):")
+        for repo_id in failed:
+            print(f"  - {repo_id}")
+    else:
+        print("All repos downloaded successfully.")
+
+
+if __name__ == "__main__":
+    main()