f

data/mimic_cxr_resized_builder.py

@@ -193,6 +193,7 @@ def build_mimic_cxr_resized_instruct_json(
     pnu_by_study: Dict[Tuple[str, str], str] = {}  # (subj, study) → PNU string (for VQA reuse)
 
     n_studies = n_missing_report = 0
+    n_missing_image = n_studies_no_image = 0
     skipped_merged_no_impression = skipped_cascade_no_findings = 0
 
     for fname, split_label in _MANIFEST_FILES:
@@ -210,10 +211,22 @@ def build_mimic_cxr_resized_instruct_json(
 
             # All views of the same study share report + CheXpert labels.
             first = rows[0]
-            rels  = [r["image_relpath"] for r in rows]
+            all_rels = [r["image_relpath"] for r in rows]
+
+            # Filter out rows whose image file is missing on disk (resize/shard
+            # can drop a few JPGs that failed to encode; manifest still lists
+            # them). Keep VQA index consistent with the SURVIVING images only.
+            rels = []
+            for r in all_rels:
+                if (root / r).is_file():
+                    rels.append(r)
+                else:
+                    n_missing_image += 1
+            if not rels:
+                n_studies_no_image += 1
+                continue
 
-            # Index every image (incl. studies with no report yet) so a VQA
-            # row that references this image can still be picked up below.
+            # Index every (surviving) image for the VQA lookup pass below.
             for r in rels:
                 image_index[r] = r
 
@@ -330,6 +343,8 @@ def build_mimic_cxr_resized_instruct_json(
     print(f"  image_mode        : {image_mode}")
     print(f"  studies indexed   : {n_studies:,}")
     print(f"  missing report    : {n_missing_report:,}")
+    print(f"  missing image rows: {n_missing_image:,}")
+    print(f"  studies w/o image : {n_studies_no_image:,}  (dropped entirely)")
     if report_mode == "merged":
         print(f"  skipped no_impr   : {skipped_merged_no_impression:,}")
     if report_mode == "split_cascade":

scripts/cxrvlm_colab_train.ipynb

@@ -68,7 +68,7 @@
     "REPORT_MODE = 'split_cascade'        # 'split' | 'merged' | 'split_cascade'\n",
     "IMAGE_MODE  = 'all_views_split'  # 'all_views_split' | 'frontal_only_split' | 'multi_image_merged'\n",
     "S1_EPOCHS   = 2\n",
-    "S2_EPOCHS   = 5\n",
+    "S2_EPOCHS   = 7\n",
     "MODE            = 'fresh'        # 'fresh' | 'resume'\n",
     "EXPLICIT_RUN_ID = None           # None  | 'IU-Xray_run_5'   (only matters when MODE='resume')\n",
     "\n",
@@ -992,7 +992,7 @@
    "outputs": [],
    "source": [
     "# ── (Optional) Stage-1 ITC alignment toggle ─────────────────────────\n",
-    "USE_ITC         = False   # True → contrastive Stage-1 (no Vicuna); False → original\n",
+    "USE_ITC         = True   # True → contrastive Stage-1 (no Vicuna); False → original\n",
     "ITC_REPO_ID     = 'hieu3636/cxr-vlm-data'\n",
     "ITC_SUBDIR      = 'cxr_bert_text_embeddings'\n",
     "ITC_TEMPERATURE = 0.07\n",
@@ -1072,7 +1072,7 @@
     "# unique images that's ~90 GB — make sure WORK has the room, or set\n",
     "# CACHE_FEATURES=False on tight quotas.\n",
     "\n",
-    "CACHE_FEATURES = True\n",
+    "CACHE_FEATURES = False\n",
     "\n",
     "if CACHE_FEATURES:\n",
     "    feature_cache_dir = WORK / 'feature_cache' / DATASET_NAME\n",
@@ -1723,4 +1723,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 5
-}
+}