Merge pull request #4 from ThuanNaN/create_dataset

normalize bbox and update train cgf

create_dataset.ipynb

@@ -2,7 +2,7 @@
  "cells": [
   {
    "cell_type": "code",
-   "execution_count": 48,
+   "execution_count": 6,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -15,7 +15,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 49,
+   "execution_count": 7,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -30,7 +30,25 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 50,
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def normalize_bbox(bbox, width, height):\n",
+    "    \"\"\"\n",
+    "    Normalize bounding box to the range [0, 100].\n",
+    "    \"\"\"\n",
+    "    return [\n",
+    "        int((bbox[0] / width) * 100),  # x_min\n",
+    "        int((bbox[1] / height) * 100),  # y_min\n",
+    "        int((bbox[2] / width) * 100),  # x_max\n",
+    "        int((bbox[3] / height) * 100),  # y_max\n",
+    "    ]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
    "metadata": {},
    "outputs": [
     {
@@ -144,7 +162,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 51,
+   "execution_count": 10,
    "metadata": {},
    "outputs": [
     {
@@ -153,7 +171,7 @@
        "5354"
       ]
      },
-     "execution_count": 51,
+     "execution_count": 10,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -164,7 +182,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 52,
+   "execution_count": 11,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -178,17 +196,22 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 53,
+   "execution_count": 12,
    "metadata": {},
    "outputs": [],
    "source": [
     "for data in mvtech_ad_data_for_regression:\n",
-    "    data[\"image_path\"] = data[\"image_path\"].replace(\"MVTEC_det\", \".\")"
+    "    data[\"image_path\"] = data[\"image_path\"].replace(\"MVTEC_det\", \".\")\n",
+    "    data[\"bbox\"] = normalize_bbox(\n",
+    "        bbox=data[\"bbox\"],\n",
+    "        width=data[\"width\"],\n",
+    "        height=data[\"height\"]\n",
+    "    )"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 54,
+   "execution_count": 13,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -198,7 +221,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 55,
+   "execution_count": 14,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -213,7 +236,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 56,
+   "execution_count": 15,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -240,7 +263,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 57,
+   "execution_count": 16,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -279,7 +302,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 58,
+   "execution_count": 17,
    "metadata": {},
    "outputs": [
     {
@@ -288,7 +311,7 @@
        "Text(0.5, 0, 'Good/Defect')"
       ]
      },
-     "execution_count": 58,
+     "execution_count": 17,
      "metadata": {},
      "output_type": "execute_result"
     },

minigpt4/configs/datasets/mvtec/default.yaml

@@ -3,4 +3,4 @@ datasets:
     data_type: images
     build_info:
       image_path: ./MVTEC_det/images
-      ann_path: ./MVTEC_det/mvtech_ad_data_for_regression.json
+      ann_path: ./MVTEC_det/train_data.json

minigpt4/datasets/datasets/mvtec_dataset.py

@@ -29,15 +29,17 @@ class MVTecDataset(Dataset):
     def __getitem__(self, index):
         info = self.ann[index]
         gt_bbox = info["bbox"]
+        ans_cls = info["class"]
 
         image_path = os.path.join(self.vis_root, info['image_path'])
         image = Image.open(image_path).convert("RGB")
         image = self.vis_processor(image)
 
-        input = "detect defect or non-defect and return the bounding box"
+        input = "a defect or not-defect object and return the bounding boxes and its label. If not, bound around the object."
         
-        ans_cls = "defect" if info["is_broken"] == True else "non-defect"
+        ans_defect = "defect" if info["is_broken"] == True else "non-defect"
-        answer = f"{ans_cls}<{gt_bbox[0]}><{gt_bbox[1]}><{gt_bbox[2]}><{gt_bbox[3]}>"
+        ans_para = f"<p>{ans_cls}-{ans_defect}</p>"
+        answer = f"{ans_para}{{<{gt_bbox[0]}><{gt_bbox[1]}><{gt_bbox[2]}><{gt_bbox[3]}>}}"
 
         instruction = random.choice(self.instruction_pool).format(input)
         instruction = "<Img><ImageHere></Img> {} ".format(instruction)

train_configs/minigptv2_finetune_mvtec.yaml

@@ -32,10 +32,10 @@ run:
   warmup_lr: 1e-6
 
   weight_decay: 0.05
-  max_epoch: 10
+  max_epoch: 5
   num_workers: 6
   warmup_steps: 1000
-  iters_per_epoch: 1000
+  iters_per_epoch: 2000
 
   seed: 42
   output_dir: "mvtec_outputs"
@@ -51,5 +51,5 @@ run:
   dist_url: "env://"
   distributed: True
 
-  wandb_log: True
+  wandb_log: False
   job_name: minigptv2_finetune