FEAT: all notebooks on all datasets run, but with poor results

README.md

@@ -5,12 +5,6 @@ Repo for TWM (Machine Vision Techniques) project @ WUT 24L semester
 [![Imports: isort](https://img.shields.io/badge/%20imports-isort-%231674b1?style=flat&labelColor=ef8336)](https://pycqa.github.io/isort/)
 
 ## TODO
-- [ ] problem konwersji danych RGB -> maska
-    - [ ] tak samo w UAVid - dane są zakodowane, tak, żeby dało się je wyświetlić, a nie do modelu
-    - [x] rozwiązać problem ze sposobem w jaki jest zakodowane gt w Dubai (kolorowe obrazki zamiast po prostu [0...5]) - tu jest chyba jakiś bug, bo w notebooku `example-masks-conversion` wychodzi inaczej niż w `dubai-no-finetune`
-    - [x] czy w AerialDrone używamy tylko maski z jednym kanałem czy kolorów z wieloma?
-    - [x] w jaki sposób, w AerialDrone, jest oznaczane to co trzeba przewidzieć (RGB classes czy to drugie)?
-    - [x] w AerialDrone, jak działa przetworzenie maski na tensor / PIL.Image (tzn. czy nie ma np. jakiegoś rescale, itd.)?
 - [ ] trening na jednym datasecie + test na jednym datasecie
 - [ ] użycie `Crop` lub `Pad` zamiast `Resize` - może będą lepsze wyniki?
 - [ ] upewnienie się, że maski nie zostały (za bardzo) zaburzone - np. bilinear i progowanie niskim progiem (będzie mniejszy latent w UNet)
@@ -34,6 +28,12 @@ Repo for TWM (Machine Vision Techniques) project @ WUT 24L semester
 - [x] zapoznanie z datasetem *Aerial Semantic Segmentation Drone Dataset*
 - [x] doinstalować torcha z CUDA (skill issue xd)
 - [x] dokończenie prezentacji
+- [x] problem konwersji danych RGB -> maska
+    - [x] tak samo w UAVid - dane są zakodowane, tak, żeby dało się je wyświetlić, a nie do modelu
+    - [x] rozwiązać problem ze sposobem w jaki jest zakodowane gt w Dubai (kolorowe obrazki zamiast po prostu [0...5]) - tu jest chyba jakiś bug, bo w notebooku `example-masks-conversion` wychodzi inaczej niż w `dubai-no-finetune`
+    - [x] czy w AerialDrone używamy tylko maski z jednym kanałem czy kolorów z wieloma?
+    - [x] w jaki sposób, w AerialDrone, jest oznaczane to co trzeba przewidzieć (RGB classes czy to drugie)?
+    - [x] w AerialDrone, jak działa przetworzenie maski na tensor / PIL.Image (tzn. czy nie ma np. jakiegoś rescale, itd.)?
 
 ## Intro
 Proponuję pójść w stronę przeglądu / ensemble różnych modeli i/lub datasetów, porównać, itd.
@@ -112,10 +112,10 @@ torch.Size([1, 1, 4000, 6016])
 *sprawdzenie tylko czy się odpalają, tzn. model prawidłowo przetwarza dane, wyniki mogą być (is zazwyczaj są) bardzo słabe na początku*
 | Model      | INRIA | UAVid | Dubai | AerialDrone |  
 | ----------- | ----------- | ----------- | ----------- | ----------- |  
-| UNet      | :heavy_check_mark:       | :heavy_check_mark:   | TBA   |  :heavy_check_mark:   | 
-| UNet++   | :heavy_check_mark:        | :heavy_check_mark:      | TBA      | :heavy_check_mark:   | 
-| DeepLabV3   | :heavy_check_mark:        | :heavy_check_mark:      | TBA      | :heavy_check_mark:   | 
-| DeepLabV3+   | :heavy_check_mark:        | :heavy_check_mark:      | TBA      | :heavy_check_mark:   | 
+| UNet      | :heavy_check_mark:       | :heavy_check_mark:   | :heavy_check_mark:   |  :heavy_check_mark:   | 
+| UNet++   | :heavy_check_mark:        | :heavy_check_mark:      | :heavy_check_mark:      | :heavy_check_mark:   | 
+| DeepLabV3   | :heavy_check_mark:        | :heavy_check_mark:      | :heavy_check_mark:      | :heavy_check_mark:   | 
+| DeepLabV3+   | :heavy_check_mark:        | :heavy_check_mark:      | :heavy_check_mark:      | :heavy_check_mark:   | 
 
 ## Wyniki
 

dubai-no-finetune.ipynb

@@ -136,8 +136,8 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 3, 640, 576])\n",
-      "torch.Size([1, 1, 640, 576])\n"
+      "torch.Size([1, 3, 576, 608])\n",
+      "torch.Size([1, 1, 576, 608])\n"
      ]
     }
    ],
@@ -249,8 +249,8 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 3, 576, 640])\n",
-      "torch.Size([1, 1, 576, 640])\n"
+      "torch.Size([1, 3, 576, 608])\n",
+      "torch.Size([1, 1, 576, 608])\n"
      ]
     },
     {
@@ -321,7 +321,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 576, 640])\n"
+      "torch.Size([1, 576, 608])\n"
      ]
     }
    ],
@@ -338,7 +338,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 576, 640])\n"
+      "torch.Size([1, 576, 608])\n"
      ]
     }
    ],
@@ -355,7 +355,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "{0: 32220, 1: 187, 2: 649, 3: 2808, 4: 5445, 5: 327331}\n"
+      "{0: 329, 1: 19714, 2: 1381, 3: 139, 4: 328433, 5: 212}\n"
      ]
     }
    ],
@@ -373,7 +373,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "{0: 127, 1: 114852, 2: 4966, 3: 591, 4: 247308, 5: 796}\n"
+      "{0: 12172, 1: 220557, 2: 72658, 3: 26781, 4: 15547, 5: 2493}\n"
      ]
     }
    ],
@@ -390,10 +390,10 @@
     {
      "data": {
       "text/plain": [
-       "{'iou': tensor(0.0056),\n",
-       " 'f1': tensor(0.0112),\n",
-       " 'accuracy': tensor(0.6704),\n",
-       " 'recall': tensor(0.0112)}"
+       "{'iou': tensor(0.0395),\n",
+       " 'f1': tensor(0.0760),\n",
+       " 'accuracy': tensor(0.6920),\n",
+       " 'recall': tensor(0.0760)}"
       ]
      },
      "execution_count": 23,
@@ -414,10 +414,10 @@
     {
      "data": {
       "text/plain": [
-       "{'iou': tensor(0.0907),\n",
-       " 'f1': tensor(0.1663),\n",
-       " 'accuracy': tensor(0.7221),\n",
-       " 'recall': tensor(0.1663)}"
+       "{'iou': tensor(0.0909),\n",
+       " 'f1': tensor(0.1667),\n",
+       " 'accuracy': tensor(0.7222),\n",
+       " 'recall': tensor(0.1667)}"
       ]
      },
      "execution_count": 24,
@@ -476,7 +476,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "{'iou': [tensor(0.0153), tensor(0.0739), tensor(0.0127), tensor(0.0089), tensor(0.0278), tensor(0.0393), tensor(0.0026), tensor(0.0387), tensor(0.0409), tensor(0.0056), tensor(0.0561), tensor(0.0366), tensor(0.0317), tensor(0.0123), tensor(0.0083), tensor(0.0042), tensor(0.0191), tensor(0.0063), tensor(0.0550), tensor(0.0150), tensor(0.0024), tensor(0.1132), tensor(0.0035), tensor(0.0411), tensor(0.0291), tensor(0.0022), tensor(0.0036), tensor(0.0146), tensor(0.0112), tensor(0.0072), tensor(0.0172), tensor(0.0024), tensor(0.0030), tensor(0.0477), tensor(0.0046), tensor(0.0019), tensor(0.0098), tensor(0.0656), tensor(0.0063), tensor(0.0204), tensor(0.0150), tensor(0.0052), tensor(0.0212), tensor(0.0126), tensor(0.0089), tensor(0.0088), tensor(0.0094), tensor(0.0100), tensor(0.0069), tensor(0.0020), tensor(0.0045), tensor(0.0452), tensor(0.0056), tensor(0.0680), tensor(0.0005), tensor(0.0034), tensor(0.0026), tensor(0.0313), tensor(0.0157), tensor(0.0175), tensor(0.0058), tensor(0.0056), tensor(0.0033), tensor(0.0474), tensor(0.0248), tensor(0.0006), tensor(0.0205), tensor(0.0323), tensor(0.0063), tensor(0.0296), tensor(0.0209), tensor(0.0018)], 'f1': [tensor(0.0301), tensor(0.1376), tensor(0.0250), tensor(0.0177), tensor(0.0541), tensor(0.0756), tensor(0.0053), tensor(0.0745), tensor(0.0786), tensor(0.0112), tensor(0.1063), tensor(0.0706), tensor(0.0614), tensor(0.0243), tensor(0.0164), tensor(0.0084), tensor(0.0376), tensor(0.0125), tensor(0.1043), tensor(0.0296), tensor(0.0049), tensor(0.2034), tensor(0.0070), tensor(0.0789), tensor(0.0565), tensor(0.0045), tensor(0.0071), tensor(0.0288), tensor(0.0221), tensor(0.0143), tensor(0.0338), tensor(0.0048), tensor(0.0061), tensor(0.0911), tensor(0.0092), tensor(0.0039), tensor(0.0194), tensor(0.1232), tensor(0.0125), tensor(0.0399), tensor(0.0295), tensor(0.0104), tensor(0.0416), tensor(0.0248), tensor(0.0177), tensor(0.0175), tensor(0.0187), tensor(0.0198), tensor(0.0137), tensor(0.0039), tensor(0.0089), tensor(0.0864), tensor(0.0112), tensor(0.1273), tensor(0.0010), tensor(0.0068), tensor(0.0052), tensor(0.0608), tensor(0.0310), tensor(0.0344), tensor(0.0116), tensor(0.0112), tensor(0.0065), tensor(0.0906), tensor(0.0485), tensor(0.0012), tensor(0.0403), tensor(0.0627), tensor(0.0125), tensor(0.0574), tensor(0.0409), tensor(0.0036)], 'accuracy': [tensor(0.6767), tensor(0.7125), tensor(0.6750), tensor(0.6726), tensor(0.6847), tensor(0.6919), tensor(0.6684), tensor(0.6915), tensor(0.6929), tensor(0.6704), tensor(0.7021), tensor(0.6902), tensor(0.6871), tensor(0.6748), tensor(0.6721), tensor(0.6695), tensor(0.6792), tensor(0.6708), tensor(0.7014), tensor(0.6765), tensor(0.6683), tensor(0.7345), tensor(0.6690), tensor(0.6930), tensor(0.6855), tensor(0.6682), tensor(0.6690), tensor(0.6763), tensor(0.6740), tensor(0.6714), tensor(0.6779), tensor(0.6683), tensor(0.6687), tensor(0.6970), tensor(0.6697), tensor(0.6680), tensor(0.6731), tensor(0.7077), tensor(0.6708), tensor(0.6800), tensor(0.6765), tensor(0.6701), tensor(0.6805), tensor(0.6749), tensor(0.6726), tensor(0.6725), tensor(0.6729), tensor(0.6733), tensor(0.6712), tensor(0.6680), tensor(0.6696), tensor(0.6955), tensor(0.6704), tensor(0.7091), tensor(0.6670), tensor(0.6689), tensor(0.6684), tensor(0.6869), tensor(0.6770), tensor(0.6781), tensor(0.6705), tensor(0.6704), tensor(0.6688), tensor(0.6969), tensor(0.6828), tensor(0.6671), tensor(0.6801), tensor(0.6876), tensor(0.6708), tensor(0.6858), tensor(0.6803), tensor(0.6679)], 'recall': [tensor(0.0301), tensor(0.1376), tensor(0.0250), tensor(0.0177), tensor(0.0541), tensor(0.0756), tensor(0.0053), tensor(0.0745), tensor(0.0786), tensor(0.0112), tensor(0.1063), tensor(0.0706), tensor(0.0614), tensor(0.0243), tensor(0.0164), tensor(0.0084), tensor(0.0376), tensor(0.0125), tensor(0.1043), tensor(0.0296), tensor(0.0049), tensor(0.2034), tensor(0.0070), tensor(0.0789), tensor(0.0565), tensor(0.0045), tensor(0.0071), tensor(0.0288), tensor(0.0221), tensor(0.0143), tensor(0.0338), tensor(0.0048), tensor(0.0061), tensor(0.0911), tensor(0.0092), tensor(0.0039), tensor(0.0194), tensor(0.1232), tensor(0.0125), tensor(0.0399), tensor(0.0295), tensor(0.0104), tensor(0.0416), tensor(0.0248), tensor(0.0177), tensor(0.0175), tensor(0.0187), tensor(0.0198), tensor(0.0137), tensor(0.0039), tensor(0.0089), tensor(0.0864), tensor(0.0112), tensor(0.1273), tensor(0.0010), tensor(0.0068), tensor(0.0052), tensor(0.0608), tensor(0.0310), tensor(0.0344), tensor(0.0116), tensor(0.0112), tensor(0.0065), tensor(0.0906), tensor(0.0485), tensor(0.0012), tensor(0.0403), tensor(0.0627), tensor(0.0125), tensor(0.0574), tensor(0.0409), tensor(0.0036)]}\n"
+      "{'iou': [tensor(0.0114), tensor(0.0618), tensor(0.0171), tensor(0.0496), tensor(0.1154), tensor(0.0141), tensor(0.1329), tensor(0.3690), tensor(0.7344), tensor(0.0732), tensor(0.2918), tensor(0.5210), tensor(0.0146), tensor(0.1446), tensor(0.0695), tensor(0.3571), tensor(0.1984), tensor(0.0307), tensor(0.1094), tensor(0.2462), tensor(0.2313), tensor(0.0449), tensor(0.0699), tensor(0.0097), tensor(0.3149), tensor(0.0784), tensor(0.1874), tensor(0.1622), tensor(0.0245), tensor(0.0357), tensor(0.1011), tensor(0.0828), tensor(0.0063), tensor(0.0360), tensor(0.0860), tensor(0.1214), tensor(0.0326), tensor(0.0735), tensor(0.0126), tensor(0.1483), tensor(0.0915), tensor(0.3903), tensor(0.0985), tensor(0.0136), tensor(0.0193), tensor(0.0362), tensor(0.0175), tensor(0.0152), tensor(0.3412), tensor(0.0298), tensor(0.6657), tensor(0.0089), tensor(0.2835), tensor(0.0226), tensor(0.0442), tensor(0.4874), tensor(0.0096), tensor(0.0437), tensor(0.0395), tensor(0.0331), tensor(0.1646), tensor(0.0127), tensor(0.0728), tensor(0.0573), tensor(0.1029), tensor(0.3060), tensor(0.0071), tensor(0.0175), tensor(0.0547), tensor(0.2606), tensor(0.4043), tensor(0.0826)], 'f1': [tensor(0.0225), tensor(0.1164), tensor(0.0335), tensor(0.0945), tensor(0.2069), tensor(0.0278), tensor(0.2346), tensor(0.5391), tensor(0.8468), tensor(0.1364), tensor(0.4518), tensor(0.6851), tensor(0.0287), tensor(0.2527), tensor(0.1299), tensor(0.5263), tensor(0.3311), tensor(0.0595), tensor(0.1972), tensor(0.3951), tensor(0.3757), tensor(0.0860), tensor(0.1306), tensor(0.0193), tensor(0.4789), tensor(0.1454), tensor(0.3156), tensor(0.2791), tensor(0.0479), tensor(0.0689), tensor(0.1837), tensor(0.1529), tensor(0.0125), tensor(0.0695), tensor(0.1584), tensor(0.2165), tensor(0.0631), tensor(0.1369), tensor(0.0249), tensor(0.2583), tensor(0.1676), tensor(0.5615), tensor(0.1794), tensor(0.0269), tensor(0.0378), tensor(0.0699), tensor(0.0344), tensor(0.0300), tensor(0.5088), tensor(0.0579), tensor(0.7993), tensor(0.0177), tensor(0.4417), tensor(0.0442), tensor(0.0847), tensor(0.6554), tensor(0.0189), tensor(0.0838), tensor(0.0760), tensor(0.0640), tensor(0.2827), tensor(0.0251), tensor(0.1357), tensor(0.1084), tensor(0.1866), tensor(0.4687), tensor(0.0141), tensor(0.0345), tensor(0.1038), tensor(0.4134), tensor(0.5758), tensor(0.1526)], 'accuracy': [tensor(0.6742), tensor(0.7055), tensor(0.6778), tensor(0.6982), tensor(0.7356), tensor(0.6759), tensor(0.7449), tensor(0.8464), tensor(0.9489), tensor(0.7121), tensor(0.8173), tensor(0.8950), tensor(0.6762), tensor(0.7509), tensor(0.7100), tensor(0.8421), tensor(0.7770), tensor(0.6865), tensor(0.7324), tensor(0.7984), tensor(0.7919), tensor(0.6953), tensor(0.7102), tensor(0.6731), tensor(0.8263), tensor(0.7151), tensor(0.7719), tensor(0.7597), tensor(0.6826), tensor(0.6896), tensor(0.7279), tensor(0.7176), tensor(0.6708), tensor(0.6898), tensor(0.7195), tensor(0.7388), tensor(0.6877), tensor(0.7123), tensor(0.6750), tensor(0.7528), tensor(0.7225), tensor(0.8538), tensor(0.7265), tensor(0.6756), tensor(0.6793), tensor(0.6900), tensor(0.6781), tensor(0.6767), tensor(0.8363), tensor(0.6860), tensor(0.9331), tensor(0.6726), tensor(0.8139), tensor(0.6814), tensor(0.6949), tensor(0.8851), tensor(0.6730), tensor(0.6946), tensor(0.6920), tensor(0.6880), tensor(0.7609), tensor(0.6750), tensor(0.7119), tensor(0.7028), tensor(0.7289), tensor(0.8229), tensor(0.6714), tensor(0.6782), tensor(0.7013), tensor(0.8045), tensor(0.8586), tensor(0.7175)], 'recall': [tensor(0.0225), tensor(0.1164), tensor(0.0335), tensor(0.0945), tensor(0.2069), tensor(0.0278), tensor(0.2346), tensor(0.5391), tensor(0.8468), tensor(0.1364), tensor(0.4518), tensor(0.6851), tensor(0.0287), tensor(0.2527), tensor(0.1299), tensor(0.5263), tensor(0.3311), tensor(0.0595), tensor(0.1972), tensor(0.3951), tensor(0.3757), tensor(0.0860), tensor(0.1306), tensor(0.0193), tensor(0.4789), tensor(0.1454), tensor(0.3156), tensor(0.2791), tensor(0.0479), tensor(0.0689), tensor(0.1837), tensor(0.1529), tensor(0.0125), tensor(0.0695), tensor(0.1584), tensor(0.2165), tensor(0.0631), tensor(0.1369), tensor(0.0249), tensor(0.2583), tensor(0.1676), tensor(0.5615), tensor(0.1794), tensor(0.0269), tensor(0.0378), tensor(0.0699), tensor(0.0344), tensor(0.0300), tensor(0.5088), tensor(0.0579), tensor(0.7993), tensor(0.0177), tensor(0.4417), tensor(0.0442), tensor(0.0847), tensor(0.6554), tensor(0.0189), tensor(0.0838), tensor(0.0760), tensor(0.0640), tensor(0.2827), tensor(0.0251), tensor(0.1357), tensor(0.1084), tensor(0.1866), tensor(0.4687), tensor(0.0141), tensor(0.0345), tensor(0.1038), tensor(0.4134), tensor(0.5758), tensor(0.1526)]}\n"
      ]
     }
    ],
@@ -540,8 +540,8 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 3, 576, 640])\n",
-      "torch.Size([1, 1, 576, 640])\n"
+      "torch.Size([1, 3, 576, 864])\n",
+      "torch.Size([1, 1, 576, 864])\n"
      ]
     }
    ],
@@ -568,7 +568,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 576, 640])\n"
+      "torch.Size([1, 576, 864])\n"
      ]
     }
    ],
@@ -585,7 +585,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 576, 640])\n"
+      "torch.Size([1, 576, 864])\n"
      ]
     }
    ],
@@ -647,7 +647,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "{1: 1, 2: 4, 3: 367997, 4: 307, 5: 331}\n"
+      "{0: 17603, 1: 375312, 2: 458, 3: 14150, 4: 88587, 5: 1554}\n"
      ]
     }
    ],
@@ -671,7 +671,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# plt.imsave(\"assets/aerial-drone-example-unet-plus-plus-output.jpeg\", np.array(outp))"
+    "plt.imsave(\"assets/dubai-example-unet-plus-plus-output.jpeg\", np.array(outp))"
    ]
   },
   {
@@ -691,10 +691,10 @@
     {
      "data": {
       "text/plain": [
-       "{'iou': tensor(0.0897),\n",
-       " 'f1': tensor(0.1647),\n",
-       " 'accuracy': tensor(0.7216),\n",
-       " 'recall': tensor(0.1647)}"
+       "{'iou': tensor(0.2857),\n",
+       " 'f1': tensor(0.4445),\n",
+       " 'accuracy': tensor(0.8148),\n",
+       " 'recall': tensor(0.4445)}"
       ]
      },
      "execution_count": 43,
@@ -762,8 +762,8 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 3, 576, 608])\n",
-      "torch.Size([1, 1, 576, 608])\n"
+      "torch.Size([1, 3, 576, 768])\n",
+      "torch.Size([1, 1, 576, 768])\n"
      ]
     }
    ],
@@ -790,7 +790,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 576, 608])\n"
+      "torch.Size([1, 576, 768])\n"
      ]
     }
    ],
@@ -807,7 +807,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "torch.Size([1, 576, 608])\n"
+      "torch.Size([1, 576, 768])\n"
      ]
     }
    ],
@@ -875,7 +875,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "plt.imsave(\"assets/aerial-drone-example-deep-lab-v3-output.jpeg\", np.array(outp))"
+    "plt.imsave(\"assets/dubai-example-deep-lab-v3-output.jpeg\", np.array(outp))"
    ]
   },
   {
@@ -895,10 +895,10 @@
     {
      "data": {
       "text/plain": [
-       "{'iou': tensor(0.0883),\n",
-       " 'f1': tensor(0.1623),\n",
-       " 'accuracy': tensor(0.7208),\n",
-       " 'recall': tensor(0.1623)}"
+       "{'iou': tensor(0.2249),\n",
+       " 'f1': tensor(0.3672),\n",
+       " 'accuracy': tensor(0.7891),\n",
+       " 'recall': tensor(0.3672)}"
       ]
      },
      "execution_count": 58,
@@ -1079,7 +1079,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "plt.imsave(\"assets/aerial-drone-example-deep-lav-v3-plus-output.jpeg\", np.array(outp))"
+    "plt.imsave(\"assets/dubai-example-deep-lab-v3-plus-output.jpeg\", np.array(outp))"
    ]
   },
   {
@@ -1099,10 +1099,10 @@
     {
      "data": {
       "text/plain": [
-       "{'iou': tensor(0.0066),\n",
-       " 'f1': tensor(0.0131),\n",
-       " 'accuracy': tensor(0.6710),\n",
-       " 'recall': tensor(0.0131)}"
+       "{'iou': tensor(0.0016),\n",
+       " 'f1': tensor(0.0031),\n",
+       " 'accuracy': tensor(0.6677),\n",
+       " 'recall': tensor(0.0031)}"
       ]
      },
      "execution_count": 73,