/Facial-Landmarks-Detection-Pytorch

Implementation of facial landmarks detection using pytorch on iBUG 300W dataset. Why Xception Net? Because it provides satisfactory accuracy, with least computational requirement and memory utilization, I mean can you believe that my model, which looks kinda massive, is only of 10MB!

Primary LanguageJupyter Notebook

Face Landmarks Detection

Inference on few celebraties

Data preprocessing

  1. Augmentations for faces:

    • Random Brightness
    • Random Contrast
    • Random Gamma
    • Random Saturation
    • Random Hue
    • Random Rotation

  2. Augmentations for landmarks:

    • Random Rotation

Model

Architecture: Xception Net

Summary:

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1         [-1, 32, 128, 128]             288
       BatchNorm2d-2         [-1, 32, 128, 128]              64
         LeakyReLU-3         [-1, 32, 128, 128]               0
            Conv2d-4         [-1, 64, 128, 128]          18,432
       BatchNorm2d-5         [-1, 64, 128, 128]             128
         LeakyReLU-6         [-1, 64, 128, 128]               0
            Conv2d-7         [-1, 64, 128, 128]             576
            Conv2d-8         [-1, 64, 128, 128]           4,096
DepthewiseSeperableConv2d-9         [-1, 64, 128, 128]               0
      BatchNorm2d-10         [-1, 64, 128, 128]             128
        LeakyReLU-11         [-1, 64, 128, 128]               0
           Conv2d-12         [-1, 64, 128, 128]             576
           Conv2d-13        [-1, 128, 128, 128]           8,192
DepthewiseSeperableConv2d-14        [-1, 128, 128, 128]               0
      BatchNorm2d-15        [-1, 128, 128, 128]             256
        MaxPool2d-16          [-1, 128, 64, 64]               0
           Conv2d-17          [-1, 128, 64, 64]           8,320
      BatchNorm2d-18          [-1, 128, 64, 64]             256
        LeakyReLU-19          [-1, 128, 64, 64]               0
           Conv2d-20          [-1, 128, 64, 64]           1,152
           Conv2d-21          [-1, 128, 64, 64]          16,384
DepthewiseSeperableConv2d-22          [-1, 128, 64, 64]               0
      BatchNorm2d-23          [-1, 128, 64, 64]             256
        LeakyReLU-24          [-1, 128, 64, 64]               0
           Conv2d-25          [-1, 128, 64, 64]           1,152
           Conv2d-26          [-1, 256, 64, 64]          32,768
DepthewiseSeperableConv2d-27          [-1, 256, 64, 64]               0
      BatchNorm2d-28          [-1, 256, 64, 64]             512
        MaxPool2d-29          [-1, 256, 32, 32]               0
           Conv2d-30          [-1, 256, 32, 32]          33,024
      BatchNorm2d-31          [-1, 256, 32, 32]             512
       EntryBlock-32          [-1, 256, 32, 32]               0
        LeakyReLU-33          [-1, 256, 32, 32]               0
           Conv2d-34          [-1, 256, 32, 32]           2,304
           Conv2d-35          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-36          [-1, 256, 32, 32]               0
      BatchNorm2d-37          [-1, 256, 32, 32]             512
        LeakyReLU-38          [-1, 256, 32, 32]               0
           Conv2d-39          [-1, 256, 32, 32]           2,304
           Conv2d-40          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-41          [-1, 256, 32, 32]               0
      BatchNorm2d-42          [-1, 256, 32, 32]             512
        LeakyReLU-43          [-1, 256, 32, 32]               0
           Conv2d-44          [-1, 256, 32, 32]           2,304
           Conv2d-45          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-46          [-1, 256, 32, 32]               0
      BatchNorm2d-47          [-1, 256, 32, 32]             512
 MiddleBasicBlock-48          [-1, 256, 32, 32]               0
        LeakyReLU-49          [-1, 256, 32, 32]               0
           Conv2d-50          [-1, 256, 32, 32]           2,304
           Conv2d-51          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-52          [-1, 256, 32, 32]               0
      BatchNorm2d-53          [-1, 256, 32, 32]             512
        LeakyReLU-54          [-1, 256, 32, 32]               0
           Conv2d-55          [-1, 256, 32, 32]           2,304
           Conv2d-56          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-57          [-1, 256, 32, 32]               0
      BatchNorm2d-58          [-1, 256, 32, 32]             512
        LeakyReLU-59          [-1, 256, 32, 32]               0
           Conv2d-60          [-1, 256, 32, 32]           2,304
           Conv2d-61          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-62          [-1, 256, 32, 32]               0
      BatchNorm2d-63          [-1, 256, 32, 32]             512
 MiddleBasicBlock-64          [-1, 256, 32, 32]               0
        LeakyReLU-65          [-1, 256, 32, 32]               0
           Conv2d-66          [-1, 256, 32, 32]           2,304
           Conv2d-67          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-68          [-1, 256, 32, 32]               0
      BatchNorm2d-69          [-1, 256, 32, 32]             512
        LeakyReLU-70          [-1, 256, 32, 32]               0
           Conv2d-71          [-1, 256, 32, 32]           2,304
           Conv2d-72          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-73          [-1, 256, 32, 32]               0
      BatchNorm2d-74          [-1, 256, 32, 32]             512
        LeakyReLU-75          [-1, 256, 32, 32]               0
           Conv2d-76          [-1, 256, 32, 32]           2,304
           Conv2d-77          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-78          [-1, 256, 32, 32]               0
      BatchNorm2d-79          [-1, 256, 32, 32]             512
 MiddleBasicBlock-80          [-1, 256, 32, 32]               0
        LeakyReLU-81          [-1, 256, 32, 32]               0
           Conv2d-82          [-1, 256, 32, 32]           2,304
           Conv2d-83          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-84          [-1, 256, 32, 32]               0
      BatchNorm2d-85          [-1, 256, 32, 32]             512
        LeakyReLU-86          [-1, 256, 32, 32]               0
           Conv2d-87          [-1, 256, 32, 32]           2,304
           Conv2d-88          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-89          [-1, 256, 32, 32]               0
      BatchNorm2d-90          [-1, 256, 32, 32]             512
        LeakyReLU-91          [-1, 256, 32, 32]               0
           Conv2d-92          [-1, 256, 32, 32]           2,304
           Conv2d-93          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-94          [-1, 256, 32, 32]               0
      BatchNorm2d-95          [-1, 256, 32, 32]             512
 MiddleBasicBlock-96          [-1, 256, 32, 32]               0
        LeakyReLU-97          [-1, 256, 32, 32]               0
           Conv2d-98          [-1, 256, 32, 32]           2,304
           Conv2d-99          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-100          [-1, 256, 32, 32]               0
     BatchNorm2d-101          [-1, 256, 32, 32]             512
       LeakyReLU-102          [-1, 256, 32, 32]               0
          Conv2d-103          [-1, 256, 32, 32]           2,304
          Conv2d-104          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-105          [-1, 256, 32, 32]               0
     BatchNorm2d-106          [-1, 256, 32, 32]             512
       LeakyReLU-107          [-1, 256, 32, 32]               0
          Conv2d-108          [-1, 256, 32, 32]           2,304
          Conv2d-109          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-110          [-1, 256, 32, 32]               0
     BatchNorm2d-111          [-1, 256, 32, 32]             512
MiddleBasicBlock-112          [-1, 256, 32, 32]               0
       LeakyReLU-113          [-1, 256, 32, 32]               0
          Conv2d-114          [-1, 256, 32, 32]           2,304
          Conv2d-115          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-116          [-1, 256, 32, 32]               0
     BatchNorm2d-117          [-1, 256, 32, 32]             512
       LeakyReLU-118          [-1, 256, 32, 32]               0
          Conv2d-119          [-1, 256, 32, 32]           2,304
          Conv2d-120          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-121          [-1, 256, 32, 32]               0
     BatchNorm2d-122          [-1, 256, 32, 32]             512
       LeakyReLU-123          [-1, 256, 32, 32]               0
          Conv2d-124          [-1, 256, 32, 32]           2,304
          Conv2d-125          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-126          [-1, 256, 32, 32]               0
     BatchNorm2d-127          [-1, 256, 32, 32]             512
MiddleBasicBlock-128          [-1, 256, 32, 32]               0
     MiddleBlock-129          [-1, 256, 32, 32]               0
          Conv2d-130          [-1, 512, 16, 16]         131,584
     BatchNorm2d-131          [-1, 512, 16, 16]           1,024
       LeakyReLU-132          [-1, 256, 32, 32]               0
          Conv2d-133          [-1, 256, 32, 32]           2,304
          Conv2d-134          [-1, 256, 32, 32]          65,536
DepthewiseSeperableConv2d-135          [-1, 256, 32, 32]               0
     BatchNorm2d-136          [-1, 256, 32, 32]             512
       LeakyReLU-137          [-1, 256, 32, 32]               0
          Conv2d-138          [-1, 256, 32, 32]           2,304
          Conv2d-139          [-1, 512, 32, 32]         131,072
DepthewiseSeperableConv2d-140          [-1, 512, 32, 32]               0
     BatchNorm2d-141          [-1, 512, 32, 32]           1,024
       MaxPool2d-142          [-1, 512, 16, 16]               0
          Conv2d-143          [-1, 512, 16, 16]           4,608
          Conv2d-144          [-1, 512, 16, 16]         262,144
DepthewiseSeperableConv2d-145          [-1, 512, 16, 16]               0
     BatchNorm2d-146          [-1, 512, 16, 16]           1,024
       LeakyReLU-147          [-1, 512, 16, 16]               0
          Conv2d-148          [-1, 512, 16, 16]           4,608
          Conv2d-149         [-1, 1024, 16, 16]         524,288
DepthewiseSeperableConv2d-150         [-1, 1024, 16, 16]               0
     BatchNorm2d-151         [-1, 1024, 16, 16]           2,048
       LeakyReLU-152         [-1, 1024, 16, 16]               0
AdaptiveAvgPool2d-153           [-1, 1024, 1, 1]               0
         Dropout-154           [-1, 1024, 1, 1]               0
       ExitBlock-155           [-1, 1024, 1, 1]               0
          Linear-156                  [-1, 136]         139,400
================================================================
Total params: 2,630,888
Trainable params: 2,630,888
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.06
Forward/backward pass size (MB): 441.02
Params size (MB): 10.04
Estimated Total Size (MB): 451.12
----------------------------------------------------------------

Training hyperparameters

  1. Objective loss : MSELoss
  2. Optimizer : Adam
  3. Learning Rate : 0.0008
  4. Epochs : 30

Training progress

Progress of the model thoughout the training

Final Results

Final outputs of the model

Read Full Article

Watch the video

Watch the video here

Author - Rishik Mourya