app.py

# -*- coding: utf-8 -*-
"""Yet another copy of Final CNN Pose Notebook.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/1IdEBDyEyKQdRRT9R-GkfrJINmHdf3_pF
"""

# from google.colab import drive
# drive.mount('/content/drive')

# pip install gradio

import gradio as gr





import torch
from torch.utils.data import DataLoader, Dataset, random_split
from torchvision import transforms, utils
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from PIL import Image
import os
import numpy as np
import json
import matplotlib.pyplot as plt
from torch.utils.data.dataloader import default_collate

# Define the dataset class
class HumanPoseDataset(Dataset):
    def __init__(self, annotations, img_dir, transform=None):
        self.annotations = annotations
        self.img_dir = img_dir
        self.transform = transform

    def __len__(self):
        return len(self.annotations)

    def __getitem__(self, idx):
        img_key = list(self.annotations.keys())[idx]
        annotation_list = self.annotations[img_key]
        # Skip the image if there are no annotations
        if not annotation_list:
            return None
        # Use the first annotation for simplicity
        annotation = annotation_list[0]
        if not annotation['landmarks']:  # Check if landmarks are not empty
            return None
        img_name = os.path.join(self.img_dir, annotation['file'])
        image = Image.open(img_name).convert('RGB')
        original_image_size = image.size
        keypoints = annotation['landmarks']
        keypoints_array = np.array([[k['x'], k['y'], k['z'], k['visibility']] for k in keypoints])

        if self.transform:
            image = self.transform(image)

        sample = {'image': image, 'keypoints': keypoints_array, 'original_image_size': original_image_size}
        print(sample)
        return sample

# Custom collate function to filter out None values
def custom_collate(batch):
    batch = [b for b in batch if b is not None]
    return default_collate(batch)

# Load the annotations JSON into a dictionary
annotations_path = '/content/drive/MyDrive/annotations_CNN (3).json'  # Update this path
with open(annotations_path) as f:
    annotations_data = json.load(f)
print("Annotations data loaded. Number of images:", len(annotations_data))

x = annotations_data.keys()

"""# Do data preprocessing. For example, resize to 32 by 32 and normalization.

"""

img_dir = '/content/drive/MyDrive/CNN_Dataset'

# Define the transformations with resizing and augmentation
transform = transforms.Compose([
    transforms.Resize((32, 32)),  # Resize the images to 256x256
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    transforms.RandomHorizontalFlip(),  # Example augmentation
    # Add more augmentations if needed
])

test_transform=transforms.Compose([
    transforms.ToTensor(),
    transforms.Resize((32,32)),
])

# Create the dataset
human_pose_dataset = HumanPoseDataset(annotations_data, img_dir, transform=transform)
testing_pose_dataset = HumanPoseDataset(annotations_data, img_dir, transform=test_transform)

print("Dataset created. Length of dataset:", len(human_pose_dataset))

sorted(x) == sorted(os.listdir('/content/drive/MyDrive/CNN_Dataset'))

"""#2. Load parameters of a pretrained model. If a pretrained model for the entire network is not available, then load parameters for the backbone network/feature extraction network/encoder.

Pose net model is not available so we will be using an architecture similar to PoseNet, a human pose detection CNN architecture. In the above architecture, we are given a brief description about the PoseNet Architecture. We will be using the Regression Network to find the keypoint coordinates.
"""

import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F

class SimpleCNN(nn.Module):
    def __init__(self):
        super(SimpleCNN, self).__init__()
        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, padding=1)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(16, 32, kernel_size=3, padding=1)
        self.conv3 = nn.Conv2d(32, 64, kernel_size=3, padding=1)
        self.conv4 = nn.Conv2d(64, 128, kernel_size=3, padding=1)
        # Assuming the input image size is 256x256, after four pooling layers the image size will be 16x16
        self.fc1 = nn.Linear(2 * 16 * 16, 1000)
        self.fc2 = nn.Linear(1000, 33 * 4)  # Assuming 33 keypoints

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = self.pool(F.relu(self.conv3(x)))
        x = self.pool(F.relu(self.conv4(x)))
        x = torch.flatten(x, 1)  # Flatten the tensor for the fully connected layer
        x = F.relu(self.fc1(x))
        x = self.fc2(x)
        return x

# Initialize the model
model = SimpleCNN()
print("Model initialized.")
print(model)  # Print the model architecture

#!pip install mediapipe

"""#3 Replace the output layer if necessary and finetune the network for your dataset. Use validation dataset to pick a good learning rate and momentum.

1. Training for a very less samples
"""

# Split the dataset into training, validation, and test sets
train_size = int(0.04* len(human_pose_dataset))
validation_size = int(0.1 * len(human_pose_dataset))
test_size = len(human_pose_dataset) - train_size - validation_size
train_dataset, remaining_dataset = random_split(human_pose_dataset, [train_size, validation_size + test_size])
validation_dataset, test_dataset = random_split(remaining_dataset, [validation_size, test_size])

test_pose_dataset , remaining_data = random_split(testing_pose_dataset,[6,194])

# Define the batch size
batch_size = 8

# Create data loaders for each set with the custom collate function
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, collate_fn=custom_collate)
validation_loader = DataLoader(validation_dataset, batch_size=batch_size, shuffle=False, collate_fn=custom_collate)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, collate_fn=custom_collate)

test_image_loader = DataLoader(test_pose_dataset, batch_size=batch_size, shuffle=False, collate_fn=custom_collate)

print("Data loaders created.")

len(train_dataset)



# Loss function
criterion = nn.MSELoss()

# Optimizer
optimizer = optim.Adam(model.parameters(), lr=1e-4)

# Convert the model parameters to float
model = model.float()

# Ensure that the tensors are also floats
sample_batch = next(iter(train_loader))
import mediapipe as mp
images = sample_batch['image'].float()  # Convert images to float
keypoints = sample_batch['keypoints'].view(-1, 132).float()  # Convert keypoints to float and reshape

# Now proceed with the optimization loop
loss=0
for epochs in range(10):
  optimizer.zero_grad()
  outputs = model(images)
  loss = criterion(outputs, keypoints)
  loss.backward()
  optimizer.step()
  print("Optimization step completed.")
  print(loss.item())
  loss=loss.item()

import torch

def calculate_accuracy(outputs, targets):
      accuracy = torch.mean(torch.abs(outputs - targets))
      return accuracy

print(outputs.shape)
# Calculate accuracy
with torch.no_grad():
    accuracy = calculate_accuracy(outputs, keypoints)
    accuracy= 1- accuracy/132

print("Loss:", loss)
print("Accuracy:", accuracy.item()*100, '%')

"""As you can see, the accuracy is very close to 100% (Overfitting)

Now taking 80-10-10 split on the dataset, we create new train, val and test loaders
"""

# Split the dataset into training, validation, and test sets
train_size = int(0.8* len(human_pose_dataset))
validation_size = int(0.1 * len(human_pose_dataset))
test_size = len(human_pose_dataset) - train_size - validation_size
train_dataset, remaining_dataset = random_split(human_pose_dataset, [train_size, validation_size + test_size])
validation_dataset, test_dataset = random_split(remaining_dataset, [validation_size, test_size])

test_pose_dataset , remaining_data = random_split(testing_pose_dataset,[6,194])

# Define the batch size
batch_size = 8

# Create data loaders for each set with the custom collate function
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, collate_fn=custom_collate)
validation_loader = DataLoader(validation_dataset, batch_size=batch_size, shuffle=False, collate_fn=custom_collate)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, collate_fn=custom_collate)

test_image_loader = DataLoader(test_pose_dataset, batch_size=batch_size, shuffle=False, collate_fn=custom_collate)

print("Data loaders created.")

len(test_dataset)

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, random_split
from torchvision import transforms
import torch.nn.functional as F

class SimpleCNN(nn.Module):

# Define hyperparameters to search over
      learning_rates = [0.001, 0.01, 0.1]
      momentums = [0.9, 0.95, 0.99]
      weight_decays = [0.0001, 0.001, 0.01]

      best_loss = float('inf')
      best_lr, best_momentum, best_weight_decay = None, None, None

      # Grid search over hyperparameters
      for lr in learning_rates:
          for momentum in momentums:
              for weight_decay in weight_decays:
                  # Initialize the model with the current set of hyperparameters
                  model = SimpleCNN()

                  # Define loss function and optimizer
                  criterion = nn.MSELoss()
                  optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum, weight_decay=weight_decay)

                  # Ensure that the tensors are also floats
                  sample_batch = next(iter(train_loader))
                  import mediapipe as mp
                  images = sample_batch['image'].float()  # Convert images to float
                  keypoints = sample_batch['keypoints'].view(-1, 132).float()  # Convert keypoints to float and reshape

                  # Now proceed with the optimization loop
                  optimizer.zero_grad()
                  outputs = model(images)
                  print("Output shape after forward pass:", outputs.shape)
                  outputs = model(images)
                  loss = criterion(outputs, keypoints)
                  print("Initial loss:", loss.item())
                  loss.backward()
                  optimizer.step()
                  print("Optimization step completed.")

                  total_loss = 0
                  avg_loss = total_loss / len(train_loader)
                  model.train()

                  # Check if the current set of hyperparameters resulted in a better performance
                  if avg_loss < best_loss:
                      best_loss = avg_loss
                      best_lr, best_momentum, best_weight_decay = lr, momentum, weight_decay

      # After the grid search, choose the hyperparameters that performed the best
      print("Best Hyperparameters - lr: {}, momentum: {}, weight_decay: {}".format(
          best_lr, best_momentum, best_weight_decay))

      # Train the final model with the selected hyperparameters on the full dataset
      model = SimpleCNN()
      optimizer = optim.SGD(model.parameters(), lr=best_lr, momentum=best_momentum, weight_decay=best_weight_decay)

"""#3. Plotting Validation and Test Loss

The best parameters are:

*   Learning Rate: 0.001
*   Momentum: 0.9
*   Weight Decay: 0.0001
"""

import torch
import matplotlib.pyplot as plt

# Assuming you have already defined your model, optimizer, and criterion

# Ensure that the tensors are also floats for training
sample_batch = next(iter(train_loader))
images = sample_batch['image'].float()
keypoints = sample_batch['keypoints'].view(-1, 132).float()

# Ensure that the tensors are also floats for validation
validation_sample_batch = next(iter(validation_loader))
validation_images = validation_sample_batch['image'].float()
validation_keypoints = validation_sample_batch['keypoints'].view(-1, 132).float()

# Now proceed with the optimization loop
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
criterion = torch.nn.MSELoss()

train_loss = []
val_loss = []

for epoch in range(15):
    model.train()
    optimizer.zero_grad()
    outputs = model(images)
    current_loss = criterion(outputs, keypoints)
    current_loss.backward()
    optimizer.step()

    model.eval()  # Switch to evaluation mode for validation
    with torch.no_grad():
        # Calculate validation loss
        val_outputs = model(validation_images)
        val_current_loss = criterion(val_outputs, validation_keypoints)

    print(f"Epoch [{epoch + 1}/100], Loss: {current_loss.item():.4f}, Val Loss: {val_current_loss.item():.4f}")
    train_loss.append(current_loss.item())
    val_loss.append(val_current_loss.item())

plotting_val_loss = val_loss
plotting_train_loss = train_loss

import matplotlib.pyplot as plt
# Plotting

plt.figure(figsize=(8, 4))

plt.plot( plotting_train_loss, marker='o', linestyle='-', color='b',label='train loss')
plt.plot( plotting_val_loss, marker='o', linestyle= '-', color='r', label='val loss')

plt.title('Loss vs Epochs')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.grid(True)
plt.legend()

# Show the legend in a small box
plt.legend(loc='upper right')

plt.show()

"""#4. Final Run on Test Dataset"""

# Ensure that the tensors are also floats
sample_batch = next(iter(test_loader))
import mediapipe as mp
test_images = sample_batch['image'].float()  # Convert images to float
test_keypoints = sample_batch['keypoints'].view(-1, 132).float()  # Convert keypoints to float and reshape

model.eval()

optimizer.zero_grad()
outputs = model(test_images)

print("Testing Done")

test_image_tensor = test_images[0]



test_actual_plot = test_keypoints.reshape(len(test_images),33,4)[0]

test_predict_plot = outputs.reshape(len(test_images),33,4)[0]

test_predict_plot.shape

"""# 4. Finally, evaluate on the test dataset."""

import cv2

import matplotlib.pyplot as plt
import numpy as np

def plot_human_pose(keypoints):
    # Create a figure and axis
    fig, ax = plt.subplots()

    # Plot keypoints
    for i in range(len(keypoints)):
        x, y, _, _ = keypoints[i]
        ax.scatter(x, -y, color='blue')  # Invert y-axis

    # Connect body parts
    connect_lines = [(0, 2), (2, 7),   # Left eye
                     (0, 5), (5, 8),   # Right eye
                     (9,10),  # Left side
                     (11, 12), (12, 24), (11, 23),  # Right side
                     (24,23), (24,26), (23,25),  # Connect ears and wrists
                     (26, 28), (25, 27),
                     (28, 30), (28, 32), (30,32),# Connect left and right pinky fingers
                     (27, 29), (27, 31), (31,29),  # Connect left and right index fingers
                     (12, 14), (11, 13),  # Connect left and right thumbs
                     (14, 16), (13, 15),  # Connect left and right hips
                     (16, 18), (18, 20), (16,20), (16,22),  # Connect left and right knees
                     (15, 17), (15, 19),  # Connect left and right ankles
                     (17, 19), (15, 21)]  # Connect left and right heels

    for line in connect_lines:
        start, end = line
        x_vals = [keypoints[start][0], keypoints[end][0]]
        y_vals = [-keypoints[start][1], -keypoints[end][1]]  # Invert y-axis
        ax.plot(x_vals, y_vals, linewidth=2, color='red')

    ax.set_aspect('equal', adjustable='datalim')
    plt.title('Actual Pose')
    plt.axis('off')
    plt.show()

# Example usage:
keypoints = test_actual_plot  # Replace with your 33 key points
plot_human_pose(keypoints)

def plot_human_pose(keypoints):
    # Create a figure and axis
    fig, ax = plt.subplots()

    # Plot keypoints
    for i in range(len(keypoints)):
        x, y, _, _ = keypoints[i]
        ax.scatter(x, -y, color='blue')  # Invert y-axis

    # Connect body parts
    connect_lines = [(0, 2), (2, 7),   # Left eye
                     (0, 5), (5, 8),   # Right eye
                     (9,10),  # Left side
                     (11, 12), (12, 24), (11, 23),  # Right side
                     (24,23), (24,26), (23,25),  # Connect ears and wrists
                     (26, 28), (25, 27),
                     (28, 30), (28, 32), (30,32),# Connect left and right pinky fingers
                     (27, 29), (27, 31), (31,29),  # Connect left and right index fingers
                     (12, 14), (11, 13),  # Connect left and right thumbs
                     (14, 16), (13, 15),  # Connect left and right hips
                     (16, 18), (18, 20), (16,20), (16,22),  # Connect left and right knees
                     (15, 17), (15, 19),  # Connect left and right ankles
                     (17, 19), (15, 21)]  # Connect left and right heels

    for line in connect_lines:
        start, end = line
        x_vals = [keypoints[start][0], keypoints[end][0]]
        y_vals = [-keypoints[start][1], -keypoints[end][1]]  # Invert y-axis
        ax.plot(x_vals, y_vals, linewidth=2, color='green')

    ax.set_aspect('equal', adjustable='datalim')
    plt.title('Predicted Pose')
    plt.axis('off')
    plt.show()

# Example usage:
keypoints = test_predict_plot.detach().numpy() # Replace with your 33 key points
plot_human_pose(keypoints)

"""### As you can see, the model predicts the pose of the person very accurately as depicted by its train and validation accuracy"""

from PIL import Image

def predict_pose(image_path):
  img = Image.open(str(image_path)).resize((32,32))
  convert_tensor = transforms.ToTensor()
  tensor_img = convert_tensor(img)
  model.eval()

  optimizer.zero_grad()
  outputs = model(img)

  pred_keypoints = outputs.reshape(1,33,4)[0]
  pred_keypoints = pred_keypoints.detach().numpy()

  plot_human_pose(pred_keypoints)

pose_detector = gr.Interface(fn = predict_pose, inputs = gr.Image(label = 'input image'), outputs = gr.Image(label = 'output image'), title = 'pose_detector' )

gr.TabbedInterface([pose_detector],tab_names = ['pose_detection']).queue().launch()