SIREN paper implementation – Nipun Batra Blog

TLDR: Sine activation function is better than ReLU for reconstructing images

Animation of the training process

Code

import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import numpy as np
import torch.nn as nn
import torch.nn.functional as F

# Remove all the warnings
import warnings
warnings.filterwarnings('ignore')

# Set env CUDA_LAUNCH_BLOCKING=1
import os
os.environ['CUDA_LAUNCH_BLOCKING'] = '1'

Code

!wget https://segment-anything.com/assets/gallery/AdobeStock_94274587_welsh_corgi_pembroke_CD.jpg -O dog.jpg

--2023-04-27 17:21:53--  https://segment-anything.com/assets/gallery/AdobeStock_94274587_welsh_corgi_pembroke_CD.jpg
Resolving segment-anything.com (segment-anything.com)... 108.138.128.23, 108.138.128.8, 108.138.128.34, ...
Connecting to segment-anything.com (segment-anything.com)|108.138.128.23|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 221810 (217K) [image/jpeg]
Saving to: ‘dog.jpg’

dog.jpg             100%[===================>] 216.61K   400KB/s    in 0.5s    

2023-04-27 17:21:55 (400 KB/s) - ‘dog.jpg’ saved [221810/221810]

Code

# Read in a image from torchvision
img = torchvision.io.read_image("dog.jpg")

Code

plt.imshow(img.permute(1, 2, 0))

Code

# Normalize the image
img = img / 255.0

Code

img.shape

torch.Size([3, 1365, 2048])

Code

# Take a random 224x224 crop of the image
crop = torchvision.transforms.functional.crop(img, 600, 750, 400, 400)


# Plot the crop
plt.imshow(crop.permute(1, 2, 0))

Code

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

Code

# Get the dimensions of the image tensor
num_channels, height, width = crop.shape

# Create a 2D grid of (x,y) coordinates
x_coords = torch.arange(width).repeat(height, 1)
y_coords = torch.arange(height).repeat(width, 1).t()
x_coords = x_coords.reshape(-1)
y_coords = y_coords.reshape(-1)

# Combine the x and y coordinates into a single tensor
X = torch.stack([x_coords, y_coords], dim=1).float()

# Move X to GPU if available
X = X.to(device)

Code

num_xy = height * width
num_xy

Code

X.shape, X

(torch.Size([160000, 2]),
 tensor([[  0.,   0.],
         [  1.,   0.],
         [  2.,   0.],
         ...,
         [397., 399.],
         [398., 399.],
         [399., 399.]], device='cuda:0'))

Code

# Extract pixel values from image tensor
pixel_values = crop.reshape(num_channels, -1).float().to(device)

# Transpose the pixel values to be (num_xy, num_channels)
pixel_values = pixel_values.transpose(0, 1)

y = pixel_values.to(device)

Code

# Create a MLP with 5 hidden layers with 256 neurons each and ReLU activations.
# Input is (x, y) and output is (r, g, b)

class MLP(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(2, 256)
        self.fc2 = nn.Linear(256, 256)
        self.fc3 = nn.Linear(256, 256)
        self.fc4 = nn.Linear(256, 256)
        self.fc5 = nn.Linear(256, 256)
        self.fc6 = nn.Linear(256, 3)

    def forward(self, x):
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = F.relu(self.fc3(x))
        x = F.relu(self.fc4(x))
        x = F.relu(self.fc5(x))
        return self.fc6(x)

Code

# Training loop function to train the model
# X: (num_xy, 2) tensor of (x, y) coordinates
# y: (num_xy, 3) tensor of (r, g, b) pixel values
# model: MLP model
# lr: learning rate
# epochs: number of epochs to train for
# bs: batch size
# print_every: print loss every print_every epochs
# Logs losses
# Saves the prediction frmo model every print_every epochs

def train(X, y, model, lr=0.01, epochs=1000, bs=1000, print_every=100):
    losses = []
    imgs = []
    optimizer = torch.optim.Adam(model.parameters(), lr=lr)
    criterion = nn.MSELoss()
    for epoch in range(epochs):
        # Get a random batch of (x, y) coordinates
        idxs = torch.randperm(num_xy)[:bs]
        batch_X = X[idxs]
        batch_y = y[idxs]

        # Predict the (r, g, b) values
        pred_y = model(batch_X)

        # Compute the loss
        loss = criterion(pred_y, batch_y)

        # Zero gradients, perform a backward pass, and update the weights
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        losses.append(loss.item())

        # Print loss every print_every epochs
        if epoch % print_every == 0:
            print(f"Epoch {epoch} loss: {loss.item()}")
            with torch.no_grad():
                # Predict the (r, g, b) values
                pred_y = model(X)

                # Reshape the predictions to be (3, height, width)
                pred_y = pred_y.transpose(0, 1).reshape(num_channels, height, width)
                imgs.append(pred_y.permute(1, 2, 0).detach().cpu())
                
    return losses, imgs

Code

m1 = MLP()
m1 = m1.to(device)
losses_mlp, imgs = train(X, y, m1, lr=0.001, epochs=4000, bs=2000, print_every=100)

Epoch 0 loss: 1.5234602689743042
Epoch 100 loss: 0.0640626773238182
Epoch 200 loss: 0.04388527199625969
Epoch 300 loss: 0.03277464583516121
Epoch 400 loss: 0.03183111175894737
Epoch 500 loss: 0.02485758438706398
Epoch 600 loss: 0.023289738222956657
Epoch 700 loss: 0.024606380611658096
Epoch 800 loss: 0.023782318457961082
Epoch 900 loss: 0.026350615546107292
Epoch 1000 loss: 0.025088826194405556
Epoch 1100 loss: 0.023389095440506935
Epoch 1200 loss: 0.02370390295982361
Epoch 1300 loss: 0.023111725226044655
Epoch 1400 loss: 0.023864751681685448
Epoch 1500 loss: 0.021725382655858994
Epoch 1600 loss: 0.021787280216813087
Epoch 1700 loss: 0.021760988980531693
Epoch 1800 loss: 0.021614212542772293
Epoch 1900 loss: 0.020562106743454933
Epoch 2000 loss: 0.019880816340446472
Epoch 2100 loss: 0.01901845820248127
Epoch 2200 loss: 0.018372364342212677
Epoch 2300 loss: 0.01828525774180889
Epoch 2400 loss: 0.018451901152729988
Epoch 2500 loss: 0.01738181710243225
Epoch 2600 loss: 0.01698809117078781
Epoch 2700 loss: 0.01643018051981926
Epoch 2800 loss: 0.01669265516102314
Epoch 2900 loss: 0.01664060726761818
Epoch 3000 loss: 0.01606595516204834
Epoch 3100 loss: 0.01667209528386593
Epoch 3200 loss: 0.015133237466216087
Epoch 3300 loss: 0.014814447611570358
Epoch 3400 loss: 0.01538220327347517
Epoch 3500 loss: 0.01484852284193039
Epoch 3600 loss: 0.01589234732091427
Epoch 3700 loss: 0.014897373504936695
Epoch 3800 loss: 0.014240250922739506
Epoch 3900 loss: 0.015261288732290268

Code

def plot_image(model, name=None):
    # Predict the (r, g, b) values
    pred_y = model(X)

    # Reshape the predictions to be (3, height, width)
    pred_y = pred_y.transpose(0, 1).reshape(num_channels, height, width)

    # plot the image
    plt.imshow(pred_y.permute(1, 2, 0).detach().cpu())
    if name:
        plt.savefig(name)

Code

plot_image(m1, "mlp_dog.png")

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

Code

# Create the animation from imgs and save it as a gif

import imageio
imageio.mimsave('mlp.gif', imgs, fps=10)

Lossy conversion from float32 to uint8. Range [-13.466928482055664, 2.713646650314331]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.18658676743507385, 1.3069090843200684]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.18308542668819427, 1.0001248121261597]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.07874367386102676, 1.0167515277862549]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.09477106481790543, 1.0060935020446777]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.033188510686159134, 1.0109848976135254]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.0989738255739212, 1.0007272958755493]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.04943906515836716, 1.0269501209259033]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.02097826451063156, 1.0289174318313599]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.029821299016475677, 1.0194318294525146]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.016834549605846405, 1.0527536869049072]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.008144930005073547, 1.0191292762756348]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.009020708501338959, 1.0909096002578735]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.017141804099082947, 1.0371521711349487]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.013367637991905212, 1.0438421964645386]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0.0005456805229187012, 1.0179295539855957]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.011109575629234314, 1.0290166139602661]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.020140215754508972, 1.078523874282837]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.0396433025598526, 1.0415352582931519]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.015714898705482483, 1.0283904075622559]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.04321514815092087, 1.0413591861724854]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.04679575562477112, 1.067355990409851]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.003602549433708191, 1.0755447149276733]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.007610529661178589, 1.052262306213379]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.033921219408512115, 1.0815953016281128]. Convert image to uint8 prior to saving to suppress this warning.

Code

# Create a MLP with 5 hidden layers with 256 neurons each and sine activations.
# Input is (x, y) and output is (r, g, b)

class MLP_sin(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(2, 256)
        self.fc2 = nn.Linear(256, 256)
        self.fc3 = nn.Linear(256, 256)
        self.fc4 = nn.Linear(256, 256)
        self.fc5 = nn.Linear(256, 256)
        self.fc6 = nn.Linear(256, 3)

    def forward(self, x):
        x = torch.sin(self.fc1(x))
        x = torch.sin(self.fc2(x))
        x = torch.sin(self.fc3(x))
        x = torch.sin(self.fc4(x))
        x = torch.sin(self.fc5(x))
        return self.fc6(x)

Code

m2 = MLP_sin()
m2 = m2.to(device)
losses_mlp_sin, imgs = train(X, y, m2, lr=0.001, epochs=4000, bs=1000, print_every=100)

Epoch 0 loss: 0.40150442719459534
Epoch 100 loss: 0.03298206627368927
Epoch 200 loss: 0.033279214054346085
Epoch 300 loss: 0.03175220638513565
Epoch 400 loss: 0.03205806389451027
Epoch 500 loss: 0.03196191042661667
Epoch 600 loss: 0.02972976118326187
Epoch 700 loss: 0.029925711452960968
Epoch 800 loss: 0.02968132309615612
Epoch 900 loss: 0.028653116896748543
Epoch 1000 loss: 0.02474542148411274
Epoch 1100 loss: 0.020879685878753662
Epoch 1200 loss: 0.019819265231490135
Epoch 1300 loss: 0.016965048387646675
Epoch 1400 loss: 0.013934656977653503
Epoch 1500 loss: 0.011689499020576477
Epoch 1600 loss: 0.010081701911985874
Epoch 1700 loss: 0.007140354719012976
Epoch 1800 loss: 0.006480662152171135
Epoch 1900 loss: 0.005266484338790178
Epoch 2000 loss: 0.004757172428071499
Epoch 2100 loss: 0.003453798359259963
Epoch 2200 loss: 0.0032651633955538273
Epoch 2300 loss: 0.0028410402592271566
Epoch 2400 loss: 0.0026403532829135656
Epoch 2500 loss: 0.0019292739452794194
Epoch 2600 loss: 0.0021367412991821766
Epoch 2700 loss: 0.0020427301060408354
Epoch 2800 loss: 0.0017756932647898793
Epoch 2900 loss: 0.0016549285501241684
Epoch 3000 loss: 0.0016728530172258615
Epoch 3100 loss: 0.001471961266361177
Epoch 3200 loss: 0.0014844941906630993
Epoch 3300 loss: 0.0014798615593463182
Epoch 3400 loss: 0.0012664658715948462
Epoch 3500 loss: 0.0012708695139735937
Epoch 3600 loss: 0.0012460555881261826
Epoch 3700 loss: 0.0012855605455115438
Epoch 3800 loss: 0.001190435141324997
Epoch 3900 loss: 0.0011714434949681163

Code

plot_image(m2, "mlp_sin_dog.png")

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

Code

imageio.mimsave('mlp_sin.gif', imgs, fps=10)

Lossy conversion from float32 to uint8. Range [-0.1441832184791565, 0.3080734610557556]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [0, 1]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.056166477501392365, 0.9270500540733337]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.04645712673664093, 0.9617018103599548]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.08092432469129562, 0.9469475746154785]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.13254448771476746, 1.0228846073150635]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.18537408113479614, 1.0271779298782349]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.15940740704536438, 1.069307804107666]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.1629665046930313, 1.0901581048965454]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.17787247896194458, 1.164113163948059]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.23600360751152039, 1.1689845323562622]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.1829279065132141, 1.1432479619979858]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.12739746272563934, 1.1281737089157104]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.11645704507827759, 1.1141674518585205]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.11797109246253967, 1.1277530193328857]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.09862736612558365, 1.0859858989715576]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.1146015003323555, 1.099491834640503]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.09405502676963806, 1.1023061275482178]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.132747620344162, 1.0877472162246704]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.11511929333209991, 1.0887328386306763]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.11015606671571732, 1.0807398557662964]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.09713895618915558, 1.087331771850586]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.0733504444360733, 1.0549205541610718]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.07674040645360947, 1.0766404867172241]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.07997756451368332, 1.0550076961517334]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.09363748133182526, 1.056591510772705]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.08970168232917786, 1.0528484582901]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.08736599236726761, 1.04934561252594]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.08859498053789139, 1.0708154439926147]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.08006224036216736, 1.0856648683547974]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.08170387893915176, 1.071043610572815]. Convert image to uint8 prior to saving to suppress this warning.
Lossy conversion from float32 to uint8. Range [-0.06969650834798813, 1.0583616495132446]. Convert image to uint8 prior to saving to suppress this warning.

Code

# Audio
!wget https://www.vincentsitzmann.com/siren/img/audio/gt_bach.wav

--2023-04-28 14:24:10--  https://www.vincentsitzmann.com/siren/img/audio/gt_bach.wav
Resolving www.vincentsitzmann.com (www.vincentsitzmann.com)... 185.199.111.153, 185.199.108.153, 185.199.110.153, ...
Connecting to www.vincentsitzmann.com (www.vincentsitzmann.com)|185.199.111.153|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 1232886 (1.2M) [audio/wav]
Saving to: ‘gt_bach.wav.3’

gt_bach.wav.3       100%[===================>]   1.17M  --.-KB/s    in 0.06s   

2023-04-28 14:24:10 (19.7 MB/s) - ‘gt_bach.wav.3’ saved [1232886/1232886]

Code

# CLear CUDA cache
torch.cuda.empty_cache()

Code

from IPython.display import Audio
Audio('gt_bach.wav')

Code

# Read the audio file
import torchaudio
audio, sr = torchaudio.load('gt_bach.wav')

Code

sr

Code

audio.shape
audio = audio[0]
audio = audio.to(device)

Code

# use last 2 seconds of audio
audio = audio[-2 * sr:]
X = torch.arange(0, len(audio)).unsqueeze(1).float().to(device)

# Rescale X between -10 and 10
X = X / X.max() * 20 - 10
X.min(), X.max()

(tensor(-10., device='cuda:0'), tensor(10., device='cuda:0'))

Code

X.shape, audio.shape, X

(torch.Size([88200, 1]),
 torch.Size([88200]),
 tensor([[-10.0000],
         [ -9.9998],
         [ -9.9995],
         ...,
         [  9.9995],
         [  9.9998],
         [ 10.0000]], device='cuda:0'))

Code

Audio(audio.cpu(), rate=sr)

Code

class SinActivation(torch.nn.Module):
    def __init__(self):
        super(SinActivation, self).__init__()
        return
    def forward(self, x):
        return torch.sin(x)
    

class SinActivation30(torch.nn.Module):
    def __init__(self):
        super(SinActivation30, self).__init__()
        return
    def forward(self, x):
        return torch.sin(30*x)

Code

import torch.nn as nn

def create_mlp(n, m, f):
    """
    n: number of hidden layers
    m: number of neurons in each hidden layer
    f: activation function
    ---
    Weighing initialization: 
    uniform distribution between -30/input_dim and 30/input_dim for first layer
    -sqrt(6/input_dim) and sqrt(6/input_dim) for the rest

    Weight init is done in the forward pass
    """

    layers = []
    layer1 = nn.Linear(1, m)
    torch.nn.init.uniform_(layer1.weight, a=-1/1, b=1/1)
    #torch.nn.init.uniform_(layer1.bias, a=-1/1, b=1/1)
    layers.append(layer1)
    layers.append(SinActivation30())
    for i in range(n):
        layer_i = nn.Linear(m, m)
        # Uniform distribution between -sqrt(6/input_dim) and sqrt(6/input_dim)
        torch.nn.init.uniform_(layer_i.weight, a=-np.sqrt(6/m), b=np.sqrt(6/m))
        torch.nn.init.uniform_(layer_i.bias, a=-np.sqrt(6/m), b=np.sqrt(6/m))
        layers.append(layer_i)
        layers.append(f)
    layers.append(nn.Linear(m, 1))

    return nn.Sequential(*layers)

Code

mlp_audio_sin_5_256 = create_mlp(5, 256, SinActivation()).to(device)
#mlp_audio_sin_8_512 = create_mlp(8, 512, SinActivation()).to(device)
#mlp_audio_sin_3_128 = create_mlp(3, 128, SinActivation()).to(device)

Code

mlp_audio_sin_5_128

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb Cell 39 in <module>
----> <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X52sdnNjb2RlLXJlbW90ZQ%3D%3D?line=0'>1</a> mlp_audio_sin_5_128

NameError: name 'mlp_audio_sin_5_128' is not defined

Code

def train_audio(X, y, model, lr=0.01, epochs=1000, bs=1000, print_every=100):
    losses = []
    optimizer = torch.optim.Adam(model.parameters(), lr=lr)
    criterion = nn.MSELoss()
    for epoch in range(epochs):
        num_rows = X.shape[0]
        idx = torch.randperm(num_rows)[:bs]
        batch_X = X[idx]
        batch_y = y[idx]
        pred_y = model(batch_X)

        # Compute the loss
        loss = criterion(pred_y, batch_y)

        # Zero gradients, perform a backward pass, and update the weights
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        losses.append(loss.item())

        # Print loss every print_every epochs
        if epoch % print_every == 0:
            print(f"Epoch {epoch} loss: {loss.item()}")

    return losses

Code

#losses_mlp_sin_3_128 = train_audio(X, audio, mlp_audio_sin_3_128, lr=0.0001,
#                                    epochs=5000, bs=len(X)//2, print_every=100)

losses_mlp_sin_5_256 = train_audio(X, audio, mlp_audio_sin_5_256, lr=0.0001,
                                    epochs=5000, bs=len(X)//2, print_every=100)

Epoch 0 loss: 0.210729718208313

---------------------------------------------------------------------------
OutOfMemoryError                          Traceback (most recent call last)
/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb Cell 41 in <module>
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=0'>1</a> #losses_mlp_sin_3_128 = train_audio(X, audio, mlp_audio_sin_3_128, lr=0.0001,
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=1'>2</a> #                                    epochs=5000, bs=len(X)//2, print_every=100)
----> <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=3'>4</a> losses_mlp_sin_5_256 = train_audio(X, audio, mlp_audio_sin_5_256, lr=0.0001,
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=4'>5</a>                                     epochs=5000, bs=len(X)//2, print_every=100)

/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb Cell 41 in train_audio(X, y, model, lr, epochs, bs, print_every)
     <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=9'>10</a> pred_y = model(batch_X)
     <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=11'>12</a> # Compute the loss
---> <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=12'>13</a> loss = criterion(pred_y, batch_y)
     <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=14'>15</a> # Zero gradients, perform a backward pass, and update the weights
     <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X54sdnNjb2RlLXJlbW90ZQ%3D%3D?line=15'>16</a> optimizer.zero_grad()

File ~/miniforge3/lib/python3.9/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/miniforge3/lib/python3.9/site-packages/torch/nn/modules/loss.py:536, in MSELoss.forward(self, input, target)
    535 def forward(self, input: Tensor, target: Tensor) -> Tensor:
--> 536     return F.mse_loss(input, target, reduction=self.reduction)

File ~/miniforge3/lib/python3.9/site-packages/torch/nn/functional.py:3295, in mse_loss(input, target, size_average, reduce, reduction)
   3292     reduction = _Reduction.legacy_get_string(size_average, reduce)
   3294 expanded_input, expanded_target = torch.broadcast_tensors(input, target)
-> 3295 return torch._C._nn.mse_loss(expanded_input, expanded_target, _Reduction.get_enum(reduction))

OutOfMemoryError: CUDA out of memory. Tried to allocate 7.25 GiB (GPU 0; 79.18 GiB total capacity; 63.06 GiB already allocated; 7.88 MiB free; 74.24 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Code

tensor([[-1.7320],
        [-1.7320],
        [-1.7319],
        ...,
        [ 1.7319],
        [ 1.7320],
        [ 1.7320]], device='cuda:0')

Code

import time

Code

a = time.time()
losses_mlp_sin_8_512 = train_audio(X, audio, mlp_audio_sin_8_512, 
                                   lr=0.0001, epochs=10, bs=len(X), print_every=1)
b = time.time()
print(b-a)

---------------------------------------------------------------------------
OutOfMemoryError                          Traceback (most recent call last)
/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb Cell 43 in <module>
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=0'>1</a> a = time.time()
----> <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=1'>2</a> losses_mlp_sin_8_512 = train_audio(X, audio, mlp_audio_sin_8_512, 
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=2'>3</a>                                    lr=0.0001, epochs=10, bs=len(X), print_every=1)
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=3'>4</a> b = time.time()
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=4'>5</a> print(b-a)

/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb Cell 43 in train_audio(X, y, model, lr, epochs, bs, print_every)
     <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=9'>10</a> pred_y = model(batch_X)
     <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=11'>12</a> # Compute the loss
---> <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=12'>13</a> loss = criterion(pred_y, batch_y)
     <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=14'>15</a> # Zero gradients, perform a backward pass, and update the weights
     <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X56sdnNjb2RlLXJlbW90ZQ%3D%3D?line=15'>16</a> optimizer.zero_grad()

File ~/miniforge3/lib/python3.9/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/miniforge3/lib/python3.9/site-packages/torch/nn/modules/loss.py:536, in MSELoss.forward(self, input, target)
    535 def forward(self, input: Tensor, target: Tensor) -> Tensor:
--> 536     return F.mse_loss(input, target, reduction=self.reduction)

File ~/miniforge3/lib/python3.9/site-packages/torch/nn/functional.py:3295, in mse_loss(input, target, size_average, reduce, reduction)
   3292     reduction = _Reduction.legacy_get_string(size_average, reduce)
   3294 expanded_input, expanded_target = torch.broadcast_tensors(input, target)
-> 3295 return torch._C._nn.mse_loss(expanded_input, expanded_target, _Reduction.get_enum(reduction))

OutOfMemoryError: CUDA out of memory. Tried to allocate 28.98 GiB (GPU 0; 79.18 GiB total capacity; 33.40 GiB already allocated; 14.51 GiB free; 59.74 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Code

mlp_audio_sin_8_512 = torch.compile(mlp_audio_sin_8_512)

Code

a = time.time()
losses_mlp_sin_8_512 = train_audio(X, audio, mlp_audio_sin_8_512, 
                                   lr=0.0001, epochs=10, bs=len(X), print_every=1)
b = time.time()
print(b-a)

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb Cell 44 in <module>
----> <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X61sdnNjb2RlLXJlbW90ZQ%3D%3D?line=0'>1</a> a = time.time()
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X61sdnNjb2RlLXJlbW90ZQ%3D%3D?line=1'>2</a> losses_mlp_sin_8_512 = train_audio(X, audio, mlp_audio_sin_8_512, 
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X61sdnNjb2RlLXJlbW90ZQ%3D%3D?line=2'>3</a>                                    lr=0.0001, epochs=10, bs=len(X), print_every=1)
      <a href='vscode-notebook-cell://ssh-remote%2B10.0.62.168/home/nipun.batra/git/blog/posts/siren-paper-implementation.ipynb#X61sdnNjb2RlLXJlbW90ZQ%3D%3D?line=3'>4</a> b = time.time()

NameError: name 'time' is not defined

Code

# Plot the reconstruction
with torch.no_grad():
    #pred_y_5_256 = mlp_audio_sin_5_256(X)
    #pred_y_8_512 = mlp_audio_sin_8_512(X)
    pred_y_3_128 = mlp_audio_sin_3_128(X)
    plt.plot(audio.cpu().numpy(), label="Ground truth")
    #plt.plot(pred_y_5_256.cpu().numpy(), label="MLP 5 layers 256 neurons")
    plt.plot(pred_y_3_128.cpu().numpy(), label="MLP 8 layers 512 neurons")
    plt.legend()

Code

import pandas as pd

df = pd.DataFrame({"GT audio": audio.cpu().numpy(), 
                   "MLP 5 layers 256 neurons": pred_y_5_256.cpu().numpy().flatten(), 
                   "MLP 8 layers 512 neurons": pred_y_8_512.cpu().numpy().flatten()})

Code

df.describe()

	GT audio	MLP 5 layers 256 neurons	MLP 8 layers 512 neurons
count	88200.000000	88200.000000	88200.000000
mean	0.000127	-0.013929	-0.010819
std	0.208728	0.025773	0.156109
min	-0.868308	-0.083747	-0.710084
25%	-0.130095	-0.030821	-0.116540
50%	-0.002093	-0.011080	-0.010339
75%	0.130701	0.002974	0.094733
max	1.000000	0.051832	0.658187

Code

audio.shape, pred_y_8_512.shape

(torch.Size([88200]), torch.Size([88200, 1]))

Code

# Play the reconstruction
Audio(pred_y_8_512.cpu().T, rate=sr)

TODO

Show the gradient of the reconstructed image for different activation functions