diffusers/tools

1#!/usr/bin/env python32import torch3import torch.nn as nn4import torch.nn.functional as F5 6 7class SuperConv(nn.Conv2d):8 9    def __init__(self, *args, is_lora=False, **kwargs):10        super().__init__(*args, **kwargs)11 12        self.is_lora = is_lora13 14    def forward(self, *args, **kwargs):15        if self.is_lora:16            return 3 + super().forward(*args, **kwargs)17        else:18            return super().forward(*args, **kwargs)19 20# Define a simple Convolutional Neural Network21class SimpleCNN(nn.Module):22    def __init__(self):23        super(SimpleCNN, self).__init__()24        self.conv1 = SuperConv(3, 6, 5) # Assuming input images are RGB, so 3 input channels25        self.pool = nn.MaxPool2d(2, 2)26        self.conv2 = SuperConv(6, 16, 5)27        self.fc1 = nn.Linear(16 * 5 * 5, 120)28        self.fc2 = nn.Linear(120, 84)29        self.fc3 = nn.Linear(84, 10)30 31    def forward(self, x):32        x = self.pool(F.relu(self.conv1(x)))33        x = self.pool(F.relu(self.conv2(x)))34        x = x.view(-1, 16 * 5 * 5)35        x = F.relu(self.fc1(x))36        x = F.relu(self.fc2(x))37        x = self.fc3(x)38        return x39 40# Create the network41net = SimpleCNN()42 43# Initialize weights with dummy values44for m in net.modules():45    if isinstance(m, nn.Conv2d):46        nn.init.constant_(m.weight, 0.1)47        nn.init.constant_(m.bias, 0.1)48    elif isinstance(m, nn.Linear):49        nn.init.constant_(m.weight, 0.1)50        nn.init.constant_(m.bias, 0.1)51 52# Perform inference53input = torch.randn(1, 3, 32, 32).to("cuda")54net = net.to("cuda")55output = net(input)56 57print(output)58 59net = torch.compile(net, mode="reduce-overhead", fullgraph=True)60 61output = net(input)62 63print(output)64