Import pytorch and use hdl coder to create system C code

I created a deep learning mode on using pytorch using the following script

class RegressionANN(nn.Module):

def __init__(self):

super().__init__()

self.linear1 = nn.Linear(8, 64) # Increased neurons to 64

self.linear2 = nn.Linear(64, 1) # Changed input to 64

def forward(self, x):

x = torch.relu(self.linear1(x))

x = self.linear2(x)

return x

model = RegressionANN()

loss_fn = nn.MSELoss()

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

# Forward pass

y_pred = model(X_train_tensor)

# Calculate initial loss

initial_loss = loss_fn(y_pred, y_train_tensor)

print(f"Initial Loss: {initial_loss.item()}")

epochs = 100

for epoch in range(epochs):

# Zero gradients

optimizer.zero_grad()

# Forward pass

y_pred = model(X_train_tensor)

# Compute loss

loss = loss_fn(y_pred, y_train_tensor)

# Backpropagation

loss.backward()

# Update parameters

optimizer.step()

if (epoch + 1) % 10 == 0:

print(f'Epoch [{epoch+1}/{epochs}], Loss: {loss.item():.4f}')

print(f"Loss after {epochs} iterations: {loss.item():.4f}")

Following that I did the following

X = torch.rand(1, 8)

traced_model = torch.jit.trace(model.forward, X)

traced_model.save("traced_linear_model.pt")

Model was saved successfully. I tried importing the model using

importNetworkFromPyTorch

How do i proceed with HDL code generation ?

Importing created a matlab .m file for each layer.

Which matlab file should I use for code generation ?

Answer 1

[Dev]

Hi @Amogh,

To generate System C code using the HDL Coder application in MATLAB R2024b, this application requires two basic files to proceed-

• The top-level function containing the model functionality.
• A test bench MATLAB script file to drive the above function.

For more information on the same, please refer to the HDL code generation using HDL Coder documentation below-https://www.mathworks.com/help/releases/r2024b/visionhdl/ug/generate-hdl-code-from-matlab.html

The code provided in the question uses the ‘importNetworkFromPyTorch’ function to import the Pytorch model, which also generates the .M files for each custom layer in the model. However, this function does not generate dedicated top-level function or test bench files for the model, a necessary requirement in the HDL Coder workflow. You can refer more on this function below-

https://www.mathworks.com/help/releases/R2024b/deeplearning/ref/importnetworkfrompytorch.html

A possible workaround to generate the System C code for this Pytorch model is to use a combination of the Deep Network Designer and the MATLAB Coder applications available in MATLAB R2024b. The Deep Network Designer app can help us generate the corresponding MATLAB code for the functionality of our model. You can refer more on the usage of this application at the link below-

https://www.mathworks.com/help/releases/R2024b/deeplearning/ref/deepnetworkdesigner-app.html

Next, the generated MATLAB code can then be provided to the MATLAB Coder app to generate the corresponding C code. For more information on the usage of this application, please refer to the link below-https://www.mathworks.com/help/releases/r2024b/coder/ref/matlabcoder-app.html

I hope the above workaround resolves the query.

[Amogh]

OK. Thanks.

Btw is it possible to connect to cadence stratus running on a server? Or should it compulsorily be on my local machine ?