# A tool created by ChatGPT to generate PyTorch tensors and perform operations, then print them as C++ 1D float arrays. import torchimport argparse # Define the argument parserparser = argparse.ArgumentParser(description="Generate PyTorch tensors and perform operations, then print them as C++ 1D float arrays.")parser.add_argument("expressions", type=str, nargs='+', help='Tensor expressions like "a(2, 3, 2)" "b(2, 4, 2)" "c(2, 3, 4)" "a @ b + c"') # Parse the command-line argumentsargs = parser.parse_args() # Function to parse tensor definitions and operationsdef parse_expression(expression):    if '(' in expression:        name, dims = expression.split('(')        dims = tuple(map(int, dims.rstrip(')').split(',')))        return name, dims    return expression # Create tensorstensors = {}dimensions = {}for expr in args.expressions:    parsed = parse_expression(expr)    if isinstance(parsed, tuple):        name, dims = parsed        tensors[name] = torch.rand(dims, requires_grad=True) * 2 - 1  # Random values between -1 and 1        tensors[name].retain_grad()        dimensions[name] = dims # Perform operationsdef evaluate_operation(operation):    if '@' in operation and '+' in operation:        left, rest = operation.split('@')        right, addend = rest.split('+')        left = left.strip()        right = right.strip()        addend = addend.strip()        result = torch.matmul(tensors[left], tensors[right].transpose(-2, -1)) + tensors[addend]    elif '+' in operation:        left, right = operation.split('+')        left = left.strip()        right = right.strip()        result = tensors[left] + tensors[right]    elif '-' in operation:        left, right = operation.split('-')        left = left.strip()        right = right.strip()        result = tensors[left] - tensors[right]    elif '*' in operation:        left, right = operation.split('*')        left = left.strip()        right = right.strip()        result = tensors[left] * tensors[right]    elif '/' in operation:        left, right = operation.split('/')        left = left.strip()        right = right.strip()        result = tensors[left] / tensors[right]    elif '@' in operation:        left, right = operation.split('@')        left = left.strip()        right = right.strip()        result = torch.matmul(tensors[left], tensors[right].transpose(-2, -1))    result.retain_grad()    return result operations = [expr for expr in args.expressions if any(op in expr for op in ['+', '-', '*', '/', '@'])]for operation in operations:    result = evaluate_operation(operation)    result_name = "result"    tensors[result_name] = result    dimensions[result_name] = result.shape # Perform backward passtensors["result"].sum().backward() # Print tensors and gradients as C++ 1D float arraysfor name, tensor in tensors.items():    tensor_1d = tensor.view(-1).detach().numpy()    dims = dimensions[name]    if name == "result":        print(f"std::vector<int> result_dims = {{{', '.join(map(str, dims))}}};")        print(f"float result_data[] = {{")    else:        print(f"std::vector<int> {name}_dims = {{{', '.join(map(str, dims))}}};")        print(f"float {name}_data[] = {{")    print(", ".join(f"{x:.6f}f" for x in tensor_1d))    print("};")     # Print gradients if available    if tensor.grad is not None:        grad_1d = tensor.grad.view(-1).detach().numpy()        print(f"float {name}_grad_data[] = {{")        print(", ".join(f"{x:.6f}f" for x in grad_1d))        print("};")