Compute the policy and value heads from a shared PyTorch feature tensor F of shape (B, C, H, W), where B is batch size, C is channel count, and H and W are spatial dimensions.

Here, Wpc and Wvc are policy_conv_weight and value_conv_weight. Wp and bp are policy_linear_weight and policy_linear_bias; Wh and bh are value_hidden_weight and value_hidden_bias; Wv and bv are value_output_weight and value_output_bias. L contains raw action logits and v contains current-player values. Each convolution uses stride 1 with no padding or bias. Flattening preserves the batch dimension and orders entries by channel, row, then column.

For an activation z in one channel, batch normalization is

Here, mu and sigma squared are that channel's mean and population variance over the batch and spatial dimensions. Gamma and beta are its supplied scale and shift, and epsilon is the supplied positive normalization constant.

The policy branch uses policy_scale and policy_shift; the value branch uses value_scale and value_shift. Return a tuple of two PyTorch float32 tensors: policy logits of shape (B, H times W plus one) and values of shape (B,). Both outputs stay on the feature device; the final policy entry is pass.