To attain the most straightforward representation for the input, we represent object trajectories and maps using sequences of line segments and type.
1# model
2dim = 96 #5 #in vector space
3n_layers = 6
4n_heads = 12 #5
5n_kv_heads = 12 #5
6multiple_of = 32
7dropout = 0.0
8
9# data
10batch_size = 64 # if gradient_accumulation_steps > 1, this is the micro-batch size
11input_length = 244
12input_width = 5 # input 3->30 # the Llama 2 tokenizer has 32K tokens #output size
Source code to run this model can be found at
argo_llama
