An free and open source community implementation of Chai-1 in PyTorch. Paper is here
Join our discord to help us implement this paper!
pip3 install chai-one######### example.py
import torch
from loguru import logger
from chai_one.model import ChaiOne
# Set up model parameters
dim_single = 128
dim_pairwise = 128
dim_msa = 128
dim_msa_input = 134 # Adjusted to match the expected input dimension
dim_additional_msa_feats = 2
window_size = 25
# Initialize the model
logger.info("Initializing ChaiOne model")
model = ChaiOne(
dim_single=dim_single,
dim_pairwise=dim_pairwise,
msa_depth=4,
dim_msa=dim_msa,
dim_msa_input=dim_msa_input, # Set to 134
dim_additional_msa_feats=0,
msa_pwa_heads=8,
msa_pwa_dim_head=32,
layerscale_output=False,
heads=8,
window_size=window_size,
num_memory_kv=0,
attn_layers=48,
)
# Create dummy input tensors
batch_size = 1
seq_length = 100
num_msa = 4
logger.info(
f"Creating input tensors with shape: batch_size={batch_size}, seq_length={seq_length}, num_msa={num_msa}"
)
single_repr = torch.randn(batch_size, seq_length, dim_single)
pairwise_repr = torch.randn(
batch_size, seq_length, seq_length, dim_pairwise
)
# Create msa tensor with matching input size for msa_init_proj (134 features)
msa = torch.randn(
batch_size, num_msa, seq_length, dim_msa_input
) # Adjusted to 134
# Forward pass
logger.info("Performing forward pass")
output = model(
single_repr=single_repr,
pairwise_repr=pairwise_repr,
msa=msa,
)
logger.info(f"Output shape: {output.shape}")MIT