About devices for training.

[MSunDYY]

Can you give your experimental devices? And how much GPU memory needed for training?

[yuyang-cloud]

Hi, We train our model on 8*A100 GPUs for approximately 2~3 days. Each GPU requires around 70 GB of memory. You may also try to tune the following hyperparameters to achieve a balance between memory usage and performance:

queue_length = 2  # Number of historical frames as input
future_pred_frame_num_train = 4  # Number of future frames to predict
future_decoder_layer_num
pts_bbox_head.transformer.num_layers