Problem in reproduce

[HITKJ]

Thans for your excellent work. I have attempted to reproduce your results, and below are the comparison results between the reproduced model and the model you provided.There is a big gap between my results and yours. Hope you can provide some guidance.

I collected data on four different maps (town01, town03, town04, town06) at 10 FPS for 25 episodes on each map. In each episode, I collected 3000 images, resulting in nearly 300,000 images for training data. During training, I used a batch size of 64 for a total of 50,000 iterations at 10 FPS and also tested at 10 FPS. Additionally, I used 1000 batches validation set at every 3000 iteration. The validation data setup was consistent with the simulation test, and the validation metrics showed a decreasing trend during the iterations.

Another question, the test time is too long, can we use multiple gpus for acceleration.