bridgetower-newyorker-gaudi2-8x

Disclaimer: Those are the results I obtained with an older version of Optimum Habana. With v1.7, it is possible to fit batches of 48 samples and to get better throughput as mentioned in this blog post: https://huggingface.co/blog/bridgetower

This model is a fine-tuned version of BridgeTower/bridgetower-large-itm-mlm-itc on the jmhessel/newyorker_caption_contest matching dataset. It achieves the following results on the evaluation set:

• Loss: 0.1147
• Memory Allocated (gb): 20.01
• Max Memory Allocated (gb): 83.39
• Total Memory Available (gb): 93.74

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 40
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 320
• total_eval_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-06
• lr_scheduler_type: linear
• num_epochs: 5.0

Training results

Framework versions

• Transformers 4.28.1
• Pytorch 2.0.1a0+git37b7ddc
• Datasets 2.13.1
• Tokenizers 0.13.3