sonoisa
/

clip-vit-b-32-japanese-v1

Feature Extraction Transformers PyTorch Japanese bert clip sentence-similarity Inference Endpoints

Model card Files Files and versions Community

clip-vit-b-32-japanese-v1 / training_args-1.json

sonoisa's picture

Update text encoder model

6e08a41 about 2 years ago

raw history blame contribute delete

No virus

751 Bytes

	{
	"adam_epsilon": 1e-08,
	"data_dir": "/content/data",
	"early_stop_callback": false,
	"en_ja_ratio": 4,
	"eval_batch_size": 32,
	"fp_16": false,
	"freeze_bert": true,
	"from_pretrained_bert": true,
	"gradient_accumulation_steps": 1,
	"learning_rate": 0.001,
	"lr_initial_epoch": 0,
	"lr_total_epochs": 4,
	"lr_warmup_epochs": 0.05,
	"max_cls_depth": 6,
	"max_grad_norm": 1.0,
	"max_input_length": 64,
	"model_name_or_path": "sonoisa/sentence-bert-base-ja-mean-tokens-v2",
	"n_gpu": 1,
	"num_train_epochs": 4,
	"seed": 42,
	"shuffle_buffer_size": 65536,
	"tokenizer_name_or_path": "sonoisa/sentence-bert-base-ja-mean-tokens-v2",
	"train_batch_size": 2560,
	"weight_decay": 0.0
	}