microsoft

/

phi-2

like 3.23k

Hello! I want to train phi2 with AutoModelForSequenceClassification. But there is error like following code.

base_model = AutoModelForSequenceClassification.from_pretrained(
    "microsoft/phi-2", 
    num_labels=2,
    device_map={"":0},
    trust_remote_code=True,
)
base_model.config.pretraining_tp = 1 # 1 is 7b
base_model.config.pad_token_id = tokenizer.pad_token_id

# Output
ValueError: Unrecognized configuration class <class 'transformers_modules.phi-2.configuration_phi.PhiConfig'> for this kind of AutoModel: AutoModelForSequenceClassification.
Model type should be one of AlbertConfig, BartConfig, BertConfig, BigBirdConfig, BigBirdPegasusConfig, BioGptConfig, BloomConfig, CamembertConfig, CanineConfig, LlamaConfig, ConvBertConfig, CTRLConfig, Data2VecTextConfig, DebertaConfig, DebertaV2Config, DistilBertConfig, ElectraConfig, ErnieConfig, ErnieMConfig, EsmConfig, FalconConfig, FlaubertConfig, FNetConfig, FunnelConfig, GPT2Config, GPT2Config, GPTBigCodeConfig, GPTNeoConfig, GPTNeoXConfig, GPTJConfig, IBertConfig, LayoutLMConfig, LayoutLMv2Config, LayoutLMv3Config, LEDConfig, LiltConfig, LlamaConfig, LongformerConfig, LukeConfig, MarkupLMConfig, MBartConfig, MegaConfig, MegatronBertConfig, MistralConfig, MixtralConfig, MobileBertConfig, MPNetConfig, MptConfig, MraConfig, MT5Config, MvpConfig, NezhaConfig, NystromformerConfig, OpenLlamaConfig, OpenAIGPTConfig, OPTConfig, PerceiverConfig, PersimmonConfig, PhiConfig, PLBartConfig, QDQBertConfig, ReformerConfig, RemBertConfig, RobertaConfig, RobertaPreLayerNormConfig, RoCBertConfig, RoFormerConfig, SqueezeBertConfig, T5Config, TapasConfig, TransfoXLConfig, UMT5Config, XLMConfig, XLMRobertaConfig, XLMRobertaXLConfig, XLNetConfig, XmodConfig, YosoConfig.

So i tried train phi2 with PhiForSequenceClassification. Use with following code.

base_model = PhiForSequenceClassification.from_pretrained(
    "microsoft/phi-2", 
    num_labels=2,
    device_map={"":0},
    trust_remote_code=True,
)
base_model.config.pretraining_tp = 1 # 1 is 7b
base_model.config.pad_token_id = tokenizer.pad_token_id

But the model output is nan. and then eval loss na too.

# Output
{'eval_loss': nan, 'eval_accuracy': 0.3618421052631579, 'eval_roc_auc': 0.5, 'eval_runtime': 19.2035, 'eval_samples_per_second': 39.576, 'eval_steps_per_second': 19.788, 'epoch': 0.03}                                                        
{'eval_loss': nan, 'eval_accuracy': 0.3618421052631579, 'eval_roc_auc': 0.5, 'eval_runtime': 19.2699, 'eval_samples_per_second': 39.44, 'eval_steps_per_second': 19.72, 'epoch': 0.04}       

I tried change TrainingArguments about fp16 argument and model arguments about torch_dtype. But It still same problem.
Can you give me some advice for training PhiForSequenceClassification?
Thank you :)

# Trial 1
training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    learning_rate=1e-6,
    ...
    fp16=True
    # bf16=True,
)

# Trial 2
base_model = PhiForSequenceClassification.from_pretrained(
    "microsoft/phi-2", 
    num_labels=2,
    device_map={"":0},
    trust_remote_code=True,
    torch_dtype=torch.float16
)