Error: name 'pos_weight' is not defined (multi-label classification)

[zrajabi]

I want to train multi-label classifier, however, I get the following error, although from what I have read, we should set pos_weight and set multi_label to True, Could you please help me what I should do?

NameError Traceback (most recent call last) in () 93 94 args = parser.parse_args() ---> 95 train(args)

in train(args) 70 multi_label=True, 71 ---> 72 logging_steps=10) 73 74 #learner.lr_find(start_lr=1e-5,optimizer_type='lamb')

/home/ec2-user/anaconda3/envs/pytorch_latest_p36/lib/python3.6/site-packages/fast_bert/learner_cls.py in from_pretrained_model(dataBunch, pretrained_path, output_dir, metrics, device, logger, finetuned_wgts_path, multi_gpu, is_fp16, loss_scale, warmup_steps, fp16_opt_level, grad_accumulation_steps, multi_label, max_grad_norm, adam_epsilon, logging_steps, freeze_transformer_layers, pos_weight, weight) 194 195 model = load_model( --> 196 dataBunch, pretrained_path, finetuned_wgts_path, device, multi_label 197 ) 198

/home/ec2-user/anaconda3/envs/pytorch_latest_p36/lib/python3.6/site-packages/fast_bert/learner_cls.py in load_model(dataBunch, pretrained_path, finetuned_wgts_path, device, multi_label) 144 config_class, model_class, _ = MODEL_CLASSES[model_type] 145 --> 146 model_class[1].pos_weight = pos_weight 147 model_class[1].weight = weight 148

NameError: name 'pos_weight' is not defined

Here is my code:

def train(args): if args.is_onepanel: args.out_dir = os.path.join("/onepanel/output/",args.out_dir) if not os.path.exists(args.out_dir): os.mkdir(args.out_dir)

pos_weight = torch.ones([11])

device_cuda = torch.device("cuda")
if torch.cuda.device_count() > 1:
    multi_gpu = True
else:
    multi_gpu = False

logger = logging.getLogger()

databunch = BertDataBunch(".", ".",
    tokenizer=args.pretrained_model,
    train_file = TRAIN_DATA_FILE,
    label_file= label_file,
    val_file= VALIDATION_DATA_FILE,
    text_col='Tweet',
    label_col=labels,
                          
    pos_weight = pos_weight,
                          
    batch_size_per_gpu = 32, #args.batch_size,
    max_seq_length=512,
    multi_gpu = multi_gpu,
    multi_label=True,
    model_type='bert')

metrics = [{'name': 'accuracy', 'function': accuracy}]

learner = BertLearner.from_pretrained_model(
    databunch,
    pretrained_path=args.pretrained_model,
    metrics=metrics,
    device=device_cuda,
    logger=logger,
    finetuned_wgts_path=None,
    output_dir=args.out_dir,
    pos_weight = pos_weight,
    warmup_steps=200,
    multi_gpu = multi_gpu,
    is_fp16= True,
    multi_label=True,
    logging_steps=10)


learner.fit(epochs=args.epochs,
        lr=2e-3,
        pos_weight = pos_weight,
        schedule_type="warmup_cosine_hard_restarts",
        optimizer_type="lamb")

learner.save_model()

[washcycle]

Just applied a fix for this in #272