RuntimeError: The size of tensor a (2) must match the size of tensor b (9833) at non-singleton dimension

[dpappas]

I followed your instructions using my data. Since the batch_size was too big for my data i changed it to 6.

Then i got this error during evaluation:

08/23/2019 17:50:14 - INFO - root - Running evaluation---------------------------------------------------------| 0.82% [49/5955 00:37<1:15:53] 08/23/2019 17:50:14 - INFO - root - Num examples = 9833 08/23/2019 17:50:14 - INFO - root - Batch size = 6 Traceback (most recent call last): File "train_fast_bert_doc_rerank.py", line 81, in <module> optimizer_type="lamb" File "/usr/local/lib/python3.6/site-packages/fast_bert/learner_cls.py", line 295, in fit results = self.validate() File "/usr/local/lib/python3.6/site-packages/fast_bert/learner_cls.py", line 382, in validate validation_scores[metric['name']] = metric['function'](all_logits, all_labels) File "/usr/local/lib/python3.6/site-packages/fast_bert/metrics.py", line 31, in accuracy_thresh return ((y_pred > thresh) == y_true.byte()).float().mean().item() RuntimeError: The size of tensor a (2) must match the size of tensor b (9833) at non-singleton dimension Could you help me? Thank you in advance

[kaushaltrivedi]

are you trying out multi-label classification?

[dpappas]

I am using binary classification as in you article in medium. Kudos by the way.

I managed to create my own functions selecting always the second index. So wright now i am ok.

[ghost]

I'm getting a similar error when using the F1 score in a binary classification problem:

RuntimeError                              Traceback (most recent call last)
<ipython-input-9-eea47d3f12a7> in <module>
      3             validate=True,      # Evaluate the model after each epoch
      4             schedule_type="warmup_cosine",
----> 5             optimizer_type="lamb")

/disk/no_backup/kd_challenge/fast-bert/fast_bert/learner_cls.py in fit(self, epochs, lr, validate, schedule_type, optimizer_type)
    405                         if validate:
    406                             # evaluate model
--> 407                             results = self.validate()
    408                             for key, value in results.items():
    409                                 tb_writer.add_scalar(

/disk/no_backup/kd_challenge/fast-bert/fast_bert/learner_cls.py in validate(self)
    527         for metric in self.metrics:
    528             validation_scores[metric["name"]] = metric["function"](
--> 529                 all_logits, all_labels
    530             )
    531 

/disk/no_backup/kd_challenge/fast-bert/fast_bert/metrics.py in F1(y_pred, y_true, threshold)
     79 
     80 def F1(y_pred: Tensor, y_true: Tensor, threshold: float = CLASSIFICATION_THRESHOLD):
---> 81     return fbeta(y_pred, y_true, thresh=threshold, beta=1)

/disk/no_backup/kd_challenge/fast-bert/fast_bert/metrics.py in fbeta(y_pred, y_true, thresh, beta, eps, sigmoid)
     40     y_pred = (y_pred > thresh).float()
     41     y_true = y_true.float()
---> 42     TP = (y_pred*y_true).sum(dim=1)
     43     prec = TP/(y_pred.sum(dim=1)+eps)
     44     rec = TP/(y_true.sum(dim=1)+eps)

RuntimeError: The size of tensor a (2) must match the size of tensor b (4671) at non-singleton dimension 1

The code used looks as follows:

import sys
import torch
import logging
import datetime
from pathlib import Path
from transformers import BertTokenizer
from fast_bert.data_cls import BertDataBunch
from fast_bert.learner_cls import BertLearner
from fast_bert.metrics import accuracy, F1

DATA_PATH = Path('./bert/data/')
LABEL_PATH = Path('./bert/labels/')
MODEL_PATH=Path('./bert/models/')
LOG_PATH=Path('./bert/logs/')
OUTPUT_PATH = MODEL_PATH/'output'
MODEL_PATH.mkdir(exist_ok=True)
LOG_PATH.mkdir(exist_ok=True)
OUTPUT_PATH.mkdir(exist_ok=True)
model_state_dict = None
FINETUNED_PATH = None
FINETUNED_PATH = Path('./bert/lm_model_bert/model_out/pytorch_model.bin')
model_state_dict = torch.load(FINETUNED_PATH)
BERT_PRETRAINED_PATH = Path('./bert/wwm_uncased_L-24_H-1024_A-16/')

run_start_time = datetime.datetime.today().strftime('%Y-%m-%d_%H-%M-%S')
logfile = str(LOG_PATH/'log-{}-{}.txt'.format(run_start_time, "roberta_bs64"))
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
    datefmt='%m/%d/%Y %H:%M:%S',
    handlers=[
        logging.FileHandler(logfile),
        logging.StreamHandler(sys.stdout)])
logger = logging.getLogger()

databunch = BertDataBunch(DATA_PATH, LABEL_PATH,
                          tokenizer='roberta-base',
                          train_file='train.csv',
                          val_file='val.csv',
                          label_file='labels.csv',
                          text_col='text',
                          label_col='label',
                          batch_size_per_gpu=64,
                          max_seq_length=64,
                          multi_gpu=True,
                          multi_label=False,
                          model_type='roberta')

device_cuda = torch.device("cuda")
metrics = [{'name': 'f1_score', 'function': F1}]

learner = BertLearner.from_pretrained_model(
    databunch,
    pretrained_path='roberta-base',
    metrics=metrics,
    device=device_cuda,
    logger=logger,
    output_dir=OUTPUT_PATH,
    finetuned_wgts_path=None,
    warmup_steps=500,
    multi_gpu=False,
    is_fp16=False,
    multi_label=False,
    logging_steps=50)

learner.fit(epochs=10,
            lr=6e-5,
            validate=True,  # Evaluate the model after each epoch
            schedule_type="warmup_cosine",
            optimizer_type="lamb")

Any suggestions?