# ValueError: could not broadcast input array from shape (30,512,32128) into shape (30,512) **URL:** \ **Category:** 🤗Transformers\ **Created:** [February 13, 2024, 6:25pm UTC](https://discuss.huggingface.co/t/valueerror-could-not-broadcast-input-array-from-shape-30-512-32128-into-shape-30-512/73040 "2024-02-13T18:25:48Z")\ **Posts on this page:** 3\ **Page:** 1
**Author:** ![llmtrails](https://avatars.discourse-cdn.com/v4/letter/l/e9bcb4/32.png) [@llmtrails](https://discuss.huggingface.co/u/llmtrails)\ **Post date:** [February 13, 2024, 6:25pm UTC](https://discuss.huggingface.co/t/valueerror-could-not-broadcast-input-array-from-shape-30-512-32128-into-shape-30-512/73040/1 "2024-02-13T18:25:48Z")
I am getting this error when I run the training on flan-t5-base llm. Using Hugging face transformers. I tried multiple options to fix this compute\_matrics but I am hitting a brick wall. I am a high school user of transformer for a science project. I am using google colab pro version of my parent. Error I am getting is ValueError Traceback (most recent call last) in \() 1 # Start training ----\> 2 trainer.train() 9 frames /usr/local/lib/python3.10/dist-packages/transformers/trainer.py in train(self, resume\_from\_checkpoint, trial, ignore\_keys\_for\_eval, \*\*kwargs) 1537 hf\_hub\_utils.enable\_progress\_bars() 1538 else: → 1539 return inner\_training\_loop( 1540 args=args, 1541 resume\_from\_checkpoint=resume\_from\_checkpoint, /usr/local/lib/python3.10/dist-packages/transformers/trainer.py in \_inner\_training\_loop(self, batch\_size, args, resume\_from\_checkpoint, trial, ignore\_keys\_for\_eval) 1942 1943 self.control = self.callback\_handler.on\_epoch\_end(args, self.state, self.control) → 1944 self.\_maybe\_log\_save\_evaluate(tr\_loss, model, trial, epoch, ignore\_keys\_for\_eval) 1945 1946 if DebugOption.TPU\_METRICS\_DEBUG in self.args.debug: /usr/local/lib/python3.10/dist-packages/transformers/trainer.py in \_maybe\_log\_save\_evaluate(self, tr\_loss, model, trial, epoch, ignore\_keys\_for\_eval) 2289 metrics = None 2290 if self.control.should\_evaluate: → 2291 metrics = self.evaluate(ignore\_keys=ignore\_keys\_for\_eval) 2292 self.\_report\_to\_hp\_search(trial, self.state.global\_step, metrics) 2293 /usr/local/lib/python3.10/dist-packages/transformers/trainer.py in evaluate(self, eval\_dataset, ignore\_keys, metric\_key\_prefix) 3093 3094 eval\_loop = self.prediction\_loop if self.args.use\_legacy\_prediction\_loop else self.evaluation\_loop → 3095 output = eval\_loop( 3096 eval\_dataloader, 3097 description=“Evaluation”, /usr/local/lib/python3.10/dist-packages/transformers/trainer.py in evaluation\_loop(self, dataloader, description, prediction\_loss\_only, ignore\_keys, metric\_key\_prefix) 3384 ) 3385 else: → 3386 metrics = self.compute\_metrics(EvalPrediction(predictions=all\_preds, label\_ids=all\_labels)) 3387 else: 3388 metrics = {} in compute\_metrics(eval\_pred) 24 logits, labels = eval\_pred 25 # Convert logits to predicted token IDs —\> 26 predictions = np.argmax(logits, axis=-1) 27 # Flatten the predictions and labels for comparison 28 true\_labels = labels.flatten() /usr/local/lib/python3.10/dist-packages/numpy/core/overrides.py in argmax(\*args, \*\*kwargs) /usr/local/lib/python3.10/dist-packages/numpy/core/fromnumeric.py in argmax(a, axis, out, keepdims) 1214 “”" 1215 kwds = {‘keepdims’: keepdims} if keepdims is not np.\_NoValue else {} → 1216 return \_wrapfunc(a, ‘argmax’, axis=axis, out=out, \*\*kwds) 1217 1218 /usr/local/lib/python3.10/dist-packages/numpy/core/fromnumeric.py in \_wrapfunc(obj, method, \*args, \*\*kwds) 52 bound = getattr(obj, method, None) 53 if bound is None: —\> 54 return \_wrapit(obj, method, \*args, \*\*kwds) 55 56 try: /usr/local/lib/python3.10/dist-packages/numpy/core/fromnumeric.py in \_wrapit(obj, method, \*args, \*\*kwds) 41 except AttributeError: 42 wrap = None —\> 43 result = getattr(asarray(obj), method)(\*args, \*\*kwds) 44 if wrap: 45 if not isinstance(result, mu.ndarray): ## ValueError: could not broadcast input array from shape (30,512,32128) into shape (30,512) My code is this. # Import necessary libraries import pandas as pd from sklearn.model\_selection import train\_test\_split from datasets import Dataset, DatasetDict, load\_metric from transformers import T5Tokenizer, T5ForConditionalGeneration, TrainingArguments, Trainer, AutoTokenizer from sklearn.metrics import accuracy\_score, recall\_score, confusion\_matrix import numpy as np import matplotlib.pyplot as plt “”"# Load and prepare datasets **Load Datasets** load the CSV files and you will need to combine the seperate files into a single file and then split it into Train, Validate and Test “”" # Load datasets def load\_dataset(file\_path): df = pd.read\_csv(file\_path, sep = ‘|’) df.columns = [‘text’, ‘labels’] # Assuming two columns: ‘input’ and ‘output’ return df “”“Split the data set into Train, validate and test”“” # Split dataset def split\_dataset(df): train\_df, temp\_df = train\_test\_split(df, test\_size=0.2, random\_state=42) val\_df, test\_df = train\_test\_split(temp\_df, test\_size=0.5, random\_state=42) return map(Dataset.from\_pandas, (train\_df, val\_df, test\_df)) “”“Now we load the data and complete the split”“” # Load and split each dataset sThread\_train, sThread\_val, sThread\_test = split\_dataset(load\_dataset(‘tags.csv’)) dThread\_train, dThread\_val, dThread\_test = split\_dataset(load\_dataset(‘tagd.csv’)) tThread\_train, tThread\_val, tThread\_test = split\_dataset(load\_dataset(‘tagn.csv’)) display(sThread\_train) “”“Combine the split data sets”“” combined\_dataset = DatasetDict({ ‘train’: Dataset.from\_dict(pd.concat([sThread\_train.to\_pandas(), dThread\_train.to\_pandas(), tThread\_train.to\_pandas()])), ‘validation’: Dataset.from\_dict(pd.concat([sThread\_val.to\_pandas(), dThread\_val.to\_pandas(), tThread\_val.to\_pandas()])), ‘test’: Dataset.from\_dict(pd.concat([sThread\_test.to\_pandas(), dThread\_test.to\_pandas(), tThread\_test.to\_pandas()])) }) display(combined\_dataset) “”" def filter\_non\_strings(dataset): return dataset.filter(lambda example: isinstance(example[‘text’], str) and isinstance(example[‘labels’], str)) “”" # Apply the filtering function to each split in the dataset #filtered\_dataset = DatasetDict({split: filter\_non\_strings(combined\_dataset[split]) for split in combined\_dataset.keys()}) #display(filtered\_dataset) “”"# Setup Tokenizer and Initiatlization Initialize the tokenizer and model “”" # Initialize the tokenizer and model tokenizer = AutoTokenizer.from\_pretrained(‘google/flan-t5-base’) model = T5ForConditionalGeneration.from\_pretrained(‘google/flan-t5-base’) “”“Tokenization function”“” # Tokenization function def tokenize\_function(examples): model\_inputs = tokenizer(examples[‘text’], padding=‘max\_length’, truncation=True, max\_length=512) with tokenizer.as\_target\_tokenizer(): labels = tokenizer(examples[‘labels’], padding=‘max\_length’, truncation=True, max\_length=512) model\_inputs[‘labels’] = labels[‘input\_ids’] return model\_inputs “”" def tokenize\_function(examples): model\_inputs = tokenizer(examples[‘text’], padding=‘max\_length’, truncation=True, max\_length=512) labels = tokenizer(examples[‘labels’], padding=‘max\_length’, truncation=True, max\_length=512) return model\_inputs, labels def tokenize\_function(examples): assert isinstance(examples[‘text’], list), “Text data is not in list format” assert all(isinstance(text, str) for text in examples[‘text’]), “Not all text entries are strings” # Tokenize the text and labels together in the model\_inputs dictionary model\_inputs = tokenizer(examples[‘text’], padding=‘max\_length’, truncation=True, max\_length=512) # Tokenize labels and add them to model\_inputs under the key ‘labels’ labels = tokenizer(examples[‘labels’], padding=‘max\_length’, truncation=True, max\_length=512) model\_inputs[‘labels’] = labels[‘input\_ids’] return model\_inputs “”" “”“Tokenize data sets”“” # Tokenize datasets - combined does not work tokenized\_datasets = combined\_dataset.map(tokenize\_function, batched=True) # Then, apply the tokenize\_function to the filtered dataset #tokenized\_datasets = filtered\_dataset.map(tokenize\_function, batched=True) “”“Training arguments”“” # Training arguments training\_args = TrainingArguments( output\_dir=“./results”, evaluation\_strategy=“epoch”, learning\_rate=2e-5, per\_device\_train\_batch\_size=4, per\_device\_eval\_batch\_size=4, num\_train\_epochs=3, weight\_decay=0.01, save\_strategy=“epoch”, load\_best\_model\_at\_end=True, ) “”“Define a function to compute accuracy”“” # prompt: how to find the shape of the shape of the input array of the training set display(tokenized\_datasets) # Define a function to compute accuracy def compute\_metrics(eval\_pred): logits, labels = eval\_pred predictions = np.argmax(logits, axis=2) # Fix the axis accuracy = accuracy\_score(labels, predictions) recall = recall\_score(labels, predictions, average=‘weighted’) # Use ‘binary’ for binary classification return {“accuracy”: accuracy, “recall”: recall} “”“Initialize Trainer”“” # Initialize Trainer trainer = Trainer( model=model, args=training\_args, train\_dataset=tokenized\_datasets[‘train’], eval\_dataset=tokenized\_datasets[‘validation’], compute\_metrics=compute\_metrics ) “”“Start Training”“” # Start training trainer.train() ---
**Author:** ![llmtrails](https://avatars.discourse-cdn.com/v4/letter/l/e9bcb4/32.png) [@llmtrails](https://discuss.huggingface.co/u/llmtrails)\ **Post date:** [February 13, 2024, 10:11pm UTC](https://discuss.huggingface.co/t/valueerror-could-not-broadcast-input-array-from-shape-30-512-32128-into-shape-30-512/73040/2 "2024-02-13T22:11:33Z")
This got solved by changing the compute function to this one. I found that the logits was a tuple. And the first element was the logits, Once that was identified then I could handle it. This fixed my problem and the code completed successfully… def compute\_metrics(eval\_pred): logits, labels = eval\_pred # Assuming logits are the first element of the tuple logits\_shape = logits[0].shape if isinstance(logits, tuple) else logits.shape #print(f"logits shape: {logits\_shape}“) # This will print the shape of the logits #print(f"labels shape: {labels.shape}”) # Continue to print labels shape ``` # Proceed with your metrics calculation predictions = np.argmax(logits[0], axis=-1) if isinstance(logits, tuple) else np.argmax(logits, axis=-1) labels = labels.flatten() predictions = predictions.flatten() accuracy = accuracy_score(labels, predictions) recall = recall_score(labels, predictions, average='weighted', zero_division=0) return {"accuracy": accuracy, "recall": recall} ``` ---
**Author:** ![system](https://us1.discourse-cdn.com/hellohellohello/original/2X/d/de4155eb4aa4108ecb32a1389d7cc37ae69f88b7.png) [@system](https://discuss.huggingface.co/u/system)\ **Post date:** [February 14, 2024, 10:12am UTC](https://discuss.huggingface.co/t/valueerror-could-not-broadcast-input-array-from-shape-30-512-32128-into-shape-30-512/73040/3 "2024-02-14T10:12:06Z")
Closed on February 14, 2024, 10:12am UTC This topic was automatically closed 12 hours after the last reply. New replies are no longer allowed.