FileNotFoundError: [Errno 2] No such file or directory: "model.pt" about zero-shot-bert-adapters

[gilpasternak]

Got same issue. model.pt is not automatically downloaded by library for some odd reason. Working in a venv with only package dependencies installed. Retried several times. Not a permissions issue.

Code that yielded error:

def zberta_predict(texts: list) -> list:
    """"
    Slight adjustment of given prediction starter code to get model to output predictions
    """
    model_name = "bert-base-uncased"
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    berta = model.instantiate_model(2, 2, device, model_name,
                                    {"entailment": 0, "contradiction": 1}, path=None, dict=True)
    zberta = ZBERTA(berta, model_name)
    return zberta.find_new_intents(texts)](url)

Update / Fix

I had wget installed in my local venv, but the code makes a system called to wget. For mac users, this means wget should be installed with brew. Updating this fixed my issue.

[gilpasternak]

An additional note: inference didn't work for me with multiple messages. This is because in the following code:

for i in tqdm(range(0, len(self.dataset if utterances is None else utterances))):
            pred_classes = []
            if utterances is not None:
                classes = unknown_intents_set([utterances[i]], lang)
            else:
                classes = self.z_classes
            for class_set in classes[i]:
                temp = class_set.replace('-', ' ').replace('None', '')
                if temp not in pred_classes and len(temp) > 1:
                    pred_classes.append(temp)

(lines 30-39 in zberta.py) the code attempts to access entry i in the list classes for an increasing position i based on the number of utterances. However, the list of classes is overwritten every time with a doubly nested list of intents, which has length 0 (ex: [['None-None', 'None-Checkbook']]). This yields an indexing error.

fix

Since the list of intents is overwritten every time, you can simply always access index 0 in the doubly nested list to obtain the list of intents. So functional code is:

for i in tqdm(range(0, len(self.dataset if utterances is None else utterances))):
            pred_classes = []
            if utterances is not None:
                classes = unknown_intents_set([utterances[i]], lang)
            else:
                classes = self.z_classes
            for class_set in classes[i]:
                temp = class_set.replace('-', ' ').replace('None', '')
                if temp not in pred_classes and len(temp) > 1:
                    pred_classes.append(temp)

Please let me know if I am missing something! If not, I've turned this into a PR!

[comidan]

Thank you for reporting this! Will look into this in a few days.

[comidan]

Hello @ToonLunk, I checked your issue and tried to reproduce it on my end on your same environment. I'm not seeing any error regarding the file model.pt not being downloaded correctly or anything related. Did you check if this may be related to not having writing permission on a certain directory?

Did you also retry? Just to check if it wasn't a temporary downloading issue on the cloud server hosting the model loaded by default.

[abhoi]

+1 to this issue. I'm getting the same error with a clean install. Where is the model file/weights supposed to be located?

[comidan]

Hello @abhoi the model should be located in the current working directory and it's automatically downloaded by the library. Will check into your situation.

Did you checked the points I raised in my previous comment above?

Can you better describe to me your working environment? Thank you for raising this issue and for your collaboration!

[comidan]

@gilpasternak thank you for your PR! This should have already been fixed in the past in a previous version, sorry for the inconvenience, and thanks for the help!

Regarding wget yes it's requred as written in the README. On MacOS usually it should be already been installed so in the README I just point out to make sure the installation is present for Windows users, if that is no more the case I will update the README! Thank you!