Inference with self trained langid model

[paulthemagno]

Hi, I trained a langid model with my dataset following these steps and ending with this method:

python -m stanza.models.lang_identifier --data-dir data  --eval-length 10 --randomize --save-name model.pt --num-epochs 100

At there is the .pt saved in the directory.

How I can test this new model, making some inferences on some inputs? I saw in the doc how to do that with the standard model, but not with new trained ones. Thank you!

[paulthemagno]

Probably the best way would be:

import stanza
nlp = stanza.Pipeline("multilingual", langid_model_path="model_stanza.pt")

am I right?

[AngledLuffa]

If it's loading, then I think that must be right...

Let us know if that works, and we'll update the docs!

On Thu, Jul 7, 2022 at 3:46 AM Paolo Magnani @.***> wrote:

Probably the best way would be:

import stanzanlp = stanza.Pipeline("multilingual", langid_model_path="model_stanza.pt")

am I right?

— Reply to this email directly, view it on GitHub https://github.com/stanfordnlp/stanza/issues/1072#issuecomment-1177388286, or unsubscribe https://github.com/notifications/unsubscribe-auth/AA2AYWLHTE26E7ET72DA6A3VS2YRHANCNFSM52Z37DSQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[J38]

Yes that is the proper way to load a custom model in Python code.

[J38]

You can do a more comprehensive eval with this command:

python -m stanza.models.lang_identifier --data-dir data --load-model model.pt --mode eval --eval-length 50 --save-name model-results.jsonl

Full documentation here: https://stanfordnlp.github.io/stanza/langid.html

[paulthemagno]

It works! Is there also easy way to extract the accuracy for each language? In order to have something like {"en": 0.8, "de": 0.1, etc..}

[paulthemagno]

I've found this way to achieve this task. I wanted to arrive to _model output and have percentages that sum to 1 with a softmax layer.

To do that I overrided _process_list of LangIDProcessor class and prediction_scores of LangIDBiLSTM class. I'm not sure this it's the cleanest way to do that. Do you agree @AngledLuffa ?

Also for the config I have to pass to LangIDProcessor, I'm not sure to have build the object correctly.

import torch
from stanza.pipeline.langid_processor import LangIDProcessor
from stanza.models.common.doc import Document

#taken from https://discuss.pytorch.org/t/apply-mask-softmax/14212/13
def masked_softmax(vec, mask, dim=1):
    masked_vec = vec * mask.float()
    max_vec = torch.max(masked_vec, dim=dim, keepdim=True)[0]
    exps = torch.exp(masked_vec-max_vec)
    masked_exps = exps * mask.float()
    masked_sums = masked_exps.sum(dim, keepdim=True)
    zeros=(masked_sums == 0)
    masked_sums += zeros.float()
    return masked_exps/masked_sums

def get_predictions_scores(text, pipeline, k):
    print(f"Text: {text}")
    print(f"Output of the pipeline directly: {pipeline(text).lang}")

    config = {}
    for key in pipeline.config:
        if key.startswith("langid_"):
            config[key.split("langid_")[1]] = pipeline.config[key]
        else:
            config[key] = pipeline.config[key]
    processor = LangIDProcessor(config = config, pipeline = pipeline, use_gpu = True)

    docs = [text]

    #override of _process_list of LangIDProcessor
    if isinstance(docs[0], str):
        docs = [Document([], text) for text in docs]
    docs_by_length = {}
    for doc in docs:
        text = processor.clean_text(doc.text) if processor._clean_text else doc.text
        doc_length = len(text)
        if doc_length not in docs_by_length:
            docs_by_length[doc_length] = []
        docs_by_length[doc_length].append((doc, text))

    for doc_length in docs_by_length:
        inputs = [doc[1] for doc in docs_by_length[doc_length]]
        #override of prediction_scores of LangIDBiLSTM to get the predictions
        x = processor._text_to_tensor(inputs)
        prediction_probs = processor._model(x)
        if processor._model.lang_subset:
            prediction_batch_size = prediction_probs.size()[0]
            batch_mask = torch.stack([processor._model.lang_mask for _ in range(prediction_batch_size)])
            prediction_probs = prediction_probs * batch_mask
            prediction_probs = masked_softmax(vec = prediction_probs, mask = batch_mask)
        else:
            softmax = torch.nn.Softmax(dim = 1)
            prediction_probs = softmax(prediction_probs)
        topk = torch.topk(prediction_probs, k)
        pred_scores = {}
        for i,pred in enumerate(topk.indices[0]):
            print(f"Language: {processor._model.idx_to_tag[pred]}: {topk.values[0][i]}")
            pred_scores[processor._model.idx_to_tag[pred]] = topk.values[0][i].item()

    return pred_scores

model_name = "my_model"
pipeline = stanza.Pipeline("multilingual", langid_model_path=model_name+".pt")
pred_scores = get_predictions_scores(text = "hello how are you?", pipeline = pipeline, k=3)
pred_scores

Anyway now I have this output:

{'en': 0.9999321699142456,
 'nn': 2.8831263989559375e-05,
 'nl': 1.2516763490566518e-05}

EDIT I used a masked_softmax in case of processor._model.lang_subset is set, taking it from here: https://discuss.pytorch.org/t/apply-mask-softmax/14212/13

without doing that it seems that the percentages were wrong. I don't know if this is the most correct way to do that. This seems to solve also this issue I found https://github.com/stanfordnlp/stanza/issues/1076

[AngledLuffa]

Ah, I misunderstood your previous question. You had said you want the accuracy for all languages, but what you want is the predictions for all languages. I should be able to add that functionality to the processor. Even better, would you be up for turning your code in this message into a pull request against the dev branch, stanza/models/langid/model.py and/or stanza/pipeline/langid_processor.py ?

On Wed, Jul 13, 2022 at 3:48 AM Paolo Magnani @.***> wrote:

I've found this way to achieve this task. I wanted to arrive to _model output and have percentages that sum to 1 with a softmax layer.

To do that I overrided _process_list of LangIDProcessor class and prediction_scores of LangIDBiLSTM class. I'm not sure this it's the cleanest way to do that. Do you agree @AngledLuffa https://github.com/AngledLuffa ?

Also for the config I have to pass to LangIDProcessor, I'm not sure to have build the object correctly.

import torchfrom stanza.pipeline.langid_processor import LangIDProcessorfrom stanza.models.common.doc import Document def get_predictions_scores(text, pipeline, k): print(f"Text: {text}") print(f"Output of the pipeline directly: {pipeline(text).lang}")

config = {}
for key in pipeline.config:
    if key.startswith("langid_"):
        config[key.split("langid_")[1]] = pipeline.config[key]
    else:
        config[key] = pipeline.config[key]
processor = LangIDProcessor(config = config, pipeline = pipeline, use_gpu = True)

docs = [text]

#override of _process_list of LangIDProcessor
if isinstance(docs[0], str):
    docs = [Document([], text) for text in docs]
docs_by_length = {}
for doc in docs:
    text = processor.clean_text(doc.text) if processor._clean_text else doc.text
    doc_length = len(text)
    if doc_length not in docs_by_length:
        docs_by_length[doc_length] = []
    docs_by_length[doc_length].append((doc, text))

for doc_length in docs_by_length:
    inputs = [doc[1] for doc in docs_by_length[doc_length]]
    #override of prediction_scores of LangIDBiLSTM to get the predictions
    x = processor._text_to_tensor(inputs)
    prediction_probs = processor._model(x)
    if processor._model.lang_subset:
        prediction_batch_size = prediction_probs.size()[0]
        batch_mask = torch.stack([processor._model.lang_mask for _ in range(prediction_batch_size)])
        prediction_probs = prediction_probs * batch_mask
    softmax = torch.nn.Softmax(dim = 1)
    prediction_probs = softmax(prediction_probs)
    topk = torch.topk(prediction_probs, k)
    pred_scores = {}
    for i,pred in enumerate(topk.indices[0]):
        print(f"Language: {processor._model.idx_to_tag[pred]}: {topk.values[0][i]}")
        pred_scores[processor._model.idx_to_tag[pred]] = topk.values[0][i].item()

return pred_scores
        model_name = "my_model"pipeline = stanza.Pipeline("multilingual", langid_model_path=model_name+".pt")pred_scores = get_predictions_scores(text = "hello how are you?", pipeline = pipeline, k=3)pred_scores

Anyway now I have this output:

{'en': 0.9999321699142456, 'nn': 2.8831263989559375e-05, 'nl': 1.2516763490566518e-05}

— Reply to this email directly, view it on GitHub https://github.com/stanfordnlp/stanza/issues/1072#issuecomment-1183068227, or unsubscribe https://github.com/notifications/unsubscribe-auth/AA2AYWLSIOCD7PD23YDZPHLVT2NJRANCNFSM52Z37DSQ . You are receiving this because you were mentioned.Message ID: @.***>

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[AngledLuffa]

Ping regarding this - are you interested in making this block of code into a PR?