app.py

import datasets
import numpy as np
import torch
import transformers
from config import epochs, batch_size, learning_rate
from model import tokenizer, multitask_model
from mtm import MultitaskTrainer, NLPDataCollator, DataLoaderWithTaskname
import pandas as pd
# from data_5_LT23 import features_dict,extra_feature_dict
from data_predict import convert_to_stsb_features,convert_to_features

features_dict = {}
extra_feature_dict = {}
sentinews_location = ""

df_document_croatian_test = pd.read_csv(sentinews_location+"textlabel.tsv", sep="\t")
df_document_croatian_test = df_document_croatian_test[["content"]]

# gather everyone if you want to have a single DatasetDict
document = DatasetDict({
    # "train": Dataset.from_pandas(df_document_sl_hr_train),
    # "valid": Dataset.from_pandas(df_document_sl_hr_valid),
    "test": Dataset.from_pandas(df_document_croatian_test)
})

dataset_dict = {
    "document": document,
}

for task_name, dataset in dataset_dict.items():
    print(task_name)
    print(dataset_dict[task_name]["test"][0])
    print()


convert_func_dict = {
    "document": convert_to_stsb_features,
    # "paragraph": convert_to_stsb_features,
    # "sentence": convert_to_stsb_features,
}

features_dict = convert_to_features(dataset_dict, convert_func_dict)



# multitask_model.from_pretrained(, config="/media/gaurish/angela/projects/CroatianSlovenEnglishBert/i-got-u-brother-cleopatra-workshop/src/models/multitask_model_3ep/config.json")
multitask_model.load_state_dict(torch.load(
    "multitask_model_3ep/pytorch_model.bin"
    ))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
multitask_model.to(device)
predictions = []
for _, batch in enumerate(features_dict["document"]['test']):
    for key, value in batch.items():
        batch[key] = batch[key].to(device)
    
    task_model = multitask_model.get_model("document")
    classifier_output = task_model.forward(
            torch.unsqueeze(batch["input_ids"], 0),
            torch.unsqueeze(batch["attention_mask"], 0),)
    
    print(tokenizer.decode(batch["input_ids"],skip_special_tokens=True))
    prediction =torch.max(classifier_output.logits, axis=1)
    predictions.append(prediction.indices.item())

pd.DataFrame({"original_predictions":predictions}).to_csv("eacl_slavic.tsv")


trainer = MultitaskTrainer(
    model=multitask_model,
    args=transformers.TrainingArguments(
        learning_rate=learning_rate,
        output_dir="/tmp",
        do_train=False,
        do_eval=True,
        # evaluation_strategy ="steps",
        # num_train_epochs=epochs,
        fp16=True,
        # Adjust batch size if this doesn't fit on the Colab GPU
        per_device_train_batch_size=batch_size,
        per_device_eval_batch_size=batch_size,
        save_steps=3000,
        # eval_steps=50,
        load_best_model_at_end=True,
        
    ),
    data_collator=NLPDataCollator(tokenizer=tokenizer),
    callbacks=[],
    
)
print(features_dict["document"]["test"])
tests_dict = {}
for task_name in ["document"]: # "paragraph", "sentence"
    test_dataloader = DataLoaderWithTaskname(
        task_name,
        trainer.get_eval_dataloader(features_dict[task_name]["test"])
    )
    print(len(trainer.get_eval_dataloader(features_dict[task_name]["test"])))
    print(test_dataloader.data_loader.collate_fn)
    print(len(test_dataloader.data_loader))
    tests_dict[task_name] = trainer.prediction_loop(
        test_dataloader,
        description=f"Testing: {task_name}"
    )
print(tests_dict)
for task_name in ["document",  ]: #"paragraph","sentence"
    for metric in ["precision", "recall", "f1"]:
        print("test {} {}:".format(metric, task_name),
              datasets.load_metric(metric,
                                   name="dev {} {}".format(metric, task_name)).compute(
                  predictions=np.argmax(
                      tests_dict[task_name].predictions, axis=1),
                  references=tests_dict[task_name].label_ids, average="macro"
              ))
print()