thak123's picture
Update app.py
75ce96c
raw
history blame
4.42 kB
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 datasets import Dataset, DatasetDict
from data_predict import convert_to_stsb_features,convert_to_features
from huggingface_hub import hf_hub_download,snapshot_download
# 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"]]
def predict():
# 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_dict({"content":["Volim ti"]})
})
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)
return features_dict
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#model_link = snapshot_download(repo_id="FFZG-cleopatra/Croatian-News-Classifier")
model_link = hf_hub_download(repo_id="FFZG-cleopatra/Croatian-News-Classifier",filename = "pytorch_model.bin")
# 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(model_link, map_location=device))
# multitask_model.to(device)
predictions = []
features_dict = predict()
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()