app

app.py

@@ -0,0 +1,50 @@
+import streamlit as st
+import torch
+import numpy as np
+import json
+import typing as tp
+
+import torch.nn.functional as F
+from torch import Tensor
+
+from datasets import ClassLabel
+import transformers
+from transformers import BertForSequenceClassification
+from transformers import BertForSequenceClassification, AutoTokenizer
+
+st.markdown("## Portuguese European and Brazilian dialect classifier")
+st.markdown("[You can see the difference between dialects here](https://en.wikipedia.org/wiki/Portuguese_language#Writing_system)")
+text = st.text_input('## Text:')
+
+tokenizer = AutoTokenizer.from_pretrained('adalbertojunior/distilbert-portuguese-cased', do_lower_case=False)
+
+classes = ['pt', 'pt_br']
+
+class_label = ClassLabel(names=classes)
+
+
+@st.cache_data
+def get_model():
+    return BertForSequenceClassification.from_pretrained(
+   './pt_br_model',
+    num_labels = 2,
+    output_attentions = False,
+    output_hidden_states = False,
+    )
+
+
+model = get_model()
+
+
+@torch.inference_mode()
+def print_results():
+    input_tensor = tokenizer(text, padding=True, truncation=True, max_length=256, add_special_tokens=True, return_tensors="pt")
+    logits = model(**input_tensor).logits
+    probabilities = F.softmax(logits, dim=1).flatten().tolist()
+    maxidx = np.argmax(probabilities)
+    results = f"### {classes[maxidx]} score: {probabilities[maxidx]*100}%"
+    st.markdown('## Results:')
+    st.markdown(results)
+
+if text:
+    print_results()

requirements.txt

@@ -1,2 +1,5 @@
 torch
-transformers
+transformers
+streamlit
+numpy
+datasets

test_inference.py

@@ -0,0 +1,35 @@
+import torch
+
+import json
+import typing as tp
+
+import torch.nn.functional as F
+from torch import Tensor
+
+from datasets import ClassLabel
+import transformers
+from transformers import BertForSequenceClassification
+from transformers import BertForSequenceClassification, AutoTokenizer
+import numpy as np
+
+tokenizer = AutoTokenizer.from_pretrained('adalbertojunior/distilbert-portuguese-cased', do_lower_case=False)
+
+classes = ['pt','pt_br']
+class_label = ClassLabel(names=classes)
+
+def get_model():
+    return BertForSequenceClassification.from_pretrained(
+   './pt_br_model',
+    num_labels = 2,
+    output_attentions = False,
+    output_hidden_states = False,
+    )
+
+model = get_model()
+text = 'hello'
+input_tensor = tokenizer(text, padding=True, truncation=True, max_length=256, add_special_tokens=True, return_tensors="pt")
+
+logits=model(**input_tensor).logits
+probabilities = F.softmax(logits, dim=1).flatten().tolist()
+maxidx = np.argmax(probabilities)
+print(classes[maxidx], probabilities[maxidx])