Spaces:
Running
Running
| import gradio as gr | |
| import os | |
| import torch | |
| import numpy as np | |
| from transformers import AutoModelForSequenceClassification | |
| from transformers import AutoTokenizer | |
| from huggingface_hub import HfApi | |
| from label_dicts import EMOTION9_LABEL_NAMES, EMOTION9_V2_LABEL_NAMES | |
| from .utils import is_disk_full, release_model | |
| HF_TOKEN = os.environ["hf_read"] | |
| languages = [ | |
| "Czech", "English", "German", "Hungarian", "Polish", "Slovak" | |
| ] | |
| domains = { | |
| "parliamentary speech": "parlspeech", | |
| } | |
| def build_huggingface_path(language: str): | |
| language = language.lower() | |
| base_model_id = f"poltextlab/xlm-roberta-large-pooled-{language}-emotions9" | |
| if language in ["slovak", "polish", "czech", "hungarian"]: | |
| return base_model_id + "-v2" | |
| return base_model_id | |
| def predict(text, model_id, tokenizer_id): | |
| device = torch.device("cpu") | |
| # Load JIT-traced model | |
| jit_model_path = f"/data/jit_models/{model_id.replace('/', '_')}.pt" | |
| model = torch.jit.load(jit_model_path).to(device) | |
| model.eval() | |
| # Load tokenizer (still regular HF) | |
| tokenizer = AutoTokenizer.from_pretrained(tokenizer_id) | |
| # Tokenize input | |
| inputs = tokenizer( | |
| text, | |
| max_length=64, | |
| truncation=True, | |
| padding=True, | |
| return_tensors="pt" | |
| ) | |
| inputs = {k: v.to(device) for k, v in inputs.items()} | |
| with torch.no_grad(): | |
| output = model(inputs["input_ids"], inputs["attention_mask"]) | |
| print(output) # debug | |
| logits = output["logits"] | |
| release_model(model, model_id) | |
| probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten() | |
| if "v2" in model_id: | |
| label_names = EMOTION9_V2_LABEL_NAMES | |
| else: | |
| label_names = EMOTION9_LABEL_NAMES | |
| output_pred = {f"[{i}] {label_names[i]}": probs[i] for i in np.argsort(probs)[::-1]} | |
| output_info = f'<p style="text-align: center; display: block">Prediction was made using the <a href="https://huggingface.co/{model_id}">{model_id}</a> model.</p>' | |
| return output_pred, output_info | |
| def predict_e6(text, language, domain): | |
| model_id = build_huggingface_path(language) | |
| tokenizer_id = "xlm-roberta-large" | |
| if is_disk_full(): | |
| os.system('rm -rf /data/models*') | |
| os.system('rm -r ~/.cache/huggingface/hub') | |
| return predict(text, model_id, tokenizer_id) | |
| demo = gr.Interface( | |
| title="Emotions (9) Babel Demo", | |
| fn=predict_e6, | |
| inputs=[gr.Textbox(lines=6, label="Input"), | |
| gr.Dropdown(languages, label="Language", value=languages[1]), | |
| gr.Dropdown(domains.keys(), label="Domain", value=list(domains.keys())[0])], | |
| outputs=[gr.Label(num_top_classes=5, label="Output"), gr.Markdown()]) | |