Update app.py

app.py

@@ -1,7 +1,7 @@
 import gradio as gr
 import torch 
 import torch.nn as nn
-import torch.nn.functional as F
+import torch.nn.functional as FU
 from torchvision import transforms
 import numpy as np 
 from ultralytics import YOLO
@@ -10,6 +10,8 @@ from pypinyin import pinyin
 from PIL import Image, ImageDraw, ImageEnhance
 import json
 import os
+import torchvision.transforms.functional as F
+
 
 """yolo model"""
 user_device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -22,7 +24,7 @@ detection_model = YOLO("model/yolo_chinese_m.pt").to(user_device).eval()
 class HSwish(nn.Module):
     """Hard Swish activation function"""
     def forward(self, x):
-        out = x * F.relu6(x + 3, inplace=True) / 6
+        out = x * FU.relu6(x + 3, inplace=True) / 6
         return out
 
 class MV2_Block(nn.Module):
@@ -194,7 +196,7 @@ classifier.load_state_dict(torch.load("model/chinese_Character_Classification_ha
 classifier.to(user_device).eval()
 
 """load classes dict"""
-with open("model/LW_Vit/chinese_classes.json", "r", encoding="utf-8") as f:
+with open("model/chinese_classes.json", "r", encoding="utf-8") as f:
     classes_dict = json.load(f)
 
 """transforms"""
@@ -223,11 +225,11 @@ class ToSquare(object):
         pad_w_extra = (max_dim - w) % 2
         pad_h_extra = (max_dim - h) % 2
         
-        # Create padding tuple (left, top, right, bottom)
-        padding = (pad_w, pad_h, pad_w + pad_w_extra, pad_h + pad_h_extra)
+        # Create padding list (left, top, right, bottom) - FIX #1: Use list instead of tuple
+        padding = [pad_w, pad_h, pad_w + pad_w_extra, pad_h + pad_h_extra]
         
         # Create new padded image
-        padded_img = transforms.functional.pad(img, padding, fill=self.fill)
+        padded_img = F.pad(img, padding, self.fill)
         
         return padded_img
 
@@ -259,13 +261,12 @@ test_transforms = transforms.Compose([
 """import llm"""
 llm = Llama.from_pretrained(
     repo_id="IncreasingLoss/FineTunedTranslation_Yi-1.5-9B-Chat-Q6_K",
-    filename="*.gguf",       
+    filename="*.gguf",       # or specify exact file name if only one GGUF in repo
     n_ctx=4096,
     n_gpu_layers=-1,
     verbose=False
 )
 
-
 """app functions"""
 def convert_to_pil(inputimg):
         # convert input to pil image
@@ -340,7 +341,7 @@ def crop_bboxes_by_reading_order(img, bboxes, written_vertical=False):
     
     for idx in order[1:]:
         c = coords[idx]
-        # if within threshold of this cluster’s mean, add; else start new
+        # if within threshold of this cluster's mean, add; else start new
         if abs(c - current_mean) <= threshold:
             current.append(idx)
             # update cluster mean incrementally
@@ -386,6 +387,9 @@ def crop_bboxes_by_reading_order(img, bboxes, written_vertical=False):
 
 def move_slider(threshhold_slider, input_img):
     pil_img = convert_to_pil(input_img)
+    # FIX #2: Add null check for pil_img
+    if pil_img is None:
+        return None
     with torch.inference_mode():
         results = detection_model(source=pil_img, conf=threshhold_slider)
         draw_img = draw_bboxes(pil_img, results)
@@ -400,6 +404,9 @@ def select_image(evt: gr.SelectData):
 def translate_text(threshhold_slider, is_vertical, input_img):
     # convert input to pil image
     pil_img = convert_to_pil(input_img)
+    # FIX #2: Add null check for pil_img
+    if pil_img is None:
+        return "", "", ""
 
     with torch.inference_mode():
         results = detection_model(source=pil_img, conf=threshhold_slider)
@@ -413,7 +420,6 @@ def translate_text(threshhold_slider, is_vertical, input_img):
             chinese_charcter = classes_dict[class_idx]
             chinese_text += chinese_charcter
 
-    #chinese_text = "这是一句测试语句，用来检验我的深度翻译计算机语言模型。"
     pinyin_text = pinyin(chinese_text)
     pinyin_sentence = " "
     for pin in pinyin_text:
@@ -426,7 +432,7 @@ def translate_text(threshhold_slider, is_vertical, input_img):
         2. Respond only with the translated English text.
         English translation: """
 
-    #try:
+    # FIX #3: Disable streaming to get direct dict access
     output = llm(
         prompt, 
         max_tokens=min(len(chinese_text)*2, 512),  # More reasonable token limit with safety cap
@@ -435,7 +441,8 @@ def translate_text(threshhold_slider, is_vertical, input_img):
         temperature=0.3,  # Slightly higher temperature for more natural language
         frequency_penalty=.5,  # Penalize repetition of tokens - higher values discourage repetition
         presence_penalty=.5,  # Penalize tokens that have appeared at all - prevents topic repetition
-        top_p=0.9  # Nucleus sampling - restricts to top 90% most likely tokens for more coherent output
+        top_p=0.9,  # Nucleus sampling - restricts to top 90% most likely tokens for more coherent output
+        stream=False  # Explicitly disable streaming to get a dict response
     )
     
     # Extract the translation
@@ -470,7 +477,7 @@ css = """
   max-height: none     !important;  /* cancel the global max-height */ 
   height: auto         !important;  /* allow auto height */ 
 }
-/* 2) Same-query override: mirror Gradio’s media query exactly */
+/* 2) Same-query override: mirror Gradio's media query exactly */
 @media (min-width: 1280px) {
   /* target the exact same class chain inside the breakpoint */
   .fixed-height.svelte-842rpi.svelte-842rpi {