Upload app.py

app.py

@@ -4,20 +4,27 @@ import gradio as gr
 from PIL import Image
 import pickle
 import json
-from huggingface_hub import hf_hub_download
+from huggingface_hub import hf_hub_download, login
 from model import image_captioning_model
 
+# Аутентификация (если репозиторий приватный)
+# login(token="hf_YOUR_API_TOKEN")  # Замените на ваш токен или уберите, если публичный
+
 # Загрузка файлов
-# https://huggingface.co/spaces/Pumponchik/ImageCaption/tree/main
 repo_id = "Pumponchik/ImageCaption"
-model_path = hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin")
-config_path = hf_hub_download(repo_id=repo_id, filename="config.json")
-tokenizer_path = hf_hub_download(repo_id=repo_id, filename="tokenizer.pkl")
-transform_path = hf_hub_download(repo_id=repo_id, filename="image_transforms.pkl")
+repo_type = "space"
+model_path = hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin", repo_type=repo_type)
+config_path = hf_hub_download(repo_id=repo_id, filename="config.json", repo_type=repo_type)
+tokenizer_path = hf_hub_download(repo_id=repo_id, filename="tokenizer.pkl", repo_type=repo_type)
+transform_path = hf_hub_download(repo_id=repo_id, filename="image_transforms.pkl", repo_type=repo_type)
+
+print(f"Model path: {model_path}")  # Отладка
+print(f"Config path: {config_path}")  # Отладка
 
 # Загрузка конфигурации
 with open(config_path, "r") as f:
     config = json.load(f)
+print(f"Loaded config: {config}")  # Отладка
 
 # Загрузка модели
 model = image_captioning_model(**config)
@@ -54,9 +61,9 @@ def generate(model, image, tokenizer, max_seq_len=20, top_p=None, top_k=1):
     batch_size = image.size(0)
 
     with torch.no_grad():
-        img_emb = model.img_features(image)  # [batch_size, 512]
-        h_0 = model.text_features.img_to_hidden(img_emb)  # [batch_size, hidden_size]
-        h_0 = h_0.unsqueeze(0).repeat(model.text_features.num_layers, 1, 1)  # [num_layers, batch_size, hidden_size]
+        img_emb = model.img_features(image)
+        h_0 = model.text_features.img_to_hidden(img_emb)
+        h_0 = h_0.unsqueeze(0).repeat(model.text_features.num_layers, 1, 1)
         c_0 = h_0.clone()
 
         input_token = torch.full((batch_size,), tokenizer.tok_to_ind['<BOS>'], dtype=torch.long, device=device)
@@ -64,11 +71,11 @@ def generate(model, image, tokenizer, max_seq_len=20, top_p=None, top_k=1):
         mask = torch.ones(batch_size, dtype=torch.bool, device=device)
 
         for _ in range(max_seq_len):
-            input_emb = model.text_features.embed(input_token).unsqueeze(1)  # [batch_size, 1, embed_dim]
+            input_emb = model.text_features.embed(input_token).unsqueeze(1)
             input_emb = model.text_features.text_to_hidden(input_emb)
-            output, (h_0, c_0) = model.text_features.rnn(input_emb, (h_0, c_0))  # [batch_size, 1, hidden_size]
-            logits = model.fc_out(output.squeeze(1))  # [batch_size, vocab_size]
-            probs = F.softmax(logits, dim=-1)  # [batch_size, vocab_size]
+            output, (h_0, c_0) = model.text_features.rnn(input_emb, (h_0, c_0))
+            logits = model.fc_out(output.squeeze(1))
+            probs = F.softmax(logits, dim=-1)
 
             next_tokens = torch.zeros(batch_size, dtype=torch.long, device=device)
             for i in range(batch_size):
@@ -103,7 +110,6 @@ def generate(model, image, tokenizer, max_seq_len=20, top_p=None, top_k=1):
             if not mask.any():
                 break
 
-        # Декодирование токенов в текст
         result_text = []
         for tokens in result_tokens:
             caption = " ".join([tokenizer.ind_to_tok.get(t, "<UNK>") for t in tokens