switched to non quantized mistral instruct

examples_metadata.py

@@ -263,6 +263,29 @@ bag_metadata = {
     },
  }
 
+def clean_json(json_data):
+    """
+    Extracts the English description and English tags from the given JSON data.
+
+    :param json_data: A dictionary representing the JSON data of a fashion item.
+    :return: A dictionary containing the English description and a list of English tags.
+    """
+    important_data = {}
+
+    # Extracting the English description
+    description_en = next((desc['text'] for desc in json_data['descriptions'] if desc['language'] == 'en'), None)
+    important_data['description_en'] = description_en
+
+    # Extracting the English tags
+    tags_en = [tag['tag']['languages']['en'] for tag in json_data['tagsData']['tags'] if 'en' in tag['tag']['languages']]
+    important_data['tags_en'] = tags_en
+
+    return list(important_data.values())
+
+dress_metadata = clean_json(dress_metadata)
+bomber_metadata = clean_json(bomber_metadata)
+bag_metadata = clean_json(bag_metadata)
+
 dress_example = f'''
    #### Input
    Metadata: {dress_metadata};

main.py

@@ -27,7 +27,8 @@ from examples_metadata import (
     dress_example,
     bomber_example,
 )
-#init()
+
+
 logging.basicConfig(level=logging.DEBUG)
 os.system("pip list")
 #print Cuda version
@@ -126,17 +127,19 @@ def shortenMods(generator, res):
     return res[0]["generated_text"]
 
 
-# usare "You:" evita che il modello generi samples extra ma legga l'input
 def formatMods(generator, res):
-    prompt = f"""
-    Given as input a list like:
+   prompt = f'''
+    I have a list like:
     -var 1
     -var 2
     -var 3...
 
-    Return as output a list as:
+    Rewrite the list and put it in square brackets
     [var1, var2, var3, ...]
 
+    no code, just the list
+    It must begin with "[" and end with "]".
+
     Examples:
     {bomber_format_example}
     {shirt_format_example}
@@ -148,18 +151,17 @@ def formatMods(generator, res):
 
     Output:
 
-    """
+    '''
 
-    prompt_template = f"""<s> [INST]
+   prompt_template=f'''<s> [INST]
    {prompt}
    [/INST]
-   """
-    print("before inference")
-    print_gpu_utilization()
-    with torch.no_grad():
-        res = generator(prompt_template)
-    # print(res)
-    return res[0]["generated_text"]
+   '''
+   print("before inference")
+   print_gpu_utilization()
+   res = generator(prompt_template)
+   #print(res)
+   return res[0]['generated_text']
 
 
 def initModel(model_name_or_path, revision):
@@ -208,30 +210,29 @@ def initModel(model_name_or_path, revision):
 def generateTags():
     start = time.time()
     res = generateMods(generator, bag_metadata, dress_example, bomber_example)
+    stripped_res = remove_before_word(res, "[/INST]")
     print("generation mods response:")
     print(res)
-    stripped_res = remove_before_word(res, "[/INST]")
     shorten_res = shortenMods(generator, stripped_res)
     print("shortened response:")
     print(shorten_res)
     shorten_res = remove_before_word(shorten_res, "[/INST]")
     formatted_res = formatMods(generator, shorten_res)
-    formatted_res = remove_before_word(formatted_res, "[/INST]")
     print("formatted response:")
     print(formatted_res)
+    formatted_res = remove_before_word(formatted_res, "[/INST]")
     end = time.time()
     print("time spent for generating tags:", end - start)
     return {"response": stripped_res, "shortened response:": shorten_res, "formatted response": formatted_res}
 
-
 app = FastAPI()
 
 
 @app.on_event("startup")
 def load_model():
     global generator
-    model_name_or_path = "TheBloke/Mistral-7B-Instruct-v0.1-GPTQ"
-    revision = "gptq-8bit-128g-actorder_True"
+    model_name_or_path = "mistralai/Mistral-7B-Instruct-v0.1"
+    revision = "main"
     generator = initModel(model_name_or_path, revision)
     print("Model loaded")