Merge pull request #11 from LyaaaaaGames/Develop

Release 1.2.0

.gitignore

@@ -127,3 +127,7 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+/server/miniconda/*
+/tokenizers/*
+/models/*
+server_logs.text

conda_config.yml

@@ -6,5 +6,6 @@ dependencies:
   - python=3.9.7
   - pip:
     - websockets==10.0
-    - torch==1.10.1
-    - transformers==4.15
+    - torch==1.10.2
+    - transformers==4.16.2
+    - sentencepiece

requirements.txt

@@ -2,4 +2,4 @@ pytorch==1.10.1
 torchvision==0.11.2
 cudatoolkit==11.3.1
 websockets==10.0
-transformers==4.10.2
+transformers==4.16

server/config.py

@@ -21,7 +21,11 @@
 #--     - Set HOST default value to 0.0.0.0
 #--
 #--   25/03/2022 Lyaaaaa
-#--     - Set LOG_FILEMODE default valie to 'a'
+#--     - Set LOG_FILEMODE default value to 'a'
+#--
+#--   21/05/2022 Lyaaaaa
+#--     - Set LOG_FILEMODE default value to 'w'
+#--     - Set LOG_LEVEL default value to DEBUG (Experimental branch only)
 #---------------------------------------------------------------------------
 
 import logging
@@ -32,5 +36,5 @@
 
 # Logs
 LOG_FILENAME = "server_logs.text"
-LOG_FILEMODE = "a"
-LOG_LEVEL    = logging.INFO
+LOG_FILEMODE = "w"
+LOG_LEVEL    = logging.DEBUG

server/generator.py

@@ -0,0 +1,47 @@
+#------------------------------------------------------------------------------
+#-- Copyright (c) 2021-2022 LyaaaaaGames
+#-- Copyright (c) 2022 AIdventure_Server contributors
+#--
+#-- Author : Lyaaaaaaaaaaaaaaa
+#--
+#-- Implementation Notes:
+#--  - The class handling the text generation.
+#--
+#-- Anticipated changes:
+#--  -
+#--
+#-- Changelog:
+#--  - 18/02/2022 Lyaaaaa
+#--    - Created the file.
+#--
+#--  - 22/05/2022 Lyaaaaa
+#--    - Updated generate_text to support the gpu once again. Simplified the
+#--        script by merging all the models input into a single dict "model_input".
+#------------------------------------------------------------------------------
+
+from model import Model
+
+class Generator(Model):
+
+#------------------------------------------------------------------------------
+#-- generate_text
+#------------------------------------------------------------------------------
+  def generate_text(self,
+                    p_prompt     = None,
+                    p_context    = None,
+                    p_memory     = None,
+                    p_parameters = None):
+
+    model_input    = p_memory + p_context + p_prompt
+    model_input    = self._Tokenizer(model_input, return_tensors = "pt")
+
+    if self.is_gpu_enabled:
+      model_input.to("cuda")
+
+    p_parameters["max_length"] += len(model_input[0])
+
+    model_output = self._Model.generate(**model_input,
+                                        **p_parameters)
+    generated_text = self._Tokenizer.decode(model_output[0], skip_special_tokens=True)
+
+    return generated_text

server/model.py

@@ -5,13 +5,17 @@
 #-- Author : Lyaaaaaaaaaaaaaaa
 #--
 #-- Portability Issues (feel free to remove this part if there is none):
-#--  - Cuda might not be available depending of the environment and OS.
+#--  - This class needs Pytorch installed.
 #--
 #-- Implementation Notes:
 #--  - This class will handle all the task related to the AI model.
 #--
 #-- Anticipated changes:
+#--  - Use a specific configuration.json file for each model (each model folder
+#--      already has a config file. It could be more efficient to save each config
+#--      into its own folder.
 #--  - Update generate_text to make it possible to generate multiple answers.
+#--  - Get rid of the configuration as I confused the generation and model config.
 #--
 #-- Changelog:
 #--  - 31/08/2021 Lyaaaaa
@@ -77,20 +81,12 @@
 #--    - Updated generate_text to append the input_ids' size to max_length in
 #--        p_parameters. Also added attention_mask as parameter to generate method.
 #--
-#--  - 11/02/2022 Lyaaaaa
-#--    - Removed _get_gpu_info call from generate_text method.
-#--    - Updated _enable_gpu, _empty_gpu_cache and _get_gpu_info to fix the
-#--        indentation.
-#--    - Updated _enable_gpu to add a try except finally conditions to avoid
-#--        to crash if the GPU runs out of memory.
-#--    - Added a print in __init__ to investigate the ISSUE#2.
-#--    - Replaced the prints by self._logger.info.
-#--
-#--  - 11/02/2022 Lyaaaaa
-#--    - Updated _enable_gpu to call _model.to("cpu") if the the runtime error
-#--        happens with the GPU.
-#--
 #--  - 18/02/2022 Lyaaaaa
+#--    - Extracted generate_text method to generator class.
+#--    - Added the model_type and AutoModelForSeq2SeqLM imports.
+#--    - Updated __init__ to receive the a parameter to specify the model's type.
+#--    - Updated _load and _download methods to use either AutoModelForCausalLM or
+#--        AutoModelForSeq2SeqLM depending of the _model_type attribute's value.
 #--    - Removed _Config and reload_config method (not used method)
 #--    - Updated generate_text to fix the "inputs on different devices" error.
 #--        attention_mark was not converted to cuda while model_input was.
@@ -111,27 +107,33 @@
 #--    - Replaced self._logger by logger.log.
 #------------------------------------------------------------------------------
 
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoModelForSeq2SeqLM, AutoTokenizer
+from model_type   import Model_Type
 import os
 import torch
 
 # Custom imports
 import logger
 
 class Model():
-  _Tokenizer = AutoTokenizer
-  _Model     = AutoModelForCausalLM
 
+  _Tokenizer  : AutoTokenizer
+  _model_type : Model_Type
+  _Model = None
 
 #------------------------------------------------------------------------------
 #-- __init__
 #------------------------------------------------------------------------------
-  def __init__(self, p_model_name = "EleutherAI/gpt-neo-125M", p_use_gpu = True):
+  def __init__(self,
+               p_model_name = "EleutherAI/gpt-neo-125M",
+               p_model_type = Model_Type.GENERATION.value,
+               p_use_gpu    = True,):
     self._tokenizer_path   = "tokenizers/" + p_model_name
     self._model_path       = "models/" + p_model_name
     self._model_name       = p_model_name
     self.is_cuda_available = torch.cuda.is_available()
     self.is_gpu_enabled    = False
+    self._model_type       = p_model_type
 
     if self._load() == False:
       self._download()
@@ -140,52 +142,6 @@ def __init__(self, p_model_name = "EleutherAI/gpt-neo-125M", p_use_gpu = True):
     else:
       logger.log.info("Model successfully loaded from local file")
 
-
-#------------------------------------------------------------------------------
-#-- generate_text
-#------------------------------------------------------------------------------
-  def generate_text(self,
-                    p_prompt     = None,
-                    p_context    = None,
-                    p_memory     = None,
-                    p_parameters = None):
-
-    model_input  = p_memory + p_context + p_prompt
-
-    model_input    = p_memory + p_context + p_prompt
-    tokens         = self._Tokenizer(model_input, return_tensors = "pt")
-    attention_mask = tokens.attention_mask
-    model_input    = tokens.input_ids
-
-    p_parameters["max_length"] += len(model_input[0])
-
-    if self.is_gpu_enabled == True:
-      model_input    = model_input.to("cuda")
-      attention_mask = attention_mask.to("cuda")
-      try:
-        model_output = self._Model.generate(input_ids       = model_input,
-                                            attention_mask  = attention_mask,
-                                            **p_parameters)
-      except Exception as error:
-        logger.log.error("Error while generating with the GPU:", error)
-        model_output = None
-        model_input    = model_input.to("cpu")
-        attention_mask = attention_mask.to("cpu")
-        self._disable_gpu()
-
-    if self.is_gpu_enabled == False:
-      model_output = self._Model.generate(input_ids       = model_input,
-                                          attention_mask  = attention_mask,
-                                          **p_parameters)
-
-    generated_text = self._Tokenizer.decode(model_output[0], skip_special_tokens=True)
-
-    if self.is_gpu_enabled == True:
-      self._empty_gpu_cache()
-
-    return generated_text
-
-
 #------------------------------------------------------------------------------
 #-- _load
 #------------------------------------------------------------------------------
@@ -198,9 +154,13 @@ def _load(self):
       return False
 
     try:
-      self._Model = AutoModelForCausalLM.from_pretrained(self._model_path)
+      if self._model_type == Model_Type.GENERATION.value:
+        self._Model = AutoModelForCausalLM.from_pretrained(self._model_path)
+      elif self._model_type == Model_Type.TRANSLATION.value:
+        self._Model = AutoModelForSeq2SeqLM.from_pretrained(self._model_path)
 
-    except:
+    except error:
+      logger.log.error(error)
       return False
 
     return True
@@ -224,9 +184,14 @@ def _download(self):
                                                     cache_dir       = "cache",
                                                     resume_download = True)
     logger.log.info("Trying to download the model...")
-    self._Model     = AutoModelForCausalLM.from_pretrained(model_name,
-                                                           cache_dir       = "cache",
-                                                           resume_download = True)
+    if self._model_type == Model_Type.GENERATION.value:
+      self._Model = AutoModelForCausalLM.from_pretrained(model_name,
+                                                         cache_dir       = "cache",
+                                                         resume_download = True)
+    elif self._model_type == Model_Type.TRANSLATION.value:
+      self._Model = AutoModelForSeq2SeqLM.from_pretrained(model_name,
+                                                          cache_dir       = "cache",
+                                                          resume_download = True)
     self._save()
 
 

server/model_type.py

@@ -0,0 +1,20 @@
+#------------------------------------------------------------------------------
+#-- Copyright (c) 2021-2022 LyaaaaaGames
+#-- Copyright (c) 2022 AIdventure_Server contributors
+#--
+#-- Author : Lyaaaaaaaaaaaaaaa
+#--
+#-- Implementation Notes:
+#--  -
+#-- Anticipated changes:
+#--  -
+#-- Changelog:
+#--  - 18/02/2022 Lyaaaaa
+#--    - Created the file
+#------------------------------------------------------------------------------
+
+from enum import Enum
+
+class Model_Type(Enum):
+  GENERATION  = 0
+  TRANSLATION = 1

server/request.py

@@ -27,6 +27,9 @@
 #--
 #--  - 29/12/2021 Lyaaaaa
 #--    - Added DOWNLOAD_MODEL and DOWNLOADED_MODEL requests.
+#--
+#--  - 01/03/2022 Lyaaaaa
+#--    - Added TEXT_TRANSLATION request.
 #------------------------------------------------------------------------------
 
 from enum import Enum
@@ -38,3 +41,4 @@ class Request(Enum):
   LOADED_MODEL     = 4
   DOWNLOAD_MODEL   = 5
   DOWNLOADED_MODEL = 6
+  TEXT_TRANSLATION = 7