Add probability/perplexity calculation and complete detoxified LM not…

…ebook

examples/detoxification_bias/Detoxify_LM_demo.ipynb

@@ -11,7 +11,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": null,
+   "execution_count": 2,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -115,18 +115,9 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 22,
+   "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "The autoreload extension is already loaded. To reload it, use:\n",
-      "  %reload_ext autoreload\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "from llments.eval.toxicity import ToxicityEvaluator\n",
     "# create a toxicity evaluator for text scoring\n",
@@ -366,48 +357,29 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# Perplexity evaluation on WAE vs. AAE\n",
-    "\n",
-    "### This part still awaits replacement with llment code"
+    "# Perplexity evaluation on WAE vs. AAE"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 11,
+   "execution_count": 75,
    "metadata": {},
    "outputs": [],
    "source": [
-    "from transformers import TrainingArguments, Trainer, DataCollatorForLanguageModeling\n",
-    "from datasets import load_dataset\n",
-    "import math\n",
-    "\n",
-    "# helper function to evaluate the perplexity of a fine-tuned model\n",
-    "def eval_ppl(data_dir, eval_files, model_path, model_name, training_args, output_dir):\n",
+    "# helper function to evaluate the perplexity of a detoxified model on a given dataset\n",
+    "def eval_ppl(data_dir, eval_files, model_path, model_name, output_dir):\n",
     "    \n",
-    "    checkpoint = HuggingFaceLM(model=model_path, tokenizer_path=model_name)\n",
+    "    checkpointLM = HuggingFaceLM(model=model_path, tokenizer_path=model_name)\n",
     "    res = []\n",
     "\n",
-    "    # TODO: replace the following with llments operators if available, \n",
-    "    # consider a HGLM.evaluate() operator\n",
     "    for eval_file in eval_files:\n",
     "        eval_file_path = os.path.join(data_dir, eval_file)\n",
-    "        eval_dataset = load_dataset(\"text\", data_files=eval_file_path, split=\"train\")\n",
     "        \n",
-    "        eval_dataset = eval_dataset.map(lambda examples: checkpoint.tokenizer(\n",
-    "            examples[\"text\"], truncation=True, padding=\"max_length\", max_length=128), batched=True)\n",
-    "\n",
-    "        data_collator = DataCollatorForLanguageModeling(tokenizer=checkpoint.tokenizer, mlm=False)\n",
-    "\n",
-    "        trainer = Trainer(\n",
-    "            model=checkpoint.model,\n",
-    "            args=training_args,\n",
-    "            data_collator=data_collator,\n",
-    "            eval_dataset=eval_dataset\n",
-    "        )\n",
-    "        eval_results = trainer.evaluate()\n",
-    "\n",
-    "        # calculate the perplexity\n",
-    "        ppl = math.exp(eval_results[\"eval_loss\"])\n",
+    "        with open(eval_file_path, \"r\") as f:\n",
+    "            sentences = f.readlines()\n",
+    "        \n",
+    "        # calculate the average perplexity of the samples\n",
+    "        ppl = checkpointLM.calculate_perplexity_batch(outputs=sentences, condition=None)\n",
     "        res.append((eval_file, ppl))\n",
     "\n",
     "    with open(f\"{output_dir}/{model_path.split('/')[-1]}.txt\", \"w\") as f:\n",
@@ -419,7 +391,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 12,
+   "execution_count": 71,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -485,14 +457,6 @@
     "if not os.path.exists(\"trash\"):\n",
     "    os.makedirs(\"trash\")\n",
     "\n",
-    "training_args = TrainingArguments(\n",
-    "    output_dir=\"trash\",\n",
-    "    per_device_eval_batch_size=1,\n",
-    "    do_train=False,                                     \n",
-    "    do_eval=True,                       \n",
-    "    fp16=False,\n",
-    ")\n",
-    "\n",
     "model_path_list = [\"gpt2\",\n",
     "    \"checkpoints/gpt2/checkpoint-2500\",\n",
     "    \"checkpoints/gpt2/checkpoint-5000\",\n",
@@ -505,7 +469,7 @@
     "    \"checkpoints/gpt2/checkpoint-22500\"]\n",
     "\n",
     "for model_path in model_path_list:\n",
-    "    res = eval_ppl(eval_data_dir, eval_files, model_path, \"gpt2\", training_args, \"eval_results_gpt2\")\n",
+    "    res = eval_ppl(eval_data_dir, eval_files, model_path, \"gpt2\", \"eval_results_gpt2\")\n",
     "    print(res)"
    ]
   },
@@ -553,7 +517,7 @@
     "\n",
     "for model_path in model_path_list:\n",
     "    res = eval_ppl(eval_data_dir, eval_files, model_path, \n",
-    "                   \"NousResearch/Llama-2-7b-hf\", training_args, \"eval_results_Llama2-7b\")\n",
+    "                   \"NousResearch/Llama-2-7b-hf\", \"eval_results_Llama2-7b\")\n",
     "    print(res)"
    ]
   },

llments/lm/base/hugging_face.py

@@ -184,7 +184,7 @@ def set_seed(self, seed: int) -> None:
         set_seed(seed)
 
     def calculate_probability(self, condition: str | None, output: str) -> float:
-        """Calculate the probability of an output given the language model.
+        """Calculate the log probability of an output given the language model.
 
         Args:
             condition: The conditioning sequence for the output.
@@ -193,7 +193,150 @@ def calculate_probability(self, condition: str | None, output: str) -> float:
         Returns:
             float: The probability of output x given the language model.
         """
-        raise NotImplementedError
+        try:
+            import numpy as np
+        except ImportError:
+            raise ImportError(
+                "You need to install 'numpy' package to use this function."
+            )
+
+        if condition:
+            full_input = condition + " " + output
+        else:
+            full_input = output
+
+        # Tokenize the full input (condition + output or just output)
+        inputs = self.tokenizer(
+            full_input,
+            return_tensors="pt",
+            truncation=True,
+            padding=False,  # Avoid padding unless needed
+        )
+
+        # Get model outputs (logits)
+        full_outputs = self.model(**inputs, return_dict=True)
+        logits = (
+            full_outputs.logits.detach().cpu().numpy()
+        )  # Convert logits to NumPy array
+        full_input_ids = inputs["input_ids"][0].cpu().numpy()
+
+        # define a softmax function
+        def softmax(logits: np.ndarray) -> np.ndarray:
+            exps = np.exp(
+                logits - np.max(logits, axis=-1, keepdims=True)
+            )  # Stabilize softmax
+            return exps / np.sum(exps, axis=-1, keepdims=True)
+
+        # Calculate the probability of the output
+        probs = softmax(logits[0])  # Only one sequence in the batch
+        probs = probs[:-1, :]
+
+        # calculate the num of tokens corresponding to the output
+        output_ids = self.tokenizer(output)["input_ids"]
+        output_ids = output_ids[1:]
+        full_input_ids = full_input_ids[1:]
+        start_idx = len(full_input_ids) - len(output_ids)
+
+        # take the last # of output_tokens from the log_probs
+        log_probs = np.log(probs[np.arange(start_idx, len(full_input_ids)), output_ids])
+
+        # convert the log_probs to a float
+        return float(np.sum(log_probs))
+
+    def calculate_perplexity(self, condition: str | None, output: str) -> float:
+        """Calculate the perplexity of an output given the language model.
+
+        Args:
+            condition: The conditioning sequence for the output.
+            output: The output sequence for which the probability is calculated.
+
+        Returns:
+            float: The perplexity of output x given the language model.
+        """
+        try:
+            import numpy as np
+        except ImportError:
+            raise ImportError(
+                "You need to install 'numpy' package to use this function."
+            )
+
+        log_prob = self.calculate_probability(condition, output)
+        num_tokens = len(self.tokenizer(output)["input_ids"]) - 1
+
+        return float(np.exp(-log_prob / num_tokens))
+
+    def calculate_perplexity_batch(
+        self, condition: list[str] | None, outputs: list[str]
+    ) -> float:
+        """Calculate the perplexity of multiple outputs given the language model.
+
+        Args:
+            condition: The conditioning sequence for the output.
+            outputs: The output sequences for which the probability is calculated.
+
+        Returns:
+            list[float]: The perplexity of outputs given the language model.
+        """
+        if condition:
+            full_inputs = [c + " " + o for c, o in zip(condition, outputs)]
+        else:
+            full_inputs = outputs
+
+        # check if the user have import Trainer, TrainingArguments, DataCollatorForLanguageModeling
+        try:
+            from datasets import Dataset
+            from transformers import (
+                DataCollatorForLanguageModeling,
+                Trainer,
+                TrainingArguments,
+            )
+            import numpy as np
+        except ImportError:
+            print(
+                "Naive implementation is used. This may harm the efficiency of the calculation."
+            )
+            try:
+                import numpy as np
+            except ImportError:
+                raise ImportError(
+                    "You need to install 'numpy' package to use this function."
+                )
+            return float(
+                np.mean(self.calculate_perplexity(None, o) for o in full_inputs)
+            )
+
+        # prepare the dataset
+        inputs = self.tokenizer(
+            full_inputs,
+            return_tensors="pt",
+            truncation=True,
+            padding=True,
+        )
+
+        dataset = Dataset.from_dict(inputs)
+        data_collator = DataCollatorForLanguageModeling(
+            tokenizer=self.tokenizer, mlm=False
+        )
+
+        training_arguments = TrainingArguments(
+            output_dir="trash",
+            per_device_eval_batch_size=1,
+            do_train=False,
+            do_eval=True,
+            fp16=False,
+            prediction_loss_only=True,
+        )
+
+        trainer = Trainer(
+            model=self.model,
+            data_collator=data_collator,
+            args=training_arguments,
+            eval_dataset=dataset,
+        )
+
+        eval_result = trainer.evaluate()
+
+        return float(np.exp(eval_result["eval_loss"]))
 
 
 class HuggingFaceLMFitter: