feat: add mmkd

configs/mmkd_black_box_api.json

@@ -0,0 +1,30 @@
+{
+  "job_type": "mmkd_black_box_api",
+  "dataset": {
+    "instruction_path": "train.json",
+    "labeled_path": "train_labeled.json",
+    "seed": 42
+  },
+  "inference":{
+    "base_url": "ENDPOINT",
+    "api_key": "TOKEN",
+    "system_prompt" : "You are a helpful assistant.",
+    "max_new_tokens": 512
+  },
+  "models": {
+    "student": "student/Qwen/Qwen2.5-VL-3B-Instruct/"
+  },
+  "training": {
+    "output_dir": "./result/",
+    "num_train_epochs": 3,
+    "per_device_train_batch_size": 1,
+    "gradient_accumulation_steps": 8,
+    "max_length": 512,
+    "save_steps": 1000,
+    "logging_steps": 1,
+    "learning_rate": 2e-5,
+    "weight_decay": 0.05,
+    "warmup_ratio": 0.1,
+    "lr_scheduler_type": "cosine"
+  }
+}

easydistill/mmkd/infer.py

@@ -0,0 +1,122 @@
+
+# Copyright 2024 Alibaba Group Holding Limited. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+import json
+import argparse
+import logging
+from tqdm import tqdm
+from openai import OpenAI
+
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+
+
+def read_json_field(filename):
+    try:
+        with open(filename, 'r') as file:
+            data = json.load(file)
+        outputs = []
+        for item in data:
+            text = item["instruction"]
+            image = item["image"]
+            outputs.append((text, image))
+        return outputs
+    except FileNotFoundError:
+        logging.error("The file was not found.")
+    except json.JSONDecodeError:
+        logging.error("There was an error decoding the JSON file.")
+    except Exception as e:
+        logging.error(f"An error occurred: {e}")
+
+
+def write_data_to_json_file(data, file_path):
+    try:
+        with open(file_path, 'w') as file:
+            json.dump(data, file, ensure_ascii=False, indent=4)
+        logging.info(f"Data successfully written to {file_path}")
+    except Exception as e:
+        logging.error(f"An error occurred: {e}")
+
+
+def generate_teacher_response_api(data_list, config):
+    client = OpenAI(
+        api_key = config["inference"]["api_key"],
+        base_url = config["inference"]["base_url"]
+    )
+    models = client.models.list()
+    model = models.data[0].id
+    logging.info(model)
+    system_prompt = config["inference"]["system_prompt"]
+    if system_prompt == "":
+        system_prompt = "You are a helpful assistant."
+    outcomes = []
+    for text, image in tqdm(data_list, desc="Call remote model and generating responses"):
+        messages = [
+            {
+                "role": "system",
+                "content": system_prompt
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image_url",
+                        "image_url": {
+                            "url": image
+                        },
+                    },
+                    {
+                        "type": "text",
+                        "text": text
+                    }
+                ]
+            }
+        ]
+        completion = client.chat.completions.create(
+            messages = messages,
+            model = model,
+            max_completion_tokens = config["inference"]["max_new_tokens"]
+        )
+        result = completion.choices[0].message.content
+        outcomes.append({'instruction': text, 'image': image, 'output': result})
+    write_data_to_json_file(outcomes, config["dataset"]["labeled_path"])
+
+
+def infer_with_teacher_model(config):
+    logging.info('Generating distillation data from the teacher model!')
+    data_list = read_json_field(config["dataset"]["instruction_path"])
+    try:
+        job_type =  config["job_type"]
+        if job_type == "mmkd_black_box_api":
+            generate_teacher_response_api(data_list, config)
+        else:
+            logging.error(f"Invalid job type: {job_type}")
+            raise ValueError(f"Invalid job type: {job_type}")
+    except ValueError as e:
+        logging.error(f"Training job terminated: {e}")
+        return
+
+        
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--config', type=str, required=True, help='path to the json config file')
+    args = parser.parse_args()
+    config = json.load(open(args.config))
+    infer_with_teacher_model(config)
+
+
+if __name__ == "__main__":
+    main()

easydistill/mmkd/train.py

@@ -0,0 +1,105 @@
+
+# Copyright 2024 Alibaba Group Holding Limited. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+import json
+import argparse
+import logging
+from datasets import load_dataset, Dataset
+from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLProcessor
+from qwen_vl_utils import process_vision_info
+from trl import SFTTrainer, SFTConfig
+
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+
+
+def train(config):
+    dataset = load_dataset("json", data_files=config["dataset"]["labeled_path"])
+    dataset = dataset.shuffle(seed=config["dataset"]["seed"])["train"]
+    student_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+        config["models"]["student"],
+        trust_remote_code=True
+    )
+    processor = Qwen2_5_VLProcessor.from_pretrained(config["models"]["student"])
+
+    def collate_fn(examples):
+        texts = []
+        images = []
+        for example in examples:
+            chat = [
+                {
+                    "role": "user", 
+                    "content": [
+                        {
+                            "type": "image","image": example["image"]
+                        }, 
+                        {
+                            "type": "text","text": example["instruction"]
+                        }
+                    ]
+                },
+                {
+                    "role": "assistant", 
+                    "content": example["output"]
+                }
+            ]
+            text = processor.apply_chat_template(chat, tokenize=False)
+            texts.append(text)
+            image, _ = process_vision_info(chat)
+            images.append(image)
+
+        batch = processor(text=texts, images=images, return_tensors="pt", padding=True)
+        labels = batch["input_ids"].clone()
+        labels[labels == processor.tokenizer.pad_token_id] = -100
+        
+        if isinstance(processor, Qwen2_5_VLProcessor):
+            image_tokens = [151652, 151653, 151655]
+        else:
+            image_tokens = [processor.tokenizer.convert_tokens_to_ids(processor.image_token)]
+            
+        for image_token_id in image_tokens:
+            labels[labels == image_token_id] = -100
+        batch["labels"] = labels
+        return batch
+    
+    training_arguments = SFTConfig(**config["training"])
+    training_arguments.gradient_checkpointing_kwargs = dict(use_reentrant=False)
+    training_arguments.remove_unused_columns = False
+    training_arguments.dataset_kwargs = {"skip_prepare_dataset": True}
+    
+    trainer = SFTTrainer(
+        model=student_model,
+        data_collator=collate_fn,
+        processing_class=processor.tokenizer,
+        args=training_arguments,
+        train_dataset=dataset
+    )
+
+    trainer.train()
+    trainer.save_model(config["training"]["output_dir"])
+    processor.tokenizer.save_pretrained(config["training"]["output_dir"])
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--config', type=str, required=True, help='path to the json config file')
+    args = parser.parse_args()
+    config = json.load(open(args.config))
+    train(config)
+
+
+if __name__ == "__main__":
+    main()