PraisonAI 3.0.0__py3-none-any.whl

praisonai/train_vision.py

@@ -0,0 +1,306 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+This script finetunes a vision language model using Unsloth's fast training framework.
+It supports vision tasks by converting raw image-caption samples into a conversation format, 
+adding vision-specific LoRA adapters, and training using TRL's SFTTrainer with UnslothVisionDataCollator.
+"""
+
+import os
+import sys
+import yaml
+import torch
+import shutil
+import subprocess
+import gc  # For garbage collection
+
+from datasets import load_dataset, concatenate_datasets, Dataset
+from unsloth import FastVisionModel, is_bf16_supported
+from unsloth.trainer import UnslothVisionDataCollator
+from transformers import TrainingArguments
+from trl import SFTTrainer
+from tqdm import tqdm  # Add progress bar
+
+
+class TrainVisionModel:
+    def __init__(self, config_path="config.yaml"):
+        self.load_config(config_path)
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.model = None
+        self.hf_tokenizer = None  # The underlying tokenizer
+
+    def load_config(self, path):
+        with open(path, "r") as file:
+            self.config = yaml.safe_load(file)
+        print("DEBUG: Loaded config:", self.config)
+
+    def print_system_info(self):
+        print("DEBUG: PyTorch version:", torch.__version__)
+        print("DEBUG: CUDA version:", torch.version.cuda)
+        if torch.cuda.is_available():
+            print("DEBUG: CUDA Device Capability:", torch.cuda.get_device_capability())
+        else:
+            print("DEBUG: CUDA is not available")
+        print("DEBUG: Python Version:", sys.version)
+        print("DEBUG: Python Path:", sys.executable)
+
+    def check_gpu(self):
+        gpu_stats = torch.cuda.get_device_properties(0)
+        print(f"DEBUG: GPU = {gpu_stats.name}. Max memory = {round(gpu_stats.total_memory/(1024**3),3)} GB.")
+
+    def check_ram(self):
+        from psutil import virtual_memory
+        ram_gb = virtual_memory().total / 1e9
+        print(f"DEBUG: Your runtime has {ram_gb:.1f} gigabytes of available RAM")
+        if ram_gb < 20:
+            print("DEBUG: Not using a high-RAM runtime")
+        else:
+            print("DEBUG: You are using a high-RAM runtime!")
+
+    def prepare_model(self):
+        print("DEBUG: Preparing vision model and tokenizer...")
+        self.model, original_tokenizer = FastVisionModel.from_pretrained(
+            model_name=self.config["model_name"],
+            load_in_4bit=self.config["load_in_4bit"],
+            use_gradient_checkpointing="unsloth"
+        )
+        print("DEBUG: Vision model and original tokenizer loaded.")
+        
+        # Use the full processor that supports image inputs.
+        self.hf_tokenizer = original_tokenizer
+        
+        # Set pad token if needed
+        if not hasattr(self.hf_tokenizer, 'pad_token') or self.hf_tokenizer.pad_token is None:
+            if hasattr(self.hf_tokenizer, 'eos_token'):
+                self.hf_tokenizer.pad_token = self.hf_tokenizer.eos_token
+            elif hasattr(self.hf_tokenizer, 'bos_token'):
+                self.hf_tokenizer.pad_token = self.hf_tokenizer.bos_token
+        
+        # Set max length
+        if hasattr(self.hf_tokenizer, 'model_max_length'):
+            self.hf_tokenizer.model_max_length = self.config.get("max_seq_length", 2048)
+        
+        # Add vision-specific LoRA adapters
+        self.model = FastVisionModel.get_peft_model(
+            self.model,
+            finetune_vision_layers=self.config.get("finetune_vision_layers", False),
+            finetune_language_layers=self.config.get("finetune_language_layers", True),
+            finetune_attention_modules=self.config.get("finetune_attention_modules", True),
+            finetune_mlp_modules=self.config.get("finetune_mlp_modules", True),
+            r=16,
+            lora_alpha=16,
+            lora_dropout=0,
+            bias="none",
+            random_state=3407,
+            use_rslora=False,
+            loftq_config=None
+        )
+        print("DEBUG: Vision LoRA adapters added.")
+
+    def convert_sample(self, sample):
+        
+        instruction = self.config.get(
+            "vision_instruction",
+            "You are an expert radiographer. Describe accurately what you see in this image."
+        )
+        conversation = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": instruction},
+                    {"type": "image", "image": sample["image"]}
+                ]
+            },
+            {
+                "role": "assistant",
+                "content": [
+                    {"type": "text", "text": sample["caption"]}
+                ]
+            },
+        ]
+        
+        return {"messages": conversation}
+
+    def load_datasets(self):
+        all_converted = []
+        for dataset_info in self.config["dataset"]:
+            print("\nDEBUG: Loading vision dataset:", dataset_info)
+            ds = load_dataset(
+                dataset_info["name"], 
+                split=dataset_info.get("split_type", "train")
+            )
+            print("DEBUG: Dataset size:", len(ds))
+            print("DEBUG: First raw sample:", ds[0])
+            print("DEBUG: Dataset features:", ds.features)
+            
+            print("\nDEBUG: Converting dataset to vision conversation format...")
+            converted_ds = [self.convert_sample(sample) for sample in ds]
+            
+            # Debug first converted sample
+            print("\nDEBUG: First converted sample structure:")
+            first = converted_ds[0]
+            print("DEBUG: Message keys:", first["messages"][0]["content"][1].keys())
+            print("DEBUG: Image type in converted:", type(first["messages"][0]["content"][1].get("image")))
+            
+            all_converted.extend(converted_ds)
+        
+        print("\nDEBUG: Combined vision dataset has", len(all_converted), "examples.")
+        return all_converted
+
+    def train_model(self):
+        print("DEBUG: Starting vision training...")
+        raw_dataset = self.load_datasets()
+        
+        # Build training arguments using TrainingArguments
+        training_args = TrainingArguments(
+            per_device_train_batch_size=self.config.get("per_device_train_batch_size", 1),
+            gradient_accumulation_steps=self.config.get("gradient_accumulation_steps", 4),
+            warmup_steps=self.config.get("warmup_steps", 5),
+            max_steps=self.config.get("max_steps", 30),
+            learning_rate=self.config.get("learning_rate", 2e-4),
+            fp16=self.config.get("fp16", not is_bf16_supported()),
+            bf16=self.config.get("bf16", is_bf16_supported()),
+            logging_steps=self.config.get("logging_steps", 1),
+            optim=self.config.get("optim", "adamw_8bit"),
+            weight_decay=self.config.get("weight_decay", 0.01),
+            lr_scheduler_type=self.config.get("lr_scheduler_type", "linear"),
+            seed=self.config.get("seed", 3407),
+            output_dir=self.config.get("output_dir", "outputs"),
+            report_to="none" if not os.getenv("PRAISON_WANDB") else "wandb",
+            remove_unused_columns=False,
+            # Add memory optimization settings
+            gradient_checkpointing=True,
+            max_grad_norm=1.0,
+        )
+        
+        trainer = SFTTrainer(
+            model=self.model,
+            tokenizer=self.hf_tokenizer,
+            data_collator=UnslothVisionDataCollator(self.model, self.hf_tokenizer),
+            train_dataset=raw_dataset,
+            args=training_args,
+            max_seq_length=self.config.get("max_seq_length", 2048),
+            dataset_text_field="",  # Required for vision training
+            dataset_kwargs={"skip_prepare_dataset": True},  # Required for vision training
+            packing=False  # Explicitly set packing to False
+        )
+        print("DEBUG: Beginning vision trainer.train() ...")
+        trainer.train()
+        print("DEBUG: Vision training complete. Saving model and tokenizer locally...")
+        self.model.save_pretrained("lora_vision_model")
+        self.hf_tokenizer.save_pretrained("lora_vision_model")
+        print("DEBUG: Saved vision model and tokenizer to 'lora_vision_model'.")
+
+    def vision_inference(self, instruction, image):
+        FastVisionModel.for_inference(self.model)
+        messages = [
+            {"role": "user", "content": [
+                {"type": "image"},
+                {"type": "text", "text": instruction}
+            ]}
+        ]
+        input_text = self.hf_tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+        inputs = self.hf_tokenizer(
+            image,
+            input_text,
+            add_special_tokens=False,
+            return_tensors="pt"
+        ).to("cuda")
+        outputs = self.model.generate(
+            **inputs,
+            max_new_tokens=128,
+            use_cache=True,
+            temperature=1.5,
+            min_p=0.1
+        )
+        print("DEBUG: Vision inference output:", self.hf_tokenizer.batch_decode(outputs))
+
+    def save_model_merged(self):
+        if os.path.exists(self.config["hf_model_name"]):
+            shutil.rmtree(self.config["hf_model_name"])
+        self.model.push_to_hub_merged(
+            self.config["hf_model_name"],
+            self.hf_tokenizer,
+            save_method="merged_16bit",
+            token=os.getenv("HF_TOKEN")
+        )
+
+    def push_model_gguf(self):
+        self.model.push_to_hub_gguf(
+            self.config["hf_model_name"],
+            self.hf_tokenizer,
+            quantization_method=self.config.get("quantization_method", "q4_k_m"),
+            token=os.getenv("HF_TOKEN")
+        )
+
+    def save_model_gguf(self):
+        self.model.save_pretrained_gguf(
+            self.config["hf_model_name"],
+            self.hf_tokenizer,
+            quantization_method="q4_k_m"
+        )
+
+    def prepare_modelfile_content(self):
+        output_model = self.config["hf_model_name"]
+        
+        template = '''{{- range $index, $_ := .Messages }}<|start_header_id|>{{ .Role }}<|end_header_id|>
+
+{{ .Content }}
+{{- if gt (len (slice $.Messages $index)) 1 }}<|eot_id|>
+{{- else if ne .Role "assistant" }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+{{ end }}
+{{- end }}'''
+        
+        return f"""FROM {output_model}
+TEMPLATE {template}
+PARAMETER temperature 0.6
+PARAMETER top_p 0.9
+"""
+
+    def create_and_push_ollama_model(self):
+        modelfile_content = self.prepare_modelfile_content()
+        with open("Modelfile", "w") as file:
+            file.write(modelfile_content)
+        subprocess.run(["ollama", "serve"])
+        subprocess.run(["ollama", "create", f"{self.config['ollama_model']}:{self.config['model_parameters']}", "-f", "Modelfile"])
+        subprocess.run(["ollama", "push", f"{self.config['ollama_model']}:{self.config['model_parameters']}"])
+
+    def run(self):
+        self.print_system_info()
+        self.check_gpu()
+        self.check_ram()
+        if self.config.get("train", "true").lower() == "true":
+            self.prepare_model()
+            self.train_model()
+        if self.config.get("huggingface_save", "true").lower() == "true":
+            self.save_model_merged()
+        if self.config.get("huggingface_save_gguf", "true").lower() == "true":
+            self.push_model_gguf()
+        if self.config.get("ollama_save", "true").lower() == "true":
+            self.create_and_push_ollama_model()
+
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(description="PraisonAI Vision Training Script")
+    parser.add_argument("command", choices=["train", "inference"], help="Command to execute")
+    parser.add_argument("--config", default="config.yaml", help="Path to configuration file")
+    args = parser.parse_args()
+
+    trainer_obj = TrainVisionModel(config_path=args.config)
+    if args.command == "train":
+        trainer_obj.run()
+    elif args.command == "inference":
+        # For inference, we load a sample image from the first dataset
+        instr = trainer_obj.config.get("vision_instruction", "You are an expert radiographer. Describe accurately what you see in this image.")
+        ds_info = trainer_obj.config["dataset"][0]
+        ds = load_dataset(ds_info["name"], split=ds_info.get("split_type", "train"))
+        sample_image = ds[0]["image"]
+        if trainer_obj.model is None or trainer_obj.hf_tokenizer is None:
+            trainer_obj.prepare_model()
+        trainer_obj.vision_inference(instr, sample_image)
+
+
+if __name__ == "__main__":
+    main()