VideoModelStudio / vms /tabs /train_tab.py
jbilcke-hf's picture
jbilcke-hf HF Staff
refactoring
64a70c0
raw
history blame
20.7 kB
"""
Train tab for Video Model Studio UI
"""
import gradio as gr
import logging
from typing import Dict, Any, List, Optional, Tuple
from pathlib import Path
from .base_tab import BaseTab
from ..config import TRAINING_PRESETS, MODEL_TYPES, ASK_USER_TO_DUPLICATE_SPACE, SMALL_TRAINING_BUCKETS
from ..utils import TrainingLogParser
logger = logging.getLogger(__name__)
class TrainTab(BaseTab):
"""Train tab for model training"""
def __init__(self, app_state):
super().__init__(app_state)
self.id = "train_tab"
self.title = "4️⃣ Train"
def create(self, parent=None) -> gr.TabItem:
"""Create the Train tab UI components"""
with gr.TabItem(self.title, id=self.id) as tab:
with gr.Row():
with gr.Column():
with gr.Row():
self.components["train_title"] = gr.Markdown("## 0 files available for training (0 bytes)")
with gr.Row():
with gr.Column():
self.components["training_preset"] = gr.Dropdown(
choices=list(TRAINING_PRESETS.keys()),
label="Training Preset",
value=list(TRAINING_PRESETS.keys())[0]
)
self.components["preset_info"] = gr.Markdown()
with gr.Row():
with gr.Column():
self.components["model_type"] = gr.Dropdown(
choices=list(MODEL_TYPES.keys()),
label="Model Type",
value=list(MODEL_TYPES.keys())[0]
)
self.components["model_info"] = gr.Markdown(
value=self.get_model_info(list(MODEL_TYPES.keys())[0])
)
with gr.Row():
self.components["lora_rank"] = gr.Dropdown(
label="LoRA Rank",
choices=["16", "32", "64", "128", "256", "512", "1024"],
value="128",
type="value"
)
self.components["lora_alpha"] = gr.Dropdown(
label="LoRA Alpha",
choices=["16", "32", "64", "128", "256", "512", "1024"],
value="128",
type="value"
)
with gr.Row():
self.components["num_epochs"] = gr.Number(
label="Number of Epochs",
value=70,
minimum=1,
precision=0
)
self.components["batch_size"] = gr.Number(
label="Batch Size",
value=1,
minimum=1,
precision=0
)
with gr.Row():
self.components["learning_rate"] = gr.Number(
label="Learning Rate",
value=2e-5,
minimum=1e-7
)
self.components["save_iterations"] = gr.Number(
label="Save checkpoint every N iterations",
value=500,
minimum=50,
precision=0,
info="Model will be saved periodically after these many steps"
)
with gr.Column():
with gr.Row():
self.components["start_btn"] = gr.Button(
"Start Training",
variant="primary",
interactive=not ASK_USER_TO_DUPLICATE_SPACE
)
self.components["pause_resume_btn"] = gr.Button(
"Resume Training",
variant="secondary",
interactive=False
)
self.components["stop_btn"] = gr.Button(
"Stop Training",
variant="stop",
interactive=False
)
with gr.Row():
with gr.Column():
self.components["status_box"] = gr.Textbox(
label="Training Status",
interactive=False,
lines=4
)
with gr.Accordion("See training logs"):
self.components["log_box"] = gr.TextArea(
label="Finetrainers output (see HF Space logs for more details)",
interactive=False,
lines=40,
max_lines=200,
autoscroll=True
)
return tab
def connect_events(self) -> None:
"""Connect event handlers to UI components"""
# Model type change event
def update_model_info(model):
params = self.get_default_params(MODEL_TYPES[model])
info = self.get_model_info(MODEL_TYPES[model])
return {
self.components["model_info"]: info,
self.components["num_epochs"]: params["num_epochs"],
self.components["batch_size"]: params["batch_size"],
self.components["learning_rate"]: params["learning_rate"],
self.components["save_iterations"]: params["save_iterations"]
}
self.components["model_type"].change(
fn=lambda v: self.app.update_ui_state(model_type=v),
inputs=[self.components["model_type"]],
outputs=[]
).then(
fn=update_model_info,
inputs=[self.components["model_type"]],
outputs=[
self.components["model_info"],
self.components["num_epochs"],
self.components["batch_size"],
self.components["learning_rate"],
self.components["save_iterations"]
]
)
# Training parameters change events
self.components["lora_rank"].change(
fn=lambda v: self.app.update_ui_state(lora_rank=v),
inputs=[self.components["lora_rank"]],
outputs=[]
)
self.components["lora_alpha"].change(
fn=lambda v: self.app.update_ui_state(lora_alpha=v),
inputs=[self.components["lora_alpha"]],
outputs=[]
)
self.components["num_epochs"].change(
fn=lambda v: self.app.update_ui_state(num_epochs=v),
inputs=[self.components["num_epochs"]],
outputs=[]
)
self.components["batch_size"].change(
fn=lambda v: self.app.update_ui_state(batch_size=v),
inputs=[self.components["batch_size"]],
outputs=[]
)
self.components["learning_rate"].change(
fn=lambda v: self.app.update_ui_state(learning_rate=v),
inputs=[self.components["learning_rate"]],
outputs=[]
)
self.components["save_iterations"].change(
fn=lambda v: self.app.update_ui_state(save_iterations=v),
inputs=[self.components["save_iterations"]],
outputs=[]
)
# Training preset change event
self.components["training_preset"].change(
fn=lambda v: self.app.update_ui_state(training_preset=v),
inputs=[self.components["training_preset"]],
outputs=[]
).then(
fn=self.update_training_params,
inputs=[self.components["training_preset"]],
outputs=[
self.components["model_type"],
self.components["lora_rank"],
self.components["lora_alpha"],
self.components["num_epochs"],
self.components["batch_size"],
self.components["learning_rate"],
self.components["save_iterations"],
self.components["preset_info"]
]
)
# Training control events
self.components["start_btn"].click(
fn=self.handle_training_start,
inputs=[
self.components["training_preset"],
self.components["model_type"],
self.components["lora_rank"],
self.components["lora_alpha"],
self.components["num_epochs"],
self.components["batch_size"],
self.components["learning_rate"],
self.components["save_iterations"],
self.app.tabs["manage_tab"].components["repo_id"]
],
outputs=[
self.components["status_box"],
self.components["log_box"]
]
).success(
fn=self.get_latest_status_message_logs_and_button_labels,
outputs=[
self.components["status_box"],
self.components["log_box"],
self.components["start_btn"],
self.components["stop_btn"],
self.components["pause_resume_btn"]
]
)
self.components["pause_resume_btn"].click(
fn=self.handle_pause_resume,
outputs=[
self.components["status_box"],
self.components["log_box"],
self.components["start_btn"],
self.components["stop_btn"],
self.components["pause_resume_btn"]
]
)
self.components["stop_btn"].click(
fn=self.handle_stop,
outputs=[
self.components["status_box"],
self.components["log_box"],
self.components["start_btn"],
self.components["stop_btn"],
self.components["pause_resume_btn"]
]
)
def handle_training_start(self, preset, model_type, *args):
"""Handle training start with proper log parser reset"""
# Safely reset log parser if it exists
if hasattr(self.app, 'log_parser') and self.app.log_parser is not None:
self.app.log_parser.reset()
else:
logger.warning("Log parser not initialized, creating a new one")
from ..utils import TrainingLogParser
self.app.log_parser = TrainingLogParser()
# Start training
return self.app.trainer.start_training(
MODEL_TYPES[model_type],
*args,
preset_name=preset
)
def get_model_info(self, model_type: str) -> str:
"""Get information about the selected model type"""
if model_type == "hunyuan_video":
return """### HunyuanVideo (LoRA)
- Required VRAM: ~48GB minimum
- Recommended batch size: 1-2
- Typical training time: 2-4 hours
- Default resolution: 49x512x768
- Default LoRA rank: 128 (~600 MB)"""
elif model_type == "ltx_video":
return """### LTX-Video (LoRA)
- Required VRAM: ~18GB minimum
- Recommended batch size: 1-4
- Typical training time: 1-3 hours
- Default resolution: 49x512x768
- Default LoRA rank: 128"""
return ""
def get_default_params(self, model_type: str) -> Dict[str, Any]:
"""Get default training parameters for model type"""
if model_type == "hunyuan_video":
return {
"num_epochs": 70,
"batch_size": 1,
"learning_rate": 2e-5,
"save_iterations": 500,
"video_resolution_buckets": SMALL_TRAINING_BUCKETS,
"video_reshape_mode": "center",
"caption_dropout_p": 0.05,
"gradient_accumulation_steps": 1,
"rank": 128,
"lora_alpha": 128
}
else: # ltx_video
return {
"num_epochs": 70,
"batch_size": 1,
"learning_rate": 3e-5,
"save_iterations": 500,
"video_resolution_buckets": SMALL_TRAINING_BUCKETS,
"video_reshape_mode": "center",
"caption_dropout_p": 0.05,
"gradient_accumulation_steps": 4,
"rank": 128,
"lora_alpha": 128
}
def update_training_params(self, preset_name: str) -> Tuple:
"""Update UI components based on selected preset while preserving custom settings"""
preset = TRAINING_PRESETS[preset_name]
# Load current UI state to check if user has customized values
current_state = self.app.load_ui_values()
# Find the display name that maps to our model type
model_display_name = next(
key for key, value in MODEL_TYPES.items()
if value == preset["model_type"]
)
# Get preset description for display
description = preset.get("description", "")
# Get max values from buckets
buckets = preset["training_buckets"]
max_frames = max(frames for frames, _, _ in buckets)
max_height = max(height for _, height, _ in buckets)
max_width = max(width for _, _, width in buckets)
bucket_info = f"\nMaximum video size: {max_frames} frames at {max_width}x{max_height} resolution"
info_text = f"{description}{bucket_info}"
# Return values in the same order as the output components
# Use preset defaults but preserve user-modified values if they exist
lora_rank_val = current_state.get("lora_rank") if current_state.get("lora_rank") != preset.get("lora_rank", "128") else preset["lora_rank"]
lora_alpha_val = current_state.get("lora_alpha") if current_state.get("lora_alpha") != preset.get("lora_alpha", "128") else preset["lora_alpha"]
num_epochs_val = current_state.get("num_epochs") if current_state.get("num_epochs") != preset.get("num_epochs", 70) else preset["num_epochs"]
batch_size_val = current_state.get("batch_size") if current_state.get("batch_size") != preset.get("batch_size", 1) else preset["batch_size"]
learning_rate_val = current_state.get("learning_rate") if current_state.get("learning_rate") != preset.get("learning_rate", 3e-5) else preset["learning_rate"]
save_iterations_val = current_state.get("save_iterations") if current_state.get("save_iterations") != preset.get("save_iterations", 500) else preset["save_iterations"]
return (
model_display_name,
lora_rank_val,
lora_alpha_val,
num_epochs_val,
batch_size_val,
learning_rate_val,
save_iterations_val,
info_text
)
def update_training_ui(self, training_state: Dict[str, Any]):
"""Update UI components based on training state"""
updates = {}
# Update status box with high-level information
status_text = []
if training_state["status"] != "idle":
status_text.extend([
f"Status: {training_state['status']}",
f"Progress: {training_state['progress']}",
f"Step: {training_state['current_step']}/{training_state['total_steps']}",
# Epoch information
# there is an issue with how epoch is reported because we display:
# Progress: 96.9%, Step: 872/900, Epoch: 12/50
# we should probably just show the steps
#f"Epoch: {training_state['current_epoch']}/{training_state['total_epochs']}",
f"Time elapsed: {training_state['elapsed']}",
f"Estimated remaining: {training_state['remaining']}",
"",
f"Current loss: {training_state['step_loss']}",
f"Learning rate: {training_state['learning_rate']}",
f"Gradient norm: {training_state['grad_norm']}",
f"Memory usage: {training_state['memory']}"
])
if training_state["error_message"]:
status_text.append(f"\nError: {training_state['error_message']}")
updates["status_box"] = "\n".join(status_text)
# Update button states
updates["start_btn"] = gr.Button(
"Start training",
interactive=(training_state["status"] in ["idle", "completed", "error", "stopped"]),
variant="primary" if training_state["status"] == "idle" else "secondary"
)
updates["stop_btn"] = gr.Button(
"Stop training",
interactive=(training_state["status"] in ["training", "initializing"]),
variant="stop"
)
return updates
def handle_pause_resume(self):
status, _, _ = self.get_latest_status_message_and_logs()
if status == "paused":
self.app.trainer.resume_training()
else:
self.app.trainer.pause_training()
return self.get_latest_status_message_logs_and_button_labels()
def handle_stop(self):
self.app.trainer.stop_training()
return self.get_latest_status_message_logs_and_button_labels()
def get_latest_status_message_and_logs(self) -> Tuple[str, str, str]:
"""Get latest status message, log content, and status code in a safer way"""
state = self.app.trainer.get_status()
logs = self.app.trainer.get_logs()
# Ensure log parser is initialized
if not hasattr(self.app, 'log_parser') or self.app.log_parser is None:
from ..utils import TrainingLogParser
self.app.log_parser = TrainingLogParser()
logger.info("Initialized missing log parser")
# Parse new log lines
if logs:
last_state = None
for line in logs.splitlines():
try:
state_update = self.app.log_parser.parse_line(line)
if state_update:
last_state = state_update
except Exception as e:
logger.error(f"Error parsing log line: {str(e)}")
continue
if last_state:
ui_updates = self.update_training_ui(last_state)
state["message"] = ui_updates.get("status_box", state["message"])
# Parse status for training state
if "completed" in state["message"].lower():
state["status"] = "completed"
return (state["status"], state["message"], logs)
def get_latest_status_message_logs_and_button_labels(self) -> Tuple[str, str, Any, Any, Any]:
status, message, logs = self.get_latest_status_message_and_logs()
return (
message,
logs,
*self.update_training_buttons(status).values()
)
def update_training_buttons(self, status: str) -> Dict:
"""Update training control buttons based on state"""
is_training = status in ["training", "initializing"]
is_paused = status == "paused"
is_completed = status in ["completed", "error", "stopped"]
return {
"start_btn": gr.Button(
interactive=not is_training and not is_paused,
variant="primary" if not is_training else "secondary",
),
"stop_btn": gr.Button(
interactive=is_training or is_paused,
variant="stop",
),
"pause_resume_btn": gr.Button(
value="Resume Training" if is_paused else "Pause Training",
interactive=(is_training or is_paused) and not is_completed,
variant="secondary",
)
}