satyapolepalle's picture
Update app.py
d4e93cc verified
Raw History Blame Contribute Delete
9.29 kB
import gradio as gr
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import torchvision.transforms as transforms
from PIL import Image
import requests
import json
import os
from typing import List, Dict, Any
from io import BytesIO
import logging
# Add FastAPI imports
from fastapi import FastAPI
from pydantic import BaseModel
import uvicorn
# Add HuggingFace Hub import
from huggingface_hub import HfApi
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# Create FastAPI app
fastapi_app = FastAPI()
# Set up authentication - the Space needs a token to upload to itself
HF_TOKEN = os.getenv("HF_TOKEN") # Repository secret
if not HF_TOKEN:
logger.warning("No HF_TOKEN found - model uploads may fail")
# Initialize HF API with token
hf_api = HfApi(token=HF_TOKEN) if HF_TOKEN else HfApi()
# Pydantic models for API
class TrainingData(BaseModel):
imageUrl: str
groundTruthText: str
sourceScript: str
imageType: str
difficultyLevel: str
class TrainingConfig(BaseModel):
epochs: int
batchSize: int
learningRate: float
validationSplit: float
saveInterval: int
modelName: str
sampleLimit: int
class TrainingRequest(BaseModel):
trainingData: List[TrainingData]
config: TrainingConfig
sessionId: str
class TrainingMetrics(BaseModel):
finalLoss: float
finalAccuracy: float
epochs: int
sampleCount: int
class TrainingResponse(BaseModel):
success: bool
jobId: str = None
modelUrl: str = None
metrics: TrainingMetrics = None
error: str = None
# Your existing classes
class OCRDataset(Dataset):
def __init__(self, training_data: List[Dict], transform=None):
self.data = training_data
self.transform = transform or transforms.Compose([
transforms.Grayscale(),
transforms.Resize((64, 256)),
transforms.ToTensor(),
transforms.Normalize((0.5,), (0.5,))
])
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
item = self.data[idx]
try:
response = requests.get(item['imageUrl'])
image = Image.open(BytesIO(response.content))
if self.transform:
image = self.transform(image)
text = item['groundTruthText']
return image, text
except Exception as e:
logger.error(f"Error loading image {item['imageUrl']}: {e}")
return torch.zeros((1, 64, 256)), ""
class SimpleCRNN(nn.Module):
def __init__(self, vocab_size=200, hidden_size=256):
super(SimpleCRNN, self).__init__()
self.cnn = nn.Sequential(
nn.Conv2d(1, 32, 3, padding=1),
nn.ReLU(),
nn.MaxPool2d(2),
nn.Conv2d(32, 64, 3, padding=1),
nn.ReLU(),
nn.MaxPool2d(2),
nn.Conv2d(64, 128, 3, padding=1),
nn.ReLU(),
nn.MaxPool2d((2, 1)),
)
self.rnn = nn.LSTM(128 * 8, hidden_size, bidirectional=True, batch_first=True)
self.classifier = nn.Linear(hidden_size * 2, vocab_size)
def forward(self, x):
conv_out = self.cnn(x)
b, c, h, w = conv_out.size()
conv_out = conv_out.view(b, c * h, w).permute(0, 2, 1)
rnn_out, _ = self.rnn(conv_out)
output = self.classifier(rnn_out)
return output
# Updated training function with proper HF Hub upload
def train_model(training_request: Dict[str, Any]) -> Dict[str, Any]:
try:
training_data = training_request['trainingData']
config = training_request['config']
session_id = training_request['sessionId']
logger.info(f"Starting training session {session_id} with {len(training_data)} samples")
dataset = OCRDataset(training_data)
dataloader = DataLoader(dataset, batch_size=config['batchSize'], shuffle=True)
model = SimpleCRNN()
optimizer = torch.optim.Adam(model.parameters(), lr=config['learningRate'])
model.train()
total_loss = 0.5
total_accuracy = 0.85
for epoch in range(config['epochs']):
epoch_loss = 0.5 - (epoch * 0.05) # Mock improving loss
epoch_accuracy = 0.7 + (epoch / config['epochs']) * 0.2 # Mock improving accuracy
logger.info(f"Epoch {epoch+1}/{config['epochs']} - Loss: {epoch_loss:.4f}, Accuracy: {epoch_accuracy:.4f}")
total_loss = epoch_loss
total_accuracy = epoch_accuracy
# FIXED: Use HF Hub API with proper authentication
model_filename = f"model_{session_id}.pth"
temp_model_path = f"/tmp/{model_filename}"
# Save model to temp directory first
torch.save(model.state_dict(), temp_model_path)
logger.info(f"Model saved to temp path: {temp_model_path}")
# Upload to HF Hub using API with authentication
try:
if HF_TOKEN:
hf_api.upload_file(
path_or_fileobj=temp_model_path,
path_in_repo=model_filename,
repo_id="rysatech/rysa-ocr-trainer",
repo_type="space",
commit_message=f"Add trained model {session_id}",
token=HF_TOKEN # Explicit token
)
logger.info(f"Model uploaded to HF Hub: {model_filename}")
else:
logger.error("No HF_TOKEN available for upload")
raise Exception("Authentication required for model upload")
except Exception as upload_error:
logger.error(f"Failed to upload model to HF Hub: {upload_error}")
raise upload_error # Don't continue if upload fails
# Clean up temp file
try:
os.remove(temp_model_path)
except:
pass
model_url = f"https://huggingface.co/spaces/rysatech/rysa-ocr-trainer/resolve/main/{model_filename}"
return {
'success': True,
'modelUrl': model_url,
'metrics': {
'finalLoss': total_loss,
'finalAccuracy': total_accuracy,
'epochs': config['epochs'],
'sampleCount': len(training_data)
}
}
except Exception as e:
logger.error(f"Training failed: {e}")
return {
'success': False,
'error': str(e)
}
# REST API ENDPOINTS
@fastapi_app.get("/api/health")
async def health_endpoint():
return {"status": "healthy", "gpu_available": torch.cuda.is_available()}
@fastapi_app.post("/api/train")
async def train_endpoint(request: TrainingRequest):
try:
# Convert Pydantic model to dict for your existing function
training_request_dict = {
'trainingData': [item.dict() for item in request.trainingData],
'config': request.config.dict(),
'sessionId': request.sessionId
}
result = train_model(training_request_dict)
if result['success']:
return TrainingResponse(
success=True,
jobId=request.sessionId,
modelUrl=result['modelUrl'],
metrics=TrainingMetrics(**result['metrics'])
)
else:
return TrainingResponse(
success=False,
error=result['error']
)
except Exception as e:
logger.error(f"API training failed: {e}")
return TrainingResponse(
success=False,
error=str(e)
)
# Gradio functions
def train_api(training_request_json: str) -> str:
try:
training_request = json.loads(training_request_json)
result = train_model(training_request)
return json.dumps(result)
except Exception as e:
return json.dumps({'success': False, 'error': str(e)})
def health_check() -> str:
return json.dumps({'status': 'healthy', 'gpu_available': torch.cuda.is_available()})
# Gradio interface
with gr.Blocks(title="Rysa OCR Training Service") as gradio_app:
gr.Markdown("# Rysa OCR Training Service")
gr.Markdown("Multi-script OCR model training for Sanskrit and Indian scripts.")
with gr.Tab("Training API"):
training_input = gr.Textbox(
label="Training Request JSON",
placeholder="Paste training request JSON here",
lines=10
)
train_button = gr.Button("Start Training")
training_output = gr.Textbox(label="Training Result", lines=10)
train_button.click(train_api, inputs=training_input, outputs=training_output)
with gr.Tab("Health Check"):
health_button = gr.Button("Check Health")
health_output = gr.Textbox(label="Health Status")
health_button.click(health_check, outputs=health_output)
# Mount Gradio app to FastAPI
app = gr.mount_gradio_app(fastapi_app, gradio_app, path="/")
if __name__ == "__main__":
uvicorn.run(app, host="0.0.0.0", port=7860)