temp_hf_repo/export_onnx.py

import sys
import os
import torch
import json
from safetensors.torch import load_file

# Add src to path
sys.path.append(os.path.join(os.getcwd(), 'DA-2-repo/src'))

try:
    from da2.model.spherevit import SphereViT
except ImportError as e:
    print(f"Error importing SphereViT: {e}")
    sys.exit(1)

# Load config
config_path = 'DA-2-repo/configs/infer.json'
with open(config_path, 'r') as f:
    config = json.load(f)

# Adjust config for fixed size export
# Using 1092x546 (multiples of 14: 1092=78*14, 546=39*14)
# This is closer to the original config's ~600k pixels
H, W = 546, 1092
config['inference']['min_pixels'] = H * W
config['inference']['max_pixels'] = H * W

print(f"Initializing model with input size {W}x{H}...")
# Instantiate model
model = SphereViT(config)

# Load weights
print("Loading weights from model.safetensors...")
try:
    weights = load_file('model.safetensors')
    missing, unexpected = model.load_state_dict(weights, strict=False)
    if missing:
        print(f"Missing keys: {len(missing)}")
        # print(missing[:5])
    if unexpected:
        print(f"Unexpected keys: {len(unexpected)}")
        # print(unexpected[:5])
except Exception as e:
    print(f"Error loading weights: {e}")
    sys.exit(1)

print("Exporting model in FP32 (full precision)...")
model.eval()

# Dummy input (float32)
dummy_input = torch.randn(1, 3, H, W)

# Export
output_file = "model.onnx"
print(f"Exporting to {output_file}...")
try:
    torch.onnx.export(
        model,
        dummy_input,
        output_file,
        opset_version=17,
        input_names=["input_image"],
        output_names=["depth_map"],
        dynamic_axes={
            "input_image": {0: "batch_size"}, 
            "depth_map": {0: "batch_size"}
        },
        export_params=True,
        do_constant_folding=True,
        verbose=False
    )
    print(f"Successfully exported to {output_file}")
except Exception as e:
    print(f"Error exporting to ONNX: {e}")
    import traceback
    traceback.print_exc()