File size: 2,446 Bytes
e4634c2 1a260cd e4634c2 1a260cd e4634c2 ee9d4b8 e027211 ee9d4b8 e027211 ee9d4b8 e4634c2 c76cec1 e4634c2 c76cec1 e4634c2 c76cec1 e4634c2 c76cec1 e4634c2 8ac071c e4634c2 c76cec1 e4634c2 1a260cd e4634c2 c76cec1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 |
import torch
from PIL import Image
import torchvision.transforms as T
import numpy as np
import os
from modeling_my_segformer import MySegformerForSemanticSegmentation
from mix_vision_transformer_config import MySegformerConfig
# Gerät auswählen
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")
# Modell laden
model_name_or_path = "TimM77/SegformerPlusPlus"
print("Starte config_load")
config = MySegformerConfig.from_pretrained(model_name_or_path)
print("Starte Model_load")
model = MySegformerForSemanticSegmentation.from_pretrained(model_name_or_path, config=config)
model.to(device).eval()
# Bild laden
image_path = "segformer_plusplus/cityscape/berlin_000543_000019_leftImg8bit.png"
image = Image.open(image_path).convert("RGB")
# --- Start of changes to match cityscape_benchmark ---
# Define the target image size as in cityscape_benchmark's default
# cityscape_benchmark uses (3, 1024, 1024), so spatial size is 1024x1024
target_image_height = 1024
target_image_width = 1024
# Calculate mean and std dynamically from the image as done in cityscape_benchmark
# Note: This is usually done over the entire training dataset for consistent normalization
# For a single image, this just normalizes to its own mean/std.
img_tensor_temp = T.ToTensor()(image)
mean = img_tensor_temp.mean(dim=(1, 2)).tolist()
std = img_tensor_temp.std(dim=(1, 2)).tolist()
print(f"Calculated Mean (for this image): {mean}")
print(f"Calculated Std (for this image): {std}")
# Preprocessing - Adjusted to match cityscape_benchmark's T.Resize and T.Normalize
transform = T.Compose([
T.Resize((target_image_height, target_image_width)), # Resize to 1024x1024
T.ToTensor(),
T.Normalize(mean=mean, std=std) # Use dynamically calculated mean/std
])
input_tensor = transform(image).unsqueeze(0).to(device)
# --- End of changes ---
print("Modell geladen, Bild geladen, Preprocessing abgeschlossen")
# Inferenz
with torch.no_grad():
output = model(input_tensor)
# This ensures you're always getting the raw logits if the model returns an object
logits = output.logits if hasattr(output, "logits") else output
pred = torch.argmax(logits, dim=1).squeeze(0).cpu().numpy()
# Ergebnis als Textdatei speichern
output_path = os.path.join("segformer_plusplus", "cityscapes_prediction_output_overHF.txt")
np.savetxt(output_path, pred, fmt="%d")
print(f"Prediction saved as {output_path}") |