Spaces:
Running
on
Zero
Running
on
Zero
Upload 4 files
Browse files- app.py +280 -0
- gen2seg_mae_pipeline.py +132 -0
- gen2seg_sd_pipeline.py +454 -0
- requirements.txt +10 -0
app.py
ADDED
@@ -0,0 +1,280 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from PIL import Image
|
4 |
+
import numpy as np
|
5 |
+
import time
|
6 |
+
import os
|
7 |
+
|
8 |
+
# --- Import Custom Pipelines ---
|
9 |
+
# Ensure these files are in the same directory or accessible in PYTHONPATH
|
10 |
+
try:
|
11 |
+
from gen2seg_sd_pipeline import gen2segSDPipeline
|
12 |
+
from gen2seg_mae_pipeline import gen2segMAEInstancePipeline
|
13 |
+
except ImportError as e:
|
14 |
+
print(f"Error importing pipeline modules: {e}")
|
15 |
+
print("Please ensure gen2seg_sd_pipeline.py and gen2seg_mae_pipeline.py are in the same directory.")
|
16 |
+
# Optionally, raise an error or exit if pipelines are critical at startup
|
17 |
+
# raise ImportError("Could not import custom pipeline modules. Check file paths.") from e
|
18 |
+
|
19 |
+
from transformers import ViTMAEForPreTraining, AutoImageProcessor
|
20 |
+
|
21 |
+
# --- Configuration ---
|
22 |
+
MODEL_IDS = {
|
23 |
+
"SD": "reachomk/gen2seg-sd",
|
24 |
+
"MAE-H": "reachomk/gen2seg-mae-h"
|
25 |
+
}
|
26 |
+
|
27 |
+
# Check if a GPU is available and set the device accordingly
|
28 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
29 |
+
print(f"Using device: {DEVICE}")
|
30 |
+
|
31 |
+
# --- Global Variables for Caching Pipelines ---
|
32 |
+
sd_pipe_global = None
|
33 |
+
mae_pipe_global = None
|
34 |
+
|
35 |
+
# --- Model Loading Functions ---
|
36 |
+
def get_sd_pipeline():
|
37 |
+
"""Loads and caches the gen2seg Stable Diffusion pipeline."""
|
38 |
+
global sd_pipe_global
|
39 |
+
if sd_pipe_global is None:
|
40 |
+
model_id_sd = MODEL_IDS["SD"]
|
41 |
+
print(f"Attempting to load SD pipeline from Hugging Face Hub: {model_id_sd}")
|
42 |
+
try:
|
43 |
+
sd_pipe_global = gen2segSDPipeline.from_pretrained(
|
44 |
+
model_id_sd,
|
45 |
+
use_safetensors=True,
|
46 |
+
# torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32, # Optional: use float16 on GPU
|
47 |
+
).to(DEVICE)
|
48 |
+
print(f"SD Pipeline loaded successfully from {model_id_sd} on {DEVICE}.")
|
49 |
+
except Exception as e:
|
50 |
+
print(f"Error loading SD pipeline from Hugging Face Hub ({model_id_sd}): {e}")
|
51 |
+
sd_pipe_global = None # Ensure it remains None on failure
|
52 |
+
# Do not raise gr.Error here; let the main function handle it.
|
53 |
+
return sd_pipe_global
|
54 |
+
|
55 |
+
def get_mae_pipeline():
|
56 |
+
"""Loads and caches the gen2seg MAE-H pipeline."""
|
57 |
+
global mae_pipe_global
|
58 |
+
if mae_pipe_global is None:
|
59 |
+
model_id_mae = MODEL_IDS["MAE-H"]
|
60 |
+
print(f"Loading MAE-H pipeline with model {model_id_mae} on {DEVICE}...")
|
61 |
+
try:
|
62 |
+
model = ViTMAEForPreTraining.from_pretrained(model_id_mae)
|
63 |
+
model.to(DEVICE)
|
64 |
+
model.eval() # Set to evaluation mode
|
65 |
+
|
66 |
+
# Load the official MAE-H image processor
|
67 |
+
# Using "facebook/vit-mae-huge" as per the original app_mae.py
|
68 |
+
image_processor = AutoImageProcessor.from_pretrained("facebook/vit-mae-huge")
|
69 |
+
|
70 |
+
mae_pipe_global = gen2segMAEInstancePipeline(model=model, image_processor=image_processor)
|
71 |
+
# The custom MAE pipeline's model is already on the DEVICE.
|
72 |
+
print(f"MAE-H Pipeline with model {model_id_mae} loaded successfully on {DEVICE}.")
|
73 |
+
except Exception as e:
|
74 |
+
print(f"Error loading MAE-H model or pipeline from Hugging Face Hub ({model_id_mae}): {e}")
|
75 |
+
mae_pipe_global = None # Ensure it remains None on failure
|
76 |
+
# Do not raise gr.Error here; let the main function handle it.
|
77 |
+
return mae_pipe_global
|
78 |
+
|
79 |
+
# --- Unified Prediction Function ---
|
80 |
+
def segment_image(input_image: Image.Image, model_choice: str) -> Image.Image:
|
81 |
+
"""
|
82 |
+
Takes a PIL Image and model choice, performs segmentation, and returns the segmented image.
|
83 |
+
"""
|
84 |
+
if input_image is None:
|
85 |
+
raise gr.Error("No image provided. Please upload an image.")
|
86 |
+
|
87 |
+
print(f"Model selected: {model_choice}")
|
88 |
+
# Ensure image is in RGB format
|
89 |
+
image_rgb = input_image.convert("RGB")
|
90 |
+
original_resolution = image_rgb.size # (width, height)
|
91 |
+
seg_array = None
|
92 |
+
|
93 |
+
try:
|
94 |
+
if model_choice == "SD":
|
95 |
+
pipe_sd = get_sd_pipeline()
|
96 |
+
if pipe_sd is None:
|
97 |
+
raise gr.Error("The SD segmentation pipeline could not be loaded. "
|
98 |
+
"Please check the Space logs for more details, or try again later.")
|
99 |
+
|
100 |
+
print(f"Running SD inference with image size: {image_rgb.size}")
|
101 |
+
start_time = time.time()
|
102 |
+
with torch.no_grad():
|
103 |
+
# The gen2segSDPipeline expects a single image or a list
|
104 |
+
# The pipeline's __call__ method handles preprocessing internally
|
105 |
+
seg_output = pipe_sd(image_rgb, match_input_resolution=False).prediction # Output is before resize
|
106 |
+
|
107 |
+
# seg_output is expected to be a numpy array (N,H,W,1) or (N,1,H,W) or tensor
|
108 |
+
# Based on gen2seg_sd_pipeline.py, if output_type="np" (default), it's [N,H,W,1]
|
109 |
+
# If output_type="pt", it's [N,1,H,W]
|
110 |
+
# The original app_sd.py converted tensor to numpy and squeezed.
|
111 |
+
if isinstance(seg_output, torch.Tensor):
|
112 |
+
seg_output = seg_output.cpu().numpy()
|
113 |
+
|
114 |
+
if seg_output.ndim == 4 and seg_output.shape[0] == 1: # Batch size 1
|
115 |
+
if seg_output.shape[1] == 1: # Grayscale, (1, 1, H, W)
|
116 |
+
seg_array = seg_output.squeeze(0).squeeze(0).astype(np.uint8)
|
117 |
+
elif seg_output.shape[-1] == 1: # Grayscale, (1, H, W, 1)
|
118 |
+
seg_array = seg_output.squeeze(0).squeeze(-1).astype(np.uint8)
|
119 |
+
elif seg_output.shape[1] == 3: # RGB, (1, 3, H, W) -> (H, W, 3)
|
120 |
+
seg_array = np.transpose(seg_output.squeeze(0), (1, 2, 0)).astype(np.uint8)
|
121 |
+
elif seg_output.shape[-1] == 3: # RGB, (1, H, W, 3)
|
122 |
+
seg_array = seg_output.squeeze(0).astype(np.uint8)
|
123 |
+
else: # Fallback for unexpected shapes
|
124 |
+
seg_array = seg_output.squeeze().astype(np.uint8)
|
125 |
+
|
126 |
+
elif seg_output.ndim == 3: # (H, W, C) or (C, H, W)
|
127 |
+
seg_array = seg_output.astype(np.uint8)
|
128 |
+
elif seg_output.ndim == 2: # (H,W)
|
129 |
+
seg_array = seg_output.astype(np.uint8)
|
130 |
+
else:
|
131 |
+
raise TypeError(f"Unexpected SD segmentation output type/shape: {type(seg_output)}, {seg_output.shape}")
|
132 |
+
end_time = time.time()
|
133 |
+
print(f"SD Inference completed in {end_time - start_time:.2f} seconds.")
|
134 |
+
|
135 |
+
|
136 |
+
elif model_choice == "MAE-H":
|
137 |
+
pipe_mae = get_mae_pipeline()
|
138 |
+
if pipe_mae is None:
|
139 |
+
raise gr.Error("The MAE-H segmentation pipeline could not be loaded. "
|
140 |
+
"Please check the Space logs for more details, or try again later.")
|
141 |
+
|
142 |
+
print(f"Running MAE-H inference with image size: {image_rgb.size}")
|
143 |
+
start_time = time.time()
|
144 |
+
with torch.no_grad():
|
145 |
+
# The gen2segMAEInstancePipeline expects a list of images
|
146 |
+
# output_type="np" returns a NumPy array
|
147 |
+
pipe_output = pipe_mae([image_rgb], output_type="np")
|
148 |
+
# Prediction is (batch_size, height, width, 3) for MAE
|
149 |
+
prediction_np = pipe_output.prediction[0] # Get the first (and only) image prediction
|
150 |
+
|
151 |
+
end_time = time.time()
|
152 |
+
print(f"MAE-H Inference completed in {end_time - start_time:.2f} seconds.")
|
153 |
+
|
154 |
+
if not isinstance(prediction_np, np.ndarray):
|
155 |
+
# This case should ideally not be reached if output_type="np"
|
156 |
+
prediction_np = prediction_np.cpu().numpy()
|
157 |
+
|
158 |
+
# Ensure it's in the expected (H, W, C) format and uint8
|
159 |
+
if prediction_np.ndim == 3 and prediction_np.shape[-1] == 3: # Expected (H, W, 3)
|
160 |
+
seg_array = prediction_np.astype(np.uint8)
|
161 |
+
else:
|
162 |
+
# Attempt to handle other shapes if necessary, or raise error
|
163 |
+
raise gr.Error(f"Unexpected MAE-H prediction shape: {prediction_np.shape}. Expected (H, W, 3).")
|
164 |
+
|
165 |
+
# The MAE pipeline already does gamma correction and scaling to 0-255.
|
166 |
+
# It also ensures 3 channels.
|
167 |
+
|
168 |
+
else:
|
169 |
+
raise gr.Error(f"Invalid model choice: {model_choice}. Please select a valid model.")
|
170 |
+
|
171 |
+
if seg_array is None:
|
172 |
+
raise gr.Error("Segmentation array was not generated. An unknown error occurred.")
|
173 |
+
|
174 |
+
print(f"Segmentation array generated with shape: {seg_array.shape}, dtype: {seg_array.dtype}")
|
175 |
+
|
176 |
+
# Convert numpy array to PIL Image
|
177 |
+
# Handle grayscale or RGB based on seg_array channels
|
178 |
+
if seg_array.ndim == 2: # Grayscale
|
179 |
+
segmented_image_pil = Image.fromarray(seg_array, mode='L')
|
180 |
+
elif seg_array.ndim == 3 and seg_array.shape[-1] == 3: # RGB
|
181 |
+
segmented_image_pil = Image.fromarray(seg_array, mode='RGB')
|
182 |
+
elif seg_array.ndim == 3 and seg_array.shape[-1] == 1: # Grayscale with channel dim
|
183 |
+
segmented_image_pil = Image.fromarray(seg_array.squeeze(-1), mode='L')
|
184 |
+
else:
|
185 |
+
raise gr.Error(f"Cannot convert seg_array with shape {seg_array.shape} to PIL Image.")
|
186 |
+
|
187 |
+
# Resize back to original image resolution using LANCZOS for high quality
|
188 |
+
segmented_image_pil = segmented_image_pil.resize(original_resolution, Image.Resampling.LANCZOS)
|
189 |
+
|
190 |
+
print(f"Segmented image processed. Output size: {segmented_image_pil.size}, mode: {segmented_image_pil.mode}")
|
191 |
+
return segmented_image_pil
|
192 |
+
|
193 |
+
except Exception as e:
|
194 |
+
print(f"Error during segmentation with {model_choice}: {e}")
|
195 |
+
# Re-raise as gr.Error for Gradio to display, if not already one
|
196 |
+
if not isinstance(e, gr.Error):
|
197 |
+
# It's often helpful to include the type of the original exception
|
198 |
+
error_type = type(e).__name__
|
199 |
+
raise gr.Error(f"An error occurred during segmentation: {error_type} - {str(e)}")
|
200 |
+
else:
|
201 |
+
raise e # Re-raise if it's already a gr.Error
|
202 |
+
|
203 |
+
# --- Gradio Interface ---
|
204 |
+
title = "gen2seg: Generative Models Enable Generalizable Instance Segmentation Demo (SD & MAE-H)"
|
205 |
+
description = f"""
|
206 |
+
<div style="text-align: center; font-family: 'Arial', sans-serif;">
|
207 |
+
<p>Upload an image and choose a model architecture to see the instance segmentation result generated by the respective model. </p>
|
208 |
+
<p>
|
209 |
+
Currently, inference is running on CPU.
|
210 |
+
Performance will be significantly better on GPU.
|
211 |
+
</p>
|
212 |
+
<ul>
|
213 |
+
<li><strong>SD</strong>: Based on Stable Diffusion 2.
|
214 |
+
<a href="https://huggingface.co/{MODEL_IDS['SD']}" target="_blank">Model Link</a>.
|
215 |
+
<em>Approx. CPU inference time: ~1-2 minutes per image.</em>
|
216 |
+
</li>
|
217 |
+
<li><strong>MAE-H</strong>: Based on Masked Autoencoder (Huge).
|
218 |
+
<a href="https://huggingface.co/{MODEL_IDS['MAE-H']}" target="_blank">Model Link</a>.
|
219 |
+
<em>Approx. CPU inference time: ~15-45 seconds per image.</em>
|
220 |
+
If you experience tokenizer artifacts or very dark images, you can use gamma correction to handle this.
|
221 |
+
</li>
|
222 |
+
</ul>
|
223 |
+
<p>
|
224 |
+
For faster inference, please check out our GitHub to run the models locally on a GPU:
|
225 |
+
<a href="https://github.com/UCDvision/gen2seg" target="_blank">https://github.com/UCDvision/gen2seg</a>
|
226 |
+
</p>
|
227 |
+
<p>If the demo experiences issues, please open an issue on our GitHub.</p>
|
228 |
+
<p> If you have not already, please see our webpage at <a href="https://reachomk.github.io/gen2seg" target="_blank">https://reachomk.github.io/gen2seg</a>
|
229 |
+
|
230 |
+
</div>
|
231 |
+
"""
|
232 |
+
|
233 |
+
article = """
|
234 |
+
"""
|
235 |
+
|
236 |
+
# Define Gradio inputs
|
237 |
+
input_image_component = gr.Image(type="pil", label="Input Image")
|
238 |
+
model_choice_component = gr.Dropdown(
|
239 |
+
choices=list(MODEL_IDS.keys()),
|
240 |
+
value="SD", # Default model
|
241 |
+
label="Choose Segmentation Model Architecture"
|
242 |
+
)
|
243 |
+
|
244 |
+
# Define Gradio output
|
245 |
+
output_image_component = gr.Image(type="pil", label="Segmented Image")
|
246 |
+
|
247 |
+
# Example images (ensure these paths are correct if you upload examples to your Space)
|
248 |
+
# For example, if you create an "examples" folder in your Space repo:
|
249 |
+
# example_paths = [
|
250 |
+
# os.path.join("examples", "example1.jpg"),
|
251 |
+
# os.path.join("examples", "example2.png")
|
252 |
+
# ]
|
253 |
+
# Filter out non-existent example files to prevent errors
|
254 |
+
# example_paths = [ex for ex in example_paths if os.path.exists(ex)]
|
255 |
+
example_paths = [] # Add paths to example images here if you have them
|
256 |
+
|
257 |
+
iface = gr.Interface(
|
258 |
+
fn=segment_image,
|
259 |
+
inputs=[input_image_component, model_choice_component],
|
260 |
+
outputs=output_image_component,
|
261 |
+
title=title,
|
262 |
+
description=description,
|
263 |
+
article=article,
|
264 |
+
examples=example_paths if example_paths else None, # Pass None if no examples
|
265 |
+
allow_flagging="never",
|
266 |
+
theme=gr.themes.Soft() # Using a soft theme for a slightly modern look
|
267 |
+
)
|
268 |
+
|
269 |
+
if __name__ == "__main__":
|
270 |
+
# Optional: Pre-load a default model on startup if desired.
|
271 |
+
# This can make the first inference faster but increases startup time.
|
272 |
+
# print("Attempting to pre-load default SD model on startup...")
|
273 |
+
# try:
|
274 |
+
# get_sd_pipeline() # Pre-load the default SD model
|
275 |
+
# print("Default SD model pre-loaded successfully or was already cached.")
|
276 |
+
# except Exception as e:
|
277 |
+
# print(f"Could not pre-load default SD model: {e}")
|
278 |
+
|
279 |
+
print("Launching Gradio interface...")
|
280 |
+
iface.launch()
|
gen2seg_mae_pipeline.py
ADDED
@@ -0,0 +1,132 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# gen2seg official inference pipeline code for Stable Diffusion model
|
2 |
+
#
|
3 |
+
# Please see our project website at https://reachomk.github.io/gen2seg
|
4 |
+
#
|
5 |
+
# Additionally, if you use our code please cite our paper, along with the two works above.
|
6 |
+
|
7 |
+
from dataclasses import dataclass
|
8 |
+
from typing import Union, List, Optional
|
9 |
+
|
10 |
+
import torch
|
11 |
+
import numpy as np
|
12 |
+
from PIL import Image
|
13 |
+
from einops import rearrange
|
14 |
+
|
15 |
+
from diffusers import DiffusionPipeline
|
16 |
+
from diffusers.utils import BaseOutput, logging
|
17 |
+
from transformers import AutoImageProcessor
|
18 |
+
|
19 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
20 |
+
|
21 |
+
|
22 |
+
@dataclass
|
23 |
+
class gen2segMAEInstanceOutput(BaseOutput):
|
24 |
+
"""
|
25 |
+
Output class for the ViTMAE Instance Segmentation Pipeline.
|
26 |
+
|
27 |
+
Args:
|
28 |
+
prediction (`np.ndarray` or `torch.Tensor`):
|
29 |
+
Predicted instance segmentation maps. The output has shape
|
30 |
+
`(batch_size, 3, height, width)` with pixel values scaled to [0, 255].
|
31 |
+
"""
|
32 |
+
prediction: Union[np.ndarray, torch.Tensor]
|
33 |
+
|
34 |
+
|
35 |
+
class gen2segMAEInstancePipeline(DiffusionPipeline):
|
36 |
+
r"""
|
37 |
+
Pipeline for Instance Segmentation using a fine-tuned ViTMAEForPreTraining model.
|
38 |
+
|
39 |
+
This pipeline takes one or more input images and returns an instance segmentation
|
40 |
+
prediction for each image. The model is assumed to have been fine-tuned using an instance
|
41 |
+
segmentation loss, and the reconstruction is performed by rearranging the model’s
|
42 |
+
patch logits into an image.
|
43 |
+
|
44 |
+
Args:
|
45 |
+
model (`ViTMAEForPreTraining`):
|
46 |
+
The fine-tuned ViTMAE model.
|
47 |
+
image_processor (`AutoImageProcessor`):
|
48 |
+
The image processor responsible for preprocessing input images.
|
49 |
+
"""
|
50 |
+
def __init__(self, model, image_processor):
|
51 |
+
super().__init__()
|
52 |
+
self.register_modules(model=model, image_processor=image_processor)
|
53 |
+
self.model = model
|
54 |
+
self.image_processor = image_processor
|
55 |
+
|
56 |
+
def check_inputs(
|
57 |
+
self,
|
58 |
+
image: Union[Image.Image, np.ndarray, torch.Tensor, List[Union[Image.Image, np.ndarray, torch.Tensor]]]
|
59 |
+
) -> List:
|
60 |
+
if not isinstance(image, list):
|
61 |
+
image = [image]
|
62 |
+
# Additional input validations can be added here if desired.
|
63 |
+
return image
|
64 |
+
|
65 |
+
@torch.no_grad()
|
66 |
+
def __call__(
|
67 |
+
self,
|
68 |
+
image: Union[Image.Image, np.ndarray, torch.Tensor, List[Union[Image.Image, np.ndarray, torch.Tensor]]],
|
69 |
+
output_type: str = "np",
|
70 |
+
**kwargs
|
71 |
+
) -> gen2segMAEInstanceOutput:
|
72 |
+
r"""
|
73 |
+
The call method of the pipeline.
|
74 |
+
|
75 |
+
Args:
|
76 |
+
image (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, or a list of these):
|
77 |
+
The input image(s) for instance segmentation. For arrays/tensors, expected values are in [0, 1].
|
78 |
+
output_type (`str`, optional, defaults to `"np"`):
|
79 |
+
The format of the output prediction. Choose `"np"` for a NumPy array or `"pt"` for a PyTorch tensor.
|
80 |
+
**kwargs:
|
81 |
+
Additional keyword arguments passed to the image processor.
|
82 |
+
|
83 |
+
Returns:
|
84 |
+
[`gen2segMAEInstanceOutput`]:
|
85 |
+
An output object containing the predicted instance segmentation maps.
|
86 |
+
"""
|
87 |
+
# 1. Check and prepare input images.
|
88 |
+
images = self.check_inputs(image)
|
89 |
+
inputs = self.image_processor(images=images, return_tensors="pt", **kwargs)
|
90 |
+
pixel_values = inputs["pixel_values"].to(self.device)
|
91 |
+
|
92 |
+
# 2. Forward pass through the model.
|
93 |
+
outputs = self.model(pixel_values=pixel_values)
|
94 |
+
logits = outputs.logits # Expected shape: (B, num_patches, patch_dim)
|
95 |
+
|
96 |
+
# 3. Retrieve patch size and image size from the model configuration.
|
97 |
+
patch_size = self.model.config.patch_size # e.g., 16
|
98 |
+
image_size = self.model.config.image_size # e.g., 224
|
99 |
+
grid_size = image_size // patch_size
|
100 |
+
|
101 |
+
# 4. Rearrange logits into the reconstructed image.
|
102 |
+
# The logits are reshaped from (B, num_patches, patch_dim) to (B, 3, H, W).
|
103 |
+
reconstructed = rearrange(
|
104 |
+
logits,
|
105 |
+
"b (h w) (p1 p2 c) -> b c (h p1) (w p2)",
|
106 |
+
h=grid_size,
|
107 |
+
p1=patch_size,
|
108 |
+
p2=patch_size,
|
109 |
+
c=3,
|
110 |
+
)
|
111 |
+
|
112 |
+
# 5. Post-process the reconstructed output.
|
113 |
+
# For each sample, shift and scale the prediction to [0, 255].
|
114 |
+
predictions = []
|
115 |
+
for i in range(reconstructed.shape[0]):
|
116 |
+
sample = reconstructed[i]
|
117 |
+
min_val = torch.abs(sample.min())
|
118 |
+
max_val = torch.abs(sample.max())
|
119 |
+
sample = (sample + min_val) / (max_val + min_val + 1e-5)
|
120 |
+
# sometimes the image is very dark so we perform gamma correction to "brighten" it
|
121 |
+
# in practice we can set this value to whatever we want or disable it entirely.
|
122 |
+
sample = sample**0.6
|
123 |
+
sample = sample * 255.0
|
124 |
+
predictions.append(sample)
|
125 |
+
prediction_tensor = torch.stack(predictions, dim=0).permute(0, 2, 3, 1)
|
126 |
+
|
127 |
+
# 6. Format the output.
|
128 |
+
if output_type == "np":
|
129 |
+
prediction = prediction_tensor.cpu().numpy()
|
130 |
+
else:
|
131 |
+
prediction = prediction_tensor
|
132 |
+
return gen2segMAEInstanceOutput(prediction=prediction)
|
gen2seg_sd_pipeline.py
ADDED
@@ -0,0 +1,454 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# gen2seg official inference pipeline code for Stable Diffusion model
|
2 |
+
#
|
3 |
+
# This code was adapted from Marigold and Diffusion E2E Finetuning.
|
4 |
+
#
|
5 |
+
# Please see our project website at https://reachomk.github.io/gen2seg
|
6 |
+
#
|
7 |
+
# Additionally, if you use our code please cite our paper, along with the two works above.
|
8 |
+
|
9 |
+
from dataclasses import dataclass
|
10 |
+
from typing import List, Optional, Tuple, Union
|
11 |
+
|
12 |
+
import numpy as np
|
13 |
+
import torch
|
14 |
+
from PIL import Image
|
15 |
+
from tqdm.auto import tqdm
|
16 |
+
from transformers import CLIPTextModel, CLIPTokenizer
|
17 |
+
|
18 |
+
from diffusers.image_processor import PipelineImageInput
|
19 |
+
from diffusers.models import (
|
20 |
+
AutoencoderKL,
|
21 |
+
UNet2DConditionModel,
|
22 |
+
)
|
23 |
+
from diffusers.schedulers import (
|
24 |
+
DDIMScheduler,
|
25 |
+
)
|
26 |
+
from diffusers.utils import (
|
27 |
+
BaseOutput,
|
28 |
+
logging,
|
29 |
+
)
|
30 |
+
from diffusers import DiffusionPipeline
|
31 |
+
from diffusers.pipelines.marigold.marigold_image_processing import MarigoldImageProcessor
|
32 |
+
|
33 |
+
# add
|
34 |
+
def zeros_tensor(
|
35 |
+
shape: Union[Tuple, List],
|
36 |
+
device: Optional["torch.device"] = None,
|
37 |
+
dtype: Optional["torch.dtype"] = None,
|
38 |
+
layout: Optional["torch.layout"] = None,
|
39 |
+
):
|
40 |
+
"""
|
41 |
+
A helper function to create tensors of zeros on the desired `device`.
|
42 |
+
Mirrors randn_tensor from diffusers.utils.torch_utils.
|
43 |
+
"""
|
44 |
+
layout = layout or torch.strided
|
45 |
+
device = device or torch.device("cpu")
|
46 |
+
latents = torch.zeros(list(shape), dtype=dtype, layout=layout).to(device)
|
47 |
+
return latents
|
48 |
+
|
49 |
+
|
50 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
51 |
+
|
52 |
+
@dataclass
|
53 |
+
class Gen2SegSDSegOutput(BaseOutput):
|
54 |
+
"""
|
55 |
+
Output class for gen2seg Instance Segmentation prediction pipeline.
|
56 |
+
|
57 |
+
Args:
|
58 |
+
prediction (`np.ndarray`, `torch.Tensor`):
|
59 |
+
Predicted instance segmentation with values in the range [0, 255]. The shape is always $numimages \times 1 \times height
|
60 |
+
\times width$, regardless of whether the images were passed as a 4D array or a list.
|
61 |
+
latent (`None`, `torch.Tensor`):
|
62 |
+
Latent features corresponding to the predictions, compatible with the `latents` argument of the pipeline.
|
63 |
+
The shape is $numimages * numensemble \times 4 \times latentheight \times latentwidth$.
|
64 |
+
"""
|
65 |
+
|
66 |
+
prediction: Union[np.ndarray, torch.Tensor]
|
67 |
+
latent: Union[None, torch.Tensor]
|
68 |
+
|
69 |
+
|
70 |
+
class Gen2SegSDPipeline(DiffusionPipeline):
|
71 |
+
"""
|
72 |
+
# add
|
73 |
+
Pipeline for Instance Segmentation prediction using our Stable Diffusion model.
|
74 |
+
Implementation is built upon Marigold: https://marigoldmonodepth.github.io and E2E FThttps://gonzalomartingarcia.github.io/diffusion-e2e-ft/
|
75 |
+
|
76 |
+
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
|
77 |
+
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
|
78 |
+
|
79 |
+
Args:
|
80 |
+
unet (`UNet2DConditionModel`):
|
81 |
+
Conditional U-Net to denoise the segmentation latent, synthesized from image latent.
|
82 |
+
vae (`AutoencoderKL`):
|
83 |
+
Variational Auto-Encoder (VAE) Model to encode and decode images and predictions to and from latent
|
84 |
+
representations.
|
85 |
+
scheduler (`DDIMScheduler`):
|
86 |
+
A scheduler to be used in combination with `unet` to denoise the encoded image latent.
|
87 |
+
text_encoder (`CLIPTextModel`):
|
88 |
+
Text-encoder, for empty text embedding.
|
89 |
+
tokenizer (`CLIPTokenizer`):
|
90 |
+
CLIP tokenizer.
|
91 |
+
default_processing_resolution (`int`, *optional*):
|
92 |
+
The recommended value of the `processing_resolution` parameter of the pipeline. This value must be set in
|
93 |
+
the model config. When the pipeline is called without explicitly setting `processing_resolution`, the
|
94 |
+
default value is used. This is required to ensure reasonable results with various model flavors trained
|
95 |
+
with varying optimal processing resolution values.
|
96 |
+
"""
|
97 |
+
|
98 |
+
model_cpu_offload_seq = "text_encoder->unet->vae"
|
99 |
+
|
100 |
+
def __init__(
|
101 |
+
self,
|
102 |
+
unet: UNet2DConditionModel,
|
103 |
+
vae: AutoencoderKL,
|
104 |
+
scheduler: Union[DDIMScheduler],
|
105 |
+
text_encoder: CLIPTextModel,
|
106 |
+
tokenizer: CLIPTokenizer,
|
107 |
+
default_processing_resolution: Optional[int] = 768, # add
|
108 |
+
):
|
109 |
+
super().__init__()
|
110 |
+
|
111 |
+
self.register_modules(
|
112 |
+
unet=unet,
|
113 |
+
vae=vae,
|
114 |
+
scheduler=scheduler,
|
115 |
+
text_encoder=text_encoder,
|
116 |
+
tokenizer=tokenizer,
|
117 |
+
)
|
118 |
+
self.register_to_config(
|
119 |
+
default_processing_resolution=default_processing_resolution,
|
120 |
+
)
|
121 |
+
|
122 |
+
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
|
123 |
+
self.default_processing_resolution = default_processing_resolution
|
124 |
+
self.empty_text_embedding = None
|
125 |
+
|
126 |
+
self.image_processor = MarigoldImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
127 |
+
|
128 |
+
def check_inputs(
|
129 |
+
self,
|
130 |
+
image: PipelineImageInput,
|
131 |
+
processing_resolution: int,
|
132 |
+
resample_method_input: str,
|
133 |
+
resample_method_output: str,
|
134 |
+
batch_size: int,
|
135 |
+
output_type: str,
|
136 |
+
) -> int:
|
137 |
+
if processing_resolution is None:
|
138 |
+
raise ValueError(
|
139 |
+
"`processing_resolution` is not specified and could not be resolved from the model config."
|
140 |
+
)
|
141 |
+
if processing_resolution < 0:
|
142 |
+
raise ValueError(
|
143 |
+
"`processing_resolution` must be non-negative: 0 for native resolution, or any positive value for "
|
144 |
+
"downsampled processing."
|
145 |
+
)
|
146 |
+
if processing_resolution % self.vae_scale_factor != 0:
|
147 |
+
raise ValueError(f"`processing_resolution` must be a multiple of {self.vae_scale_factor}.")
|
148 |
+
if resample_method_input not in ("nearest", "nearest-exact", "bilinear", "bicubic", "area"):
|
149 |
+
raise ValueError(
|
150 |
+
"`resample_method_input` takes string values compatible with PIL library: "
|
151 |
+
"nearest, nearest-exact, bilinear, bicubic, area."
|
152 |
+
)
|
153 |
+
if resample_method_output not in ("nearest", "nearest-exact", "bilinear", "bicubic", "area"):
|
154 |
+
raise ValueError(
|
155 |
+
"`resample_method_output` takes string values compatible with PIL library: "
|
156 |
+
"nearest, nearest-exact, bilinear, bicubic, area."
|
157 |
+
)
|
158 |
+
if batch_size < 1:
|
159 |
+
raise ValueError("`batch_size` must be positive.")
|
160 |
+
if output_type not in ["pt", "np"]:
|
161 |
+
raise ValueError("`output_type` must be one of `pt` or `np`.")
|
162 |
+
|
163 |
+
# image checks
|
164 |
+
num_images = 0
|
165 |
+
W, H = None, None
|
166 |
+
if not isinstance(image, list):
|
167 |
+
image = [image]
|
168 |
+
for i, img in enumerate(image):
|
169 |
+
if isinstance(img, np.ndarray) or torch.is_tensor(img):
|
170 |
+
if img.ndim not in (2, 3, 4):
|
171 |
+
raise ValueError(f"`image[{i}]` has unsupported dimensions or shape: {img.shape}.")
|
172 |
+
H_i, W_i = img.shape[-2:]
|
173 |
+
N_i = 1
|
174 |
+
if img.ndim == 4:
|
175 |
+
N_i = img.shape[0]
|
176 |
+
elif isinstance(img, Image.Image):
|
177 |
+
W_i, H_i = img.size
|
178 |
+
N_i = 1
|
179 |
+
else:
|
180 |
+
raise ValueError(f"Unsupported `image[{i}]` type: {type(img)}.")
|
181 |
+
if W is None:
|
182 |
+
W, H = W_i, H_i
|
183 |
+
elif (W, H) != (W_i, H_i):
|
184 |
+
raise ValueError(
|
185 |
+
f"Input `image[{i}]` has incompatible dimensions {(W_i, H_i)} with the previous images {(W, H)}"
|
186 |
+
)
|
187 |
+
num_images += N_i
|
188 |
+
|
189 |
+
return num_images
|
190 |
+
|
191 |
+
def progress_bar(self, iterable=None, total=None, desc=None, leave=True):
|
192 |
+
if not hasattr(self, "_progress_bar_config"):
|
193 |
+
self._progress_bar_config = {}
|
194 |
+
elif not isinstance(self._progress_bar_config, dict):
|
195 |
+
raise ValueError(
|
196 |
+
f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}."
|
197 |
+
)
|
198 |
+
|
199 |
+
progress_bar_config = dict(**self._progress_bar_config)
|
200 |
+
progress_bar_config["desc"] = progress_bar_config.get("desc", desc)
|
201 |
+
progress_bar_config["leave"] = progress_bar_config.get("leave", leave)
|
202 |
+
if iterable is not None:
|
203 |
+
return tqdm(iterable, **progress_bar_config)
|
204 |
+
elif total is not None:
|
205 |
+
return tqdm(total=total, **progress_bar_config)
|
206 |
+
else:
|
207 |
+
raise ValueError("Either `total` or `iterable` has to be defined.")
|
208 |
+
|
209 |
+
@torch.no_grad()
|
210 |
+
def __call__(
|
211 |
+
self,
|
212 |
+
image: PipelineImageInput,
|
213 |
+
processing_resolution: Optional[int] = None,
|
214 |
+
match_input_resolution: bool = False,
|
215 |
+
resample_method_input: str = "bilinear",
|
216 |
+
resample_method_output: str = "bilinear",
|
217 |
+
batch_size: int = 1,
|
218 |
+
output_type: str = "np",
|
219 |
+
output_latent: bool = False,
|
220 |
+
return_dict: bool = True,
|
221 |
+
):
|
222 |
+
"""
|
223 |
+
Function invoked when calling the pipeline.
|
224 |
+
|
225 |
+
Args:
|
226 |
+
image (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`),
|
227 |
+
`List[torch.Tensor]`: An input image or images used as an input for the instance segmentation task. For
|
228 |
+
arrays and tensors, the expected value range is between `[0, 1]`. Passing a batch of images is possible
|
229 |
+
by providing a four-dimensional array or a tensor. Additionally, a list of images of two- or
|
230 |
+
three-dimensional arrays or tensors can be passed. In the latter case, all list elements must have the
|
231 |
+
same width and height.
|
232 |
+
processing_resolution (`int`, *optional*, defaults to `None`):
|
233 |
+
Effective processing resolution. When set to `0`, matches the larger input image dimension. This
|
234 |
+
produces crisper predictions, but may also lead to the overall loss of global context. The default
|
235 |
+
value `None` resolves to the optimal value from the model config.
|
236 |
+
match_input_resolution (`bool`, *optional*, defaults to `True`):
|
237 |
+
When enabled, the output prediction is resized to match the input dimensions. When disabled, the longer
|
238 |
+
side of the output will equal to `processing_resolution`.
|
239 |
+
resample_method_input (`str`, *optional*, defaults to `"bilinear"`):
|
240 |
+
Resampling method used to resize input images to `processing_resolution`. The accepted values are:
|
241 |
+
`"nearest"`, `"nearest-exact"`, `"bilinear"`, `"bicubic"`, or `"area"`.
|
242 |
+
resample_method_output (`str`, *optional*, defaults to `"bilinear"`):
|
243 |
+
Resampling method used to resize output predictions to match the input resolution. The accepted values
|
244 |
+
are `"nearest"`, `"nearest-exact"`, `"bilinear"`, `"bicubic"`, or `"area"`.
|
245 |
+
batch_size (`int`, *optional*, defaults to `1`):
|
246 |
+
Batch size; only matters passing a tensor of images.
|
247 |
+
output_type (`str`, *optional*, defaults to `"np"`):
|
248 |
+
Preferred format of the output's `prediction`. The accepted ßvalues are: `"np"` (numpy array) or `"pt"` (torch tensor).
|
249 |
+
output_latent (`bool`, *optional*, defaults to `False`):
|
250 |
+
When enabled, the output's `latent` field contains the latent codes corresponding to the predictions
|
251 |
+
within the ensemble. These codes can be saved, modified, and used for subsequent calls with the
|
252 |
+
`latents` argument.
|
253 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
254 |
+
Whether or not to return a [`Gen2SegSDSegOutput`] instead of a plain tuple.
|
255 |
+
|
256 |
+
# add
|
257 |
+
E2E FT models are deterministic single step models involving no ensembling, i.e. E=1.
|
258 |
+
"""
|
259 |
+
|
260 |
+
# 0. Resolving variables.
|
261 |
+
device = self._execution_device
|
262 |
+
dtype = self.dtype
|
263 |
+
|
264 |
+
# Model-specific optimal default values leading to fast and reasonable results.
|
265 |
+
if processing_resolution is None:
|
266 |
+
processing_resolution = self.default_processing_resolution
|
267 |
+
|
268 |
+
#print(image[0].size)
|
269 |
+
#processing_resolution = 8 * round(max(image[0].size) / 8)
|
270 |
+
|
271 |
+
# 1. Check inputs.
|
272 |
+
num_images = self.check_inputs(
|
273 |
+
image,
|
274 |
+
processing_resolution,
|
275 |
+
resample_method_input,
|
276 |
+
resample_method_output,
|
277 |
+
batch_size,
|
278 |
+
output_type,
|
279 |
+
)
|
280 |
+
|
281 |
+
# 2. Prepare empty text conditioning.
|
282 |
+
# Model invocation: self.tokenizer, self.text_encoder.
|
283 |
+
prompt = ""
|
284 |
+
text_inputs = self.tokenizer(
|
285 |
+
prompt,
|
286 |
+
padding="do_not_pad",
|
287 |
+
max_length=self.tokenizer.model_max_length,
|
288 |
+
truncation=True,
|
289 |
+
return_tensors="pt",
|
290 |
+
)
|
291 |
+
text_input_ids = text_inputs.input_ids.to(device)
|
292 |
+
self.empty_text_embedding = self.text_encoder(text_input_ids)[0] # [1,2,1024]
|
293 |
+
|
294 |
+
# 3. Preprocess input images. This function loads input image or images of compatible dimensions `(H, W)`,
|
295 |
+
# optionally downsamples them to the `processing_resolution` `(PH, PW)`, where
|
296 |
+
# `max(PH, PW) == processing_resolution`, and pads the dimensions to `(PPH, PPW)` such that these values are
|
297 |
+
# divisible by the latent space downscaling factor (typically 8 in Stable Diffusion). The default value `None`
|
298 |
+
# of `processing_resolution` resolves to the optimal value from the model config. It is a recommended mode of
|
299 |
+
# operation and leads to the most reasonable results. Using the native image resolution or any other processing
|
300 |
+
# resolution can lead to loss of either fine details or global context in the output predictions.
|
301 |
+
image, padding, original_resolution = self.image_processor.preprocess(
|
302 |
+
image, processing_resolution, resample_method_input, device, dtype
|
303 |
+
) # [N,3,PPH,PPW]
|
304 |
+
# image =(image+torch.abs(image.min()))
|
305 |
+
# image = image/(torch.abs(image.max())+torch.abs(image.min()))
|
306 |
+
# # prediction = prediction**0.5
|
307 |
+
# #prediction = torch.clip(prediction, min=-1, max=1)+1
|
308 |
+
# image = (image) * 2
|
309 |
+
# image = image - 1
|
310 |
+
# 4. Encode input image into latent space. At this step, each of the `N` input images is represented with `E`
|
311 |
+
# ensemble members. Each ensemble member is an independent diffused prediction, just initialized independently.
|
312 |
+
# Latents of each such predictions across all input images and all ensemble members are represented in the
|
313 |
+
# `pred_latent` variable. The variable `image_latent` is of the same shape: it contains each input image encoded
|
314 |
+
# into latent space and replicated `E` times. Encoding into latent space happens in batches of size `batch_size`.
|
315 |
+
# Model invocation: self.vae.encoder.
|
316 |
+
image_latent, pred_latent = self.prepare_latents(
|
317 |
+
image, batch_size
|
318 |
+
) # [N*E,4,h,w], [N*E,4,h,w]
|
319 |
+
|
320 |
+
del image
|
321 |
+
|
322 |
+
batch_empty_text_embedding = self.empty_text_embedding.to(device=device, dtype=dtype).repeat(
|
323 |
+
batch_size, 1, 1
|
324 |
+
) # [B,1024,2]
|
325 |
+
|
326 |
+
# 5. Process the denoising loop. All `N * E` latents are processed sequentially in batches of size `batch_size`.
|
327 |
+
# The unet model takes concatenated latent spaces of the input image and the predicted modality as an input, and
|
328 |
+
# outputs noise for the predicted modality's latent space.
|
329 |
+
# Model invocation: self.unet.
|
330 |
+
pred_latents = []
|
331 |
+
|
332 |
+
for i in range(0, num_images, batch_size):
|
333 |
+
batch_image_latent = image_latent[i : i + batch_size] # [B,4,h,w]
|
334 |
+
batch_pred_latent = batch_image_latent[i : i + batch_size] # [B,4,h,w]
|
335 |
+
effective_batch_size = batch_image_latent.shape[0]
|
336 |
+
text = batch_empty_text_embedding[:effective_batch_size] # [B,2,1024]
|
337 |
+
|
338 |
+
# add
|
339 |
+
# Single step inference for E2E FT models
|
340 |
+
self.scheduler.set_timesteps(1, device=device)
|
341 |
+
for t in self.scheduler.timesteps:
|
342 |
+
batch_latent = batch_image_latent # torch.cat([batch_image_latent, batch_pred_latent], dim=1) # [B,8,h,w]
|
343 |
+
noise = self.unet(batch_latent, t, encoder_hidden_states=text, return_dict=False)[0] # [B,4,h,w]
|
344 |
+
batch_pred_latent = self.scheduler.step(
|
345 |
+
noise, t, batch_image_latent
|
346 |
+
).pred_original_sample # [B,4,h,w], # add
|
347 |
+
# directly take pred_original_sample rather than prev_sample
|
348 |
+
|
349 |
+
pred_latents.append(batch_pred_latent)
|
350 |
+
|
351 |
+
pred_latent = torch.cat(pred_latents, dim=0) # [N*E,4,h,w]
|
352 |
+
|
353 |
+
del (
|
354 |
+
pred_latents,
|
355 |
+
image_latent,
|
356 |
+
batch_empty_text_embedding,
|
357 |
+
batch_image_latent,
|
358 |
+
# batch_pred_latent,
|
359 |
+
text,
|
360 |
+
batch_latent,
|
361 |
+
noise,
|
362 |
+
)
|
363 |
+
|
364 |
+
# 6. Decode predictions from latent into pixel space. The resulting `N * E` predictions have shape `(PPH, PPW)`,
|
365 |
+
# which requires slight postprocessing. Decoding into pixel space happens in batches of size `batch_size`.
|
366 |
+
# Model invocation: self.vae.decoder.
|
367 |
+
prediction = torch.cat(
|
368 |
+
[
|
369 |
+
self.decode_prediction(pred_latent[i : i + batch_size])
|
370 |
+
for i in range(0, pred_latent.shape[0], batch_size)
|
371 |
+
],
|
372 |
+
dim=0,
|
373 |
+
) # [N*E,1,PPH,PPW]
|
374 |
+
|
375 |
+
if not output_latent:
|
376 |
+
pred_latent = None
|
377 |
+
|
378 |
+
# 7. Remove padding. The output shape is (PH, PW).
|
379 |
+
prediction = self.image_processor.unpad_image(prediction, padding) # [N*E,1,PH,PW]
|
380 |
+
|
381 |
+
# 9. If `match_input_resolution` is set, the output prediction are upsampled to match the
|
382 |
+
# input resolution `(H, W)`. This step may introduce upsampling artifacts, and therefore can be disabled.
|
383 |
+
# Depending on the downstream use-case, upsampling can be also chosen based on the tolerated artifacts by
|
384 |
+
# setting the `resample_method_output` parameter (e.g., to `"nearest"`).
|
385 |
+
if match_input_resolution:
|
386 |
+
prediction = self.image_processor.resize_antialias(
|
387 |
+
prediction, original_resolution, resample_method_output, is_aa=False
|
388 |
+
) # [N,1,H,W]
|
389 |
+
|
390 |
+
# 10. Prepare the final outputs.
|
391 |
+
if output_type == "np":
|
392 |
+
prediction = self.image_processor.pt_to_numpy(prediction) # [N,H,W,1]
|
393 |
+
|
394 |
+
# 11. Offload all models
|
395 |
+
self.maybe_free_model_hooks()
|
396 |
+
|
397 |
+
if not return_dict:
|
398 |
+
return (prediction, pred_latent)
|
399 |
+
|
400 |
+
return Gen2SegSDSegOutput(
|
401 |
+
prediction=prediction,
|
402 |
+
latent=pred_latent,
|
403 |
+
)
|
404 |
+
|
405 |
+
def prepare_latents(
|
406 |
+
self,
|
407 |
+
image: torch.Tensor,
|
408 |
+
batch_size: int,
|
409 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
410 |
+
def retrieve_latents(encoder_output):
|
411 |
+
if hasattr(encoder_output, "latent_dist"):
|
412 |
+
return encoder_output.latent_dist.mode()
|
413 |
+
elif hasattr(encoder_output, "latents"):
|
414 |
+
return encoder_output.latents
|
415 |
+
else:
|
416 |
+
raise AttributeError("Could not access latents of provided encoder_output")
|
417 |
+
|
418 |
+
image_latent = torch.cat(
|
419 |
+
[
|
420 |
+
retrieve_latents(self.vae.encode(image[i : i + batch_size]))
|
421 |
+
for i in range(0, image.shape[0], batch_size)
|
422 |
+
],
|
423 |
+
dim=0,
|
424 |
+
) # [N,4,h,w]
|
425 |
+
image_latent = image_latent * self.vae.config.scaling_factor # [N*E,4,h,w]
|
426 |
+
|
427 |
+
# add
|
428 |
+
# provide zeros as noised latent
|
429 |
+
pred_latent = zeros_tensor(
|
430 |
+
image_latent.shape,
|
431 |
+
device=image_latent.device,
|
432 |
+
dtype=image_latent.dtype,
|
433 |
+
) # [N*E,4,h,w]
|
434 |
+
|
435 |
+
return image_latent, pred_latent
|
436 |
+
|
437 |
+
def decode_prediction(self, pred_latent: torch.Tensor) -> torch.Tensor:
|
438 |
+
if pred_latent.dim() != 4 or pred_latent.shape[1] != self.vae.config.latent_channels:
|
439 |
+
raise ValueError(
|
440 |
+
f"Expecting 4D tensor of shape [B,{self.vae.config.latent_channels},H,W]; got {pred_latent.shape}."
|
441 |
+
)
|
442 |
+
|
443 |
+
prediction = self.vae.decode(pred_latent / self.vae.config.scaling_factor, return_dict=False)[0] # [B,3,H,W]
|
444 |
+
#print(prediction.max())
|
445 |
+
#print(prediction.min())
|
446 |
+
|
447 |
+
prediction =(prediction+torch.abs(prediction.min()))
|
448 |
+
prediction = prediction/(torch.abs(prediction.max())+torch.abs(prediction.min()))
|
449 |
+
#prediction = prediction**0.5
|
450 |
+
#prediction = torch.clip(prediction, min=-1, max=1)+1
|
451 |
+
prediction = (prediction) * 255.0
|
452 |
+
#print(prediction.max())
|
453 |
+
#print(prediction.min())
|
454 |
+
return prediction # [B,1,H,W]
|
requirements.txt
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
gradio
|
2 |
+
torch
|
3 |
+
torchvision
|
4 |
+
Pillow
|
5 |
+
numpy
|
6 |
+
diffusers
|
7 |
+
transformers
|
8 |
+
einops
|
9 |
+
tqdm
|
10 |
+
safetensors
|