Spaces:
Running
Running
Droping Metrics
Browse files
app.py
CHANGED
@@ -13,11 +13,8 @@ try:
|
|
13 |
except Exception:
|
14 |
kc = None
|
15 |
|
16 |
-
from skimage.metrics import peak_signal_noise_ratio as psnr_metric
|
17 |
-
from skimage.metrics import structural_similarity as ssim_metric
|
18 |
from model import ViTUNetColorizer
|
19 |
|
20 |
-
|
21 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
22 |
|
23 |
CKPT = "checkpoints/checkpoint_epoch_017_20250810_193435.pt"
|
@@ -57,49 +54,34 @@ def pad_to_multiple(img_np, m=16):
|
|
57 |
ph, pw = math.ceil(h/m)*m, math.ceil(w/m)*m
|
58 |
return cv2.copyMakeBorder(img_np,0,ph-h,0,pw-w,cv2.BORDER_CONSTANT,value=(0,0,0)), (h,w)
|
59 |
|
60 |
-
def compute_metrics(pred, gt):
|
61 |
-
p = pred.astype(np.float32)/255.; g = gt.astype(np.float32)/255.
|
62 |
-
mae = float(np.mean(np.abs(p-g)))
|
63 |
-
psnr = float(psnr_metric(g, p, data_range=1.0))
|
64 |
-
try:
|
65 |
-
ssim = float(ssim_metric(g, p, channel_axis=2, data_range=1.0, win_size=7))
|
66 |
-
except TypeError:
|
67 |
-
ssim = float(ssim_metric(g, p, multichannel=True, data_range=1.0, win_size=7))
|
68 |
-
return round(mae,4), round(psnr,2), round(ssim,4)
|
69 |
-
|
70 |
def to_grayscale(image):
|
71 |
if image is None:
|
72 |
return None
|
73 |
return image.convert("L").convert("RGB")
|
74 |
|
75 |
-
def infer(image: Image.Image
|
76 |
if image is None:
|
77 |
-
return None, None
|
78 |
if model is None:
|
79 |
-
return None,
|
80 |
|
81 |
pil = image.convert("RGB")
|
82 |
rgb = np.array(pil)
|
83 |
-
|
84 |
proc, (oh, ow) = pad_to_multiple(rgb, 16); back = (ow, oh)
|
85 |
|
86 |
L = to_L(proc)
|
87 |
with torch.no_grad():
|
88 |
ab = model(L)
|
89 |
out = lab_to_rgb(L, ab)
|
90 |
-
|
91 |
out = out[:back[1], :back[0]]
|
92 |
|
93 |
-
mae = psnr = ssim = None
|
94 |
-
if want_metrics:
|
95 |
-
mae, psnr, ssim = compute_metrics(out, np.array(pil))
|
96 |
-
|
97 |
gray_pil = pil.convert("L").convert("RGB")
|
98 |
_, bo = cv2.imencode(".jpg", cv2.cvtColor(np.array(gray_pil), cv2.COLOR_RGB2BGR))
|
99 |
_, bc = cv2.imencode(".jpg", cv2.cvtColor(out, cv2.COLOR_RGB2BGR))
|
100 |
so = "data:image/jpeg;base64," + base64.b64encode(bo).decode()
|
101 |
sc = "data:image/jpeg;base64," + base64.b64encode(bc).decode()
|
102 |
-
|
103 |
compare_html = f"""
|
104 |
<div style="margin:auto; border-radius:14px; overflow:hidden;">
|
105 |
<img-comparison-slider>
|
@@ -109,7 +91,7 @@ def infer(image: Image.Image, want_metrics: bool):
|
|
109 |
</div>
|
110 |
"""
|
111 |
|
112 |
-
return out,
|
113 |
|
114 |
def save_for_download(image_array):
|
115 |
"""Saves a NumPy array to a temporary file and returns the path."""
|
@@ -159,7 +141,6 @@ with gr.Blocks(theme=THEME, title="Image Colorizer", head=HEAD) as demo:
|
|
159 |
sources=["upload", "webcam", "clipboard"]
|
160 |
)
|
161 |
img_in.upload(fn=to_grayscale, inputs=img_in, outputs=img_in)
|
162 |
-
show_m = gr.Checkbox(label="Show metrics", value=True)
|
163 |
with gr.Row():
|
164 |
run = gr.Button("Colorize")
|
165 |
clr = gr.Button("Clear")
|
@@ -174,33 +155,25 @@ with gr.Blocks(theme=THEME, title="Image Colorizer", head=HEAD) as demo:
|
|
174 |
|
175 |
with gr.Column(scale=7):
|
176 |
out_html = gr.HTML(label="Result", value=PLACEHOLDER_HTML)
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
ssim_box = gr.Number(label="SSIM", interactive=False, precision=4)
|
181 |
-
|
182 |
-
def _go(image, want_metrics):
|
183 |
-
out_image, mae, psnr, ssim, cmp_html = infer(image, want_metrics)
|
184 |
-
if not want_metrics:
|
185 |
-
mae = psnr = ssim = None
|
186 |
-
|
187 |
download_button_update = gr.update(visible=True) if out_image is not None else gr.update(visible=False)
|
188 |
-
|
189 |
-
return out_image, cmp_html, mae, psnr, ssim, download_button_update
|
190 |
|
191 |
run.click(
|
192 |
_go,
|
193 |
-
inputs=[img_in
|
194 |
-
outputs=[result_state, out_html,
|
195 |
)
|
196 |
|
197 |
def _clear():
|
198 |
-
return None, None, PLACEHOLDER_HTML,
|
199 |
|
200 |
clr.click(
|
201 |
_clear,
|
202 |
inputs=None,
|
203 |
-
outputs=[img_in, result_state, out_html,
|
204 |
)
|
205 |
|
206 |
download_btn.click(
|
@@ -213,4 +186,4 @@ if __name__ == "__main__":
|
|
213 |
try:
|
214 |
demo.launch(show_api=False)
|
215 |
except TypeError:
|
216 |
-
demo.launch()
|
|
|
13 |
except Exception:
|
14 |
kc = None
|
15 |
|
|
|
|
|
16 |
from model import ViTUNetColorizer
|
17 |
|
|
|
18 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
19 |
|
20 |
CKPT = "checkpoints/checkpoint_epoch_017_20250810_193435.pt"
|
|
|
54 |
ph, pw = math.ceil(h/m)*m, math.ceil(w/m)*m
|
55 |
return cv2.copyMakeBorder(img_np,0,ph-h,0,pw-w,cv2.BORDER_CONSTANT,value=(0,0,0)), (h,w)
|
56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
def to_grayscale(image):
|
58 |
if image is None:
|
59 |
return None
|
60 |
return image.convert("L").convert("RGB")
|
61 |
|
62 |
+
def infer(image: Image.Image):
|
63 |
if image is None:
|
64 |
+
return None, None
|
65 |
if model is None:
|
66 |
+
return None, "<div>Checkpoint not found.</div>"
|
67 |
|
68 |
pil = image.convert("RGB")
|
69 |
rgb = np.array(pil)
|
70 |
+
|
71 |
proc, (oh, ow) = pad_to_multiple(rgb, 16); back = (ow, oh)
|
72 |
|
73 |
L = to_L(proc)
|
74 |
with torch.no_grad():
|
75 |
ab = model(L)
|
76 |
out = lab_to_rgb(L, ab)
|
|
|
77 |
out = out[:back[1], :back[0]]
|
78 |
|
|
|
|
|
|
|
|
|
79 |
gray_pil = pil.convert("L").convert("RGB")
|
80 |
_, bo = cv2.imencode(".jpg", cv2.cvtColor(np.array(gray_pil), cv2.COLOR_RGB2BGR))
|
81 |
_, bc = cv2.imencode(".jpg", cv2.cvtColor(out, cv2.COLOR_RGB2BGR))
|
82 |
so = "data:image/jpeg;base64," + base64.b64encode(bo).decode()
|
83 |
sc = "data:image/jpeg;base64," + base64.b64encode(bc).decode()
|
84 |
+
|
85 |
compare_html = f"""
|
86 |
<div style="margin:auto; border-radius:14px; overflow:hidden;">
|
87 |
<img-comparison-slider>
|
|
|
91 |
</div>
|
92 |
"""
|
93 |
|
94 |
+
return out, compare_html
|
95 |
|
96 |
def save_for_download(image_array):
|
97 |
"""Saves a NumPy array to a temporary file and returns the path."""
|
|
|
141 |
sources=["upload", "webcam", "clipboard"]
|
142 |
)
|
143 |
img_in.upload(fn=to_grayscale, inputs=img_in, outputs=img_in)
|
|
|
144 |
with gr.Row():
|
145 |
run = gr.Button("Colorize")
|
146 |
clr = gr.Button("Clear")
|
|
|
155 |
|
156 |
with gr.Column(scale=7):
|
157 |
out_html = gr.HTML(label="Result", value=PLACEHOLDER_HTML)
|
158 |
+
|
159 |
+
def _go(image):
|
160 |
+
out_image, cmp_html = infer(image)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
161 |
download_button_update = gr.update(visible=True) if out_image is not None else gr.update(visible=False)
|
162 |
+
return out_image, cmp_html, download_button_update
|
|
|
163 |
|
164 |
run.click(
|
165 |
_go,
|
166 |
+
inputs=[img_in],
|
167 |
+
outputs=[result_state, out_html, download_btn]
|
168 |
)
|
169 |
|
170 |
def _clear():
|
171 |
+
return None, None, PLACEHOLDER_HTML, gr.update(visible=False)
|
172 |
|
173 |
clr.click(
|
174 |
_clear,
|
175 |
inputs=None,
|
176 |
+
outputs=[img_in, result_state, out_html, download_btn]
|
177 |
)
|
178 |
|
179 |
download_btn.click(
|
|
|
186 |
try:
|
187 |
demo.launch(show_api=False)
|
188 |
except TypeError:
|
189 |
+
demo.launch()
|