Spaces:
Runtime error
Runtime error
Refine code and use text instead of file
Browse files
app.py
CHANGED
@@ -1,7 +1,9 @@
|
|
1 |
-
|
|
|
2 |
import streamlit as st
|
3 |
import io
|
4 |
import gc
|
|
|
5 |
|
6 |
########################################################################################################
|
7 |
# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
|
@@ -20,6 +22,8 @@ from torchvision.transforms import functional as VF
|
|
20 |
|
21 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
22 |
|
|
|
|
|
23 |
|
24 |
class ToBinary(torch.autograd.Function):
|
25 |
|
@@ -52,9 +56,8 @@ class ResBlock(nn.Module):
|
|
52 |
|
53 |
class REncoderSmall(nn.Module):
|
54 |
|
55 |
-
def __init__(self
|
56 |
super().__init__()
|
57 |
-
self.args = args
|
58 |
dd = 8
|
59 |
self.Bxx = nn.BatchNorm2d(dd * 64)
|
60 |
|
@@ -80,10 +83,7 @@ class REncoderSmall(nn.Module):
|
|
80 |
self.C22 = nn.Conv2d(dd * 64, 256, kernel_size=3, padding=1)
|
81 |
self.C23 = nn.Conv2d(256, dd * 64, kernel_size=3, padding=1)
|
82 |
|
83 |
-
self.COUT = nn.Conv2d(dd * 64,
|
84 |
-
args.my_img_bit,
|
85 |
-
kernel_size=3,
|
86 |
-
padding=1)
|
87 |
|
88 |
def forward(self, img):
|
89 |
ACT = F.mish
|
@@ -110,14 +110,10 @@ class REncoderSmall(nn.Module):
|
|
110 |
|
111 |
class RDecoderSmall(nn.Module):
|
112 |
|
113 |
-
def __init__(self
|
114 |
super().__init__()
|
115 |
-
self.args = args
|
116 |
dd = 8
|
117 |
-
self.CIN = nn.Conv2d(
|
118 |
-
dd * 64,
|
119 |
-
kernel_size=3,
|
120 |
-
padding=1)
|
121 |
|
122 |
self.B00 = nn.BatchNorm2d(dd * 64)
|
123 |
self.C00 = nn.Conv2d(dd * 64, 256, kernel_size=3, padding=1)
|
@@ -165,9 +161,8 @@ class RDecoderSmall(nn.Module):
|
|
165 |
|
166 |
class REncoderLarge(nn.Module):
|
167 |
|
168 |
-
def __init__(self,
|
169 |
super().__init__()
|
170 |
-
self.args = args
|
171 |
self.CXX = nn.Conv2d(3, dd, kernel_size=3, padding=1)
|
172 |
self.BXX = nn.BatchNorm2d(dd)
|
173 |
self.CX0 = nn.Conv2d(dd, ee, kernel_size=3, padding=1)
|
@@ -175,10 +170,7 @@ class REncoderLarge(nn.Module):
|
|
175 |
self.R0 = ResBlock(dd * 4, ff)
|
176 |
self.R1 = ResBlock(dd * 16, ff)
|
177 |
self.R2 = ResBlock(dd * 64, ff)
|
178 |
-
self.CZZ = nn.Conv2d(dd * 64,
|
179 |
-
args.my_img_bit,
|
180 |
-
kernel_size=3,
|
181 |
-
padding=1)
|
182 |
|
183 |
def forward(self, x):
|
184 |
ACT = F.mish
|
@@ -198,13 +190,9 @@ class REncoderLarge(nn.Module):
|
|
198 |
|
199 |
class RDecoderLarge(nn.Module):
|
200 |
|
201 |
-
def __init__(self,
|
202 |
super().__init__()
|
203 |
-
self.
|
204 |
-
self.CZZ = nn.Conv2d(args.my_img_bit,
|
205 |
-
dd * 64,
|
206 |
-
kernel_size=3,
|
207 |
-
padding=1)
|
208 |
self.BZZ = nn.BatchNorm2d(dd * 64)
|
209 |
self.R0 = ResBlock(dd * 64, ff)
|
210 |
self.R1 = ResBlock(dd * 16, ff)
|
@@ -234,32 +222,22 @@ def prepare_model(model_prefix):
|
|
234 |
gc.collect()
|
235 |
|
236 |
if model_prefix == 'out-v7c_d8_256-224-13bit-OB32x0.5-745':
|
237 |
-
R_ENCODER, R_DECODER = REncoderSmall, RDecoderSmall
|
238 |
else:
|
239 |
if 'd16_512' in model_prefix:
|
240 |
dd, ee, ff = 16, 64, 512
|
241 |
elif 'd32_1024' in model_prefix:
|
242 |
dd, ee, ff = 32, 128, 1024
|
243 |
-
R_ENCODER
|
244 |
-
|
245 |
-
|
246 |
-
args = types.SimpleNamespace()
|
247 |
-
args.my_img_bit = 13
|
248 |
-
encoder = R_ENCODER(args).eval().to(device)
|
249 |
-
decoder = R_DECODER(args).eval().to(device)
|
250 |
|
251 |
-
|
252 |
-
|
253 |
|
254 |
encoder.load_state_dict(
|
255 |
-
torch.load(
|
256 |
-
cached_download(hf_hub_url(MODEL_REPO, f'{model_prefix}-E.pth'))))
|
257 |
decoder.load_state_dict(
|
258 |
-
torch.load(
|
259 |
-
cached_download(hf_hub_url(MODEL_REPO, f'{model_prefix}-D.pth'))))
|
260 |
-
|
261 |
-
encoder.eval()
|
262 |
-
decoder.eval()
|
263 |
|
264 |
return encoder, decoder
|
265 |
|
@@ -277,11 +255,23 @@ def encode(model_prefix, img):
|
|
277 |
z = encoder(img)
|
278 |
z = ToBinary.apply(z)
|
279 |
|
280 |
-
|
|
|
|
|
|
|
|
|
281 |
|
282 |
|
283 |
-
def decode(model_prefix,
|
284 |
_, decoder = prepare_model(model_prefix)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
285 |
decoded = decoder(torch.Tensor(z).to(device))
|
286 |
return VF.to_pil_image(decoded[0])
|
287 |
|
@@ -300,20 +290,14 @@ with encoder_tab:
|
|
300 |
if uploaded_file is not None:
|
301 |
image = Image.open(uploaded_file)
|
302 |
col_in.image(image, 'Input Image')
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
307 |
-
label="Download Encoded Data",
|
308 |
-
data=buffer,
|
309 |
-
file_name=uploaded_file.name + '.npy',
|
310 |
-
)
|
311 |
-
col_out.image(decode(model_prefix, z), 'Output Image preview')
|
312 |
|
313 |
with decoder_tab:
|
314 |
col_in, col_out = st.columns(2)
|
315 |
-
|
316 |
-
if
|
317 |
-
|
318 |
-
image = decode(model_prefix, z)
|
319 |
col_out.image(image, 'Output Image')
|
|
|
1 |
+
import base64
|
2 |
+
from huggingface_hub import hf_hub_download
|
3 |
import streamlit as st
|
4 |
import io
|
5 |
import gc
|
6 |
+
import json
|
7 |
|
8 |
########################################################################################################
|
9 |
# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
|
|
|
22 |
|
23 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
24 |
|
25 |
+
IMG_BITS = 13
|
26 |
+
|
27 |
|
28 |
class ToBinary(torch.autograd.Function):
|
29 |
|
|
|
56 |
|
57 |
class REncoderSmall(nn.Module):
|
58 |
|
59 |
+
def __init__(self):
|
60 |
super().__init__()
|
|
|
61 |
dd = 8
|
62 |
self.Bxx = nn.BatchNorm2d(dd * 64)
|
63 |
|
|
|
83 |
self.C22 = nn.Conv2d(dd * 64, 256, kernel_size=3, padding=1)
|
84 |
self.C23 = nn.Conv2d(256, dd * 64, kernel_size=3, padding=1)
|
85 |
|
86 |
+
self.COUT = nn.Conv2d(dd * 64, IMG_BITS, kernel_size=3, padding=1)
|
|
|
|
|
|
|
87 |
|
88 |
def forward(self, img):
|
89 |
ACT = F.mish
|
|
|
110 |
|
111 |
class RDecoderSmall(nn.Module):
|
112 |
|
113 |
+
def __init__(self):
|
114 |
super().__init__()
|
|
|
115 |
dd = 8
|
116 |
+
self.CIN = nn.Conv2d(IMG_BITS, dd * 64, kernel_size=3, padding=1)
|
|
|
|
|
|
|
117 |
|
118 |
self.B00 = nn.BatchNorm2d(dd * 64)
|
119 |
self.C00 = nn.Conv2d(dd * 64, 256, kernel_size=3, padding=1)
|
|
|
161 |
|
162 |
class REncoderLarge(nn.Module):
|
163 |
|
164 |
+
def __init__(self, dd, ee, ff):
|
165 |
super().__init__()
|
|
|
166 |
self.CXX = nn.Conv2d(3, dd, kernel_size=3, padding=1)
|
167 |
self.BXX = nn.BatchNorm2d(dd)
|
168 |
self.CX0 = nn.Conv2d(dd, ee, kernel_size=3, padding=1)
|
|
|
170 |
self.R0 = ResBlock(dd * 4, ff)
|
171 |
self.R1 = ResBlock(dd * 16, ff)
|
172 |
self.R2 = ResBlock(dd * 64, ff)
|
173 |
+
self.CZZ = nn.Conv2d(dd * 64, IMG_BITS, kernel_size=3, padding=1)
|
|
|
|
|
|
|
174 |
|
175 |
def forward(self, x):
|
176 |
ACT = F.mish
|
|
|
190 |
|
191 |
class RDecoderLarge(nn.Module):
|
192 |
|
193 |
+
def __init__(self, dd, ee, ff):
|
194 |
super().__init__()
|
195 |
+
self.CZZ = nn.Conv2d(IMG_BITS, dd * 64, kernel_size=3, padding=1)
|
|
|
|
|
|
|
|
|
196 |
self.BZZ = nn.BatchNorm2d(dd * 64)
|
197 |
self.R0 = ResBlock(dd * 64, ff)
|
198 |
self.R1 = ResBlock(dd * 16, ff)
|
|
|
222 |
gc.collect()
|
223 |
|
224 |
if model_prefix == 'out-v7c_d8_256-224-13bit-OB32x0.5-745':
|
225 |
+
R_ENCODER, R_DECODER = REncoderSmall(), RDecoderSmall()
|
226 |
else:
|
227 |
if 'd16_512' in model_prefix:
|
228 |
dd, ee, ff = 16, 64, 512
|
229 |
elif 'd32_1024' in model_prefix:
|
230 |
dd, ee, ff = 32, 128, 1024
|
231 |
+
R_ENCODER = REncoderLarge(dd, ee, ff)
|
232 |
+
R_DECODER = RDecoderLarge(dd, ee, ff)
|
|
|
|
|
|
|
|
|
|
|
233 |
|
234 |
+
encoder = R_ENCODER.eval().to(device)
|
235 |
+
decoder = R_DECODER.eval().to(device)
|
236 |
|
237 |
encoder.load_state_dict(
|
238 |
+
torch.load(hf_hub_download(MODEL_REPO, f'{model_prefix}-E.pth')))
|
|
|
239 |
decoder.load_state_dict(
|
240 |
+
torch.load(hf_hub_download(MODEL_REPO, f'{model_prefix}-D.pth')))
|
|
|
|
|
|
|
|
|
241 |
|
242 |
return encoder, decoder
|
243 |
|
|
|
255 |
z = encoder(img)
|
256 |
z = ToBinary.apply(z)
|
257 |
|
258 |
+
with io.BytesIO() as buffer:
|
259 |
+
np.save(buffer, np.packbits(z.cpu().numpy().astype('bool')))
|
260 |
+
z_b64 = base64.b64encode(buffer.getvalue()).decode()
|
261 |
+
|
262 |
+
return json.dumps({"shape": list(z.shape), "data": z_b64})
|
263 |
|
264 |
|
265 |
+
def decode(model_prefix, z_str):
|
266 |
_, decoder = prepare_model(model_prefix)
|
267 |
+
|
268 |
+
z_json = json.loads(z_str)
|
269 |
+
with io.BytesIO() as buffer:
|
270 |
+
buffer.write(base64.b64decode(z_json["data"]))
|
271 |
+
buffer.seek(0)
|
272 |
+
z = np.load(buffer)
|
273 |
+
z = np.unpackbits(z).astype('float').reshape(z_json["shape"])
|
274 |
+
|
275 |
decoded = decoder(torch.Tensor(z).to(device))
|
276 |
return VF.to_pil_image(decoded[0])
|
277 |
|
|
|
290 |
if uploaded_file is not None:
|
291 |
image = Image.open(uploaded_file)
|
292 |
col_in.image(image, 'Input Image')
|
293 |
+
z_str = encode(model_prefix, image)
|
294 |
+
col_out.write("Encoded to:")
|
295 |
+
col_out.code(z_str,language=None)
|
296 |
+
col_out.image(decode(model_prefix, z_str), 'Output Image preview')
|
|
|
|
|
|
|
|
|
|
|
297 |
|
298 |
with decoder_tab:
|
299 |
col_in, col_out = st.columns(2)
|
300 |
+
z_str = col_in.text_area('Paste encoded string here:')
|
301 |
+
if len(z_str) > 0:
|
302 |
+
image = decode(model_prefix, z_str)
|
|
|
303 |
col_out.image(image, 'Output Image')
|