Spaces:
Runtime error
Runtime error
MeetMeAt92
commited on
Commit
•
9123e7a
1
Parent(s):
de04575
Create model.h5
Browse files
model.h5
ADDED
@@ -0,0 +1,331 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import cv2
|
3 |
+
import random
|
4 |
+
import numpy as np
|
5 |
+
from glob import glob
|
6 |
+
from PIL import Image, ImageOps
|
7 |
+
import matplotlib.pyplot as plt
|
8 |
+
|
9 |
+
import tensorflow as tf
|
10 |
+
from tensorflow import keras
|
11 |
+
from tensorflow.keras import layers
|
12 |
+
|
13 |
+
from google.colab import drive
|
14 |
+
drive.mount('/content/gdrive')
|
15 |
+
|
16 |
+
|
17 |
+
random.seed(10)
|
18 |
+
|
19 |
+
IMAGE_SIZE = 128
|
20 |
+
BATCH_SIZE = 4
|
21 |
+
MAX_TRAIN_IMAGES = 300
|
22 |
+
|
23 |
+
|
24 |
+
def read_image(image_path):
|
25 |
+
image = tf.io.read_file(image_path)
|
26 |
+
image = tf.image.decode_png(image, channels=3)
|
27 |
+
image.set_shape([None, None, 3])
|
28 |
+
image = tf.cast(image, dtype=tf.float32) / 255.0
|
29 |
+
|
30 |
+
return image
|
31 |
+
|
32 |
+
|
33 |
+
def random_crop(low_image, enhanced_image):
|
34 |
+
low_image_shape = tf.shape(low_image)[:2]
|
35 |
+
low_w = tf.random.uniform(
|
36 |
+
shape=(), maxval=low_image_shape[1] - IMAGE_SIZE + 1, dtype=tf.int32
|
37 |
+
)
|
38 |
+
low_h = tf.random.uniform(
|
39 |
+
shape=(), maxval=low_image_shape[0] - IMAGE_SIZE + 1, dtype=tf.int32
|
40 |
+
)
|
41 |
+
enhanced_w = low_w
|
42 |
+
enhanced_h = low_h
|
43 |
+
low_image_cropped = low_image[
|
44 |
+
low_h : low_h + IMAGE_SIZE, low_w : low_w + IMAGE_SIZE
|
45 |
+
]
|
46 |
+
enhanced_image_cropped = enhanced_image[
|
47 |
+
enhanced_h : enhanced_h + IMAGE_SIZE, enhanced_w : enhanced_w + IMAGE_SIZE
|
48 |
+
]
|
49 |
+
return low_image_cropped, enhanced_image_cropped
|
50 |
+
|
51 |
+
|
52 |
+
def load_data(low_light_image_path, enhanced_image_path):
|
53 |
+
low_light_image = read_image(low_light_image_path)
|
54 |
+
enhanced_image = read_image(enhanced_image_path)
|
55 |
+
low_light_image, enhanced_image = random_crop(low_light_image, enhanced_image)
|
56 |
+
return low_light_image, enhanced_image
|
57 |
+
|
58 |
+
|
59 |
+
def get_dataset(low_light_images, enhanced_images):
|
60 |
+
dataset = tf.data.Dataset.from_tensor_slices((low_light_images, enhanced_images))
|
61 |
+
|
62 |
+
dataset = dataset.map(load_data, num_parallel_calls=tf.data.AUTOTUNE)
|
63 |
+
|
64 |
+
dataset = dataset.batch(BATCH_SIZE, drop_remainder=True)
|
65 |
+
return dataset
|
66 |
+
|
67 |
+
|
68 |
+
train_low_light_images = sorted(glob("/content/gdrive/MyDrive/dataset/lol_dataset/our485/low/*"))[:MAX_TRAIN_IMAGES]
|
69 |
+
train_enhanced_images = sorted(glob("/content/gdrive/MyDrive/dataset/lol_dataset/our485/high/*"))[:MAX_TRAIN_IMAGES]
|
70 |
+
|
71 |
+
val_low_light_images = sorted(glob("/content/gdrive/MyDrive/dataset/lol_dataset/our485/low/*"))[MAX_TRAIN_IMAGES:]
|
72 |
+
val_enhanced_images = sorted(glob("/content/gdrive/MyDrive/dataset/lol_dataset/our485/high/*"))[MAX_TRAIN_IMAGES:]
|
73 |
+
|
74 |
+
test_low_light_images = sorted(glob("/content/gdrive/MyDrive/dataset/lol_dataset/eval15/low/*"))
|
75 |
+
test_enhanced_images = sorted(glob("/content/gdrive/MyDrive/dataset/lol_dataset/eval15/high/*"))
|
76 |
+
|
77 |
+
|
78 |
+
train_dataset = get_dataset(train_low_light_images, train_enhanced_images)
|
79 |
+
val_dataset = get_dataset(val_low_light_images, val_enhanced_images)
|
80 |
+
|
81 |
+
|
82 |
+
print("Train Dataset:", train_dataset)
|
83 |
+
print("Val Dataset:", val_dataset)
|
84 |
+
|
85 |
+
|
86 |
+
def selective_kernel_feature_fusion(
|
87 |
+
multi_scale_feature_1, multi_scale_feature_2, multi_scale_feature_3
|
88 |
+
):
|
89 |
+
channels = list(multi_scale_feature_1.shape)[-1]
|
90 |
+
combined_feature = layers.Add()(
|
91 |
+
[multi_scale_feature_1, multi_scale_feature_2, multi_scale_feature_3]
|
92 |
+
)
|
93 |
+
gap = layers.GlobalAveragePooling2D()(combined_feature)
|
94 |
+
channel_wise_statistics = tf.reshape(gap, shape=(-1, 1, 1, channels))
|
95 |
+
compact_feature_representation = layers.Conv2D(
|
96 |
+
filters=channels // 8, kernel_size=(1, 1), activation="relu"
|
97 |
+
)(channel_wise_statistics)
|
98 |
+
feature_descriptor_1 = layers.Conv2D(
|
99 |
+
channels, kernel_size=(1, 1), activation="softmax"
|
100 |
+
)(compact_feature_representation)
|
101 |
+
feature_descriptor_2 = layers.Conv2D(
|
102 |
+
channels, kernel_size=(1, 1), activation="softmax"
|
103 |
+
)(compact_feature_representation)
|
104 |
+
feature_descriptor_3 = layers.Conv2D(
|
105 |
+
channels, kernel_size=(1, 1), activation="softmax"
|
106 |
+
)(compact_feature_representation)
|
107 |
+
feature_1 = multi_scale_feature_1 * feature_descriptor_1
|
108 |
+
feature_2 = multi_scale_feature_2 * feature_descriptor_2
|
109 |
+
feature_3 = multi_scale_feature_3 * feature_descriptor_3
|
110 |
+
aggregated_feature = layers.Add()([feature_1, feature_2, feature_3])
|
111 |
+
return aggregated_feature
|
112 |
+
|
113 |
+
|
114 |
+
|
115 |
+
|
116 |
+
def spatial_attention_block(input_tensor):
|
117 |
+
average_pooling = tf.reduce_max(input_tensor, axis=-1)
|
118 |
+
average_pooling = tf.expand_dims(average_pooling, axis=-1)
|
119 |
+
max_pooling = tf.reduce_mean(input_tensor, axis=-1)
|
120 |
+
max_pooling = tf.expand_dims(max_pooling, axis=-1)
|
121 |
+
concatenated = layers.Concatenate(axis=-1)([average_pooling, max_pooling])
|
122 |
+
feature_map = layers.Conv2D(1, kernel_size=(1, 1))(concatenated)
|
123 |
+
feature_map = tf.nn.sigmoid(feature_map)
|
124 |
+
return input_tensor * feature_map
|
125 |
+
|
126 |
+
|
127 |
+
def channel_attention_block(input_tensor):
|
128 |
+
channels = list(input_tensor.shape)[-1]
|
129 |
+
average_pooling = layers.GlobalAveragePooling2D()(input_tensor)
|
130 |
+
feature_descriptor = tf.reshape(average_pooling, shape=(-1, 1, 1, channels))
|
131 |
+
feature_activations = layers.Conv2D(
|
132 |
+
filters=channels // 8, kernel_size=(1, 1), activation="relu"
|
133 |
+
)(feature_descriptor)
|
134 |
+
feature_activations = layers.Conv2D(
|
135 |
+
filters=channels, kernel_size=(1, 1), activation="sigmoid"
|
136 |
+
)(feature_activations)
|
137 |
+
return input_tensor * feature_activations
|
138 |
+
|
139 |
+
|
140 |
+
def dual_attention_unit_block(input_tensor):
|
141 |
+
channels = list(input_tensor.shape)[-1]
|
142 |
+
feature_map = layers.Conv2D(
|
143 |
+
channels, kernel_size=(3, 3), padding="same", activation="relu"
|
144 |
+
)(input_tensor)
|
145 |
+
feature_map = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(
|
146 |
+
feature_map
|
147 |
+
)
|
148 |
+
channel_attention = channel_attention_block(feature_map)
|
149 |
+
spatial_attention = spatial_attention_block(feature_map)
|
150 |
+
concatenation = layers.Concatenate(axis=-1)([channel_attention, spatial_attention])
|
151 |
+
concatenation = layers.Conv2D(channels, kernel_size=(1, 1))(concatenation)
|
152 |
+
return layers.Add()([input_tensor, concatenation])
|
153 |
+
|
154 |
+
|
155 |
+
# Recursive Residual Modules
|
156 |
+
|
157 |
+
|
158 |
+
def down_sampling_module(input_tensor):
|
159 |
+
channels = list(input_tensor.shape)[-1]
|
160 |
+
main_branch = layers.Conv2D(channels, kernel_size=(1, 1), activation="relu")(
|
161 |
+
input_tensor
|
162 |
+
)
|
163 |
+
main_branch = layers.Conv2D(
|
164 |
+
channels, kernel_size=(3, 3), padding="same", activation="relu"
|
165 |
+
)(main_branch)
|
166 |
+
main_branch = layers.MaxPooling2D()(main_branch)
|
167 |
+
main_branch = layers.Conv2D(channels * 2, kernel_size=(1, 1))(main_branch)
|
168 |
+
skip_branch = layers.MaxPooling2D()(input_tensor)
|
169 |
+
skip_branch = layers.Conv2D(channels * 2, kernel_size=(1, 1))(skip_branch)
|
170 |
+
return layers.Add()([skip_branch, main_branch])
|
171 |
+
|
172 |
+
|
173 |
+
def up_sampling_module(input_tensor):
|
174 |
+
channels = list(input_tensor.shape)[-1]
|
175 |
+
main_branch = layers.Conv2D(channels, kernel_size=(1, 1), activation="relu")(
|
176 |
+
input_tensor
|
177 |
+
)
|
178 |
+
main_branch = layers.Conv2D(
|
179 |
+
channels, kernel_size=(3, 3), padding="same", activation="relu"
|
180 |
+
)(main_branch)
|
181 |
+
main_branch = layers.UpSampling2D()(main_branch)
|
182 |
+
main_branch = layers.Conv2D(channels // 2, kernel_size=(1, 1))(main_branch)
|
183 |
+
skip_branch = layers.UpSampling2D()(input_tensor)
|
184 |
+
skip_branch = layers.Conv2D(channels // 2, kernel_size=(1, 1))(skip_branch)
|
185 |
+
return layers.Add()([skip_branch, main_branch])
|
186 |
+
|
187 |
+
|
188 |
+
# MRB Block
|
189 |
+
def multi_scale_residual_block(input_tensor, channels):
|
190 |
+
# features
|
191 |
+
level1 = input_tensor
|
192 |
+
level2 = down_sampling_module(input_tensor)
|
193 |
+
level3 = down_sampling_module(level2)
|
194 |
+
# DAU
|
195 |
+
level1_dau = dual_attention_unit_block(level1)
|
196 |
+
level2_dau = dual_attention_unit_block(level2)
|
197 |
+
level3_dau = dual_attention_unit_block(level3)
|
198 |
+
# SKFF
|
199 |
+
level1_skff = selective_kernel_feature_fusion(
|
200 |
+
level1_dau,
|
201 |
+
up_sampling_module(level2_dau),
|
202 |
+
up_sampling_module(up_sampling_module(level3_dau)),
|
203 |
+
)
|
204 |
+
level2_skff = selective_kernel_feature_fusion(
|
205 |
+
down_sampling_module(level1_dau), level2_dau, up_sampling_module(level3_dau)
|
206 |
+
)
|
207 |
+
level3_skff = selective_kernel_feature_fusion(
|
208 |
+
down_sampling_module(down_sampling_module(level1_dau)),
|
209 |
+
down_sampling_module(level2_dau),
|
210 |
+
level3_dau,
|
211 |
+
)
|
212 |
+
# DAU 2
|
213 |
+
level1_dau_2 = dual_attention_unit_block(level1_skff)
|
214 |
+
level2_dau_2 = up_sampling_module((dual_attention_unit_block(level2_skff)))
|
215 |
+
level3_dau_2 = up_sampling_module(
|
216 |
+
up_sampling_module(dual_attention_unit_block(level3_skff))
|
217 |
+
)
|
218 |
+
# SKFF 2
|
219 |
+
skff_ = selective_kernel_feature_fusion(level1_dau_2, level2_dau_2, level3_dau_2)
|
220 |
+
conv = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(skff_)
|
221 |
+
return layers.Add()([input_tensor, conv])
|
222 |
+
|
223 |
+
|
224 |
+
|
225 |
+
|
226 |
+
def recursive_residual_group(input_tensor, num_mrb, channels):
|
227 |
+
conv1 = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(input_tensor)
|
228 |
+
for _ in range(num_mrb):
|
229 |
+
conv1 = multi_scale_residual_block(conv1, channels)
|
230 |
+
conv2 = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(conv1)
|
231 |
+
return layers.Add()([conv2, input_tensor])
|
232 |
+
|
233 |
+
|
234 |
+
def mirnet_model(num_rrg, num_mrb, channels):
|
235 |
+
input_tensor = keras.Input(shape=[None, None, 3])
|
236 |
+
x1 = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(input_tensor)
|
237 |
+
for _ in range(num_rrg):
|
238 |
+
x1 = recursive_residual_group(x1, num_mrb, channels)
|
239 |
+
conv = layers.Conv2D(3, kernel_size=(3, 3), padding="same")(x1)
|
240 |
+
output_tensor = layers.Add()([input_tensor, conv])
|
241 |
+
return keras.Model(input_tensor, output_tensor)
|
242 |
+
|
243 |
+
|
244 |
+
model = mirnet_model(num_rrg=3, num_mrb=2, channels=64)
|
245 |
+
|
246 |
+
|
247 |
+
def charbonnier_loss(y_true, y_pred):
|
248 |
+
return tf.reduce_mean(tf.sqrt(tf.square(y_true - y_pred) + tf.square(1e-3)))
|
249 |
+
|
250 |
+
|
251 |
+
def peak_signal_noise_ratio(y_true, y_pred):
|
252 |
+
return tf.image.psnr(y_pred, y_true, max_val=255.0)
|
253 |
+
|
254 |
+
|
255 |
+
optimizer = keras.optimizers.Adam(learning_rate=1e-4)
|
256 |
+
model.compile(
|
257 |
+
optimizer=optimizer, loss=charbonnier_loss, metrics=[peak_signal_noise_ratio]
|
258 |
+
)
|
259 |
+
|
260 |
+
history = model.fit(
|
261 |
+
train_dataset,
|
262 |
+
validation_data=val_dataset,
|
263 |
+
#epochs traning cycles set krna k lia
|
264 |
+
epochs=1,
|
265 |
+
callbacks=[
|
266 |
+
keras.callbacks.ReduceLROnPlateau(
|
267 |
+
monitor="val_peak_signal_noise_ratio",
|
268 |
+
factor=0.5,
|
269 |
+
patience=5,
|
270 |
+
verbose=1,
|
271 |
+
min_delta=1e-7,
|
272 |
+
mode="max",
|
273 |
+
)
|
274 |
+
],
|
275 |
+
)
|
276 |
+
|
277 |
+
plt.plot(history.history["loss"], label="train_loss")
|
278 |
+
plt.plot(history.history["val_loss"], label="val_loss")
|
279 |
+
plt.xlabel("Epochs")
|
280 |
+
plt.ylabel("Loss")
|
281 |
+
plt.title("Train and Validation Losses Over Epochs", fontsize=14)
|
282 |
+
plt.legend()
|
283 |
+
plt.grid()
|
284 |
+
plt.show()
|
285 |
+
|
286 |
+
|
287 |
+
plt.plot(history.history["peak_signal_noise_ratio"], label="train_psnr")
|
288 |
+
plt.plot(history.history["val_peak_signal_noise_ratio"], label="val_psnr")
|
289 |
+
plt.xlabel("Epochs")
|
290 |
+
plt.ylabel("PSNR")
|
291 |
+
plt.title("Train and Validation PSNR Over Epochs", fontsize=14)
|
292 |
+
plt.legend()
|
293 |
+
plt.grid()
|
294 |
+
plt.show()
|
295 |
+
|
296 |
+
|
297 |
+
|
298 |
+
|
299 |
+
def plot_results(images, titles, figure_size=(12, 12)):
|
300 |
+
fig = plt.figure(figsize=figure_size)
|
301 |
+
for i in range(len(images)):
|
302 |
+
fig.add_subplot(1, len(images), i + 1).set_title(titles[i])
|
303 |
+
_ = plt.imshow(images[i])
|
304 |
+
plt.axis("off")
|
305 |
+
plt.show()
|
306 |
+
|
307 |
+
|
308 |
+
def infer(original_image):
|
309 |
+
image = keras.preprocessing.image.img_to_array(original_image)
|
310 |
+
image = image.astype("float16") / 255.0
|
311 |
+
image = np.expand_dims(image, axis=0)
|
312 |
+
output = model.predict(image)
|
313 |
+
output_image = output[0] * 255.0
|
314 |
+
output_image = output_image.clip(0, 255)
|
315 |
+
output_image = output_image.reshape(
|
316 |
+
(np.shape(output_image)[0], np.shape(output_image)[1], 3)
|
317 |
+
)
|
318 |
+
output_image = Image.fromarray(np.uint8(output_image))
|
319 |
+
original_image = Image.fromarray(np.uint8(original_image))
|
320 |
+
return output_image
|
321 |
+
|
322 |
+
|
323 |
+
|
324 |
+
for low_light_image in random.sample(test_low_light_images, 2):
|
325 |
+
original_image = Image.open(low_light_image)
|
326 |
+
enhanced_image = infer(original_image)
|
327 |
+
plot_results(
|
328 |
+
[original_image, ImageOps.autocontrast(original_image), enhanced_image],
|
329 |
+
["Original", "PIL Autocontrast", "MIRNet Enhanced"],
|
330 |
+
(20, 12),
|
331 |
+
)
|