-
Notifications
You must be signed in to change notification settings - Fork 0
/
images.py
353 lines (306 loc) · 12.6 KB
/
images.py
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
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.utils.data import Dataset
from torchvision.io import read_image
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from datetime import datetime
import os
from PIL import Image
from sklearn.preprocessing import StandardScaler
def process_image(im, final_size, destination, index):
bg = (127, 127, 127)
w, h = im.size
if w == h:
image = im.resize(final_size)
image.save(os.path.join(destination, index + ".jpg"))
elif w > h:
image = Image.new(im.mode, (w, w), bg)
image.paste(im, (0, (w - h) // 2))
image = image.resize(final_size)
image.save(os.path.join(destination, index + ".jpg"))
elif h > w:
image = Image.new(im.mode, (h, h), bg)
image.paste(im, ((h - w) // 2, 0))
image = image.resize(final_size)
image.save(os.path.join(destination, index + ".jpg"))
def preprocess_images(labels_csv, source, destination, final_size):
print("Starting image preprocessing...")
if not os.path.isdir(destination):
os.mkdir(destination)
df = pd.read_csv(labels_csv)
indexes = df['Id']
i = 0
n = len(indexes)
t0 = datetime.now()
processed = 0
skipped = 0
for index in indexes:
try:
if (i + 1) % (n // 10) == 0:
print(f"Checking image {i + 1}/{n}")
except ZeroDivisionError:
pass
if os.path.isfile(os.path.join(destination, index + ".jpg")):
with Image.open(os.path.join(destination, index + ".jpg")) as dst_im:
if dst_im.size == final_size:
skipped += 1
pass
else:
with Image.open(os.path.join(source, index + ".jpg")) as src_im:
process_image(src_im, final_size, destination, index)
processed += 1
else:
with Image.open(os.path.join(source, index + ".jpg")) as src_im:
process_image(src_im, final_size, destination, index)
processed += 1
i += 1
print(f"Image preprocessing took {datetime.now() - t0}, processed {processed} images, skipped {skipped}")
class PawpularityDataset(Dataset):
def __init__(self, csv, img_dir, tr_test, split=0.9, transformations=None):
self.transformations = transformations
self.img_dir = img_dir
self.df = pd.read_csv(csv)
self.df = self.df.sample(frac=1).reset_index(drop=True)
if tr_test == 'train':
self.df = self.df.truncate(after=np.floor(len(self.df) * split))
else:
self.df = self.df.truncate(before=np.floor(len(self.df) * split))
self.targets = self.df['Pawpularity']
self.targets = self.targets.to_numpy(dtype='float32')
self.indexes = self.df['Id']
self.metadata = self.df.drop(columns=['Pawpularity', 'Id'])
self.metadata = self.metadata.to_numpy(dtype="float32")
self.scaler = StandardScaler()
self.metadata = self.scaler.fit_transform(self.metadata)
def __len__(self):
return len(self.df)
def __getitem__(self, item):
img_path = os.path.join(self.img_dir, self.indexes.iloc[item])
image = read_image(img_path + ".jpg")
if self.transformations:
image = self.transformations(image)
image = image.type(torch.float32)
image = (image - torch.mean(image)) / torch.std(image)
metadata = torch.from_numpy(self.metadata[item])
target = torch.tensor(self.targets[item].reshape(-1))
data = (image, metadata)
return data, target
class PawpularityModel(nn.Module):
def __init__(self, img_size):
super(PawpularityModel, self).__init__()
self.image_cnn = nn.Sequential(
nn.Conv2d(in_channels=3, out_channels=32, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(32),
nn.Conv2d(in_channels=32, out_channels=32, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(32),
nn.Conv2d(in_channels=32, out_channels=32, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(32),
nn.MaxPool2d(2),
nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(64),
nn.Conv2d(in_channels=64, out_channels=64, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(64),
nn.Conv2d(in_channels=64, out_channels=64, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(64),
nn.MaxPool2d(2),
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(128),
nn.Conv2d(in_channels=128, out_channels=128, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(128),
nn.Conv2d(in_channels=128, out_channels=128, kernel_size=(3, 3), padding='same'),
nn.ReLU(),
nn.BatchNorm2d(128),
nn.MaxPool2d(2),
nn.Flatten()
)
self.metadata_ann = nn.Sequential(
nn.Linear(12, 512),
nn.ReLU(),
nn.Dropout(),
nn.Linear(512, 512),
nn.ReLU(),
nn.Dropout()
)
self.dense = nn.Sequential(
nn.Linear(img_size[0] * img_size[1] * 2 + 512, 4096),
nn.ReLU(),
nn.Dropout(),
nn.Linear(4096, 4096),
nn.ReLU(),
nn.Dropout(),
nn.Linear(4096, 4096),
nn.ReLU(),
nn.Dropout(),
nn.Linear(4096, 1024),
nn.ReLU(),
nn.Dropout(),
nn.Linear(1024, 1)
)
def forward(self, X):
image, metadata = X
image, metadata = image.to(device), metadata.to(device)
ann_out = self.metadata_ann(metadata)
cnn_out = self.image_cnn(image)
dense_input = torch.cat((cnn_out, ann_out), dim=1)
out = self.dense(dense_input)
return out
def train(model, device, criterion, optimizer, train_batches, test_batches,
baseline_rmse, train_loader, test_loader, epochs):
train_losses = []
test_losses = []
epochs = epochs
t0 = datetime.now()
print("Starting training...")
for epoch in range(epochs):
print(f"Starting epoch {epoch + 1}.")
model.train()
train_loss = []
batch = 0
for inputs, targets in train_loader:
batch += 1
targets = targets.to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
train_loss.append(loss.item())
if batch % (train_batches // 10) == 0:
print(f"Processed train batch {batch}/{int(np.ceil(train_batches))}")
train_losses.append(np.mean(train_loss))
batch = 0
model.eval()
test_loss = []
for inputs, targets in test_loader:
batch += 1
targets = targets.to(device)
outputs = model(inputs)
loss = criterion(outputs, targets)
test_loss.append(loss.item())
if batch % (test_batches // 10) == 0:
print(f"Processed test batch {batch}/{int(np.ceil(test_batches))}")
test_losses.append(np.mean(test_loss))
dt = datetime.now() - t0
train_epoch_loss = train_losses[-1]
test_epoch_loss = test_losses[-1]
print(f"Epoch: {epoch + 1}/{epochs}\n"
f"Train loss: {train_epoch_loss:.4f} (root {np.sqrt(train_epoch_loss):.4f})\n"
f"Baseline diff: {np.sqrt(train_epoch_loss) - baseline_rmse:.4f}\n"
f"Test loss: {test_epoch_loss:.4f} (root {np.sqrt(test_epoch_loss):.4f})\n"
f"Baseline diff: {np.sqrt(test_epoch_loss) - baseline_rmse:.4f}\n"
f"Total duration: {dt}")
plt.plot(train_losses, label="Train losses")
plt.plot(test_losses, label="Test losses")
plt.show()
def grade(model, device, train_batches, test_batches, baseline_rmse, train_loader, test_loader):
print("Starting grading...")
with torch.no_grad():
train_outputs = []
train_targets = []
model.train()
batch = 0
for inputs, targets in train_loader:
batch += 1
if batch % (train_batches // 10) == 0:
print(f"Grading training batch {batch}/{int(np.ceil(train_batches))}...")
targets = targets.to(device)
outputs = model(inputs).cpu().numpy().flatten().tolist()
train_targets += targets.cpu().numpy().flatten().tolist()
train_outputs += outputs
train_outputs = np.array(train_outputs)
train_targets = np.array(train_targets)
train_diff = train_targets - train_outputs
plt.hist(train_diff, bins=range(-75, 75), label="Train diff", color='blue')
plt.show()
train_rmse = np.sqrt(((train_targets - train_outputs) ** 2).mean())
test_outputs = []
test_targets = []
model.eval()
batch = 0
for inputs, targets in test_loader:
batch += 1
if batch % (test_batches // 10) == 0:
print(f"Grading test batch {batch}/{int(np.ceil(test_batches))}...")
targets = targets.to(device)
outputs = model(inputs).cpu().numpy().flatten().tolist()
test_targets += targets.cpu().numpy().flatten().tolist()
test_outputs += outputs
test_outputs = np.array(test_outputs)
test_targets = np.array(test_targets)
test_diff = test_targets - test_outputs
plt.hist(test_diff, bins=range(-75, 75), label="Test diff", color='orange')
plt.show()
test_rmse = np.sqrt(((test_targets - test_outputs) ** 2).mean())
print(f"Train RMSE: {train_rmse:.4f}, baseline diff: {train_rmse - baseline_rmse:.4f}\n"
f"Test RMSE: {test_rmse:.4f}, baseline diff: {test_rmse - baseline_rmse:.4f}")
baseline_rmse = 20.59095133915306
# Image size, preprocessing
img_size = (128, 128)
preprocess_images('train.csv', 'train', 'train-post', img_size)
# Instantiating datasets
train_dataset = PawpularityDataset(csv='train.csv',
img_dir='train-post',
tr_test='train',
transformations=None)
test_dataset = PawpularityDataset(csv='train.csv',
img_dir='train-post',
tr_test='test')
# Training parameters
batch_sz = 96
train_batches = train_dataset.__len__() / batch_sz
test_batches = test_dataset.__len__() / batch_sz
train_loader = DataLoader(dataset=train_dataset, batch_size=batch_sz, shuffle=True)
test_loader = DataLoader(dataset=test_dataset, batch_size=batch_sz, shuffle=False)
model = PawpularityModel(img_size)
device = torch.device("cuda:0")
model.to(device)
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=3e-6)
# Training run
train(model, device, criterion, optimizer, train_batches, test_batches, baseline_rmse,
train_loader, test_loader, epochs=50)
torch.save(model.state_dict(), 'model.pth')
print("Saved model.")
# Grading
grade(model, device, train_batches, test_batches, baseline_rmse, train_loader, test_loader)
# Create submissions
with torch.no_grad():
preprocess_images('test.csv', 'test', 'test-post', img_size)
submission = pd.DataFrame(columns=['Id', 'Pawpularity'])
df = pd.read_csv('test.csv')
metadata_df = df.drop(columns=['Id'])
metadata = metadata_df.to_numpy(dtype="float32")
scaler = StandardScaler()
metadata = scaler.fit_transform(metadata)
indexes = df['Id']
ids = []
pawpularities = []
i = 0
for index in indexes:
ids.append(index)
image = read_image(os.path.join('test-post', index + ".jpg"))
image = image.type(torch.float32)
image = (image - torch.mean(image)) / torch.std(image)
image = image.reshape(1, 3, img_size[0], img_size[1])
md = metadata[i]
md = md.reshape(1, 12)
md = torch.from_numpy(md)
data = (image, md)
output = model(data).cpu().item()
pawpularities.append(output)
i += 1
submission['Id'] = ids
submission['Pawpularity'] = pawpularities
submission.to_csv('submission.csv', index=False)
print("Saved submission.")