File size: 8,740 Bytes
eec42bd |
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 |
# dataloader.py
import os
import math
import pandas as pd
import tensorflow as tf
from functools import partial
def validate_data_files(data_df):
"""Validate all files exist before starting training"""
missing_files = []
for _, row in data_df.iterrows():
if not os.path.exists(row['filename']):
missing_files.append(row['filename'])
if not os.path.exists(row['target']):
missing_files.append(row['target'])
if missing_files:
print("Missing files:")
for f in missing_files[:10]:
print(f" {f}")
if len(missing_files) > 10:
print(f" ... and {len(missing_files)-10} more")
raise FileNotFoundError(f"Found {len(missing_files)} missing files")
def _get_pupil_position(pmap, datum, x_shape):
total_mass = tf.reduce_sum(pmap)
if total_mass > 0:
shape = tf.shape(pmap)
h, w = shape[0], shape[1]
ii, jj = tf.meshgrid(tf.range(h), tf.range(w), indexing='ij')
y = tf.reduce_sum(tf.cast(ii, 'float32') * pmap) / total_mass
x = tf.reduce_sum(tf.cast(jj, 'float32') * pmap) / total_mass
return tf.stack((y, x))
if 'roi_x' in datum and 'roi_y' in datum and 'roi_w' in datum:
roi_x = tf.cast(datum['roi_x'], 'float32')
roi_y = tf.cast(datum['roi_y'], 'float32')
half = tf.cast(datum['roi_w'] / 2, 'float32')
result = tf.stack((roi_y + half, roi_x + half))
else: # fallback to center of the image
result = tf.cast(tf.stack((x_shape[0] / 2, x_shape[1] / 2)), dtype='float32')
return result
def additional_augmentations(image, mask, p=0.3):
# Keep original image with probability 1-p
if tf.random.uniform([]) > p:
return image, mask
# Random noise augmentation
if tf.random.uniform([]) < 0.3:
noise = tf.random.normal(tf.shape(image), mean=0, stddev=0.05)
image = image + noise
image = tf.clip_by_value(image, 0, 1)
# Random blur using gaussian filter
if tf.random.uniform([]) < 0.3:
kernel_size = 3
sigma = tf.random.uniform([], 0, 1.0)
x = tf.range(-kernel_size // 2 + 1, kernel_size // 2 + 1, dtype=tf.float32)
gaussian = tf.exp(-(x ** 2) / (2 * sigma ** 2))
gaussian = gaussian / tf.reduce_sum(gaussian)
gaussian = tf.reshape(gaussian, [kernel_size, 1])
gaussian_kernel = gaussian @ tf.transpose(gaussian)
gaussian_kernel = tf.reshape(gaussian_kernel, [kernel_size, kernel_size, 1, 1])
image = tf.nn.conv2d(tf.expand_dims(image, 0), gaussian_kernel,
strides=[1,1,1,1], padding='SAME')[0]
# Random sharpening
if tf.random.uniform([]) < 0.3:
kernel = tf.constant([[-1,-1,-1], [-1,9,-1], [-1,-1,-1]], dtype=tf.float32)
kernel = tf.reshape(kernel, [3, 3, 1, 1])
image = tf.nn.conv2d(tf.expand_dims(image, 0), kernel,
strides=[1,1,1,1], padding='SAME')[0]
image = tf.clip_by_value(image, 0, 1)
return image, mask
@tf.function
def load_datum(datum, x_shape=(128, 128, 1), augment=False):
try:
x = tf.io.read_file(datum['filename'])
y = tf.io.read_file(datum['target'])
# HWC [0,1] float32
x = tf.io.decode_image(x, channels=1, dtype='float32', expand_animations=False)
y = tf.io.decode_image(y, dtype='float32', expand_animations=False)
# Get image dimensions
h = tf.shape(x)[0]
w = tf.shape(x)[1]
# Extract pupil information
pupil_map = y[:, :, 0]
pupil_area = tf.reduce_sum(pupil_map)
pupil_pos_yx = _get_pupil_position(pupil_map, datum, x_shape)
# Target size we want to achieve
target_size = tf.minimum(tf.minimum(h, w), x_shape[0])
if not augment:
# Calculate center crop
h_start = (h - target_size) // 2
w_start = (w - target_size) // 2
else:
# Random crop within safe bounds considering pupil position
h_start = tf.random.uniform([], 0, h - target_size + 1, dtype=tf.int32)
w_start = tf.random.uniform([], 0, w - target_size + 1, dtype=tf.int32)
# Perform crop
x = tf.image.crop_to_bounding_box(x, h_start, w_start, target_size, target_size)
y = tf.image.crop_to_bounding_box(y, h_start, w_start, target_size, target_size)
if augment:
# Rotation with arbitrary angles
k = tf.random.uniform([], 0, 4, dtype=tf.int32)
x = tf.image.rot90(x, k=k)
y = tf.image.rot90(y, k=k)
# Flips with pupil position consideration
if tf.random.uniform([]) < 0.5:
x = tf.image.flip_left_right(x)
y = tf.image.flip_left_right(y)
if tf.random.uniform([]) < 0.5:
x = tf.image.flip_up_down(x)
y = tf.image.flip_up_down(y)
# Apply additional augmentations
x, y = additional_augmentations(x, y)
# Calculate pupil visibility after transformation
new_pupil_map = y[:, :, 0]
new_pupil_area = tf.reduce_sum(new_pupil_map)
eye = (new_pupil_area / pupil_area) if pupil_area > 0 else 0.
# Process eye and blink information
datum_eye = tf.cast(datum['eye'], 'float32')
datum_blink = tf.cast(datum['blink'], 'float32')
# Handle blink cases
if datum_eye == 0:
datum_blink = 0.
if (datum_eye == 1) & (datum_blink == 0):
datum_eye = eye
# Resize if needed
if target_size != x_shape[0]:
x = tf.image.resize(x, [x_shape[0], x_shape[0]])
y = tf.image.resize(y, [x_shape[0], x_shape[0]])
y = y[:, :, :1]
y2 = tf.stack((datum_eye, datum_blink))
return x, y, y2
except Exception as e:
print(f"Error processing datum: {str(e)}")
raise
def get_loader(dataframe, batch_size=8, shuffle=False, **kwargs):
categories = dataframe.exp.values
dataset = tf.data.Dataset.from_tensor_slices(dict(dataframe))
if shuffle:
dataset = dataset.shuffle(1000)
dataset = dataset.map(
partial(load_datum, **kwargs),
num_parallel_calls=tf.data.AUTOTUNE,
deterministic=not shuffle
)
dataset = dataset.batch(batch_size)
def _pack_targets(*ins):
inputs = ins[0]
targets = {'mask': ins[1], 'tags': ins[2]}
return [inputs, targets]
dataset = dataset.map(
_pack_targets,
num_parallel_calls=tf.data.AUTOTUNE,
deterministic=not shuffle
)
dataset = dataset.prefetch(tf.data.AUTOTUNE)
return dataset, categories
def load_datasets(dataset_dirs):
def _load_and_prepare_annotations(dataset_dir):
# Normalize path
dataset_dir = os.path.normpath(dataset_dir)
data_path = os.path.join(dataset_dir, 'annotation', 'annotations.csv')
if not os.path.exists(data_path):
raise FileNotFoundError(f"Annotations file not found: {data_path}")
data = pd.read_csv(data_path)
# Create directories if they don't exist
png_dir = os.path.join(dataset_dir, 'annotation', 'png')
full_frames_dir = os.path.join(dataset_dir, 'fullFrames')
os.makedirs(png_dir, exist_ok=True)
os.makedirs(full_frames_dir, exist_ok=True)
# Filter valid files before creating paths
valid_data = data[data.apply(lambda x: os.path.exists(os.path.join(full_frames_dir, os.path.basename(x['filename']))), axis=1)]
if len(valid_data) == 0:
raise ValueError(f"No valid image files found in {full_frames_dir}")
# Create target paths
valid_data['target'] = valid_data['filename'].apply(
lambda x: os.path.join(png_dir, os.path.splitext(os.path.basename(x))[0] + '.png')
)
valid_data['filename'] = valid_data['filename'].apply(
lambda x: os.path.join(full_frames_dir, os.path.basename(x))
)
return valid_data
all_data = []
for d in dataset_dirs:
try:
dataset = _load_and_prepare_annotations(d)
all_data.append(dataset)
except Exception as e:
print(f"Error loading dataset from {d}: {str(e)}")
continue
if not all_data:
raise ValueError("No valid datasets found in any of the provided directories")
dataset = pd.concat(all_data)
dataset['sub'] = dataset['sub'].astype(str)
print(f"Found {len(dataset)} valid image pairs")
return dataset |