I am trying to train a CNN model with an image dataset for medical image segmentation:
history = model.fit(x = train_dataset,
validation_data = val_dataset,
epochs= epochs,
steps_per_epoch= steps_per_epoch ,
callbacks = [iou_monitor,es,mc]
)
Such that the epochs =1 and batch_size = 4. The train_dataset consists of x, (y1,y2) where x is the input CT image ,and y1 and y2 are the corresponding segmentation masks ( multi-task model). Now, when I check train_dataset from the file system manually I found that the total number of samples are 948. Also, when I run the following code:
# Initialize sample counter
total_samples = 0
# Iterate over dataset to count samples
for batch in train_dataset:
x_batch, (y1_batch, y2_batch) = batch # Unpack batch
batch_size = tf.shape(x_batch)[0].numpy() # Get batch size from X
total_samples += batch_size # Accumulate count
print(f"Total samples in dataset: {total_samples}")
I get:
Total samples in dataset: 948
Which confirms my finding. This means that the steps_per_epoch is num_of_samples_in_train_dataset / batch_size = 948/4 = 237. Now, when I trian the model I get:
236/237 ━━━━━━━━━━━━━━━━━━━━ 0s 388ms/step - loss: 0.6321 - mask_1_output_loss: 0.1476 - mask_2_output_loss: 0.4845
/usr/lib/python3.10/contextlib.py:153: UserWarning: Your input ran out of data; interrupting training. Make sure that your dataset or generator can generate at least `steps_per_epoch * epochs` batches. You may need to use the `.repeat()` function when building your dataset.
self.gen.throw(typ, value, traceback)
My data pipeline is defined as follows:
# Function to read the input image
def read_image(image_path):
try:
image_path = image_path.decode()
x = plt.imread(image_path)
x = x.astype(np.float32)
# Normalize x
x = x / 255.0
if x.shape != (512, 512, 3):
raise ValueError(f"Invalid image dimensions for x: {x.shape}, expected (512, 512, 3)")
#resize the image to 256,256,3
x = cv2.resize(x, None, fx = img_size / original_size, fy = img_size / original_size, interpolation=cv2.INTER_AREA) .astype('float32')
# Add channel dimension as the first dimension
#x = np.transpose(x, (2, 0, 1))
x = tf.convert_to_tensor(x, dtype=tf.float32)
return [x]
except Exception as e:
print(f"Error reading image: {image_path}, Error: {e}")
return None
# Function to read the mask image
def read_mask(mask_path):
try:
# The mask is in the npy format
mask_path= mask_path.decode()
y = np.load(mask_path)
y = y.astype(np.float32)
if y.shape != (512, 512, 2):
raise ValueError(f"Invalid original image dimensions: {y.shape}, expected (512, 512, 2)")
# Check if the mask is binary
# resize the image to 256,256,2
y = cv2.resize(y, None, fx = img_size / original_size, fy = img_size / original_size, interpolation=cv2.INTER_NEAREST).astype('float32')
#print("Shape after resize : ",y.shape)
# Add channel dimension as the first dimension
y = np.transpose(y, (2, 0, 1))
y1,y2 = y[0,:,:] , y[1,:,:]
y1 = np.expand_dims(y1, axis=-1) # Converts (img_size, img_size) to ( img_size, img_size,1)
y2 = np.expand_dims(y2, axis=-1)
y1 = tf.convert_to_tensor(y1, dtype=tf.float32)
y2 = tf.convert_to_tensor(y2, dtype=tf.float32)
return [y1,y2]
except Exception as e:
print(f"Error reading mask: {mask_path}, Error: {e}")
return None
# Data Augmentation
def augment_image(image,mask1,mask2):
# Flip and rotate the image and mask 50% of the time
if tf.random.uniform(()) > 0.5:
image = tf.image.flip_left_right(image)
mask1 = tf.image.flip_left_right(mask1)
mask2 = tf.image.flip_left_right(mask2)
#image_channels_last_shape_2 = image_channels_last.shape
#mask_channels_last_shape_2 = mask_channels_last.shape
if tf.random.uniform(()) > 0.5:
image = tf.image.flip_up_down(image)
mask1 = tf.image.flip_up_down(mask1)
mask2 = tf.image.flip_up_down(mask2)
#image_channels_last_shape_3 = image_channels_last.shape
#mask_channels_last_shape_3 = mask_channels_last.shape
if tf.random.uniform(()) > 0.5:
# Randomly select how many 90-degree rotations to apply (0, 1, 2, or 3)
rotation_count = tf.random.uniform((), minval=0, maxval=4, dtype=tf.int32)
# Apply the same rotation to the image and masks
image = tf.image.rot90(image, k=rotation_count)
mask1 = tf.image.rot90(mask1, k=rotation_count)
mask2 = tf.image.rot90(mask2, k=rotation_count)
return image, mask1, mask2
# Preprocessing function
def preprocess(x, y ,data_augmentation = False):
#x = x.decode()
#y = y.decode()
output = tf.numpy_function(read_image, [x], [ tf.float32])
x= output [0]
output = tf.numpy_function(read_mask, [y], [ tf.float32, tf.float32])
y1, y2 = output
# convert the input image to grayscale
#x=tf.image.rgb_to_grayscale(x)
# Standardize each channel of the input image
#x= standardize_image(x)
# Apply data augmentation
if data_augmentation:
x,y1,y2 = augment_image(x,y1,y2)
# Convert to channels_first format
x = tf.transpose(x, perm=[2, 0, 1]) # [H, W, C] -> [C, H, W]
y1 = tf.transpose(y1, perm=[2, 0, 1]) # [H, W, C] -> [C, H, W]
y2 = tf.transpose(y2, perm=[2, 0, 1]) # [H, W, C] -> [C, H, W]
x.set_shape((3,img_size, img_size))
y1.set_shape((1,img_size, img_size))
y2.set_shape((1,img_size, img_size))
return x, (y1,y2)
# Create the TensorFlow dataset
def tf_dataset(x, y, batch_size=32, epochs=30, data_augmentation = False):
dataset = tf.data.Dataset.from_tensor_slices((x, y))
dataset = dataset.shuffle(buffer_size=100)
dataset = dataset.map(lambda a, b: preprocess(a, b, data_augmentation), num_parallel_calls=tf.data.experimental.AUTOTUNE)
dataset = dataset.repeat(epochs).batch(batch_size)
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
return dataset
Also, I tried to turn off the data augmentation, but still the I get the same warning. I don't know what to do now.
I am trying to train a CNN model with an image dataset for medical image segmentation:
history = model.fit(x = train_dataset,
validation_data = val_dataset,
epochs= epochs,
steps_per_epoch= steps_per_epoch ,
callbacks = [iou_monitor,es,mc]
)
Such that the epochs =1 and batch_size = 4. The train_dataset consists of x, (y1,y2) where x is the input CT image ,and y1 and y2 are the corresponding segmentation masks ( multi-task model). Now, when I check train_dataset from the file system manually I found that the total number of samples are 948. Also, when I run the following code:
# Initialize sample counter
total_samples = 0
# Iterate over dataset to count samples
for batch in train_dataset:
x_batch, (y1_batch, y2_batch) = batch # Unpack batch
batch_size = tf.shape(x_batch)[0].numpy() # Get batch size from X
total_samples += batch_size # Accumulate count
print(f"Total samples in dataset: {total_samples}")
I get:
Total samples in dataset: 948
Which confirms my finding. This means that the steps_per_epoch is num_of_samples_in_train_dataset / batch_size = 948/4 = 237. Now, when I trian the model I get:
236/237 ━━━━━━━━━━━━━━━━━━━━ 0s 388ms/step - loss: 0.6321 - mask_1_output_loss: 0.1476 - mask_2_output_loss: 0.4845
/usr/lib/python3.10/contextlib.py:153: UserWarning: Your input ran out of data; interrupting training. Make sure that your dataset or generator can generate at least `steps_per_epoch * epochs` batches. You may need to use the `.repeat()` function when building your dataset.
self.gen.throw(typ, value, traceback)
My data pipeline is defined as follows:
# Function to read the input image
def read_image(image_path):
try:
image_path = image_path.decode()
x = plt.imread(image_path)
x = x.astype(np.float32)
# Normalize x
x = x / 255.0
if x.shape != (512, 512, 3):
raise ValueError(f"Invalid image dimensions for x: {x.shape}, expected (512, 512, 3)")
#resize the image to 256,256,3
x = cv2.resize(x, None, fx = img_size / original_size, fy = img_size / original_size, interpolation=cv2.INTER_AREA) .astype('float32')
# Add channel dimension as the first dimension
#x = np.transpose(x, (2, 0, 1))
x = tf.convert_to_tensor(x, dtype=tf.float32)
return [x]
except Exception as e:
print(f"Error reading image: {image_path}, Error: {e}")
return None
# Function to read the mask image
def read_mask(mask_path):
try:
# The mask is in the npy format
mask_path= mask_path.decode()
y = np.load(mask_path)
y = y.astype(np.float32)
if y.shape != (512, 512, 2):
raise ValueError(f"Invalid original image dimensions: {y.shape}, expected (512, 512, 2)")
# Check if the mask is binary
# resize the image to 256,256,2
y = cv2.resize(y, None, fx = img_size / original_size, fy = img_size / original_size, interpolation=cv2.INTER_NEAREST).astype('float32')
#print("Shape after resize : ",y.shape)
# Add channel dimension as the first dimension
y = np.transpose(y, (2, 0, 1))
y1,y2 = y[0,:,:] , y[1,:,:]
y1 = np.expand_dims(y1, axis=-1) # Converts (img_size, img_size) to ( img_size, img_size,1)
y2 = np.expand_dims(y2, axis=-1)
y1 = tf.convert_to_tensor(y1, dtype=tf.float32)
y2 = tf.convert_to_tensor(y2, dtype=tf.float32)
return [y1,y2]
except Exception as e:
print(f"Error reading mask: {mask_path}, Error: {e}")
return None
# Data Augmentation
def augment_image(image,mask1,mask2):
# Flip and rotate the image and mask 50% of the time
if tf.random.uniform(()) > 0.5:
image = tf.image.flip_left_right(image)
mask1 = tf.image.flip_left_right(mask1)
mask2 = tf.image.flip_left_right(mask2)
#image_channels_last_shape_2 = image_channels_last.shape
#mask_channels_last_shape_2 = mask_channels_last.shape
if tf.random.uniform(()) > 0.5:
image = tf.image.flip_up_down(image)
mask1 = tf.image.flip_up_down(mask1)
mask2 = tf.image.flip_up_down(mask2)
#image_channels_last_shape_3 = image_channels_last.shape
#mask_channels_last_shape_3 = mask_channels_last.shape
if tf.random.uniform(()) > 0.5:
# Randomly select how many 90-degree rotations to apply (0, 1, 2, or 3)
rotation_count = tf.random.uniform((), minval=0, maxval=4, dtype=tf.int32)
# Apply the same rotation to the image and masks
image = tf.image.rot90(image, k=rotation_count)
mask1 = tf.image.rot90(mask1, k=rotation_count)
mask2 = tf.image.rot90(mask2, k=rotation_count)
return image, mask1, mask2
# Preprocessing function
def preprocess(x, y ,data_augmentation = False):
#x = x.decode()
#y = y.decode()
output = tf.numpy_function(read_image, [x], [ tf.float32])
x= output [0]
output = tf.numpy_function(read_mask, [y], [ tf.float32, tf.float32])
y1, y2 = output
# convert the input image to grayscale
#x=tf.image.rgb_to_grayscale(x)
# Standardize each channel of the input image
#x= standardize_image(x)
# Apply data augmentation
if data_augmentation:
x,y1,y2 = augment_image(x,y1,y2)
# Convert to channels_first format
x = tf.transpose(x, perm=[2, 0, 1]) # [H, W, C] -> [C, H, W]
y1 = tf.transpose(y1, perm=[2, 0, 1]) # [H, W, C] -> [C, H, W]
y2 = tf.transpose(y2, perm=[2, 0, 1]) # [H, W, C] -> [C, H, W]
x.set_shape((3,img_size, img_size))
y1.set_shape((1,img_size, img_size))
y2.set_shape((1,img_size, img_size))
return x, (y1,y2)
# Create the TensorFlow dataset
def tf_dataset(x, y, batch_size=32, epochs=30, data_augmentation = False):
dataset = tf.data.Dataset.from_tensor_slices((x, y))
dataset = dataset.shuffle(buffer_size=100)
dataset = dataset.map(lambda a, b: preprocess(a, b, data_augmentation), num_parallel_calls=tf.data.experimental.AUTOTUNE)
dataset = dataset.repeat(epochs).batch(batch_size)
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
return dataset
Also, I tried to turn off the data augmentation, but still the I get the same warning. I don't know what to do now.
Share Improve this question edited Mar 5 at 15:17 Christoph Rackwitz 15.9k5 gold badges39 silver badges51 bronze badges asked Jan 29 at 13:15 AAA_11AAA_11 736 bronze badges1 Answer
Reset to default 0I think your dataset is running out cause you're batching before repeating. Try replacing dataset = dataset.repeat(epochs).batch(batch_size), it should not be the other way around. Try this, and do you have a Git repo for this