我正在 Kaggle 上训练模型，这是代码，我在编码数据时遇到一些问题，经过多次尝试后我在这里
https://codefile.io/f/bl9Smlnxtp
# Define transformations
train_datagen = ImageDataGenerator(
rescale=1./255, # Normalize pixel values
rotation_range=20, # Random rotations
width_shift_range=0.2, # Random horizontal shifts
height_shift_range=0.2, # Random vertical shifts
shear_range=0.2, # Random shearing transformations
zoom_range=0.2, # Random zoom
horizontal_flip=True, # Random horizontal flipping
fill_mode='nearest' # How to fill in new pixels after transformations
)
# Define datasets based on gender and view position
datasets = {
'male_pa': df[(df['Patient Gender'] == 'M') & (df['View Position'] == 'PA')],
'male_ap': df[(df['Patient Gender'] == 'M') & (df['View Position'] == 'AP')],
'female_pa': df[(df['Patient Gender'] == 'F') & (df['View Position'] == 'PA')],
'female_ap': df[(df['Patient Gender'] == 'F') & (df['View Position'] == 'AP')],
}
# Function to process labels
def process_labels(label):
labels = label.split('|')
encoded_labels = np.zeros(num_classes, dtype=int) # Create a zero array
for idx, disease in enumerate(all_diseases):
if disease in labels:
encoded_labels[idx] = 1
return encoded_labels
# Get all unique diseases
all_diseases = set('|'.join(df['Finding Labels']).split('|'))
num_classes = len(all_diseases) # Number of classes
# Create and process generators for each dataset
train_datasets = {}
validation_datasets = {}
for name, subset_df in datasets.items():
# Determine all possible image folders based on the data
image_folders = [os.path.join(data_dir, f"images_{i:03d}", "images")
for i in range(1, 13)]
print(image_folders)
def get_image_path(image_index):
for folder in image_folders:
full_path = os.path.join(folder, image_index)
if os.path.exists(full_path):
return full_path
# Debugging print for missing images
return None
# Construct the 'image_path' column, searching for files across folders
subset_df = subset_df.copy() # Make explicit copies to ensure modification
subset_df.loc[:, 'image_path'] = subset_df['Image Index'].apply(get_image_path)
# Filter out missing images (if needed)
subset_df = subset_df[subset_df['image_path'].notnull()]
# Split data into training and test sets
train_df, test_df = train_test_split(subset_df, test_size=0.2, random_state=42)
# Training dataset generator
train_dataset_generator = train_datagen.flow_from_dataframe(
dataframe=train_df,
directory=None,
target_size=(224, 224),
batch_size=32,
class_mode='raw',
x_col="image_path",
y_col="Finding Labels",
y_col_preprocessor=process_labels,
shuffle=False,
interpolation="nearest",
validate_filenames=True,
dtype='int32' # <-- Set the data type for the labels
)
train_datasets[name] = train_dataset_generator
print("hi")
# Test dataset generator
test_datagen = ImageDataGenerator(rescale=1./255) # No augmentation for test
test_dataset_generator = test_datagen.flow_from_dataframe(
dataframe=test_df,
directory=None,
target_size=(224, 224),
batch_size=32,
class_mode='raw',
x_col="image_path",
y_col="Finding Labels",
y_col_preprocessor=process_labels,
shuffle=False,
interpolation="nearest",
validate_filenames=True,
dtype='int32' # <-- Set the data type for the labels
)
validation_datasets[name] = test_dataset_generator
# Print the number of classes
print("Number of classes:", num_classes)
for dataset_name, train_dataset in train_datasets.items():
print(f"Training model for dataset: {dataset_name}")
# Load the ResNet50 model without the top classification layer
base_model = ResNet50(weights='imagenet', include_top=False, input_shape=(224, 224, 3))
# Freeze the base model's layers
for layer in base_model.layers:
layer.trainable = False
# Add new classification layers on top of the base model
x = base_model.output
x = tf.keras.layers.GlobalAveragePooling2D()(x)
x = tf.keras.layers.Dense(1024, activation='relu')(x)
predictions = tf.keras.layers.Dense(num_classes, activation='sigmoid')(x)
# Create the final model
model = tf.keras.Model(inputs=base_model.input, outputs=predictions)
# Compile the model
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['accuracy'])
# Train the model
epochs = 10
model.fit(train_dataset,
epochs=epochs,
verbose=1,
validation_data=train_dataset,
validation_freq=1) # Validate after each epoch
# Save the trained model
model.save(f'chest_xray_model_{dataset_name}.h5')
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[91], line 28
26 # Train the model
27 epochs = 10
---> 28 model.fit(train_dataset,
29 epochs=epochs,
30 verbose=1,
31 validation_data=train_dataset,
32 validation_freq=1) # Validate after each epoch
34 # Save the trained model
35 model.save(f'chest_xray_model_{dataset_name}.h5')
File /opt/conda/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py:123, in filter_traceback.<locals>.error_handler(*args, **kwargs)
120 filtered_tb = _process_traceback_frames(e.__traceback__)
121 # To get the full stack trace, call:
122 # `keras.config.disable_traceback_filtering()`
--> 123 raise e.with_traceback(filtered_tb) from None
124 finally:
125 del filtered_tb
File /opt/conda/lib/python3.10/site-packages/tree/__init__.py:435, in map_structure(func, *structures, **kwargs)
432 for other in structures[1:]:
433 assert_same_structure(structures[0], other, check_types=check_types)
434 return unflatten_as(structures[0],
--> 435 [func(*args) for args in zip(*map(flatten, structures))])
File /opt/conda/lib/python3.10/site-packages/tree/__init__.py:435, in <listcomp>(.0)
432 for other in structures[1:]:
433 assert_same_structure(structures[0], other, check_types=check_types)
434 return unflatten_as(structures[0],
--> 435 [func(*args) for args in zip(*map(flatten, structures))])
ValueError: Invalid dtype: object
我尝试在不同的地方进行编码，从导入 csv 时的正确编码到创建数据生成时的正确编码，再到拟合模型现在仍然可以工作
“ValueError: Invalid dtype: object”表明存在包含对象类型数据的数组（可能是 numpy 数组）。 为了验证这一点，最好尝试为每个名称传递单独的特征和标签数据集。在此过程中，可以首先使用标准 TF 函数将数据集转换为张量（使用 tf.constant 或将数据集创建为 tf.data.Dataset.from_tensor_slices）。如果成功转换，问题就会消失。