ValueError: слой custom_res_net_6 никогда не вызывался и, следовательно, не имеет определенного вывода при реализации gr ⇐ Python

[Anonymous]

У меня следующая модель

Код: Выделить всё

@register_keras_serializable()
class CustomResNet(Model):
def __init__(self, num_classes, trained = False, **kwargs):
super(CustomResNet, self).__init__(**kwargs)
self.num_classes = num_classes
print(self.num_classes)
if trained:
self.resnet = tf.keras.applications.ResNet50(
include_top=False, #sem classificado
weights='imagenet',
input_shape=(100, 75, 3),
)
self.resnet.trainable = False
else:
self.resnet = tf.keras.applications.ResNet50(
include_top=False, #sem classificado
weights=None,
input_shape=(100, 75, 3),
)
self.resnet.trainable = True
self.flatten = Flatten()
self.drop1 = tf.keras.layers.Dropout(0.3)
self.bn1 = BatchNormalization()
self.dense1 = Dense(256, activation='relu', kernel_regularizer = tf.keras.regularizers.l2(0.01), name = "dense1")
self.drop2 = tf.keras.layers.Dropout(0.3)
self.bn2 = BatchNormalization()
self.dense2 = Dense(128, activation='relu', kernel_regularizer = tf.keras.regularizers.l2(0.01), name = "dense2")
self.dense_out = Dense(self.num_classes, activation='softmax', name = "dense_out")

def call(self, inputs, training=None):
x = self.resnet(inputs)
print(f"Shape after ResNet: {x.shape}")
x = self.flatten(x)
print(f"Shape after Flatten: {x.shape}")
x = self.drop1(x)
x = self.bn1(x)
x = self.dense1(x)
print(f"Shape after dense1: {x.shape}")
x = self.drop2(x)
x = self.bn2(x)
x = self.dense2(x)
print(f"Shape after dense 2: {x.shape}")
x = self.dense_out(x)
print(f"Shape after dense_out: {x.shape}")
return x

def get_config(self):
# Include `num_classes` in the config
config = super(CustomResNet, self).get_config()  # Include base class config
config['num_classes'] = self.num_classes
return config

@classmethod
def from_config(cls, config):
# Evita passar o mesmo argumento mais de uma vez
return cls(**config)

def build(self, input_shape):
self.resnet.build(input_shape)
super(CustomResNet, self).build(input_shape)
self.built = True

И следующая реализация grad-cam

Код: Выделить всё

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.keras.models import Model
from tensorflow.keras.preprocessing import image

def grad_cam(model, img_array, last_conv_layer_name, pred_index=None):
"""
Generate a Grad-CAM heatmap for a given image.

Args:
- model: The trained model
- img_array: The image to be processed (numpy array)
- last_conv_layer_name: Name of the last convolutional layer for Grad-CAM
- pred_index:  Index of the predicted class (optional)

Returns:
- A heatmap that highlights important areas of the image
"""
# Ensure the model is in inference mode
model = model if isinstance(model, tf.keras.Model) else model.model

# Convert img_array to tensor if needed and add batch dimension
img_array = np.expand_dims(img_array, axis=0)

# Retrieve the last convolutional layer
last_conv_layer = model.get_layer('resnet50').get_layer(last_conv_layer_name)

# Create a model that outputs the last convolutional layer and the final predictions
grad_model = Model(
inputs=[model.get_layer('resnet50').input],
outputs=[last_conv_layer.output, model.get_layer('dense_out').output]
)

with tf.GradientTape() as tape:
# Record the forward pass
conv_outputs, predictions = grad_model(img_array)

# If pred_index is None, use the predicted class index
if pred_index is None:
pred_index = tf.argmax(predictions[0])

# Compute the gradient of the predicted class (class index) with respect to the last convolutional layer
loss = predictions[:, pred_index]

# Compute the gradient of the loss with respect to the convolutional layer outputs
grads = tape.gradient(loss, conv_outputs)

# Average the gradients across all the axes of the feature maps to get the importance of each channel
pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2))

# Multiply each channel in the feature map by the corresponding pooled gradient
conv_outputs = conv_outputs[0]
for i in range(conv_outputs.shape[-1]):
conv_outputs[:, :, i] *= pooled_grads[i]

# Average the feature maps along the channels axis to get a 2D heatmap
heatmap = np.mean(conv_outputs, axis=-1)

# Normalize the heatmap to a range [0, 1]
heatmap = np.maximum(heatmap, 0)
heatmap /= np.max(heatmap)

# Return the heatmap
return heatmap

def display_grad_cam(img_array, heatmap, alpha=0.4):
"""
Display the Grad-CAM heatmap overlaid on the original image.

Args:
- img_array: The original image (numpy array)
- heatmap: The Grad-CAM heatmap
- alpha: Transparency level for overlaying the heatmap

Returns:
- None, but displays the image with Grad-CAM overlay
"""
# Resize heatmap to match the size of the original image
heatmap = np.uint8(255 * heatmap)
heatmap = tf.image.resize(heatmap, (img_array.shape[0], img_array.shape[1]))
heatmap = np.array(heatmap)

# Load the image into a format that can be plotted
img = image.array_to_img(img_array[0])

# Display the image with the heatmap overlay
plt.imshow(img)
plt.imshow(heatmap, cmap='jet', alpha=alpha)  # Jet colormap for better visibility
plt.axis('off')
plt.show()

# Usage example:
# Assuming `model` is your trained `CustomResNet` model and `x_test[0]` is an image from your dataset.
# Choose the last convolutional layer name from the ResNet50 model
last_conv_layer_name = "conv5_block3_out"  # Change this if necessary based on your model

# Run Grad-CAM
heatmap = grad_cam(model=resnet50_geral_ft, img_array=x_test[0:1], last_conv_layer_name=last_conv_layer_name)

# Display the result
display_grad_cam(x_test[0:1], heatmap)

При запуске я получил ошибку.

ValueError: слой Density_out никогда не вызывался и, следовательно, нет определенного вывода.

Даже после использования фиктивного ввода и вызова сборки. Сводная информация моей модели запускается, но в столбце вывода отображается только ?
Инициализация моей модели следующая:

Код: Выделить всё

resnet50_geral_ft = CustomResNet(num_classes=14)
resnet50_geral_ft.build(input_shape=(None, 100, 75, 3))
resnet50_geral_ft.compile(optimizer = optimizer,loss='categorical_crossentropy', metrics=default_metrics)
resnet50_geral_ft = load_model( checkpoint_filepath  + "resnet50_geral_ft.keras", custom_objects={'CustomResNet': CustomResNet}  )

Я не могу понять, чего мне здесь не хватает.... кто-нибудь знает?

Подробнее здесь: https://stackoverflow.com/questions/792 ... has-no-def