keract: Using an input other than the one provided by the pre-trained model fails.

But that solution should save you for now. The trick is to use the default Input tensor provided by the pre-trained model and push extra information (push num_slices into the first dimension), calling the pre-trained model and un-squeeze the first dimension to recover batch_size and num_slices.

import warnings

import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import Lambda
from tensorflow.python.keras.applications.inception_v3 import InceptionV3

from utils import print_names_and_shapes

tf.compat.v1.disable_eager_execution()
warnings.filterwarnings("ignore")
np.random.seed(40)

from tensorflow.keras.layers import Dense, Dropout, GlobalAveragePooling2D, LSTM, Bidirectional
from tensorflow.keras.models import Model
import tensorflow.keras.backend as K
from tensorflow.keras.optimizers import Adam

import keract

import os

os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = ''

IMAGE_SIZE = 300
N_SLICES = 30


def build_model():
    base_model = InceptionV3(include_top=False)
    x = base_model.output
    x = GlobalAveragePooling2D()(x)  # x.shape = (batch_size*num_slices, 2048).
    # un-squeeze the first dimension to batch_size, and num_slices. 
    # input.shape = (batch_size*num_slices, 2048)
    # output.shape = (batch_size, num_slices, 2048)
    l3 = Lambda(lambda z: K.reshape(z, (-1, N_SLICES, x.shape[-1])))(x)

    # input.shape = (batch_size, time_steps, input_dim=2048).
    bnn = Bidirectional(LSTM(1000))(l3)
    bnn = Dropout(0.5)(bnn)
    bnn = Dense(512, activation='relu')(bnn)
    bnn = Dropout(0.3)(bnn)
    preds = Dense(3, activation='softmax', name='out')(bnn)

    return Model(base_model.input, preds, name='m2')


def compile_model(model, learning_rate=1e-5):
    """

    :param model: builded model
    :return: compiled model
    """

    optimizer = Adam(lr=learning_rate)
    model.compile(loss='binary_crossentropy',
                  optimizer=optimizer)
    return model


model = build_model()
model = compile_model(model, learning_rate=1e-4)

# The first dimension contains actually two dimensions:
# - the slices dimension
# - the batch dimension
batch_size = 2
inputs = np.random.uniform(size=(N_SLICES * batch_size, IMAGE_SIZE, IMAGE_SIZE, 3))
print(model.predict_on_batch(inputs).shape)
activations = keract.get_activations(model, inputs, nested=True)
print_names_and_shapes(activations)

From this reference: https://keras.io/api/applications/

You can do it easily like that:

import numpy as np
from tensorflow.python.keras import Model
from tensorflow.python.keras.applications.resnet import ResNet50
from tensorflow.python.keras.layers import Dense

import keract


def load_and_configure_model(path):
    base_model = ResNet50(include_top=False)
    x = base_model.output
    # 250 new classes.
    x = Dense(250, activation='softmax', name='custom')(x)
    model = Model(inputs=base_model.input, outputs=x)
    if path is not None:
        model.load_weights(path, by_name=True, skip_mismatch=True)
    return model


resnet_50_crop = load_and_configure_model(path=None)

# random image.
img = np.random.uniform(size=(1, 224, 224, 3))

# (1, 7, 7, 2048).
# too many channels to display here.
activations = keract.get_activations(resnet_50_crop, img, layer_names='conv5_block3_out')
activations['conv5_block3_out'] = activations['conv5_block3_out'][..., 0:64]
keract.display_activations(activations, fig_size=(4, 4))

By the way those are the shapes of each layer output for ResNet50:

input_1 (1, 224, 224, 3)
conv1_pad (1, 230, 230, 3)
conv1_conv (1, 112, 112, 64)
conv1_bn (1, 112, 112, 64)
conv1_relu (1, 112, 112, 64)
pool1_pad (1, 114, 114, 64)
pool1_pool (1, 56, 56, 64)
conv2_block1_1_conv (1, 56, 56, 64)
conv2_block1_1_bn (1, 56, 56, 64)
conv2_block1_1_relu (1, 56, 56, 64)
conv2_block1_2_conv (1, 56, 56, 64)
conv2_block1_2_bn (1, 56, 56, 64)
conv2_block1_2_relu (1, 56, 56, 64)
conv2_block1_0_conv (1, 56, 56, 256)
conv2_block1_3_conv (1, 56, 56, 256)
conv2_block1_0_bn (1, 56, 56, 256)
conv2_block1_3_bn (1, 56, 56, 256)
conv2_block1_add (1, 56, 56, 256)
conv2_block1_out (1, 56, 56, 256)
conv2_block2_1_conv (1, 56, 56, 64)
conv2_block2_1_bn (1, 56, 56, 64)
conv2_block2_1_relu (1, 56, 56, 64)
conv2_block2_2_conv (1, 56, 56, 64)
conv2_block2_2_bn (1, 56, 56, 64)
conv2_block2_2_relu (1, 56, 56, 64)
conv2_block2_3_conv (1, 56, 56, 256)
conv2_block2_3_bn (1, 56, 56, 256)
conv2_block2_add (1, 56, 56, 256)
conv2_block2_out (1, 56, 56, 256)
conv2_block3_1_conv (1, 56, 56, 64)
conv2_block3_1_bn (1, 56, 56, 64)
conv2_block3_1_relu (1, 56, 56, 64)
conv2_block3_2_conv (1, 56, 56, 64)
conv2_block3_2_bn (1, 56, 56, 64)
conv2_block3_2_relu (1, 56, 56, 64)
conv2_block3_3_conv (1, 56, 56, 256)
conv2_block3_3_bn (1, 56, 56, 256)
conv2_block3_add (1, 56, 56, 256)
conv2_block3_out (1, 56, 56, 256)
conv3_block1_1_conv (1, 28, 28, 128)
conv3_block1_1_bn (1, 28, 28, 128)
conv3_block1_1_relu (1, 28, 28, 128)
conv3_block1_2_conv (1, 28, 28, 128)
conv3_block1_2_bn (1, 28, 28, 128)
conv3_block1_2_relu (1, 28, 28, 128)
conv3_block1_0_conv (1, 28, 28, 512)
conv3_block1_3_conv (1, 28, 28, 512)
conv3_block1_0_bn (1, 28, 28, 512)
conv3_block1_3_bn (1, 28, 28, 512)
conv3_block1_add (1, 28, 28, 512)
conv3_block1_out (1, 28, 28, 512)
conv3_block2_1_conv (1, 28, 28, 128)
conv3_block2_1_bn (1, 28, 28, 128)
conv3_block2_1_relu (1, 28, 28, 128)
conv3_block2_2_conv (1, 28, 28, 128)
conv3_block2_2_bn (1, 28, 28, 128)
conv3_block2_2_relu (1, 28, 28, 128)
conv3_block2_3_conv (1, 28, 28, 512)
conv3_block2_3_bn (1, 28, 28, 512)
conv3_block2_add (1, 28, 28, 512)
conv3_block2_out (1, 28, 28, 512)
conv3_block3_1_conv (1, 28, 28, 128)
conv3_block3_1_bn (1, 28, 28, 128)
conv3_block3_1_relu (1, 28, 28, 128)
conv3_block3_2_conv (1, 28, 28, 128)
conv3_block3_2_bn (1, 28, 28, 128)
conv3_block3_2_relu (1, 28, 28, 128)
conv3_block3_3_conv (1, 28, 28, 512)
conv3_block3_3_bn (1, 28, 28, 512)
conv3_block3_add (1, 28, 28, 512)
conv3_block3_out (1, 28, 28, 512)
conv3_block4_1_conv (1, 28, 28, 128)
conv3_block4_1_bn (1, 28, 28, 128)
conv3_block4_1_relu (1, 28, 28, 128)
conv3_block4_2_conv (1, 28, 28, 128)
conv3_block4_2_bn (1, 28, 28, 128)
conv3_block4_2_relu (1, 28, 28, 128)
conv3_block4_3_conv (1, 28, 28, 512)
conv3_block4_3_bn (1, 28, 28, 512)
conv3_block4_add (1, 28, 28, 512)
conv3_block4_out (1, 28, 28, 512)
conv4_block1_1_conv (1, 14, 14, 256)
conv4_block1_1_bn (1, 14, 14, 256)
conv4_block1_1_relu (1, 14, 14, 256)
conv4_block1_2_conv (1, 14, 14, 256)
conv4_block1_2_bn (1, 14, 14, 256)
conv4_block1_2_relu (1, 14, 14, 256)
conv4_block1_0_conv (1, 14, 14, 1024)
conv4_block1_3_conv (1, 14, 14, 1024)
conv4_block1_0_bn (1, 14, 14, 1024)
conv4_block1_3_bn (1, 14, 14, 1024)
conv4_block1_add (1, 14, 14, 1024)
conv4_block1_out (1, 14, 14, 1024)
conv4_block2_1_conv (1, 14, 14, 256)
conv4_block2_1_bn (1, 14, 14, 256)
conv4_block2_1_relu (1, 14, 14, 256)
conv4_block2_2_conv (1, 14, 14, 256)
conv4_block2_2_bn (1, 14, 14, 256)
conv4_block2_2_relu (1, 14, 14, 256)
conv4_block2_3_conv (1, 14, 14, 1024)
conv4_block2_3_bn (1, 14, 14, 1024)
conv4_block2_add (1, 14, 14, 1024)
conv4_block2_out (1, 14, 14, 1024)
conv4_block3_1_conv (1, 14, 14, 256)
conv4_block3_1_bn (1, 14, 14, 256)
conv4_block3_1_relu (1, 14, 14, 256)
conv4_block3_2_conv (1, 14, 14, 256)
conv4_block3_2_bn (1, 14, 14, 256)
conv4_block3_2_relu (1, 14, 14, 256)
conv4_block3_3_conv (1, 14, 14, 1024)
conv4_block3_3_bn (1, 14, 14, 1024)
conv4_block3_add (1, 14, 14, 1024)
conv4_block3_out (1, 14, 14, 1024)
conv4_block4_1_conv (1, 14, 14, 256)
conv4_block4_1_bn (1, 14, 14, 256)
conv4_block4_1_relu (1, 14, 14, 256)
conv4_block4_2_conv (1, 14, 14, 256)
conv4_block4_2_bn (1, 14, 14, 256)
conv4_block4_2_relu (1, 14, 14, 256)
conv4_block4_3_conv (1, 14, 14, 1024)
conv4_block4_3_bn (1, 14, 14, 1024)
conv4_block4_add (1, 14, 14, 1024)
conv4_block4_out (1, 14, 14, 1024)
conv4_block5_1_conv (1, 14, 14, 256)
conv4_block5_1_bn (1, 14, 14, 256)
conv4_block5_1_relu (1, 14, 14, 256)
conv4_block5_2_conv (1, 14, 14, 256)
conv4_block5_2_bn (1, 14, 14, 256)
conv4_block5_2_relu (1, 14, 14, 256)
conv4_block5_3_conv (1, 14, 14, 1024)
conv4_block5_3_bn (1, 14, 14, 1024)
conv4_block5_add (1, 14, 14, 1024)
conv4_block5_out (1, 14, 14, 1024)
conv4_block6_1_conv (1, 14, 14, 256)
conv4_block6_1_bn (1, 14, 14, 256)
conv4_block6_1_relu (1, 14, 14, 256)
conv4_block6_2_conv (1, 14, 14, 256)
conv4_block6_2_bn (1, 14, 14, 256)
conv4_block6_2_relu (1, 14, 14, 256)
conv4_block6_3_conv (1, 14, 14, 1024)
conv4_block6_3_bn (1, 14, 14, 1024)
conv4_block6_add (1, 14, 14, 1024)
conv4_block6_out (1, 14, 14, 1024)
conv5_block1_1_conv (1, 7, 7, 512)
conv5_block1_1_bn (1, 7, 7, 512)
conv5_block1_1_relu (1, 7, 7, 512)
conv5_block1_2_conv (1, 7, 7, 512)
conv5_block1_2_bn (1, 7, 7, 512)
conv5_block1_2_relu (1, 7, 7, 512)
conv5_block1_0_conv (1, 7, 7, 2048)
conv5_block1_3_conv (1, 7, 7, 2048)
conv5_block1_0_bn (1, 7, 7, 2048)
conv5_block1_3_bn (1, 7, 7, 2048)
conv5_block1_add (1, 7, 7, 2048)
conv5_block1_out (1, 7, 7, 2048)
conv5_block2_1_conv (1, 7, 7, 512)
conv5_block2_1_bn (1, 7, 7, 512)
conv5_block2_1_relu (1, 7, 7, 512)
conv5_block2_2_conv (1, 7, 7, 512)
conv5_block2_2_bn (1, 7, 7, 512)
conv5_block2_2_relu (1, 7, 7, 512)
conv5_block2_3_conv (1, 7, 7, 2048)
conv5_block2_3_bn (1, 7, 7, 2048)
conv5_block2_add (1, 7, 7, 2048)
conv5_block2_out (1, 7, 7, 2048)
conv5_block3_1_conv (1, 7, 7, 512)
conv5_block3_1_bn (1, 7, 7, 512)
conv5_block3_1_relu (1, 7, 7, 512)
conv5_block3_2_conv (1, 7, 7, 512)
conv5_block3_2_bn (1, 7, 7, 512)
conv5_block3_2_relu (1, 7, 7, 512)
conv5_block3_3_conv (1, 7, 7, 2048)
conv5_block3_3_bn (1, 7, 7, 2048)
conv5_block3_add (1, 7, 7, 2048)
conv5_block3_out (1, 7, 7, 2048)

Nevermind what i said earlier, i realized that i was returning the model before loading the weights.

Fixed by doing:

def load_and_configure_model(model_name, optimizer, loss, metrics, path):
    num_classes = 250
    assert model_name == 'resnet50'
    model = ResNet50(include_top=True, weights='imagenet')
    transfer_layer = model.get_layer('avg_pool')
    x = Dense(num_classes, activation='softmax')
    model = Model(inputs=model.input, outputs=transfer_layer.output)
    model.trainable = True
    for layer in model.layers:
      layer.trainable = False
    for layer in model.layers:
      trainable = ('conv5_block3' in layer.name)      
      layer.trainable = trainable
    transferred_resnet50 = Sequential()
    transferred_resnet50.add(model)
    transferred_resnet50.add(x)
    transferred_resnet50.compile(optimizer=optimizer, loss=loss, metrics=metrics)
    if path is not None:
        transferred_resnet50.load_weights(path)
    return model

When in fact after trying to get the activations from the fine-tuned model leads to the same error:

Python Invalid argument: You must feed a value for placeholder tensor 'input_1' with dtype float and shape [?,224,224,3]

Great! Keras can sometimes be a pain to use.

@FalsoMoralista great! Ideally your case should be properly handled but it’s quite a pain to handle all the cases.

You can also follow some of the examples in: https://github.com/philipperemy/keract/blob/20846aaed0da68d5c328864b56d9a07fd696a76c/tests/get_activations_test.py#L16