Image Prediction

Import Library¶

In [1]:

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

%matplotlib inline

2022-11-23 17:10:00.971077: I tensorflow/core/util/util.cc:169] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.

Load Pre-Trained Model (ResNet50)¶

In [2]:

model = tf.keras.applications.resnet50.ResNet50(include_top=True, weights='imagenet') # still doing classification using pre-trained weight

2022-11-23 17:10:03.415147: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F AVX512_VNNI FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-11-23 17:10:03.920049: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 22103 MB memory:  -> device: 0, name: NVIDIA GeForce RTX 3090, pci bus id: 0000:65:00.0, compute capability: 8.6

In [3]:

len(model.layers)

Out[3]:

177

In [4]:

model.layers

Out[4]:

[<keras.engine.input_layer.InputLayer at 0x7f26bf420820>,
 <keras.layers.reshaping.zero_padding2d.ZeroPadding2D at 0x7f26bf420f40>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bfa9a0a0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26beba86a0>,
 <keras.layers.core.activation.Activation at 0x7f27383b1be0>,
 <keras.layers.reshaping.zero_padding2d.ZeroPadding2D at 0x7f26beba8520>,
 <keras.layers.pooling.max_pooling2d.MaxPooling2D at 0x7f26bc1fd4f0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1fd160>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc2158b0>,
 <keras.layers.core.activation.Activation at 0x7f26bc215880>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc21a850>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc214c70>,
 <keras.layers.core.activation.Activation at 0x7f26bc214c40>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1fdd60>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc20d820>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1fd8e0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc22bb20>,
 <keras.layers.merging.add.Add at 0x7f26bc2079a0>,
 <keras.layers.core.activation.Activation at 0x7f26bc21ab80>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc2339d0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1bae50>,
 <keras.layers.core.activation.Activation at 0x7f26bc1bab80>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc226940>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1bdee0>,
 <keras.layers.core.activation.Activation at 0x7f26bc22b0d0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1d2a60>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1d6e50>,
 <keras.layers.merging.add.Add at 0x7f26bc233df0>,
 <keras.layers.core.activation.Activation at 0x7f26bc1d2550>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1defa0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1e3a30>,
 <keras.layers.core.activation.Activation at 0x7f26bc1c5430>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1eeee0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1f13d0>,
 <keras.layers.core.activation.Activation at 0x7f26bc1d21c0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc17bca0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc17d7f0>,
 <keras.layers.merging.add.Add at 0x7f26bc17b040>,
 <keras.layers.core.activation.Activation at 0x7f26bc1e3dc0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1de850>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1e33a0>,
 <keras.layers.core.activation.Activation at 0x7f26bc21afa0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1fd4c0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc188e20>,
 <keras.layers.core.activation.Activation at 0x7f26bc1fd790>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1d6160>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc18edc0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc21a1c0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26beba8d90>,
 <keras.layers.merging.add.Add at 0x7f26bc194be0>,
 <keras.layers.core.activation.Activation at 0x7f26bc194a30>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1ca670>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc19bf70>,
 <keras.layers.core.activation.Activation at 0x7f26bc191c70>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1a3550>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1aa7f0>,
 <keras.layers.core.activation.Activation at 0x7f26bc1aa6a0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1af400>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1b66a0>,
 <keras.layers.merging.add.Add at 0x7f26bc194ee0>,
 <keras.layers.core.activation.Activation at 0x7f26bc1a36d0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc143b50>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc14a7c0>,
 <keras.layers.core.activation.Activation at 0x7f26bc152e50>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc152c40>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc152910>,
 <keras.layers.core.activation.Activation at 0x7f26bc14a2e0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1431f0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1678b0>,
 <keras.layers.merging.add.Add at 0x7f26bc160e50>,
 <keras.layers.core.activation.Activation at 0x7f26bc16e0a0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0f7310>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0fd190>,
 <keras.layers.core.activation.Activation at 0x7f26bc15cfd0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0fd580>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc104c40>,
 <keras.layers.core.activation.Activation at 0x7f26bc1081c0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0fd9a0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0f7a00>,
 <keras.layers.merging.add.Add at 0x7f26bc1086a0>,
 <keras.layers.core.activation.Activation at 0x7f26bc1432e0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1675e0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1ca4f0>,
 <keras.layers.core.activation.Activation at 0x7f26bc160d90>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc11b160>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc11b190>,
 <keras.layers.core.activation.Activation at 0x7f26bc1203a0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1b63a0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1208e0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1aafd0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc11baf0>,
 <keras.layers.merging.add.Add at 0x7f26bc127580>,
 <keras.layers.core.activation.Activation at 0x7f26bc11bac0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1207c0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc12c9d0>,
 <keras.layers.core.activation.Activation at 0x7f26bc124640>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1350a0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc1350d0>,
 <keras.layers.core.activation.Activation at 0x7f26bc12c2b0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0c3e50>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0c9850>,
 <keras.layers.merging.add.Add at 0x7f26bc0c3fa0>,
 <keras.layers.core.activation.Activation at 0x7f26bc0c9c10>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0d72b0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0d7070>,
 <keras.layers.core.activation.Activation at 0x7f26bc0bc1f0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0df640>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0e4b80>,
 <keras.layers.core.activation.Activation at 0x7f26bc0df6a0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0ea190>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0f2ac0>,
 <keras.layers.merging.add.Add at 0x7f26bc0e48b0>,
 <keras.layers.core.activation.Activation at 0x7f26bc0809d0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc07b9d0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc08ab50>,
 <keras.layers.core.activation.Activation at 0x7f26bc090dc0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc090400>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0f2760>,
 <keras.layers.core.activation.Activation at 0x7f26bc0904f0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0cff10>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0cf8e0>,
 <keras.layers.merging.add.Add at 0x7f26bc12c4f0>,
 <keras.layers.core.activation.Activation at 0x7f26bc120340>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc1ca7c0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc152b80>,
 <keras.layers.core.activation.Activation at 0x7f26bc124790>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc097fa0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc097e80>,
 <keras.layers.core.activation.Activation at 0x7f26bc1fdc40>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc09ef10>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0a27c0>,
 <keras.layers.merging.add.Add at 0x7f26bc09efa0>,
 <keras.layers.core.activation.Activation at 0x7f26bc0a2880>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc09cdf0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0a8160>,
 <keras.layers.core.activation.Activation at 0x7f26bc09c640>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0b0880>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0b6a90>,
 <keras.layers.core.activation.Activation at 0x7f26bc0b07c0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0b6850>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0429a0>,
 <keras.layers.merging.add.Add at 0x7f26bc0b6730>,
 <keras.layers.core.activation.Activation at 0x7f26bc050370>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc057940>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc057430>,
 <keras.layers.core.activation.Activation at 0x7f26bc120760>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc06edf0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc076730>,
 <keras.layers.core.activation.Activation at 0x7f26bc067f40>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc04a1c0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26b07c4790>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc042e50>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc06e3a0>,
 <keras.layers.merging.add.Add at 0x7f26b07d2eb0>,
 <keras.layers.core.activation.Activation at 0x7f26b07d2700>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26b07c4ac0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0670d0>,
 <keras.layers.core.activation.Activation at 0x7f26b07caa00>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc067c40>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26b07d2100>,
 <keras.layers.core.activation.Activation at 0x7f26bc06ec10>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0a8850>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc188fa0>,
 <keras.layers.merging.add.Add at 0x7f26bc0a4c10>,
 <keras.layers.core.activation.Activation at 0x7f26bc0f2af0>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26bc0d76a0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26bc0d7370>,
 <keras.layers.core.activation.Activation at 0x7f26bc0bc250>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26b07de9a0>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26b07d7b20>,
 <keras.layers.core.activation.Activation at 0x7f26b07d7f10>,
 <keras.layers.convolutional.conv2d.Conv2D at 0x7f26b07e8070>,
 <keras.layers.normalization.batch_normalization.BatchNormalization at 0x7f26b07e2790>,
 <keras.layers.merging.add.Add at 0x7f26bc0c9970>,
 <keras.layers.core.activation.Activation at 0x7f26b07d7dc0>,
 <keras.layers.pooling.global_average_pooling2d.GlobalAveragePooling2D at 0x7f26b07c4850>,
 <keras.layers.core.dense.Dense at 0x7f26b07f7dc0>]

In [5]:

model.layers[0].input_shape

Out[5]:

[(None, 224, 224, 3)]

In [6]:

model.layers[-1].output_shape

Out[6]:

(None, 1000)

Load Image¶

In [7]:

image_file = 'cat.jpg'
im = plt.imread(image_file)

In [8]:

plt.imshow(im)
print(im.max())
print(im.min())

252
14

Image Pre-Processing¶

In [9]:

preprocess = tf.keras.applications.imagenet_utils.preprocess_input
im_processed = preprocess(im)

In [10]:

plt.imshow(im_processed)
print(im_processed.max())
print(im_processed.min())

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

146.061
-102.779

Image Prediction¶

In [11]:

im_input = im_processed[np.newaxis, :, :, :]
res = model.predict(im_input)
class_id = np.argmax(res)

2022-11-23 17:10:08.847055: I tensorflow/stream_executor/cuda/cuda_dnn.cc:384] Loaded cuDNN version 8401

1/1 [==============================] - 6s 6s/step

2022-11-23 17:10:12.028093: I tensorflow/stream_executor/cuda/cuda_blas.cc:1786] TensorFloat-32 will be used for the matrix multiplication. This will only be logged once.

In [12]:

print(res.shape)
print(class_id)

(1, 1000)
282

Using Custom Activation Function¶

In [13]:

@tf.custom_gradient
def custom_relu(x):
    def grad(dy):
        return tf.cast(x>0, tf.float32)*tf.cast(dy>0, tf.float32)*dy
    return tf.nn.relu(x), grad

cnt = 0
for layer in model.layers:
    if hasattr(layer, 'activation'):
        if layer.activation == tf.keras.activations.relu:
            layer.activation = custom_relu
            cnt += 1

In [14]:

cnt

Out[14]:

49

Get Gradient¶

In [15]:

im_tensor = tf.Variable(im_input)
with tf.GradientTape() as tape:
    scores = model(im_tensor)
    max_score = tf.reduce_max(scores)
grads = tape.gradient(max_score, im_tensor)

In [16]:

grads = grads.numpy()

In [17]:

grads.shape

Out[17]:

(1, 224, 224, 3)

In [18]:

grads = np.squeeze(grads)

In [19]:

plt.imshow(grads)

Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).

Out[19]:

<matplotlib.image.AxesImage at 0x7f269c3aa490>

In [20]:

plt.hist(grads.flatten())

Out[20]:

(array([5.00000e+00, 2.10000e+01, 4.90000e+01, 2.03000e+02, 1.91200e+03,
        1.47289e+05, 9.13000e+02, 1.09000e+02, 2.20000e+01, 5.00000e+00]),
 array([-0.0025214 , -0.00205292, -0.00158444, -0.00111595, -0.00064747,
        -0.00017899,  0.00028949,  0.00075797,  0.00122645,  0.00169493,
         0.00216341], dtype=float32),
 <BarContainer object of 10 artists>)

In [21]:

g = (grads-grads.min())/(grads.max()-grads.min())

In [22]:

plt.hist(g.flatten())

Out[22]:

(array([5.00000e+00, 2.10000e+01, 4.90000e+01, 2.03000e+02, 1.91200e+03,
        1.47289e+05, 9.13000e+02, 1.09000e+02, 2.20000e+01, 5.00000e+00]),
 array([0. , 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1. ],
       dtype=float32),
 <BarContainer object of 10 artists>)

In [23]:

plt.imshow(g)

Out[23]:

<matplotlib.image.AxesImage at 0x7f269c20f250>

In [ ]: