Introduction
In TensorFlow 2.x, model.summary() works directly on Keras models (Sequential, Functional, or subclassed). For TensorFlow 1.x models or raw tf.function graphs, you can use tf.debugging.set_log_device_placement, tf.profiler, or manually iterate over tf.trainable_variables() to display a model summary. The torchsummary-style output can be achieved with the tf.keras.utils.plot_model() function for visual summaries.
Keras model.summary() (Standard Approach)
1import tensorflow as tf
2
3# Sequential model
4model = tf.keras.Sequential([
5 tf.keras.layers.Dense(128, activation='relu', input_shape=(784,)),
6 tf.keras.layers.Dropout(0.2),
7 tf.keras.layers.Dense(64, activation='relu'),
8 tf.keras.layers.Dense(10, activation='softmax'),
9])
10
11model.summary()
Output:
1Model: "sequential"
2_________________________________________________________________
3 Layer (type) Output Shape Param #
4=================================================================
5 dense (Dense) (None, 128) 100480
6 dropout (Dropout) (None, 128) 0
7 dense_1 (Dense) (None, 64) 8256
8 dense_2 (Dense) (None, 10) 650
9=================================================================
10Total params: 109386 (427.29 KB)
11Trainable params: 109386 (427.29 KB)
12Non-trainable params: 0 (0.00 Byte)
Functional API Summary
1import tensorflow as tf
2
3inputs = tf.keras.Input(shape=(224, 224, 3))
4x = tf.keras.layers.Conv2D(32, 3, activation='relu')(inputs)
5x = tf.keras.layers.MaxPooling2D()(x)
6x = tf.keras.layers.Conv2D(64, 3, activation='relu')(x)
7x = tf.keras.layers.GlobalAveragePooling2D()(x)
8x = tf.keras.layers.Dense(128, activation='relu')(x)
9outputs = tf.keras.layers.Dense(10, activation='softmax')(x)
10
11model = tf.keras.Model(inputs=inputs, outputs=outputs, name='my_cnn')
12model.summary()
13
14# Show with more detail (including layer connectivity)
15model.summary(show_trainable=True, expand_nested=True)
Subclassed Model Summary
Subclassed models require calling the model once with sample input before summary() works:
1import tensorflow as tf
2
3class MyModel(tf.keras.Model):
4 def __init__(self):
5 super().__init__()
6 self.dense1 = tf.keras.layers.Dense(128, activation='relu')
7 self.dense2 = tf.keras.layers.Dense(64, activation='relu')
8 self.output_layer = tf.keras.layers.Dense(10, activation='softmax')
9
10 def call(self, inputs):
11 x = self.dense1(inputs)
12 x = self.dense2(x)
13 return self.output_layer(x)
14
15model = MyModel()
16
17# This fails before the model is built:
18# model.summary() # ValueError: This model has not yet been built
19
20# Option 1: Call with sample input
21model(tf.zeros((1, 784)))
22model.summary()
23
24# Option 2: Explicitly build
25model.build(input_shape=(None, 784))
26model.summary()
Visual Model Plot
1import tensorflow as tf
2
3model = tf.keras.Sequential([
4 tf.keras.layers.Conv2D(32, 3, activation='relu', input_shape=(28, 28, 1)),
5 tf.keras.layers.MaxPooling2D(),
6 tf.keras.layers.Flatten(),
7 tf.keras.layers.Dense(128, activation='relu'),
8 tf.keras.layers.Dense(10, activation='softmax'),
9])
10
11# Save model architecture as an image
12tf.keras.utils.plot_model(
13 model,
14 to_file='model_architecture.png',
15 show_shapes=True, # Show input/output shapes
16 show_layer_names=True, # Show layer names
17 show_layer_activations=True, # Show activation functions
18 show_dtype=False,
19 rankdir='TB', # Top to bottom (or 'LR' for left to right)
20 dpi=150,
21)
Requires graphviz and pydot:
1pip install pydot graphviz
2# Also install system graphviz:
3# Ubuntu: sudo apt install graphviz
4# macOS: brew install graphviz
Custom Summary Function
1import tensorflow as tf
2
3def detailed_summary(model):
4 """Print a detailed summary including output shapes and parameter counts."""
5 print(f"Model: {model.name}")
6 print(f"{'Layer':<30} {'Type':<25} {'Output Shape':<20} {'Params':<10} {'Trainable'}")
7 print("=" * 110)
8
9 total_params = 0
10 trainable_params = 0
11
12 for layer in model.layers:
13 output_shape = str(layer.output_shape) if hasattr(layer, 'output_shape') else 'N/A'
14 params = layer.count_params()
15 trainable = "Yes" if layer.trainable and params > 0 else "No"
16 total_params += params
17 if layer.trainable:
18 trainable_params += params
19
20 print(f"{layer.name:<30} {layer.__class__.__name__:<25} {output_shape:<20} {params:<10} {trainable}")
21
22 print("=" * 110)
23 print(f"Total params: {total_params:,}")
24 print(f"Trainable params: {trainable_params:,}")
25 print(f"Non-trainable params: {total_params - trainable_params:,}")
26
27detailed_summary(model)
Printing Summary to a String or File
1import tensorflow as tf
2import io
3
4model = tf.keras.Sequential([
5 tf.keras.layers.Dense(64, input_shape=(100,)),
6 tf.keras.layers.Dense(10),
7])
8
9# Capture summary as a string
10string_buffer = io.StringIO()
11model.summary(print_fn=lambda x: string_buffer.write(x + '\n'))
12summary_string = string_buffer.getvalue()
13print(summary_string)
14
15# Write summary to a file
16with open('model_summary.txt', 'w') as f:
17 model.summary(print_fn=lambda x: f.write(x + '\n'))
18
19# Log summary to TensorBoard
20import datetime
21log_dir = "logs/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
22writer = tf.summary.create_file_writer(log_dir)
23with writer.as_default():
24 tf.summary.text("model_summary", summary_string, step=0)
TensorFlow 1.x: Manual Variable Summary
1# TF1 — no model.summary(), iterate variables manually
2import tensorflow.compat.v1 as tf
3tf.disable_v2_behavior()
4
5# After building the graph
6total_params = 0
7for var in tf.trainable_variables():
8 shape = var.get_shape()
9 params = 1
10 for dim in shape:
11 params *= dim
12 total_params += params
13 print(f"{var.name:<50} {str(shape):<20} {params}")
14
15print(f"\nTotal trainable parameters: {total_params:,}")
Common Pitfalls
Calling summary() on an unbuilt subclassed model: Subclassed tf.keras.Model does not know its input shape until it processes data. Call model.build(input_shape=...) or model(sample_input) before calling summary(), otherwise it raises ValueError.
Expecting summary() to show nested model details: By default, summary() shows nested models (like a pretrained backbone) as a single line. Use model.summary(expand_nested=True) to recursively expand nested models and show all internal layers.
Missing graphviz for plot_model: tf.keras.utils.plot_model() requires both the Python pydot package and the system graphviz binary. Installing only the Python package without the system binary causes FileNotFoundError: dot not found.
Incorrect parameter count for shared layers: Layers used multiple times in a Functional model share parameters. summary() shows the correct total, but counting parameters by iterating model.layers can double-count shared layers. Use model.count_params() for the accurate total.
Summary showing "multiple" for output shapes: When a layer has multiple outputs or is used in multiple places, the output shape column shows "multiple" instead of a concrete shape. This is expected — use model.summary(expand_nested=True) or inspect individual layers with layer.output_shape for details.
Summary
Use model.summary() on any Keras model (Sequential, Functional, or subclassed) to display layer names, output shapes, and parameter counts
For subclassed models, call model.build(input_shape=...) or pass sample data before calling summary()
Use tf.keras.utils.plot_model() to generate a visual diagram of the model architecture
Capture summary output as a string with print_fn=lambda x: buffer.write(x)
Use expand_nested=True to show internal layers of nested models
For TF1 graphs, manually iterate over tf.trainable_variables() to display variable shapes and parameter counts
