{ "id": 267288, "title": "[Info] Keras : layer_range to plot model", "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/267288", "author_name": "Innat", "post_date": "2021-08-22T15:48:06.865000", "votes": 7, "comment_count": 0, "views": 0, "content": "

Sorry, it's off-topic but we thought it would be useful to others in the modeling approach.

\n
\n

Many times plotting big models is so messy as they're too big to visually inspect. To remedy this, tf.keras.utils.plot_model has now layer_range parameter to sub-plot the internal part only of deep neural network. It would be very convenient to debug the model (shape, layer flow, etc), especially if we like to integrate any custom layers inside a big network.

\n
import tensorflow as tf \nmodel = tf.keras.applications.EfficientNetB1(\n    include_top=True,\n    weights=\"imagenet\",\n    input_tensor=None,\n    input_shape=None,\n    pooling=None,\n    classes=1000,\n    classifier_activation=\"softmax\"\n)\n
\n

Now, let's plot only layer from block5a_expand_conv to block5a_project_bn,

\n
tf.keras.utils.plot_model(\n    model,\n    show_shapes=True,\n    show_dtype=True,\n    show_layer_names=True,\n    expand_nested=True,\n    layer_range=['block5a_expand_conv', 'block5a_project_bn'],\n)\n
\n

\"download\"

\n
\n

layer_range: input of list containing two str items, which is the starting layer name and ending layer name (both inclusive) indicating the range of layers for which the pydot.Dot will be generated. It also accepts regex patterns instead of the exact names. In such a case, start predicate will be the first element it matches to layer_range[0] and the end predicate will be the last element it matches to layer_range[1]. By default None which considers all layers of the model. Note that you must pass range such that the resultant subgraph must be complete.

\n
\n

It's a small feature but I found it very effective to debug the customized model. Thanks to the Ashutosh Hathidara for quick implementation, PR discussion.

", "messages": [ { "id": 1486032, "postDate": "2021-08-22T15:48:06.867Z", "content": "

Sorry, it's off-topic but we thought it would be useful to others in the modeling approach.

\n
\n

Many times plotting big models is so messy as they're too big to visually inspect. To remedy this, tf.keras.utils.plot_model has now layer_range parameter to sub-plot the internal part only of deep neural network. It would be very convenient to debug the model (shape, layer flow, etc), especially if we like to integrate any custom layers inside a big network.

\n
import tensorflow as tf \nmodel = tf.keras.applications.EfficientNetB1(\n    include_top=True,\n    weights=\"imagenet\",\n    input_tensor=None,\n    input_shape=None,\n    pooling=None,\n    classes=1000,\n    classifier_activation=\"softmax\"\n)\n
\n

Now, let's plot only layer from block5a_expand_conv to block5a_project_bn,

\n
tf.keras.utils.plot_model(\n    model,\n    show_shapes=True,\n    show_dtype=True,\n    show_layer_names=True,\n    expand_nested=True,\n    layer_range=['block5a_expand_conv', 'block5a_project_bn'],\n)\n
\n

\"download\"

\n
\n

layer_range: input of list containing two str items, which is the starting layer name and ending layer name (both inclusive) indicating the range of layers for which the pydot.Dot will be generated. It also accepts regex patterns instead of the exact names. In such a case, start predicate will be the first element it matches to layer_range[0] and the end predicate will be the last element it matches to layer_range[1]. By default None which considers all layers of the model. Note that you must pass range such that the resultant subgraph must be complete.

\n
\n

It's a small feature but I found it very effective to debug the customized model. Thanks to the Ashutosh Hathidara for quick implementation, PR discussion.

", "rawMarkdown": "**Sorry, it's off-topic but we thought it would be useful to others in the modeling approach.**\n\n---\n\nMany times plotting big models is so messy as they're too big to visually inspect. To remedy this, **`tf.keras.utils.plot_model`** has now [`layer_range`](https://keras.io/api/utils/model_plotting_utils/) parameter to **sub-plot** the **internal part only** of deep neural network. It would be very convenient to debug the model (shape, layer flow, etc), especially if we like to integrate any **custom layers** inside a big network. \n\n```\nimport tensorflow as tf \nmodel = tf.keras.applications.EfficientNetB1(\n include_top=True,\n weights=\"imagenet\",\n input_tensor=None,\n input_shape=None,\n pooling=None,\n classes=1000,\n classifier_activation=\"softmax\"\n)\n```\n\nNow, let's plot only layer from `block5a_expand_conv` to `block5a_project_bn`,\n\n```\ntf.keras.utils.plot_model(\n model,\n show_shapes=True,\n show_dtype=True,\n show_layer_names=True,\n expand_nested=True,\n layer_range=['block5a_expand_conv', 'block5a_project_bn'],\n)\n```\n\n![download](https://user-images.githubusercontent.com/17668390/130361327-59e498b1-30a1-41c5-a2a8-810ead6fa889.png)\n\n\n> **layer_range**: input of list containing two str items, which is the starting layer name and ending layer name (**both inclusive**) indicating the range of layers for which the pydot.Dot will be generated. It also accepts **regex patterns** instead of the exact names. In such a case, start predicate will be the first element it matches to **layer_range[0]** and the end predicate will be the last element it matches to layer_range[1]. By default None which considers all layers of the model. **Note that you must pass range such that the resultant subgraph must be complete.**\n\nIt's a small feature but I found it very effective to **debug the customized model**. Thanks to the [Ashutosh Hathidara](https://github.com/ashutosh1919) for quick implementation, [PR discussion](https://github.com/tensorflow/tensorflow/pull/49860). ", "votes": 7 } ], "comments": [], "raw_markdown_by_id": { "1486032": "**Sorry, it's off-topic but we thought it would be useful to others in the modeling approach.**\n\n---\n\nMany times plotting big models is so messy as they're too big to visually inspect. To remedy this, **`tf.keras.utils.plot_model`** has now [`layer_range`](https://keras.io/api/utils/model_plotting_utils/) parameter to **sub-plot** the **internal part only** of deep neural network. It would be very convenient to debug the model (shape, layer flow, etc), especially if we like to integrate any **custom layers** inside a big network. \n\n```\nimport tensorflow as tf \nmodel = tf.keras.applications.EfficientNetB1(\n include_top=True,\n weights=\"imagenet\",\n input_tensor=None,\n input_shape=None,\n pooling=None,\n classes=1000,\n classifier_activation=\"softmax\"\n)\n```\n\nNow, let's plot only layer from `block5a_expand_conv` to `block5a_project_bn`,\n\n```\ntf.keras.utils.plot_model(\n model,\n show_shapes=True,\n show_dtype=True,\n show_layer_names=True,\n expand_nested=True,\n layer_range=['block5a_expand_conv', 'block5a_project_bn'],\n)\n```\n\n![download](https://user-images.githubusercontent.com/17668390/130361327-59e498b1-30a1-41c5-a2a8-810ead6fa889.png)\n\n\n> **layer_range**: input of list containing two str items, which is the starting layer name and ending layer name (**both inclusive**) indicating the range of layers for which the pydot.Dot will be generated. It also accepts **regex patterns** instead of the exact names. In such a case, start predicate will be the first element it matches to **layer_range[0]** and the end predicate will be the last element it matches to layer_range[1]. By default None which considers all layers of the model. **Note that you must pass range such that the resultant subgraph must be complete.**\n\nIt's a small feature but I found it very effective to **debug the customized model**. Thanks to the [Ashutosh Hathidara](https://github.com/ashutosh1919) for quick implementation, [PR discussion](https://github.com/tensorflow/tensorflow/pull/49860). " } }