This is a simple helper for:
- Visualizing a saved checkpoint in Tensorboar
- Selecting the nodes that will be on the freezed model
- Freezing a Tensorflow Graph to serve on GCP
This module exports the graph to be visualized on Tensorboard
optional arguments:
-h, --help Show this help message and exit
-d, --dir Tensorflow models's directory
-s, --savedir Logdir of tensorboard
Generate a pickled nodes list from the provided node
optional arguments:
-h, --help show this help message and exit
-l [LIST [LIST ...]], --list [LIST [LIST ...]]
List of words in nodes names to be excluded. If
omited, all nodes will be kept
-d DIR, --dir DIR Directory to save the nodes list
-s SAVE_DIR, --save_dir SAVE_DIR
Save dir for the generated list. If 'None' is
provided, no list will be saved
A simple Graph freezer for Tensorflow on GCP
optional arguments:
-h, --help show this help message and exit
-d DIR, --dir DIR Tensorflow domel's directory
-s SAVE_DIR, --save_dir SAVE_DIR
Directory where the frozen graph will be saved
-i [INPUTS [INPUTS ...]], --inputs [INPUTS [INPUTS ...]]
Input tensor list
-o [OUTPUTS [OUTPUTS ...]], --outputs [OUTPUTS [OUTPUTS ...]]
Output tensor list
-n NODES, --nodes NODES
Pickled file wich contains the nodes to be kept
Step 1: Visualize the tensorboard to see wich nodes will be kept and wich are the input and output nodes
Extract the model to tensorboard:
$ python model_to_tensorboard.py -d {model_dir} -s {save_dir}
Run Tensorboard
$ tensorboard --logdir {save_dir}
Create the list passing the node names or node names keywords to remove them from the graph
$ python nodes_extract.py -l {list} -d {model_dir} -s {save_dir}
$ python freeze_graph.py -d {model_dir} -i {input_node} -o {outout_node} -n {node_list} -s {save_dir}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent