Trained model into a pb file

[gulingfengze]

I want to turn the trained model into a pb file, but I don't know how to do it. Any good methods or suggestions?

[sercant]

Hey @gulingfengze, check out this answer on stackoverflow.

[dtchuink]

Hi @gulingfengze

Did you succeed on generating the pb file? I am still struggling to do it.

Thanks.

[gulingfengze]

@dtchuink I'm sorry, I tried many times, but failed to generate pb file successfully.

[snehasusan2142]

I was able to generate pb file and perform the inference with it.Output looks very good for me. i used front end as Mobilenetv2 and SUPPORTED_MODEL mobileunet. **Note:**while downloading the codes rename it to .py format to run in Python.

Step1: During training you have to use an extra line after loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=network, labels=net_output)) New line is: new_network = tf.nn.softmax(network, name="softmax_output")#you can provide suitable output node name

Step2: Generate checkpoint file. it should contain 4 files. checkpoint,meta file,index file and a data file

Step3: Generate Pb file using below code:ckpt2pb.py ckpt2pb.txt

step4: You can optimize this graph using graph transformation tool:

bazel-bin/tensorflow/tools/graph_transforms/transform_graph
--in_graph=generated_graph.pb
--out_graph=optimized_graph.pb
--inputs='Placeholder'
--outputs='softmax_output'
--transforms=' add_default_attributes strip_unused_nodes(type=float, shape="1,299,299,3") remove_nodes(op=Identity, op=CheckNumerics) fold_constants(ignore_errors=true) fold_batch_norms fold_old_batch_norms quantize_weights quantize_nodes strip_unused_nodes sort_by_execution_order'

step5: use this optimised pb file for inference using the below code:(slight modification made in the predict.py code):predictwithpb.py predictwithpb.txt

with the help of tensorboard, we can visualize this graph.

[cena001plus]

@snehasusan2142 hello, i found one question, i kan convert ckpt mokel to pb model, but the pb model is small ,just 700k, but the ckpt model is 10M , the ckpt model can give good result , but i load tflite model , the result is wrong, i can't find the reason. It's good if you can give me some sugestion. @GeorgeSeif