Pooya Davoodi

Isn't the evaluation done the same way as training, i.e. feeding all the target tokens to the decoder and comparing the prediction with the target? Whereas inference is done by...

You are right. The seq2seq algorithm with encoder-decoder is a little different from typical classification algorithms. For both training and validation, prediction is done based on a given target (besides...

I am not familiar with this code, but doesn't this use the target sequences (ground truth) as an input to the decoder: https://github.com/google/seq2seq/blob/master/seq2seq/models/basic_seq2seq.py#L82-L83

Something like below shows how to access the graph: ``` graph_def = func.graph.as_graph_def() for node in graph_def.node: if node.op == "TRTEngineOp": # Do something for func in graph_def.library.function: for node...

But that only tells you what ops are in the graph. TRTEngineOp doesn't necessarily run TensorRT, it could sometimes fallback to native TF depending on issues arising during runtime. I...

I have seen an issue related to running int8 calibration in the same process that previously ran fp32/fp16 conversion. But if you run each conversion once per process, I expect...

I just tried TF-TRT FP32 and it worked. I got the following perf on a P100: ``` Step 0: 10.8ms Step 50: 10.8ms Step 100: 10.8ms Step 150: 10.8ms Step...

Looks like the output of the graph (i.e. `graph_func(batch_images)`) is a list instead of a tensor. Check how many items are in the list. My guess is that the output...

This started to happen to me in the last few weeks. I had seen it months ago too and it got fixed, but now it seems to be back.

NMS (non-maximum--suppression) is quite expensive. There is an NMS op called combined_nms that you can use for SSD. TF-TRT can optimize that op quite well. When you build the model...