How can I get the log during training the Yolov5 model?

[gganduu]

Pytorch Yolov5 training loss will return two parameters, one is for the log, but there is only one return in Analytics Zoo loss, how can I get the training log then?

[yangw1234]

Hi @gganduu I do not understand your question.

Could you provide more details like code snippets?

[gganduu]

Yes, we sent the code by email already.(To wang yang and qiu xin)

[yangw1234]

Synced offline, here are the brief summary:

Feature Requests:

Automatically printing loss and other metrics of every epoch or iteration when calling Estimator.fit

Current workaround we provides:

fitting multiple epochs in a loop. E.g.

change

est = Estimator.from_pytorch(...)
est.fit(data, epochs=epochs)

to

est = Estimator.from_pytorch(...)
for i in range(num_epochs):
    result = est.fit(data, epochs=1)
    model = est.get_model()
    print(f"epoch {i}: {result}")

[jason-dai]

Synced offline, here are the brief summary:

Feature Requests:

Automatically printing loss and other metrics of every epoch or iteration when calling Estimator.fit

Current workaround we provides:

fitting multiple epochs in a loop. E.g.

change

est = Estimator.from_pytorch(...)
est.fit(data, epochs=epochs)

to

est = Estimator.from_pytorch(...)
for i in range(num_epochs):
    result = est.fit(data, epochs=1)
    model = est.get_model()
    print(f"epoch {i}: {result}")

What should be the desired behavior?

[yangw1234]

What should be the desired behavior?

I think it is a trade off between flexibility and usability.

Our current behavior is more flexible, like pytorch, as it can allow users to checkpointing in their prefered frequency and printing log in their desired format. The downside is they have to write some code.

If we are tageting usability, I think it also make sense to implement a fixed number of checkpointing and logging stategies and ask user to configure the one closest to their needs.

It is a judgment call.

[jason-dai]

What should be the desired behavior?

I think it is a trade off between flexibility and usability.

Our current behavior is more flexible, like pytorch, as it can allow users to checkpointing in their prefered frequency and printing log in their desired format. The downside is they have to write some code.

If we are tageting usability, I think it also make sense to implement a fixed number of checkpointing and logging stategies and ask user to configure the one closest to their needs.

It is a judgment call.

What's the keras behavior?

[yangw1234]

What's the keras behavior?

Keras will automatically output training loss, metrics and speed. Model checkpoint can be configured using model checkpoint callback.
Stick to Keras?

[jason-dai]

What's the keras behavior?

Keras will automatically output training loss, metrics and speed. Model checkpoint can be configured using model checkpoint callback. Stick to Keras?

Either Keras or PyTorch Lightening style?

[jason-dai]

Do you need to backport to AZ?

[yangw1234]

Do you need to backport to AZ?

I'll backport. This is closed automatically by github.