Add NN options, simplify code, minor fixes #189

arcutright · 2021-03-16T03:49:41Z

Fix terminal colors not rendering under Windows cmd.exe
Fix uncommon issue where 'date' column would be named 'Date', breaking fbprophet predictions. Didn't bother investigating, just lowercased the column names.
Add documentation for GPU acceleration in TensorFlow for GamestonkTerminal (under pred)
Refactor out common code from NN definitions (made the file less than half as long)
Print exception traces for NNs
Add a few new flags to NN (documented)
Add some missing loss functions to NNs (there are more, but I chose some common ones)
Add documentation links to --help menu for NNs within pred

'date' was sometimes 'Date'

- add more loss models - add batch_size for model training - add documentation links in '-h' for NN models - add flags for XLA settings

DidierRLopes

Amazing contribution here!! This is looking really nice!

Can you remove the line that fixed the "Date" - "date" since this should now have been taken care of? Once you do that I'll merge this!

And when I checkout your branch and used NN, although everything worked smoothly I got a message:
WARNING:tensorflow:6 out of the last 6 calls to <function Model.make_predict_function.<locals>.predict_function at 0x14fe37048> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for more details.

Is this expected?

Great work again, really happy about this!

DidierRLopes · 2021-03-16T08:24:45Z

gamestonk_terminal/prediction_techniques/fbprophet.py

@@ -95,6 +95,9 @@ def fbprophet(l_args, s_ticker, df_stock):

        df_stock = df_stock.sort_index(ascending=True)
        df_stock.reset_index(level=0, inplace=True)
+        df_stock.columns = map(
+            str.lower, df_stock.columns


I think https://github.com/DidierRLopes/GamestonkTerminal/pull/155 fixed this. This should be no longer necessary.

DidierRLopes · 2021-03-16T08:33:03Z

gamestonk_terminal/prediction_techniques/neural_networks.py

-        default=None,
-        help="The end date (format YYYY-MM-DD) to select - Backtesting",
-    )
+# ----------------------------------------------------------------------------------------------------


This shouldn't be needed anymore

DidierRLopes · 2021-03-16T08:34:17Z

gamestonk_terminal/prediction_techniques/neural_networks.py

        print("")

+    finally:


Didn't know about this keyword! This is cool!

SubtleMuffin · 2021-03-16T20:52:14Z

gamestonk_terminal/prediction_techniques/README-gpu-accel.md

+
+### CUDA 11.1, cuDNN 8.0.5 - not tested
+
+### CUDA 11.2.1, cuDNN 8.1 - does not work


Haha -- try this. Tried on CUDA 11.2.

I actually tried this but I was still running into out-of-memory issues fairly often despite having far more VRAM than needed for the models I was using. May have been due to cuDNN 8.1, may have been easy to resolve, I didn't fight it for long.

Hmmm... Here is my info for the system:

I used this version to install on an archlinux system. The command is sudo pacman -S cuda cudnn python-tensorflow-opt-cuda.

The output of nvidia-smi gives:

Tue Mar 16 19:33:17 2021 +-----------------------------------------------------------------------------+ | NVIDIA-SMI 460.56 Driver Version: 460.56 CUDA Version: 11.2 |

The output of nvcc --version gives:

nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2021 NVIDIA Corporation Built on Thu_Jan_28_19:32:09_PST_2021 Cuda compilation tools, release 11.2, V11.2.142 Build cuda_11.2.r11.2/compiler.29558016_0

And finally the cat of /usr/include/cudnn_version.h gives

#define CUDNN_MAJOR 8 #define CUDNN_MINOR 1 #define CUDNN_PATCHLEVEL 0

So I think CUDA 11.2.1, and cuDNN 8.1 is tested and proved to be working.

arcutright · 2021-03-16T23:41:48Z

Amazing contribution here!! This is looking really nice!

Can you remove the line that fixed the "Date" - "date" since this should now have been taken care of? Once you do that I'll merge this!

Thanks, changes have been made.

And when I checkout your branch and used NN, although everything worked smoothly I got a message:
WARNING:tensorflow:6 out of the last 6 calls to <function Model.make_predict_function.<locals>.predict_function at 0x14fe37048> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for more details.

Is this expected?

Great work again, really happy about this!

So this issue has to do with the input shape of x used in the calls to model.predict(x) (see snippet below). It could be fixed but I don't think it's worthwhile for an application like this since GST isn't likely to make many predictions at a time. It's just a warning that repeated calls to predict like this would be inefficient. See tensorflow/tensorflow#34025 for more details.

# Prediction
yhat = model.predict(
    stock_train_data[-ns_parser.n_inputs :].reshape(1, ns_parser.n_inputs, 1), # this input shape
    verbose=0,
)

DidierRLopes · 2021-03-16T23:50:28Z

Amazing contribution here!! This is looking really nice!
Can you remove the line that fixed the "Date" - "date" since this should now have been taken care of? Once you do that I'll merge this!

Thanks, changes have been made.

And when I checkout your branch and used NN, although everything worked smoothly I got a message:
WARNING:tensorflow:6 out of the last 6 calls to <function Model.make_predict_function.<locals>.predict_function at 0x14fe37048> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for more details.
Is this expected?
Great work again, really happy about this!

So this issue has to do with the input shape of x used in the calls to model.predict(x) (see snippet below). It could be fixed but I don't think it's worthwhile for an application like this since GST isn't likely to make many predictions at a time. It's just a warning that repeated calls to predict like this would be inefficient. See tensorflow/tensorflow#34025 for more details.
# Prediction
yhat = model.predict(
    stock_train_data[-ns_parser.n_inputs :].reshape(1, ns_parser.n_inputs, 1), # this input shape
    verbose=0,
)

Got you. Is there a way to not output this warning?

arcutright added 8 commits March 15, 2021 20:48

Fix terminal colors not rendering under Windows

ab24742

Fix uncommon column case issue for fbprophet

bfad63e

'date' was sometimes 'Date'

Add documentation for GPU acceleration

2059675

Reuse common NN code, Add more options, Remove invalid optimizer

ba6e214

- add more loss models - add batch_size for model training - add documentation links in '-h' for NN models - add flags for XLA settings

Add documentation for previous commits

a1bf6ef

Try to please black

7e3b15b

Try to please flake8

70c5a79

Add linter notes for future contributors

4518c1b

DidierRLopes requested changes Mar 16, 2021

View reviewed changes

DidierRLopes mentioned this pull request Mar 16, 2021

[IMPROVE] Automatically adjust the figure window size according to screen size #180

Closed

SubtleMuffin reviewed Mar 16, 2021

View reviewed changes

arcutright added 2 commits March 16, 2021 19:24

Remove unnecessary bugfix

444b53f

Minor changes to documentation, comments

0aa06f0

DidierRLopes approved these changes Mar 16, 2021

View reviewed changes

Merge branch 'main' into feature/pred-xla-refactor

138cb21

DidierRLopes merged commit d87a99d into OpenBB-finance:main Mar 17, 2021

aia added the feat S Small T-Shirt size Feature label Mar 18, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add NN options, simplify code, minor fixes #189

Add NN options, simplify code, minor fixes #189

arcutright commented Mar 16, 2021

DidierRLopes left a comment

DidierRLopes Mar 16, 2021

DidierRLopes Mar 16, 2021

DidierRLopes Mar 16, 2021

SubtleMuffin Mar 16, 2021

arcutright Mar 16, 2021

SubtleMuffin Mar 16, 2021

arcutright commented Mar 16, 2021

DidierRLopes commented Mar 16, 2021


		### CUDA 11.1, cuDNN 8.0.5 - not tested

		### CUDA 11.2.1, cuDNN 8.1 - does not work

Add NN options, simplify code, minor fixes #189

Add NN options, simplify code, minor fixes #189

Conversation

arcutright commented Mar 16, 2021

DidierRLopes left a comment

Choose a reason for hiding this comment

DidierRLopes Mar 16, 2021

Choose a reason for hiding this comment

DidierRLopes Mar 16, 2021

Choose a reason for hiding this comment

DidierRLopes Mar 16, 2021

Choose a reason for hiding this comment

SubtleMuffin Mar 16, 2021

Choose a reason for hiding this comment

arcutright Mar 16, 2021

Choose a reason for hiding this comment

SubtleMuffin Mar 16, 2021

Choose a reason for hiding this comment

arcutright commented Mar 16, 2021

DidierRLopes commented Mar 16, 2021