Added hyperparameter guidelines for DELF training (#9219)

* Merged commit includes the following changes:
326369548  by Andre Araujo:

    Fix import issues.

--
326159826  by Andre Araujo:

    Changed the implementation of the cosine weights from Keras layer to tf.Variable to manually control for L2 normalization.

--
326139082  by Andre Araujo:

    Support local feature matching using ratio test.

    To allow for easily choosing which matching type to use, we rename a flag/argument and modify all related files to avoid breakages.

    Also include a small change when computing nearest neighbors for geometric matching, to parallelize computation, which saves a little bit of time during execution (argument "n_jobs=-1").

--
326119848  by Andre Araujo:

    Option to measure DELG latency taking binarization into account.

--
324316608  by Andre Araujo:

    DELG global features training.

--
323693131  by Andre Araujo:

    PY3 conversion for delf public lib.

--
32104...

research/delf/delf/python/training/README.md

@@ -143,6 +143,8 @@ curl -Os http://storage.googleapis.com/delf/resnet50_imagenet_weights.tar.gz
 tar -xzvf resnet50_imagenet_weights.tar.gz
 ```
 
+### Training with Local Features
+
 Assuming the TFRecord files were generated in the `gldv2_dataset/tfrecord/`
 directory, running the following command should start training a model and
 output the results in the `gldv2_training` directory:
@@ -156,13 +158,7 @@ python3 train.py \
   --logdir=gldv2_training/
 ```
 
-On a multi-GPU machine the batch size can be increased to speed up the training
-using the `--batch_size` parameter. On a 8 Tesla P100 GPUs machine you can set
-the batch size to `256`:
-
-```
---batch_size=256
-```
+### Training with Local and Global Features
 
 It is also possible to train the model with an improved global features head as
 introduced in the [DELG paper](https://arxiv.org/abs/2001.05027). To do this,
@@ -179,6 +175,15 @@ python3 train.py \
   --delg_global_features
 ```
 
+### Hyperparameter Guidelines
+
+In order to improve the convergence of the training, the following
+hyperparameter values have been tested and validated on the following
+infrastructures, the remaining `train.py` flags keeping their **default 
+values**:
+* 8 Tesla P100 GPUs: `--batch_size=256`, `--initial_lr=0.01`
+* 4 Tesla P100 GPUs: `--batch_size=128`, `--initial_lr=0.005`
+
 *NOTE*: We are currently working on adding the autoencoder described in the DELG
 paper to this codebase. Currently, it is not yet implemented here. Stay tuned!