retinanet model model.predict() return only boxes with -1 after training

[hugopi]

Dear keras-cv Team,

Firstly, I would like to express my sincere appreciation for the excellent work you have done on Keras-cv. Your commitment and dedication to the project are evident in the high-quality code and comprehensive documentation that you have provided.

I have been studying your repository and I have some questions that I was hoping you could help me with. I am currently working on a ship detection model. So I am trying to implement your retinaNet model. After suiting my data to the specific keras-cv input I tried to fit the model. However I am experiencing some troubles with predictions.

PROBLEM :

When I use model.fit(image_test) the predictions returned are all -1 :

{'boxes': array([[[-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.],
        [-1., -1., -1., -1.]]], dtype=float32), 'confidence': array([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0.]], dtype=float32), 'classes': array([[-1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,
        -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,
        -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,
        -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,
        -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,
        -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,
        -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1., -1.,
        -1., -1., -1., -1., -1., -1., -1., -1., -1.]], dtype=float32), 'num_detections': array([0], dtype=int32)}```

With model(image_test) I get : 

```predictions =
{'box': <tf.Tensor: shape=(1, 49104, 4), dtype=float32, numpy=
array([[[-0.09781634,  0.19049369, -0.04102969, -0.21385634],
        [-0.2612432 ,  0.17355151, -0.15664971,  0.14730531],
        [-0.19195703,  0.3431507 ,  0.14365467,  0.06342052],
        ...,
        [ 0.00151845,  0.0044898 , -0.01494532,  0.00200025],
        [ 0.01669142, -0.01205819,  0.01728174,  0.00686904],
        [ 0.02757294, -0.00464585, -0.01185186, -0.00704384]]],
      dtype=float32)>, 'classification': <tf.Tensor: shape=(1, 49104, 2), dtype=float32, numpy=
array([[[-4.491832 , -4.6013255],
        [-4.2144737, -4.557256 ],
        [-4.6533785, -4.6464787],
        ...,
        [-4.5967946, -4.599116 ],
        [-4.6118917, -4.59214  ],
        [-4.5627737, -4.5775194]]], dtype=float32)>}

DATA : 
image_test = 
tf.Tensor(
[[[[0.3839975  0.27996504 0.33198124]
   [0.40149528 0.29109997 0.34629762]
   [0.44508073 0.27996504 0.36252287]
   ...
   [0.37477133 0.22842602 0.30159864]
   [0.3979957  0.25674066 0.32736814]
   [0.3760439  0.21379144 0.29491767]]

  [[0.39481428 0.29269072 0.34375247]
   [0.394178   0.30605265 0.35011533]
   [0.40149528 0.30764335 0.35456935]
   ...
   [0.42312893 0.26342162 0.34327528]
   [0.43203694 0.25801322 0.34502506]
   [0.40181342 0.22365387 0.31273365]]

  [[0.41390282 0.30923405 0.36156845]
   [0.37699834 0.31591508 0.34645668]
   [0.3101885  0.2993717  0.3047801 ]
   ...
   [0.44571704 0.2828283  0.36427265]
   [0.4425356  0.25801322 0.35027438]
   [0.41485724 0.24178797 0.3283226 ]]

  ...

  [[0.3585461  0.27837428 0.31846023]
   [0.29746282 0.2882367  0.29284978]
   [0.29173627 0.28251016 0.2871232 ]
   ...
   [0.29587212 0.29841724 0.29714468]
   [0.31336996 0.2993717  0.3063708 ]
   [0.36809036 0.30064425 0.3343673 ]]

  [[0.36109126 0.29396328 0.32752725]
   [0.32768634 0.29841724 0.31305176]
   [0.3079615  0.27392033 0.2909409 ]
   ...
   [0.3525014  0.3079615  0.33023146]
   [0.34295717 0.30191684 0.322437  ]
   [0.37063548 0.3025531  0.3365943 ]]

  [[0.35568285 0.30191684 0.32879984]
   [0.33182216 0.2936451  0.31273365]
   [0.32705003 0.27551103 0.30128053]
   ...
   [0.42853734 0.29269072 0.360614  ]
   [0.40213156 0.28537342 0.34375247]
   [0.3820886  0.2847371  0.33341286]]]], shape=(1, 512, 512, 3), dtype=float32)```
   
my model is trained with images and labels with folowing pattern : 

```<tf.Tensor: shape=(1, 512, 512, 3), dtype=float32, numpy=
array([[[[0.33106965, 0.24011646, 0.285593  ],
         [0.32864422, 0.24666505, 0.2876546 ],
         [0.31190884, 0.24545236, 0.2786806 ],
         ...,
         [0.27674025, 0.22459374, 0.250667  ],
         [0.2527286 , 0.21925785, 0.2359932 ],
         [0.27334464, 0.2211982 , 0.2472714 ]],

        [[0.31215134, 0.25709438, 0.28462285],
         [0.31554696, 0.25782198, 0.28668448],
         [0.30657288, 0.25054574, 0.27855933],
         ...,
         [0.2791657 , 0.22895953, 0.2540626 ],
         [0.2660684 , 0.24423967, 0.25515404],
         [0.29008004, 0.23041478, 0.2602474 ]],

        [[0.30196458, 0.25685182, 0.2794082 ],
         [0.3087558 , 0.2532137 , 0.28098473],
         [0.30705798, 0.24690762, 0.2769828 ],
         ...,
         [0.26364297, 0.21974294, 0.24169298],
         [0.23914629, 0.2226534 , 0.23089983],
         [0.2786806 , 0.22071311, 0.24969682]],

        ...,

        [[0.29759887, 0.21877277, 0.2581858 ],
         [0.30341986, 0.19936942, 0.25139466],
         [0.3060878 , 0.18869756, 0.2473927 ],
         ...,
         [0.21101144, 0.23381034, 0.22241089],
         [0.20227993, 0.2231385 , 0.2127092 ],
         [0.21295173, 0.22507884, 0.21901529]],

        [[0.32670388, 0.22895953, 0.27783167],
         [0.32888675, 0.20422022, 0.2665535 ],
         [0.3233083 , 0.19500364, 0.259156  ],
         ...,
         [0.21683241, 0.23672086, 0.22677661],
         [0.21343681, 0.22071311, 0.21707493],
         [0.21780258, 0.226049  , 0.2219258 ]],

        [[0.30633035, 0.21537717, 0.26085377],
         [0.32112542, 0.21343681, 0.26728114],
         [0.32476354, 0.21295173, 0.26885766],
         ...,
         [0.2231385 , 0.24254183, 0.23284017],
         [0.21489207, 0.23041478, 0.2226534 ],
         [0.20470531, 0.23235507, 0.2185302 ]]]], dtype=float32)>, {'boxes': <tf.RaggedTensor [[[22.0, 285.0, 4.0, 6.0],
  [203.0, 301.0, 4.0, 4.0],
  [237.0, 146.0, 25.0, 20.0]]]>, 'classes': <tf.RaggedTensor [[1.0, 1.0, 0.0]]>})```



CODE : 

``from GPU_manager.gpu_manager import GPU_manager
from Configs.configs import CFG
from Data_factory.data_generator import Data_generator

import os.path
import pandas as pd

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import optimizers
import keras_cv
import datetime
from sklearn.model_selection import train_test_split

# Workaround to silent tf INFO messages (but not WARNING and ERROR messages)
# Must happen before importing tensorflow
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

# GLOBALS
MODEL = 'retinaNET'
NB_SAMPLES = CFG['Model']["dataset"]["NB_SAMPLES"]
BATCH_SIZE = CFG['Model']["train"]["BATCH_SIZE"]
EPOCHS = CFG['Model']["train"]["EPOCHS"]
SPLIT_PATH = os.path.join(CFG["Resource"]["Dataset"]["Xview_radar"]["ANNOTATIONS_DIR"], "validation_scaled.csv")
CHIP_DIR = CFG["Resource"]["Dataset"]["Xview_radar"]["CHIP_DIR"]
TIME = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
CHECKPOINT_PATH = os.path.join(CFG["Resource"]["Savings"]["SAVE_DIR"], MODEL, "checkpoint", "fit" + TIME, "cp.ckpt")
LOGS_DIR = os.path.join(CFG["Resource"]["Savings"]["SAVE_DIR"], MODEL, "logs", "fit" + TIME)

# gpu configuration
gpu = GPU_manager()
gpu.setup_gpu_env()

# Dataset
with tf.device('/CPU:0'):
    # load the data
    patch_paths = Data_generator.parse_data_directory(CHIP_DIR)

    train, remain = train_test_split(patch_paths, train_size=0.8, random_state=0, shuffle=True)
    valid, test = train_test_split(remain, test_size=0.5, random_state=0, shuffle=True)

    train_generator = Data_generator(train, BATCH_SIZE, split='train', shuffle_on_epoch_end=True, shuffle_at_init=True)
    #train_dataset = train_generator.data_loader()

    valid_generator = Data_generator(valid, BATCH_SIZE, split='valid', shuffle_on_epoch_end=True, shuffle_at_init=True)
    #valid_dataset = valid_generator.data_loader()

    test_generator = Data_generator(test, batch_size=1, split='test', shuffle_on_epoch_end=True, shuffle_at_init=True)
    #test_dataset = test_generator.data_loader()

# Model
print("\nModel")
print(f"-- {MODEL}")
print(f"-- BATCH_SIZE : {BATCH_SIZE}")
print(f"-- EPOCHS : {EPOCHS}")

model = keras_cv.models.RetinaNet(
    num_classes=2,
    bounding_box_format="xywh",
    backbone=keras_cv.models.ResNet50Backbone.from_preset(
        "resnet50_imagenet"
    )
)
model.backbone.trainable = False

for layer in model.layers:
    if hasattr(layer, 'rate'):
        layer.rate = 0.5


optimizer = tf.keras.optimizers.SGD(
    learning_rate=0.01, momentum=0.9, global_clipnorm=10.0)

# Train model
model.compile(
    classification_loss='focal',
    box_loss='huber',
    optimizer=optimizer,
    jit_compile=False
)

callbacks = [
    keras.callbacks.TensorBoard(log_dir=LOGS_DIR,
                                histogram_freq=1,
                                write_graph=True,
                                write_images=False,
                                update_freq='epoch',
                                profile_batch='1,2',
                                embeddings_freq=1),
    keras.callbacks.ModelCheckpoint(CHECKPOINT_PATH, save_weights_only=True)
]

print("\nTraining")

fit_history = model.fit(
    train_dataset,
    validation_data=valid_dataset,
    epochs=EPOCHS,
    callbacks=callbacks
)

print("-- history:")
print(pd.DataFrame(fit_history.history))

print(">> Done")```


CONFIG : 

cuda-nvcc                 11.3.58              
cudatoolkit               11.2.2              
cudnn                     8.1.0.77
tensorflow		  2.11
keras-cv                  0.4.2  (unreleased version, github version of the day)



PROBLEM 2 : 

Generator seems not working on model.fit(train_generator)

ERROR : 

Traceback (most recent call last):
  File "/snap/pycharm-community/327/plugins/python-ce/helpers/pydev/pydevconsole.py", line 364, in runcode
    coro = func()
  File "<input>", line 1, in <module>
  File "/snap/pycharm-community/327/plugins/python-ce/helpers/pydev/_pydev_bundle/pydev_umd.py", line 198, in runfile
    pydev_imports.execfile(filename, global_vars, local_vars)  # execute the script
  File "/snap/pycharm-community/327/plugins/python-ce/helpers/pydev/_pydev_imps/_pydev_execfile.py", line 18, in execfile
    exec(compile(contents+"\n", file, 'exec'), glob, loc)
  File "/home/hugo/devWorkingPlace/src/rd-ship-detection/models/keras_cv/RetinaNet/retinaNet_train.py", line 94, in <module>
    fit_history = model.fit(
  File "/home/hugo/.local/lib/python3.10/site-packages/keras/utils/traceback_utils.py", line 70, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/tmp/__autograph_generated_filey_3qvtq1.py", line 15, in tf__train_function
    retval_ = ag__.converted_call(ag__.ld(step_function), (ag__.ld(self), ag__.ld(iterator)), None, fscope)
  File "/home/hugo/miniconda3/envs/sinbad/lib/python3.10/site-packages/keras_cv/models/object_detection/retina_net/retina_net.py", line 448, in train_step
    y_for_label_encoder = bounding_box.convert_format(
  File "/home/hugo/miniconda3/envs/sinbad/lib/python3.10/site-packages/keras_cv/bounding_box/converters.py", line 366, in convert_format
    boxes["boxes"] = convert_format(
  File "/home/hugo/miniconda3/envs/sinbad/lib/python3.10/site-packages/keras_cv/bounding_box/converters.py", line 377, in convert_format
    raise ValueError(
ValueError: in user code:
    File "/home/hugo/.local/lib/python3.10/site-packages/keras/engine/training.py", line 1249, in train_function  *
        return step_function(self, iterator)
    File "/home/hugo/.local/lib/python3.10/site-packages/keras/engine/training.py", line 1233, in step_function  **
        outputs = model.distribute_strategy.run(run_step, args=(data,))
    File "/home/hugo/.local/lib/python3.10/site-packages/keras/engine/training.py", line 1222, in run_step  **
        outputs = model.train_step(data)
    File "/home/hugo/miniconda3/envs/sinbad/lib/python3.10/site-packages/keras_cv/models/object_detection/retina_net/retina_net.py", line 448, in train_step
        y_for_label_encoder = bounding_box.convert_format(
    File "/home/hugo/miniconda3/envs/sinbad/lib/python3.10/site-packages/keras_cv/bounding_box/converters.py", line 366, in convert_format
        boxes["boxes"] = convert_format(
    File "/home/hugo/miniconda3/envs/sinbad/lib/python3.10/site-packages/keras_cv/bounding_box/converters.py", line 377, in convert_format
        raise ValueError(
    ValueError: Expected `boxes` to be a Tensor with a final dimension of `4`. Instead, got `boxes.shape=(None, None, None)`.




Thank you again for your hard work and dedication to this project. I appreciate your willingness to share your knowledge and expertise with the community, and I look forward to hearing back from you.

Best regards,
Hugo PIRES

[ianstenbit]

Hi @hugopi -- could you post some TensorBoard logs? It may be useful to look at whether the loss has converged.

The -1 boxes mean that no predictions were made by the model. This could be because the default prediction decoder has a high NMS threshold. You can try using a lower threshold to see if you get any boxes like this:

model.prediction_decoder = keras_cv.layers.MultiClassNonMaxSuppression(bounding_box_format="xywh", from_logits=True,
                                                                confidence_threshold=0.05, # This is the thing to tune. Higher means fewer boxes are predicted.
                                                                iou_threshold=0.5,
                                                                max_detections=100,
                                                                max_detections_per_class=25)

It's also possible that there's something wrong with the input format of your boxes. I'd strongly recommend visualizing your boxes before training to verify that they are in the right format ("xywh" in this case)

[ianstenbit]

Hi Hugo -- based on the tensorboard logs, it seems possibile that there's something other than overfitting at play.

It could be overfitting, which I would recommend addressing by adding some data augmentation to your training script if you don't have it already.

I'd also strongly recommend making your backbone fine-tunable by updating

model.backbone.trainable = False

to

model.backbone.trainable = True

And if you haven't already -- visualize the inputs before training to make sure they're formatted correctly, and visualize the output boxes to see if they are coherent at all

[hugopi]

Hi @ianstenbit !

Thanks for getting back to me. I appreciate your advice on addressing the overfitting by adding some data augmentation to the training script. I will work on that.

I also followed your suggestion to make the backbone fine-tunable by updating the code as you mentioned. But I am limited by my GPU, the code is complaining about memory. I am using a NVIDIA GeForce GTX 1070 of 8Go. Is There any minimal configuration to be able to set backbone trainable ? (Sorry, transfer learning is new for me... )

I visualized the inputs before training to verify that they are formatted correctly. And actually, there are some outliers. You were right.

Happy to see your are implementing yolov8! I will definitely try it.

Thanks again for your help!

[jbischof]

@hugopi your best bet with memory issues is to reduce batch size

[hugopi]

Hi @jbischof,

Thanks for the tip, I had already reduce the batch size but it result in NaN losses even if i set global_clipnorm to 10.0.

Moreover, I have a more global question but it needs some contextualization.
As a fifth-year french student engineer working on a boat detection project using SAR imagery, I chose to work with keras-cv for its promising potential and the flexibility it provides for model implementation.

I am particularly interested in implementing fasterRCNN, RetinaNet, and Yolo models, and your library has been instrumental in making this possible. However, as a newbie to the field, I have found myself struggling without tutorials. Despite this, I have been learning a lot by reading through your source code.

I have several questions like NaN losses or input differences between fasterRCNN and RetinaNet, among others. I would like to continue seeking your guidance as I navigate through these challenges without flooding discussions of my problems. Please let me know if it would be better to open a discussion for each topic or to continue our exchange on this feed.

Once again, thank you for your invaluable contributions to the field and for your continued support in my learning journey.

Hugo

[jbischof]

Thanks for the kind words @hugopi! I'd recommend filing Issues for any bugs and feature requests you find.

Our best advice overall is to visualize both the labels and predictions in your data using our visualization tools. Most bugs are either from issues with the labels or configuration of the model (bounding box format, image format).