Custom Dataset Low Map and No Convergence(with images!)

I am using custom data. I have modified the callbacks in train.py so that it is now monitoring validation loss from the validation set, as opposed to the training set. Original learning rate was 1e-4
`def create_callbacks(saved_weights_name, tensorboard_logs, model_to_save):
makedirs(tensorboard_logs)

early_stop = EarlyStopping(
    monitor     = 'val_loss', 
    min_delta   = 0.001, 
    patience    = 50, 
    mode        = 'min', 
    verbose     = 1
)
checkpoint = CustomModelCheckpoint(
    model_to_save   = model_to_save,
    filepath        = saved_weights_name,# + '{epoch:02d}.h5', 
    monitor         = 'val_loss', 
    verbose         = 1, 
    save_best_only  = True, 
    mode            = 'min', 
    period          = 1
)
reduce_on_plateau = ReduceLROnPlateau(
    monitor  = 'val_loss',
    factor   = 0.1,
    patience = 10,
    verbose  = 1,
    mode     = 'min',
    epsilon  = 0.01,
    cooldown = 0,
)
tensorboard = CustomTensorBoard(
    log_dir                = tensorboard_logs,
    write_graph            = True,
    write_images           = True,
)    
return [early_stop, checkpoint, reduce_on_plateau, tensorboard]`

history = train_model.fit_generator(
          generator        = train_generator, 
          validation_data = valid_generator,
          steps_per_epoch  = len(train_generator) * config['train']['train_times'], 
          epochs           = config['train']['nb_epochs'] + config['train']['warmup_epochs'], 
          verbose          = 2 if config['train']['debug'] else 1,
          callbacks        = callbacks, 
          workers          = 4,
          max_queue_size   = 8
      )

Here is a plot of loss, blue being validation loss and red being training loss:

And here is a sample image of my dataset:

where same numbers are top left/bottom right of a bounding box. This example was pretty messy, but other people who have used this dataset claim to reach 70%+ map on a default single shot detector with no modifications. My maximum was 0.2. I am using the default yolov3 and haven't made any configurations.
Here is my config:
`{
"model" : {
"min_input_size": 352,
"max_input_size": 352,
"anchors": [13,58, 18,14, 30,31, 47,55, 50,128, 75,24, 95,69, 126,121, 161,208],
"labels": ["nlb"]
},

"train": {
    "train_image_folder":   "/content/drive/MyDrive/yolo/nlb_train_image/",
    "train_annot_folder":   "/content/drive/MyDrive/yolo/nlb_train_annot/",
      "cache_name":		"nlb_train.pkl",
      "pretrained_weights":   "",

    "train_times":          1,
    "batch_size":           16,
    "learning_rate":        1e-4,
    "nb_epochs":            200,
    "warmup_epochs":        3,
    "ignore_thresh":        0.4,
    "gpus":                 "0",

    "grid_scales":          [1,1,1],
    "obj_scale":            5,
    "noobj_scale":          1,
    "xywh_scale":           1,
    "class_scale":          1,

    "tensorboard_dir":      "logs",
    "saved_weights_name":   "nlb.h5",
    "debug":                true
},

"valid": {
    "valid_image_folder":   "/content/drive/MyDrive/yolo/nlb_valid_image/",
    "valid_annot_folder":   "/content/drive/MyDrive/yolo/nlb_valid_image/",
    "cache_name":		"nlb_valid.pkl",

    "valid_times":          1
}

}
`

Data augmented from 612 -> 7140 for training set, valid_set = 152, test_set = 141. I have played around with increasing validation set but still get poor results. I changed initial learning rate from 1e-3 to 1e-5 but again, poor results. If anyone is willing to test it themselves, here are links to the dataset:

train images
https://drive.google.com/drive/folders/1qmkCFNNuAtsOtFzQpmEgZyaN-DeSOuyr?usp=sharing
train annot
https://drive.google.com/drive/folders/1-5SZAcuOXz1Rt_eZ19RtVvgMXr8aId6j?usp=sharing
valid images
https://drive.google.com/drive/folders/1Xxz28iVaxXMIDA_UVQPKOAs0KfU4KtVT?usp=sharing
valid annot
https://drive.google.com/drive/folders/1Ml7DjEBwS50eUaMO378zYGdZTQsdfUNB?usp=sharing
test images
https://drive.google.com/drive/folders/1tEiW-Xx74kWua810sR5e0OyWvPABhJ1z?usp=sharing
test annot
https://drive.google.com/drive/folders/12jvZpb06tRopSpYLUSoiONOlvBWPV1ud?usp=sharing
train.pkl file
https://drive.google.com/file/d/1-5a1JGGYr6xP8ydCJOBOfhTPOym93u62/view?usp=sharing
valid.pkl file
https://drive.google.com/file/d/1-86kPCtYQKmYVzxyFZfQoNIf03_YdwDA/view?usp=sharing

I have spent an embarrassing amount of time trying to solve this problem.

[lexuansanh]

Because you put valid_generator in Model.fit_generator, Then data of valid_generator is augmented same as train_generator. This may affect the valid_generator evaluation. You can use the evaluate() function to evaluate on valid_generator. or tweak in _aug_image() function in BatchGenerator Class so that it doesn't modify the original image of valid data

[lexuansanh]

I see few problems here:

1. If you don't aug_image. Training can be overfitting. Because train_Generator is not augmentated. The method is to add 1 argument to the Generator Class. Ex: self.aug = True: Do aug_image (using for train_generator), if self.aug = False: Do not aug_image (Return original image, using for valid_generator).
2. Maybe your Anchors Box does not match the data. Public data with anchors has been carefully calculated by the author through gen_anchors.py. If necessary, the anchor box can be recalculated.
3. It is possible that the evaluate() function does not match your own data. You can check the evaluate() function again and edit it to suit your data.Ex: Check line 83, 84 of evaluate():
   detections = all_detections[i][label]
annotations = all_annotations[i][label]

[lexuansanh]

I see yours iou quite low. I usually get a value of >= 0.8. I usually create anchor box after Generator, because then the bounding boxs are really right to create anchors. you can create many times and choose the one with the biggest iou.
Your data is augmented quite bit. you can view and remove some augmentation that you think are unnecessary or worsen your train data.

Would you advise increasing number of anchors per detection stage such as changing from the default 3 to 4? For my dataset, I would like to have a fast detection speed, but accuracy is more important because it is supposed to be looking for a disease. Perhaps yolo was the wrong algorithm to use all along?

[lexuansanh]

Aha. I understand what you mean. In my opinion, there are 2 ways:

1. Increase the number of Anchor Boxes for each output. Normally each output will have 3 featured anchors, now we will add 1 more anchors box, then it will be 4 anchors for 1 output. Do the same for the other 2 outputs, for a total of 12 anchor boxes. I find this way quite suitable for you. You would edit the following:
   ex:
   Layer 80 => 82: replace 4*(5+nb_class) for 3*(5+nb_class)
   pred_yolo_1 = _conv_block(x, [{'filter': 1024, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 80}, {'filter': (4*(5+nb_class)), 'kernel': 1, 'stride': 1, 'bnorm': False, 'leaky': False, 'layer_idx': 81}], do_skip=False)
   Do the same for # Layer 92 => 94, # Layer 99 => 106
   Then calculate the dimensions in yololayer accordingly:
   ex:
   replace [batch_size, 1, 1, 4, 1] for [batch_size, 1, 1, 3, 1]
   self.cell_grid = tf.tile(tf.concat([cell_x,cell_y],-1), [batch_size, 1, 1, 4, 1])
   adjust the shape of the y_predict [batch, grid_h, grid_w, 4, 4+1+nb_class]
   ....
2. Modify Model. This way is complicated and requires computing power of the computer and much larger memory, due to the increased number of parameters:
   Specifically:
   input_image = 832, 832
   add more _conv_block and do the same to:
   output1: batch,104,104,3,(4+1+nb_class)
   output2: batch,52,52,3,(4+1+nb_class)
   output3: batch,26,26,3,(4+1+nb_class)
   output4: batch,13,13,3,(4+1+nb_class)

Or any input value to get 4 outputs whose grid cells are integers and all decrement by base 2. ex (448 -> 56 ---> 28 -> 14 -> 7), ,.. .
These methods are just my ideas and I'm not sure if it will work for you :D

Thank you very much. I tried implementing the first approach, and tried to use 5 anchors per detection layer by printing out 15 anchors in generate_anchors.py instead of the default 9. iou increased from 0.54 (using 9 anchors) to 0.62. Increasing # of anchors increases iou. In the code below, I marked the lines I changed in yolo.py with the comment "#CHANGED HERE"

from keras.layers import Conv2D, Input, BatchNormalization, LeakyReLU, ZeroPadding2D, UpSampling2D, Lambda
from keras.layers.merge import add, concatenate
from keras.models import Model
from keras.engine.topology import Layer
import tensorflow as tf
import keras


debug = False

class YoloLayer(Layer):
    def __init__(self, anchors, max_grid, batch_size, warmup_batches, ignore_thresh, 
                    grid_scale, obj_scale, noobj_scale, xywh_scale, class_scale, 
                    **kwargs):
        # make the model settings persistent
        self.ignore_thresh  = ignore_thresh
        self.warmup_batches = warmup_batches


# CHANGED HERE self.anchors to [1, 1, 1, 5, 2] from [1, 1, 1, 3, 2]. Possibly not needed?

        self.anchors        = tf.constant(anchors, dtype='float', shape=[1,1,1,5,2])
        self.grid_scale     = grid_scale
        self.obj_scale      = obj_scale
        self.noobj_scale    = noobj_scale
        self.xywh_scale     = xywh_scale
        self.class_scale    = class_scale        

        # make a persistent mesh grid
        max_grid_h, max_grid_w = max_grid

        cell_x = tf.to_float(tf.reshape(tf.tile(tf.range(max_grid_w), [max_grid_h]), (1, max_grid_h, max_grid_w, 1, 1)))
        cell_y = tf.transpose(cell_x, (0,2,1,3,4))


# CHANGED HERE [batch_size, 1, 1, 5, 1] changed from the [batch_size, 1, 1, 3, 1] like suggested

        self.cell_grid = tf.tile(tf.concat([cell_x,cell_y],-1), [batch_size, 1, 1, 5, 1])

        super(YoloLayer, self).__init__(**kwargs)

    def build(self, input_shape):
        super(YoloLayer, self).build(input_shape)  # Be sure to call this somewhere!

    def call(self, x):
        input_image, y_pred, y_true, true_boxes = x

        # adjust the shape of the y_predict [batch, grid_h, grid_w, 3, 4+1+nb_class]


# CHANGED HERE I am assuming when you say adjust y_predict you have to change tf.constants([3,-1]) to tf.constants([5,-1])

        y_pred = tf.reshape(y_pred, tf.concat([tf.shape(y_pred)[:3], tf.constant([5, -1])], axis=0))
        
        # initialize the masks
        object_mask     = tf.expand_dims(y_true[..., 4], 4)

        # the variable to keep track of number of batches processed
        batch_seen = tf.Variable(0.)        

        # compute grid factor and net factor
        grid_h      = tf.shape(y_true)[1]
        grid_w      = tf.shape(y_true)[2]
        grid_factor = tf.reshape(tf.cast([grid_w, grid_h], tf.float32), [1,1,1,1,2])

        net_h       = tf.shape(input_image)[1]
        net_w       = tf.shape(input_image)[2]            
        net_factor  = tf.reshape(tf.cast([net_w, net_h], tf.float32), [1,1,1,1,2])
        
        """
        Adjust prediction
        """
        pred_box_xy    = (self.cell_grid[:,:grid_h,:grid_w,:,:] + tf.sigmoid(y_pred[..., :2]))  # sigma(t_xy) + c_xy
        pred_box_wh    = y_pred[..., 2:4]                                                       # t_wh
        pred_box_conf  = tf.expand_dims(tf.sigmoid(y_pred[..., 4]), 4)                          # adjust confidence
        pred_box_class = y_pred[..., 5:]                                                        # adjust class probabilities      

        """
        Adjust ground truth
        """
        true_box_xy    = y_true[..., 0:2] # (sigma(t_xy) + c_xy)
        true_box_wh    = y_true[..., 2:4] # t_wh
        true_box_conf  = tf.expand_dims(y_true[..., 4], 4)
        true_box_class = tf.argmax(y_true[..., 5:], -1)         

        """
        Compare each predicted box to all true boxes
        """        
        # initially, drag all objectness of all boxes to 0
        conf_delta  = pred_box_conf - 0 

        # then, ignore the boxes which have good overlap with some true box
        true_xy = true_boxes[..., 0:2] / grid_factor
        true_wh = true_boxes[..., 2:4] / net_factor
        
        true_wh_half = true_wh / 2.
        true_mins    = true_xy - true_wh_half
        true_maxes   = true_xy + true_wh_half
        
        pred_xy = tf.expand_dims(pred_box_xy / grid_factor, 4)
        pred_wh = tf.expand_dims(tf.exp(pred_box_wh) * self.anchors / net_factor, 4)
        
        pred_wh_half = pred_wh / 2.
        pred_mins    = pred_xy - pred_wh_half
        pred_maxes   = pred_xy + pred_wh_half    

        intersect_mins  = tf.maximum(pred_mins,  true_mins)
        intersect_maxes = tf.minimum(pred_maxes, true_maxes)

        intersect_wh    = tf.maximum(intersect_maxes - intersect_mins, 0.)
        intersect_areas = intersect_wh[..., 0] * intersect_wh[..., 1]
        
        true_areas = true_wh[..., 0] * true_wh[..., 1]
        pred_areas = pred_wh[..., 0] * pred_wh[..., 1]

        union_areas = pred_areas + true_areas - intersect_areas
        iou_scores  = tf.truediv(intersect_areas, union_areas)

        best_ious   = tf.reduce_max(iou_scores, axis=4)        
        conf_delta *= tf.expand_dims(tf.to_float(best_ious < self.ignore_thresh), 4)

        """
        Compute some online statistics
        """            
        true_xy = true_box_xy / grid_factor
        true_wh = tf.exp(true_box_wh) * self.anchors / net_factor

        true_wh_half = true_wh / 2.
        true_mins    = true_xy - true_wh_half
        true_maxes   = true_xy + true_wh_half

        pred_xy = pred_box_xy / grid_factor
        pred_wh = tf.exp(pred_box_wh) * self.anchors / net_factor 
        
        pred_wh_half = pred_wh / 2.
        pred_mins    = pred_xy - pred_wh_half
        pred_maxes   = pred_xy + pred_wh_half      

        intersect_mins  = tf.maximum(pred_mins,  true_mins)
        intersect_maxes = tf.minimum(pred_maxes, true_maxes)
        intersect_wh    = tf.maximum(intersect_maxes - intersect_mins, 0.)
        intersect_areas = intersect_wh[..., 0] * intersect_wh[..., 1]
        
        true_areas = true_wh[..., 0] * true_wh[..., 1]
        pred_areas = pred_wh[..., 0] * pred_wh[..., 1]

        union_areas = pred_areas + true_areas - intersect_areas
        iou_scores  = tf.truediv(intersect_areas, union_areas)
        iou_scores  = object_mask * tf.expand_dims(iou_scores, 4)
        
        count       = tf.reduce_sum(object_mask)
        count_noobj = tf.reduce_sum(1 - object_mask)
        detect_mask = tf.to_float((pred_box_conf*object_mask) >= 0.5)
        class_mask  = tf.expand_dims(tf.to_float(tf.equal(tf.argmax(pred_box_class, -1), true_box_class)), 4)
        recall50    = tf.reduce_sum(tf.to_float(iou_scores >= 0.5 ) * detect_mask  * class_mask) / (count + 1e-3)
        recall75    = tf.reduce_sum(tf.to_float(iou_scores >= 0.75) * detect_mask  * class_mask) / (count + 1e-3)    
        avg_iou     = tf.reduce_sum(iou_scores) / (count + 1e-3)
        avg_obj     = tf.reduce_sum(pred_box_conf  * object_mask)  / (count + 1e-3)
        avg_noobj   = tf.reduce_sum(pred_box_conf  * (1-object_mask))  / (count_noobj + 1e-3)
        avg_cat     = tf.reduce_sum(object_mask * class_mask) / (count + 1e-3) 

        """
        Warm-up training
        """
        batch_seen = tf.assign_add(batch_seen, 1.)
        
        true_box_xy, true_box_wh, xywh_mask = tf.cond(tf.less(batch_seen, self.warmup_batches+1), 
                              lambda: [true_box_xy + (0.5 + self.cell_grid[:,:grid_h,:grid_w,:,:]) * (1-object_mask), 
                                       true_box_wh + tf.zeros_like(true_box_wh) * (1-object_mask), 
                                       tf.ones_like(object_mask)],
                              lambda: [true_box_xy, 
                                       true_box_wh,
                                       object_mask])

        """
        Compare each true box to all anchor boxes
        """      
        wh_scale = tf.exp(true_box_wh) * self.anchors / net_factor
        wh_scale = tf.expand_dims(2 - wh_scale[..., 0] * wh_scale[..., 1], axis=4) # the smaller the box, the bigger the scale

        xy_delta    = xywh_mask   * (pred_box_xy-true_box_xy) * wh_scale * self.xywh_scale
        wh_delta    = xywh_mask   * (pred_box_wh-true_box_wh) * wh_scale * self.xywh_scale
        conf_delta  = object_mask * (pred_box_conf-true_box_conf) * self.obj_scale + (1-object_mask) * conf_delta * self.noobj_scale
        class_delta = object_mask * \
                      tf.expand_dims(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=true_box_class, logits=pred_box_class), 4) * \
                      self.class_scale

        loss_xy    = tf.reduce_sum(tf.square(xy_delta),       list(range(1,5)))
        loss_wh    = tf.reduce_sum(tf.square(wh_delta),       list(range(1,5)))
        loss_conf  = tf.reduce_sum(tf.square(conf_delta),     list(range(1,5)))
        loss_class = tf.reduce_sum(class_delta,               list(range(1,5)))

        loss = loss_xy + loss_wh + loss_conf + loss_class

        if debug:
            loss = tf.Print(loss, [grid_h, avg_obj], message='avg_obj \t\t', summarize=1000)
            loss = tf.Print(loss, [grid_h, avg_noobj], message='avg_noobj \t\t', summarize=1000)
            loss = tf.Print(loss, [grid_h, avg_iou], message='avg_iou \t\t', summarize=1000)
            loss = tf.Print(loss, [grid_h, avg_cat], message='avg_cat \t\t', summarize=1000)
            loss = tf.Print(loss, [grid_h, recall50], message='recall50 \t', summarize=1000)
            loss = tf.Print(loss, [grid_h, recall75], message='recall75 \t', summarize=1000)   
            loss = tf.Print(loss, [grid_h, count], message='count \t', summarize=1000)     
            loss = tf.Print(loss, [grid_h, tf.reduce_sum(loss_xy), 
                                        tf.reduce_sum(loss_wh), 
                                        tf.reduce_sum(loss_conf), 
                                        tf.reduce_sum(loss_class)],  message='loss xy, wh, conf, class: \t',   summarize=1000)   


        return loss*self.grid_scale

    def compute_output_shape(self, input_shape):
        return [(None, 1)]

def _conv_block(inp, convs, do_skip=True):
    x = inp
    count = 0
    
    for conv in convs:
        if count == (len(convs) - 2) and do_skip:
            skip_connection = x
        count += 1
        
        if conv['stride'] > 1: x = ZeroPadding2D(((1,0),(1,0)))(x) # unlike tensorflow darknet prefer left and top paddings
        x = Conv2D(conv['filter'], 
                   conv['kernel'], 
                   strides=conv['stride'], 
                   padding='valid' if conv['stride'] > 1 else 'same', # unlike tensorflow darknet prefer left and top paddings
                   name='conv_' + str(conv['layer_idx']), 
                   use_bias=False if conv['bnorm'] else True)(x)
        if conv['bnorm']: x = BatchNormalization(epsilon=0.001, name='bnorm_' + str(conv['layer_idx']))(x)
        if conv['leaky']: x = LeakyReLU(alpha=0.1, name='leaky_' + str(conv['layer_idx']))(x)

    return add([skip_connection, x]) if do_skip else x    


def create_yolov3_model(
    nb_class, 
    anchors, 
    max_box_per_image, 
    max_grid, 
    batch_size, 
    warmup_batches,
    ignore_thresh,
    grid_scales,
    obj_scale,
    noobj_scale,
    xywh_scale,
    class_scale
):
    input_image = Input(shape=(None, None, 3)) # net_h, net_w, 3
    true_boxes  = Input(shape=(1, 1, 1, max_box_per_image, 4))
    print(len(anchors))


# CHANGED HERE well, not really, but I am assuming you keep the len(anchors)//6 the same becasue 15(# of anchors generated) 
* 2 = 30, and 30 // 6 equals 5, the number of anchors for each detection layer that I want
* 
    true_yolo_1 = Input(shape=(None, None, len(anchors)//6, 4+1+nb_class)) # grid_h, grid_w, nb_anchor, 5+nb_class
    true_yolo_2 = Input(shape=(None, None, len(anchors)//6, 4+1+nb_class)) # grid_h, grid_w, nb_anchor, 5+nb_class
    true_yolo_3 = Input(shape=(None, None, len(anchors)//6, 4+1+nb_class)) # grid_h, grid_w, nb_anchor, 5+nb_class

    # Layer  0 => 4
    x = _conv_block(input_image, [{'filter': 32, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 0},
                                  {'filter': 64, 'kernel': 3, 'stride': 2, 'bnorm': True, 'leaky': True, 'layer_idx': 1},
                                  {'filter': 32, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 2},
                                  {'filter': 64, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 3}])

    # Layer  5 => 8
    x = _conv_block(x, [{'filter': 128, 'kernel': 3, 'stride': 2, 'bnorm': True, 'leaky': True, 'layer_idx': 5},
                        {'filter':  64, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 6},
                        {'filter': 128, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 7}])

    # Layer  9 => 11
    x = _conv_block(x, [{'filter':  64, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 9},
                        {'filter': 128, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 10}])

    # Layer 12 => 15
    x = _conv_block(x, [{'filter': 256, 'kernel': 3, 'stride': 2, 'bnorm': True, 'leaky': True, 'layer_idx': 12},
                        {'filter': 128, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 13},
                        {'filter': 256, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 14}])

    # Layer 16 => 36
    for i in range(7):
        x = _conv_block(x, [{'filter': 128, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 16+i*3},
                            {'filter': 256, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 17+i*3}])
        
    skip_36 = x
        
    # Layer 37 => 40
    x = _conv_block(x, [{'filter': 512, 'kernel': 3, 'stride': 2, 'bnorm': True, 'leaky': True, 'layer_idx': 37},
                        {'filter': 256, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 38},
                        {'filter': 512, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 39}])

    # Layer 41 => 61
    for i in range(7):
        x = _conv_block(x, [{'filter': 256, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 41+i*3},
                            {'filter': 512, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 42+i*3}])
        
    skip_61 = x
        
    # Layer 62 => 65
    x = _conv_block(x, [{'filter': 1024, 'kernel': 3, 'stride': 2, 'bnorm': True, 'leaky': True, 'layer_idx': 62},
                        {'filter':  512, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 63},
                        {'filter': 1024, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 64}])

    # Layer 66 => 74
    for i in range(3):
        x = _conv_block(x, [{'filter':  512, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 66+i*3},
                            {'filter': 1024, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 67+i*3}])

    # Layer 75 => 79
    x = _conv_block(x, [{'filter':  512, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 75},
                        {'filter': 1024, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 76},
                        {'filter':  512, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 77},
                        {'filter': 1024, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 78},
                        {'filter':  512, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 79}], do_skip=False)

    # Layer 80 => 82


# CHANGED HERE changed the 'filter': (5*(5+nb_class)) from 'filter': (3*(5+nb_class))

    pred_yolo_1 = _conv_block(x, [{'filter': 1024, 'kernel': 3, 'stride': 1, 'bnorm': True,  'leaky': True,  'layer_idx': 80},
                             {'filter': (5*(5+nb_class)), 'kernel': 1, 'stride': 1, 'bnorm': False, 'leaky': False, 'layer_idx': 81}], do_skip=False)


# CHANGED HERE original was anchors[12:], but because I generated 15 anchors in generate_anchors.py I adjusted to [20:]

    loss_yolo_1 = YoloLayer(anchors[20:], 
                            [1*num for num in max_grid], 
                            batch_size, 
                            warmup_batches, 
                            ignore_thresh, 
                            grid_scales[0],
                            obj_scale,
                            noobj_scale,
                            xywh_scale,
                            class_scale)([input_image, pred_yolo_1, true_yolo_1, true_boxes])

    # Layer 83 => 86
    x = _conv_block(x, [{'filter': 256, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 84}], do_skip=False)
    x = UpSampling2D(2)(x)
    x = concatenate([x, skip_61])

    # Layer 87 => 91
    x = _conv_block(x, [{'filter': 256, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 87},
                        {'filter': 512, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 88},
                        {'filter': 256, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 89},
                        {'filter': 512, 'kernel': 3, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 90},
                        {'filter': 256, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True, 'layer_idx': 91}], do_skip=False)

    # Layer 92 => 94

# CHANGED HERE changed the 'filter': (5*(5+nb_class)) from 'filter': (3*(5+nb_class))

    pred_yolo_2 = _conv_block(x, [{'filter': 512, 'kernel': 3, 'stride': 1, 'bnorm': True,  'leaky': True,  'layer_idx': 92},
                             {'filter': (5*(5+nb_class)), 'kernel': 1, 'stride': 1, 'bnorm': False, 'leaky': False, 'layer_idx': 93}], do_skip=False)

# CHANGED HERE original was anchors[6:12], but because I generated 15 anchors in generate_anchors.py I adjusted to [10:20]

    loss_yolo_2 = YoloLayer(anchors[10:20], 
                            [2*num for num in max_grid], 
                            batch_size, 
                            warmup_batches, 
                            ignore_thresh, 
                            grid_scales[1],
                            obj_scale,
                            noobj_scale,
                            xywh_scale,
                            class_scale)([input_image, pred_yolo_2, true_yolo_2, true_boxes])

    # Layer 95 => 98
    x = _conv_block(x, [{'filter': 128, 'kernel': 1, 'stride': 1, 'bnorm': True, 'leaky': True,   'layer_idx': 96}], do_skip=False)
    x = UpSampling2D(2)(x)
    x = concatenate([x, skip_36])

    # Layer 99 => 106

# CHANGED HERE changed the 'filter': (5*(5+nb_class)) from 'filter': (3*(5+nb_class))

    pred_yolo_3 = _conv_block(x, [{'filter': 128, 'kernel': 1, 'stride': 1, 'bnorm': True,  'leaky': True,  'layer_idx': 99},
                             {'filter': 256, 'kernel': 3, 'stride': 1, 'bnorm': True,  'leaky': True,  'layer_idx': 100},
                             {'filter': 128, 'kernel': 1, 'stride': 1, 'bnorm': True,  'leaky': True,  'layer_idx': 101},
                             {'filter': 256, 'kernel': 3, 'stride': 1, 'bnorm': True,  'leaky': True,  'layer_idx': 102},
                             {'filter': 128, 'kernel': 1, 'stride': 1, 'bnorm': True,  'leaky': True,  'layer_idx': 103},
                             {'filter': 256, 'kernel': 3, 'stride': 1, 'bnorm': True,  'leaky': True,  'layer_idx': 104},
                             {'filter': (5*(5+nb_class)), 'kernel': 1, 'stride': 1, 'bnorm': False, 'leaky': False, 'layer_idx': 105}], do_skip=False)


# CHANGED HERE original was anchors[:6], but because I generated 15 anchors in generate_anchors.py I adjusted to [:10]

    loss_yolo_3 = YoloLayer(anchors[:10], 
                            [4*num for num in max_grid], 
                            batch_size, 
                            warmup_batches, 
                            ignore_thresh, 
                            grid_scales[2],
                            obj_scale,
                            noobj_scale,
                            xywh_scale,
                            class_scale)([input_image, pred_yolo_3, true_yolo_3, true_boxes]) 

    train_model = Model([input_image, true_boxes, true_yolo_1, true_yolo_2, true_yolo_3], [loss_yolo_1, loss_yolo_2, loss_yolo_3])
    infer_model = Model(input_image, [pred_yolo_1, pred_yolo_2, pred_yolo_3])

    return [train_model, infer_model]

def dummy_loss(y_true, y_pred):
    return tf.sqrt(tf.reduce_sum(y_pred))

This portion seems to work, but there is an error when I actually start training on epochs. Not sure if it was either something wrong that I did in yolo or generator. It says:

yolo = yolos[max_index//3]
IndexError: list index out of range

Only thing I can think of is possibly changing something in these lines?

# initialize the inputs and the outputs
        yolo_1 = np.zeros((r_bound - l_bound, 1*base_grid_h,  1*base_grid_w, len(self.anchors)//3, 4+1+len(self.labels))) # desired network output 1
        yolo_2 = np.zeros((r_bound - l_bound, 2*base_grid_h,  2*base_grid_w, len(self.anchors)//3, 4+1+len(self.labels))) # desired network output 2
        yolo_3 = np.zeros((r_bound - l_bound, 4*base_grid_h,  4*base_grid_w, len(self.anchors)//3, 4+1+len(self.labels))) # desired network output 3
        yolos = [yolo_3, yolo_2, yolo_1]

I didn't touch generator.

[lexuansanh]

yolo = yolos[max_index//5]
Because max_index = 14 (0:14 idx of Anchor boxs). max_index//3 may be = 4 for yolos --> IndexError: list index out of range.
And you should also more change in Generator class same as yololayer class

Thank you again. It compiles and runs, but now it seems like performance gets worse. This might be an error in my implementation, or maybe more anchors may not be a good idea?

import cv2
import copy
import numpy as np
from keras.utils import Sequence
from utils.bbox import BoundBox, bbox_iou
from utils.image import apply_random_scale_and_crop, random_distort_image, random_flip, correct_bounding_boxes

class BatchGenerator(Sequence):
    def __init__(self, 
        instances, 
        anchors,   
        labels,        
        downsample=32, # ratio between network input's size and network output's size, 32 for YOLOv3
        max_box_per_image=30,
        batch_size=1,
        min_net_size=320,
        max_net_size=320,    
        shuffle=True, 
        jitter=True, 
        norm=None
    ):
        self.instances          = instances
        self.batch_size         = batch_size
        self.labels             = labels
        self.downsample         = downsample
        self.max_box_per_image  = max_box_per_image
        self.min_net_size       = (min_net_size//self.downsample)*self.downsample
        self.max_net_size       = (max_net_size//self.downsample)*self.downsample
        self.shuffle            = shuffle
        self.jitter             = jitter
        self.norm               = norm
        self.anchors            = [BoundBox(0, 0, anchors[2*i], anchors[2*i+1]) for i in range(len(anchors)//2)]
        self.net_h              = 416  
        self.net_w              = 416

        if shuffle: np.random.shuffle(self.instances)
            
    def __len__(self):
        return int(np.ceil(float(len(self.instances))/self.batch_size))           

    def __getitem__(self, idx):
        # get image input size, change every 10 batches
        net_h, net_w = self._get_net_size(idx)
        base_grid_h, base_grid_w = net_h//self.downsample, net_w//self.downsample

        # determine the first and the last indices of the batch
        l_bound = idx*self.batch_size
        r_bound = (idx+1)*self.batch_size

        if r_bound > len(self.instances):
            r_bound = len(self.instances)
            l_bound = r_bound - self.batch_size

        x_batch = np.zeros((r_bound - l_bound, net_h, net_w, 3))             # input images
        t_batch = np.zeros((r_bound - l_bound, 1, 1, 1,  self.max_box_per_image, 4))   # list of groundtruth boxes

        # initialize the inputs and the outputs

# CHANGED HERE Actually didn't, but this may be a casue for an issue. Reasoning behind self.anchoirs/3 is becasue 15(number of anchors I chose) // 3 is 5, which is number of anchors I want per detection layer
        yolo_1 = np.zeros((r_bound - l_bound, 1*base_grid_h,  1*base_grid_w, len(self.anchors)//3, 4+1+len(self.labels))) # desired network output 1
        yolo_2 = np.zeros((r_bound - l_bound, 2*base_grid_h,  2*base_grid_w, len(self.anchors)//3, 4+1+len(self.labels))) # desired network output 2
        yolo_3 = np.zeros((r_bound - l_bound, 4*base_grid_h,  4*base_grid_w, len(self.anchors)//3, 4+1+len(self.labels))) # desired network output 3
        yolos = [yolo_3, yolo_2, yolo_1]

        dummy_yolo_1 = np.zeros((r_bound - l_bound, 1))
        dummy_yolo_2 = np.zeros((r_bound - l_bound, 1))
        dummy_yolo_3 = np.zeros((r_bound - l_bound, 1))
        
        instance_count = 0
        true_box_index = 0

        # do the logic to fill in the inputs and the output
        for train_instance in self.instances[l_bound:r_bound]:
            # augment input image and fix object's position and size
            img, all_objs = self._aug_image(train_instance, net_h, net_w)
            
            for obj in all_objs:
                # find the best anchor box for this object
                max_anchor = None                
                max_index  = -1
                max_iou    = -1

                shifted_box = BoundBox(0, 
                                       0,
                                       obj['xmax']-obj['xmin'],                                                
                                       obj['ymax']-obj['ymin'])    
                
                for i in range(len(self.anchors)):
                    anchor = self.anchors[i]
                    iou    = bbox_iou(shifted_box, anchor)

                    if max_iou < iou:
                        max_anchor = anchor
                        max_index  = i
                        max_iou    = iou                
                
                # determine the yolo to be responsible for this bounding box

# CHANGED HERE Solved out of index error! 
                yolo = yolos[max_index//5]
                grid_h, grid_w = yolo.shape[1:3]
                
                # determine the position of the bounding box on the grid
                center_x = .5*(obj['xmin'] + obj['xmax'])
                center_x = center_x / float(net_w) * grid_w # sigma(t_x) + c_x
                center_y = .5*(obj['ymin'] + obj['ymax'])
                center_y = center_y / float(net_h) * grid_h # sigma(t_y) + c_y
                
                # determine the sizes of the bounding box
                w = np.log((obj['xmax'] - obj['xmin']) / float(max_anchor.xmax)) # t_w
                h = np.log((obj['ymax'] - obj['ymin']) / float(max_anchor.ymax)) # t_h

                box = [center_x, center_y, w, h]

                # determine the index of the label
                obj_indx = self.labels.index(obj['name'])  

                # determine the location of the cell responsible for this object
                grid_x = int(np.floor(center_x))
                grid_y = int(np.floor(center_y))

                # assign ground truth x, y, w, h, confidence and class probs to y_batch

# Am I supposed to change  %3 to %5? I tried both, but performance is still pretty bad for both on racoon dataset.

                yolo[instance_count, grid_y, grid_x, max_index%3]      = 0
                yolo[instance_count, grid_y, grid_x, max_index%3, 0:4] = box
                yolo[instance_count, grid_y, grid_x, max_index%3, 4  ] = 1.
                yolo[instance_count, grid_y, grid_x, max_index%3, 5+obj_indx] = 1

                # assign the true box to t_batch
                true_box = [center_x, center_y, obj['xmax'] - obj['xmin'], obj['ymax'] - obj['ymin']]
                t_batch[instance_count, 0, 0, 0, true_box_index] = true_box

                true_box_index += 1
                true_box_index  = true_box_index % self.max_box_per_image    

            # assign input image to x_batch
            if self.norm != None: 
                x_batch[instance_count] = self.norm(img)
            else:
                # plot image and bounding boxes for sanity check
                for obj in all_objs:
                    cv2.rectangle(img, (obj['xmin'],obj['ymin']), (obj['xmax'],obj['ymax']), (255,0,0), 3)
                    cv2.putText(img, obj['name'], 
                                (obj['xmin']+2, obj['ymin']+12), 
                                0, 1.2e-3 * img.shape[0], 
                                (0,255,0), 2)
                
                x_batch[instance_count] = img

            # increase instance counter in the current batch
            instance_count += 1                 
                
        return [x_batch, t_batch, yolo_1, yolo_2, yolo_3], [dummy_yolo_1, dummy_yolo_2, dummy_yolo_3]

    def _get_net_size(self, idx):
        if idx%10 == 0:
            net_size = self.downsample*np.random.randint(self.min_net_size/self.downsample, \
                                                         self.max_net_size/self.downsample+1)
            # print("resizing: ", net_size, net_size)
            self.net_h, self.net_w = net_size, net_size
        return self.net_h, self.net_w
    
    def _aug_image(self, instance, net_h, net_w):
        image_name = instance['filename']
        image = cv2.imread(image_name) # RGB image
        
        if image is None: print('Cannot find ', image_name)
        image = image[:,:,::-1] # RGB image
            
        image_h, image_w, _ = image.shape
        
        # determine the amount of scaling and cropping
        dw = self.jitter * image_w;
        dh = self.jitter * image_h;

        new_ar = (image_w + np.random.uniform(-dw, dw)) / (image_h + np.random.uniform(-dh, dh));
        # scale = np.random.uniform(0.25, 2);
        scale = 1

        if (new_ar < 1):
            new_h = int(scale * net_h);
            new_w = int(net_h * new_ar);
        else:
            new_w = int(scale * net_w);
            new_h = int(net_w / new_ar);
            
        #dx = int(np.random.uniform(0, net_w - new_w));
        #dy = int(np.random.uniform(0, net_h - new_h));
        dx = 0
        dy = 0
        
        # apply scaling and cropping
        # print(image.shape)
        im_sized = apply_random_scale_and_crop(image, new_w, new_h, net_w, net_h, dx, dy)
        
        # randomly distort hsv space
        # im_sized = random_distort_image(im_sized)
        
        # randomly flip
        # flip = np.random.randint(2)
        flip = 1
        im_sized = random_flip(im_sized, flip)
            
        # correct the size and pos of bounding boxes
        all_objs = correct_bounding_boxes(instance['object'], new_w, new_h, net_w, net_h, dx, dy, flip, image_w, image_h)
        

        return im_sized, all_objs   

    def on_epoch_end(self):
        if self.shuffle: np.random.shuffle(self.instances)
            
    def num_classes(self):
        return len(self.labels)

    def size(self):
        return len(self.instances)    

    def get_anchors(self):
        anchors = []

        for anchor in self.anchors:
            anchors += [anchor.xmax, anchor.ymax]

        return anchors

    def load_annotation(self, i):
        annots = []

        for obj in self.instances[i]['object']:
            annot = [obj['xmin'], obj['ymin'], obj['xmax'], obj['ymax'], self.labels.index(obj['name'])]
            annots += [annot]

        if len(annots) == 0: annots = [[]]

        return np.array(annots)

    def load_image(self, i):
        return cv2.imread(self.instances[i]['filename'])    

In this generator.py, I only changed like 3 lines in total. However, I now get severe overfitting on racoon dataset. For example, training loss = 1 and validation loss = 30. Of course, this may be because racoon dataset is pretty simple, so more anchors may be overkill. Testing it on the plant dataset still results in poor performance, altough interestingly, I did notice that training loss decreases much faster than previously.

[lexuansanh]

1. Adding anchor boxes is not a bad idea. It is shown by your training on raccoon that has converged and has low loss. Adding anchors will increase the calculation time of the computer and can do Out of Memory if the computer memory is not large enough. Specifically, the number of bounding boxes = anchors * (13 * 13 + 26 * 26 + 52 * 52). For 3 anchors = 10647 boxes. For 5 anchors this value is 17745 boxes. Calculations will take longer, non-max suppression also longer.

2. I found valid loss = 30 not a bad number. To accurately evaluate your model, evoluate() should be used to calculate the MAP. If this value is low for both train_generator and valid_generator, the training may fail. However, the AP of train_generator is high and the AP of valid_generator is low, indicating that we have overfitting. In the raccoon set, it's probably low on data and you didn't augment the image via the _aug_image() function.

3. Just 1 data in your data set error (ex: wrong bounding box, or wrongly labeled, ..) is enough for the training to fail.

4. In addition, if your computer is powerful enough and memory is large, increase batch_size to 32, this is very powerful to optimize the loss function with your large training data.

5. Or maybe your valid data doesn't really match the augmented training data. Let's augment from the beginning for all data and then divide the data into train and valid data.

6. edit the code in:
   yolo[instance_count, grid_y, grid_x, max_index%3] = 0
yolo[instance_count, grid_y, grid_x, max_index%3, 0:4] = box
yolo[instance_count, grid_y, grid_x, max_index%3, 4 ] = 1
yolo[instance_count, grid_y, grid_x, max_index%3, 5+obj_indx] = 1
   to:
   yolo[instance_count, grid_y, grid_x, max_index%5] = 0
yolo[instance_count, grid_y, grid_x, max_index%5, 0:4] = box
yolo[instance_count, grid_y, grid_x, max_index%5, 4 ] = 1
yolo[instance_count, grid_y, grid_x, max_index%5, 5+obj_indx] = 1