Loading

MABe Task 3: Learning New Behavior

[Task 3] Learning New Behaviour [Baseline]

Baseline notebook for MABe Learning New Behavior Task

ashivani

AIcrowd-Logo

Join the communty!
chat on Discord

🐀🐀🐀🐀🐀🐀🐀🐀🐀🐀🐀🐁🐁🐁🐁🐁🐁🐁🐁🐁🐁
🐀 MABe Learning New Behaviors: Baseline 🐁
🐀🐀🐀🐀🐀🐀🐀🐀🐀🐀🐀🐁🐁🐁🐁🐁🐁🐁🐁🐁🐁

How to use this notebook 📝

  1. Copy the notebook. This is a shared template and any edits you make here will not be saved. You should copy it into your own drive folder. For this, click the "File" menu (top-left), then "Save a Copy in Drive". You can edit your copy however you like.
  2. Link it to your AIcrowd account. In order to submit your predictions to AIcrowd, you need to provide your account's API key.

Setup AIcrowd Utilities 🛠

In [ ]:
!pip install -U aicrowd-cli

Install packages 🗃

Please add all pacakages installations in this section

In [ ]:
!pip install tensorflow-addons

Import necessary modules and packages 📚

In [ ]:
import numpy as np
import os

from tensorflow import keras
import tensorflow as tf
from keras.models import Sequential
import keras.layers as layers
import tensorflow_addons as tfa

import sklearn
import matplotlib.pyplot as plt
import pandas as pd
from sklearn.model_selection import train_test_split
from copy import deepcopy
from tqdm.auto  import tqdm
import gc

Download the dataset 📲

Please get your API key from https://www.aicrowd.com/participants/me

In [ ]:
API_KEY = "8d3a0fe18526544898ce7e12c244d400"
!aicrowd login --api-key $API_KEY
In [ ]:
!aicrowd  dataset download --challenge mabe-task-3-learning-new-behavior

Extract the downloaded dataset to data directory

In [ ]:
!rm -rf data
!mkdir data
 
!mv train.npy data/train.npy
!mv test-release.npy data/test.npy
!mv sample-submission.npy data/sample_submission.npy

Load Data

The dataset files are python dictionaries, this is a descirption of how the data is organized.

In [ ]:
train = np.load('data/train.npy',allow_pickle=True).item()
test = np.load('data/test.npy',allow_pickle=True).item()
sample_submission = np.load('data/sample_submission.npy',allow_pickle=True).item()

Dataset Specifications 💾

  • train.npy - Training set for the task, which follows the following schema:

           

  • test-release.npy - Test set for the task, which follows the following schema :

           

  • sample_submission.npy - Template for a sample submission which follows the following schema
{
    "<sequence_id-1>" : [0, 0, 1, 2, ...],
    "<sequence_id-2>" : [0, 1, 2, 0, ...]
}

Each key in the dictionary here refers to the unique sequence id obtained for the sequences in the test set. The value for each of the keys is expected to hold a list of corresponing annotations. The annotations are represented by the index of the corresponding annotation words in the vocabular provided in the test set.

How does the data look like? 🔍

Task 3 has 7 sets of new behaviors, all binary classifications The test set is combined for all behaviors, you need to output behavior labels for all sequences in test set.

In [ ]:
print("Dataset keys - ", train.keys())
print()
for behavior in train:
  print("Vocabulary - ", train[behavior]['vocabulary'])
  print("Number of train Sequences - ", len(train[behavior]['sequences']))
  print()
print("Number of test Sequences - ", len(test['sequences']))

Submission Format

Test set 458 sequences, so you need to make 7*458 sets of predictions.

In [ ]:
print("Sample Submission keys - ", sample_submission.keys())
for beh in sample_submission:
  print(f"Test videos for {beh}", len(sample_submission[beh]))

Sample overview

In [ ]:
behavior_0 = train['behavior-0']
sequence_names = list(behavior_0["sequences"].keys())
sequence_key = sequence_names[0]
single_sequence = behavior_0["sequences"][sequence_key]
print("Sequence name - ", sequence_key)
print("Single Sequence keys ", single_sequence.keys())
print(f"Number of Frames in {sequence_key} - ", len(single_sequence['annotations']))
print(f"Keypoints data shape of {sequence_key} - ", single_sequence['keypoints'].shape)
print(f"annotator_id of {sequence_key} - ", single_sequence['annotator_id'])

Helper function for visualization 💁

Don't forget to run the cell 😉

In [ ]:
import matplotlib.pyplot as plt
from matplotlib import animation
from matplotlib import colors
from matplotlib import rc
 
rc('animation', html='jshtml')
 
# Note: Image processing may be slow if too many frames are animated.                
 
#Plotting constants
FRAME_WIDTH_TOP = 1024
FRAME_HEIGHT_TOP = 570
 
RESIDENT_COLOR = 'lawngreen'
INTRUDER_COLOR = 'skyblue'
 
PLOT_MOUSE_START_END = [(0, 1), (0, 2), (1, 3), (2, 3), (3, 4),
                        (3, 5), (4, 6), (5, 6), (1, 2)]
 
class_to_color = {'other': 'white', 'behavior-0' : 'red'}
 
class_to_number = {s: i for i, s in enumerate(behavior_0['vocabulary'])}
 
number_to_class = {i: s for i, s in enumerate(behavior_0['vocabulary'])}
 
def num_to_text(anno_list):
  return np.vectorize(number_to_class.get)(anno_list)
 
def set_figax():
    fig = plt.figure(figsize=(6, 4))
 
    img = np.zeros((FRAME_HEIGHT_TOP, FRAME_WIDTH_TOP, 3))
 
    ax = fig.add_subplot(111)
    ax.imshow(img)
 
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
 
    return fig, ax
 
def plot_mouse(ax, pose, color):
    # Draw each keypoint
    for j in range(7):
        ax.plot(pose[j, 0], pose[j, 1], 'o', color=color, markersize=5)
 
    # Draw a line for each point pair to form the shape of the mouse
 
    for pair in PLOT_MOUSE_START_END:
        line_to_plot = pose[pair, :]
        ax.plot(line_to_plot[:, 0], line_to_plot[
                :, 1], color=color, linewidth=1)
 
def animate_pose_sequence(video_name, keypoint_sequence, start_frame = 0, stop_frame = 100, 
                          annotation_sequence = None):
    # Returns the animation of the keypoint sequence between start frame
    # and stop frame. Optionally can display annotations.
    seq = keypoint_sequence.transpose((0,1,3,2))
 
    image_list = []
    
    counter = 0
    for j in range(start_frame, stop_frame):
        if counter%20 == 0:
          print("Processing frame ", j)
        fig, ax = set_figax()
        plot_mouse(ax, seq[j, 0, :, :], color=RESIDENT_COLOR)
        plot_mouse(ax, seq[j, 1, :, :], color=INTRUDER_COLOR)
        
        if annotation_sequence is not None:
          annot = annotation_sequence[j]
          annot = number_to_class[annot]
          plt.text(50, -20, annot, fontsize = 16, 
                   bbox=dict(facecolor=class_to_color[annot], alpha=0.5))
 
        ax.set_title(
            video_name + '\n frame {:03d}.png'.format(j))
 
        ax.axis('off')
        fig.tight_layout(pad=0)
        ax.margins(0)
 
        fig.canvas.draw()
        image_from_plot = np.frombuffer(fig.canvas.tostring_rgb(),
                                        dtype=np.uint8)
        image_from_plot = image_from_plot.reshape(
            fig.canvas.get_width_height()[::-1] + (3,)) 
 
        image_list.append(image_from_plot)
 
        plt.close()
        counter = counter + 1
 
    # Plot animation.
    fig = plt.figure()
    plt.axis('off')
    im = plt.imshow(image_list[0])
 
    def animate(k):
        im.set_array(image_list[k])
        return im,
    ani = animation.FuncAnimation(fig, animate, frames=len(image_list), blit=True)
    return ani
 
def plot_annotation_strip(annotation_sequence, start_frame = 0, stop_frame = 100, title="Behavior Labels"):
  # Plot annotations as a annotation strip.
 
  # Map annotations to a number.
  annotation_num = []
  for item in annotation_sequence[start_frame:stop_frame]:
    annotation_num.append(class_to_number[item])
 
  all_classes = list(set(annotation_sequence[start_frame:stop_frame]))
 
  cmap = colors.ListedColormap(['white', 'red'])
  bounds=[-0.5,0.5,1.5]
  norm = colors.BoundaryNorm(bounds, cmap.N)
 
  height = 200
  arr_to_plot = np.repeat(np.array(annotation_num)[:,np.newaxis].transpose(),
                                                  height, axis = 0)
  
  fig, ax = plt.subplots(figsize = (16, 3))
  ax.imshow(arr_to_plot, interpolation = 'none',cmap=cmap, norm=norm)
 
  ax.set_yticks([])
  ax.set_xlabel('Frame Number')
  plt.title(title)
 
  import matplotlib.patches as mpatches
 
  legend_patches = []
  for item in all_classes:
    legend_patches.append(mpatches.Patch(color=class_to_color[item], label=item))
 
  plt.legend(handles=legend_patches,loc='center left', bbox_to_anchor=(1, 0.5))
 
  plt.tight_layout()

Visualize the mouse movements🎥

Sample visualization for plotting pose gifs.

In [ ]:
keypoint_sequence = single_sequence['keypoints']
annotation_sequence = single_sequence['annotations']

ani = animate_pose_sequence(sequence_key,
                            keypoint_sequence, 
                            start_frame = 3000,
                            stop_frame = 3100,
                            annotation_sequence = annotation_sequence)

# Display the animaion on colab
ani

Showing a section of the validation data (Index needs to be selected for a full video)

In [ ]:
annotation_sequence = single_sequence['annotations']
text_sequence = num_to_text(annotation_sequence)
 
plot_annotation_strip(
    text_sequence,
    start_frame=0,
    stop_frame=len(annotation_sequence) + 1000
)

Basic EDA 🤓

There are 7 new behaviors in task3, each has different frequency of occurence per video.

Each sequence has different amounts of each behavior, here we get the percentage of frames of each behavior in each sequence. We can use this to split the dataset for validation in a stratified way.

In [ ]:
# Function for showing dataframes nicely on jupyter
from IPython.display import display, HTML
def pretty_print_dataframe(df):
  display(HTML(df.to_html()))
In [ ]:
def get_behavior_percentage_frames(behavior_ds): 
  vocabulary = behavior_ds['vocabulary']
  def get_percentage(sequence_key):
    anno_seq = behavior_ds['sequences'][sequence_key]['annotations']
    counts = {k: np.mean(np.array(anno_seq) == v) for k,v in vocabulary.items()}
    return counts

  anno_percentages = {k: get_percentage(k) for k in behavior_ds['sequences']}
  anno_perc_df = pd.DataFrame(anno_percentages).T
  return anno_perc_df

print("Percentage of frames in every sequence for every class")
for behavior in train:
  pretty_print_dataframe( get_behavior_percentage_frames(train[behavior]) )
In [ ]:
percentages = []
for behavior in sorted(list(train.keys())):
  beh_sequences = train[behavior]
  all_annotations = []
  for sk in beh_sequences['sequences']:
    anno = beh_sequences['sequences'][sk]['annotations']
    all_annotations.extend(list(anno))
  percentages.append(np.mean(np.array(all_annotations) == 1))
pd.DataFrame({"Behavior": sorted(list(train.keys())),
              "Percentage Frames": percentages})

Training The Model 🏋️‍♂️

The given MABe dataset contain many sequences of time series data, each frame has its own behavior label. Training on just a single frame does not give good results due to less information.

So here past and future frames are also added to each input. But also all the frames are not concatenated as as the boundaries of the past and future frames need to stay separate for each video.

Seeding helper

Its good practice to seed before every run, that way its easily reproduced.

In [ ]:
def seed_everything(seed):
  np.random.seed(seed)
  os.environ['PYTHONHASHSEED'] = str(seed)
  tf.random.set_seed(seed)

seed=2021
seed_everything(seed)

Generator 🔌

The generator is used to take input winodws from each sequence after randomly sampling frames.

It also provides code for augmentations

  1. Random rotation
  2. Random translate

🚧 Note that these augmentations are applied in the same across all frames in a selected window, e.g - Random rotation by 10 degrees will rotate all frames in the input window by the same angle.

In [ ]:
class MABe_Generator(keras.utils.Sequence):
    def __init__(self, pose_dict, 
                 batch_size, dim, 
                 use_conv, num_classes, augment=False,
                 class_to_number=None,
                 past_frames=0, future_frames=0, 
                 frame_gap=1, shuffle=False,
                 mode='fit'):
        self.batch_size = batch_size
        self.video_keys = list(pose_dict.keys())
        self.dim = dim
        self.use_conv = use_conv
        self.past_frames = past_frames
        self.future_frames = future_frames
        self.frame_gap = frame_gap
        self.shuffle = shuffle
        self.num_classes=num_classes
        self.augment = augment
        self.mode = mode

        self.class_to_number = class_to_number

        self.video_indexes = []
        self.frame_indexes = []
        self.X = {}
        if self.mode == 'fit':
          self.y = []
        self.pad = self.past_frames * self.frame_gap
        future_pad = self.future_frames * self.frame_gap
        pad_width = (self.pad, future_pad), (0, 0), (0, 0), (0, 0)
        self.seq_lengths = {}
        for vc, key in enumerate(self.video_keys):
          if self.mode == 'fit':
            anno = pose_dict[key]['annotations']
            self.y.extend(anno)
          nframes = len(pose_dict[key]['keypoints'])
          self.video_indexes.extend([vc for _ in range(nframes)])
          self.frame_indexes.extend(range(nframes))
          self.X[key] = np.pad(pose_dict[key]['keypoints'], pad_width)
          self.seq_lengths[key] = nframes
        
        if self.mode == 'fit':
          self.y = np.array(self.y)
        
        self.X_dtype = self.X[key].dtype

        self.indexes = list(range(len(self.frame_indexes)))

        if self.mode == 'predict':
          extra_predicts = -len(self.indexes) % self.batch_size # So that last part is not missed
          self.indexes.extend(self.indexes[:extra_predicts])
          self.indexes = np.array(self.indexes)
        
        self.on_epoch_end()

    def __len__(self):
        return len(self.indexes) // self.batch_size

    def augment_fn(self, x):
      # Rotate
      angle = (np.random.rand()-0.5) * (np.pi * 2)
      c, s = np.cos(angle), np.sin(angle)
      rot = np.array([[c, -s], [s, c]])
      x = np.dot(x, rot)

      # Shift - All get shifted together
      shift = (np.random.rand(2)-0.5) * 2 * 0.25
      x = x + shift
      return x

    def __getitem__(self, index):
        bs = self.batch_size
        indexes = self.indexes[index*bs:(index+1)*bs]
        X = np.empty((bs, *self.dim), self.X_dtype)
        if self.mode == 'predict':
          vkey_fi_list = []
        for bi, idx in enumerate(indexes):
          vkey = self.video_keys[self.video_indexes[idx]]
          fi = self.frame_indexes[idx]
          if self.mode == 'predict':
            vkey_fi_list.append((vkey, fi))
          fi = fi + self.pad
          start = fi - self.past_frames*self.frame_gap
          stop = fi + (self.future_frames + 1)*self.frame_gap
          assert start >= 0

          Xi = self.X[vkey][start:stop:self.frame_gap].copy()
          
          if self.augment:
            Xi = self.augment_fn(Xi)
          X[bi] = np.reshape(Xi, self.dim)
          

        if self.mode == 'fit':
          y_vals = self.y[indexes]
          # Converting to one hot because F1 callback needs one hot
          y = np.zeros( (bs,self.num_classes), np.float32)
          y[np.arange(bs), y_vals] = 1
          return X, y

        elif self.mode == 'predict':
          return X, vkey_fi_list

    def on_epoch_end(self):
        if self.shuffle == True:
            np.random.shuffle(self.indexes)

Trainer 🏋️

The trainer class implements a unified interface for using the datagenerator.

It supports fully connected or 1D convolutional networks, as well as other hyperparameters for the model and the generator.

In [ ]:
class Trainer:
  def __init__(self, *,
               train_data,
               val_data,
               test_data,
               feature_dim, 
               batch_size, 
               num_classes,
               augment=False,
               class_to_number=None,
               past_frames=0, 
               future_frames=0,
               frame_gap=1, 
               use_conv=False):
    flat_dim = np.prod(feature_dim)
    if use_conv:
      input_dim = ((past_frames + future_frames + 1), flat_dim,)
    else:
      input_dim = (flat_dim * (past_frames + future_frames + 1),)

    self.input_dim = input_dim
    self.use_conv=use_conv
    self.num_classes=num_classes

    c2n = {'other': 0,'investigation': 1,
                'attack' : 2, 'mount' : 3}
    self.class_to_number = class_to_number or c2n

    self.train_generator = MABe_Generator(train_data, 
                                      batch_size=batch_size, 
                                      dim=input_dim,
                                      num_classes=num_classes, 
                                      past_frames=past_frames, 
                                      future_frames=future_frames,
                                      class_to_number=self.class_to_number,
                                      use_conv=use_conv,
                                      frame_gap=frame_gap,
                                      augment=augment,
                                      shuffle=True,
                                      mode='fit')

    self.val_generator = MABe_Generator(val_data, 
                                        batch_size=batch_size, 
                                        dim=input_dim, 
                                        num_classes=num_classes, 
                                        past_frames=past_frames,
                                        future_frames=future_frames,
                                        use_conv=use_conv,
                                        class_to_number=self.class_to_number,
                                        frame_gap=frame_gap,
                                        augment=False,
                                        shuffle=False,
                                        mode='fit')
    
    self.test_generator = MABe_Generator(test_data, 
                                        batch_size=8192, 
                                        dim=input_dim, 
                                        num_classes=num_classes, 
                                        past_frames=past_frames,
                                        future_frames=future_frames,
                                        use_conv=use_conv,
                                        class_to_number=self.class_to_number,
                                        frame_gap=frame_gap,
                                        augment=False,
                                        shuffle=False,
                                        mode='predict')
  
  def delete_model(self):
    self.model = None
  
  def initialize_model(self, layer_channels=(512, 256), dropout_rate=0., 
                       learning_rate=1e-3, conv_size=5):

    def add_dense_bn_activate(model, out_dim, activation='relu', drop=0.):
      model.add(layers.Dense(out_dim))
      model.add(layers.BatchNormalization())
      model.add(layers.Activation('relu'))
      if drop > 0:
        model.add(layers.Dropout(rate=drop))
      return model
    
    def add_conv_bn_activate(model, out_dim, activation='relu', conv_size=3, drop=0.):
      model.add(layers.Conv1D(out_dim, conv_size))
      model.add(layers.BatchNormalization())
      model.add(layers.Activation('relu'))
      model.add(layers.MaxPooling1D(2, 2))
      if drop > 0:
        model.add(layers.Dropout(rate=drop))
      return model

    model = Sequential()
    model.add(layers.Input(self.input_dim))
    model.add(layers.BatchNormalization())
    for ch in layer_channels:
      if self.use_conv:
        model = add_conv_bn_activate(model, ch, conv_size=conv_size,
                                     drop=dropout_rate)
      else:
        model = add_dense_bn_activate(model, ch, drop=dropout_rate)
    model.add(layers.Flatten())
    model.add(layers.Dense(self.num_classes, activation='softmax'))

    metrics = [tfa.metrics.F1Score(num_classes=self.num_classes)]
    optimizer = keras.optimizers.Adam(lr=learning_rate)
    model.compile(loss='categorical_crossentropy', 
                  optimizer=optimizer,
                  metrics=metrics)

    self.model = model

  def _set_model(self, model):
      """ Set an external, provide initialized and compiled keras model """
      self.model = model

  def train(self, epochs=20, class_weight=None):
    if self.model is None:
      print("Please Call trainer.initialize_model first")
      return
    self.model.fit(self.train_generator,
          validation_data=self.val_generator,
          epochs=epochs,
          class_weight=class_weight)
        
  def get_validation_labels(self, on_test_set=False):
    y_val = []
    for _, y in self.val_generator:
      y_val.extend(list(y))
    y_val = np.argmax(np.array(y_val), axis=-1)
    return y_val

  def get_validation_predictions(self):
    y_val_pred = self.model.predict(self.val_generator)
    y_val_pred = np.argmax(y_val_pred, axis=-1)
    return y_val_pred

  def get_validation_metrics(self):
    y_val = self.get_validation_labels()
    y_val_pred = self.get_validation_predictions()

    f1_scores = sklearn.metrics.f1_score(y_val, y_val_pred,average=None)
    rec_scores = sklearn.metrics.precision_score(y_val, y_val_pred,average=None)
    prec_scores = sklearn.metrics.recall_score(y_val, y_val_pred,average=None)
    classes = list(self.class_to_number.keys())
    metrics = pd.DataFrame({"Class": classes, "F1": f1_scores, "Precision": prec_scores, "Recall": rec_scores})
    return metrics
  
  def get_test_predictions(self):
    all_test_preds = {}
    for vkey in self.test_generator.video_keys:
      nframes = self.test_generator.seq_lengths[vkey]
      all_test_preds[vkey] = np.zeros(nframes, dtype=np.int32)

    for X, vkey_fi_list in tqdm(self.test_generator):
      test_pred = self.model.predict(X)
      test_pred = np.argmax(test_pred, axis=-1)

      for p, (vkey, fi) in zip(test_pred, vkey_fi_list):
        all_test_preds[vkey][fi] = p
    return all_test_preds

Preprocess

We'll normalize the data based on the information that the frame size is 1024x570

The original data is of shape (sequence length, mouse, x y coordinate, keypoint) = (length, 2, 2, 7)

We'll swap the x y and the keypoint axis, which will help in rotation augmentation.

In [ ]:
def normalize_data(orig_pose_dictionary):
  for key in orig_pose_dictionary:
    X = orig_pose_dictionary[key]['keypoints']
    X = X.transpose((0,1,3,2)) #last axis is x, y coordinates
    X[..., 0] = X[..., 0]/1024
    X[..., 1] = X[..., 1]/570
    orig_pose_dictionary[key]['keypoints'] = X
  return orig_pose_dictionary

Dataset split

Since MABe has multiple sequences, it is sensible to split the dataset based on different sequences rather than randomly sampling frames, which may leak information.

About half the sequences don't have "attack" behavior, hence we'll stratify based on whether "attack" behavior is present or absent.

This function only does a single split, but you can also do multiple splits for cross validation.

For Task 2 and 3 there are very few sequences, hence we split the sequences in half for validation.

In [ ]:
def split_validation(orig_pose_dictionary, seed=2021, 
                       test_size=0.5, split_videos=False):
  if split_videos:
    pose_dictionary = {}
    for key in orig_pose_dictionary:
      key_pt1 = key + '_part1'
      key_pt2 = key + '_part2'
      anno_len = len(orig_pose_dictionary[key]['annotations'])
      split_idx = anno_len//2
      pose_dictionary[key_pt1] = {
          'annotations': orig_pose_dictionary[key]['annotations'][:split_idx],
          'keypoints': orig_pose_dictionary[key]['keypoints'][:split_idx]}
      pose_dictionary[key_pt2] = {
          'annotations': orig_pose_dictionary[key]['annotations'][split_idx:],
          'keypoints': orig_pose_dictionary[key]['keypoints'][split_idx:]}
  else:
    pose_dictionary = orig_pose_dictionary
  

  all_anno = []
  for key in pose_dictionary:
    all_anno.extend(pose_dictionary[key]['annotations'])
  keys = np.unique(all_anno)
  def get_percentage(task_key):
    anno_seq = pose_dictionary[task_key]['annotations']
    counts = {k: np.mean(np.array(anno_seq) == k) for k in keys}
    return counts

  anno_percentages = {tk: get_percentage(tk) for tk in pose_dictionary}
  anno_perc_df = pd.DataFrame(anno_percentages).T

  rng_state = np.random.RandomState(seed)
  try:
    idx_train, idx_val = train_test_split(anno_perc_df.index,
                                      stratify=anno_perc_df['attack'] > 0, 
                                      test_size=test_size,
                                      random_state=rng_state)
  except:
    idx_train, idx_val = train_test_split(anno_perc_df.index,
                                      test_size=test_size,
                                      random_state=rng_state)
    
  train_data = {k : pose_dictionary[k] for k in idx_train}
  val_data = {k : pose_dictionary[k] for k in idx_val}
  return train_data, val_data, anno_perc_df

Train function and inference

This below function is specific for Task 3, it has a set of hyperparameters we found with some tuning. Though results can be improved with further tuning.

  • It trains a binary classifier for each new behavior class.

  • It has the option to use a pretrained model from Task1.

  • It also generates the submission dictionary after training is completed.

In [ ]:
def run_task3(results_dir, dataset, test_data, behavior, class_weight, pretrained_file=None):
  HPARAMS = {}
  val_size = HPARAMS["val_size"] = 0.5
  normalize = HPARAMS["normalize"] = True
  HPARAMS["seed"] = seed
  split_videos = HPARAMS["split_videos"] = True

  if normalize:
    dataset = normalize_data(deepcopy(dataset))
    test_data = normalize_data(deepcopy(test_data))

  train_data, val_data, anno_perc_df = split_validation(dataset, 
                                                        seed=seed, 
                                                        test_size=val_size, 
                                                        split_videos=split_videos)                            
  num_classes = len(anno_perc_df.keys())
  feature_dim = HPARAMS["feature_dim"] = (2,7,2)

  # Generator parameters
  past_frames = HPARAMS["past_frames"] = 50
  future_frames = HPARAMS["future_frames"] = 50
  frame_gap = HPARAMS["frame_gap"] = 1
  use_conv = HPARAMS["use_conv"] = True
  batch_size = HPARAMS["batch_size"] = 128

  # Model parameters
  dropout_rate = HPARAMS["dropout_rate"] = 0.5
  learning_rate = HPARAMS["learning_rate"] = 5e-5
  layer_channels = HPARAMS["layer_channels"] = (128, 64, 32)
  conv_size = HPARAMS["conv_size"] = 5
  augment = HPARAMS["augment"] = True
  epochs = HPARAMS["epochs"] = 20
  class_to_number = HPARAMS['class_to_number'] = {"other": 0, behavior: 1}
  HPARAMS['class_weight'] = class_weight

  trainer = Trainer(train_data=train_data,
                    val_data=val_data,
                    test_data=test_data,
                    feature_dim=feature_dim, 
                    batch_size=batch_size, 
                    num_classes=num_classes,
                    augment=augment,
                    class_to_number=class_to_number,
                    past_frames=past_frames, 
                    future_frames=future_frames,
                    frame_gap=frame_gap,
                    use_conv=use_conv)

  trainer.initialize_model(layer_channels=layer_channels,
                          dropout_rate=dropout_rate,
                          learning_rate=learning_rate,
                          conv_size=conv_size)
  
  if pretrained_file and os.path.exists(pretrained_file):
    HPARAMS['pretrained_file'] = pretrained_file
    
    # Load encoder weights Freeze all except top layer
    pretrained_model = keras.models.load_model(pretrained_file)
    for idx, layer in enumerate(pretrained_model.layers[:-1]):
      trainer.model.layers[idx].set_weights(layer.get_weights())
      if not isinstance(layer, layers.BatchNormalization):
        trainer.model.layers[idx].trainable = False
    
    linear_probe_epochs = HPARAMS['linear_probe_epochs'] = 30
    linear_probe_lr = HPARAMS['linear_probe_lr'] = 1e-3
    trainer.model.optimizer.learning_rate.assign(linear_probe_lr)
    trainer.train(epochs=linear_probe_epochs, class_weight=class_weight)

    # Unfreeze layers
    for idx, layer in enumerate(pretrained_model.layers[:-1]):
      trainer.model.layers[idx].trainable = True
    trainer.model.optimizer.learning_rate.assign(learning_rate)
  
  trainer.train(epochs=epochs, class_weight=class_weight)
  trainer.model.save(f'{results_dir}/task3_{behavior}.h5')
  np.save(f"{results_dir}/task3_{behavior}_hparams", HPARAMS)

  val_metrics = trainer.get_validation_metrics()
  test_results = trainer.get_test_predictions()

  np.save(f"{results_dir}/task3_{behavior}_test_results", test_results) 
  val_metrics.to_csv(f"{results_dir}/task3_{behavior}_metrics_val.csv", index=False)
  del trainer # clear ram as the test dataset is large
  del dataset # clear ram as the test dataset is large
  del test_data # clear ram as the test dataset is large
  del train_data, val_data # clear ram as the test dataset is large
  gc.collect() 
  return test_results

Class Weights for loss function

  • Since there is huge imbalance in each behavior class, increasing class weight of behaviors helps.

  • These are set based on the percentage of frames found in the EDA section.

  • There is not clear formula for this, but 1/(fraction of frames) is a good starting point and can be further tuned from there.

In [ ]:
class_weights_task3 = {
    "behavior-0": {0: 1, 1: 20},
    "behavior-1": {0: 1, 1: 50},
    "behavior-2": {0: 1, 1: 5},
    "behavior-3": {0: 1, 1: 3},
    "behavior-4": {0: 1, 1: 100},
    "behavior-5": {0: 1, 1: 20},
    "behavior-6": {0: 1, 1: 10},
}
In [ ]:
from google.colab import drive
drive.mount('/content/drive')
In [ ]:
results_dir = '.'
# Need to take this file from task 1, also make sure all network params and generator parameters are same
pretrained_file = "/content/drive/MyDrive/aicrowd_mabe_models/task1_augmented.h5" 
# pretrained_file = None # If you want to skip the pretrained model part - uncomment this
submission = {}
for behavior in train:
  class_weight = class_weights_task3[behavior]
  behavior_results= run_task3(results_dir, 
                              train[behavior]['sequences'],
                              test['sequences'], 
                              behavior, 
                              class_weight, 
                              pretrained_file)
  submission[behavior] = behavior_results
  gc.collect() # clear RAM as the test dataset is large

Validate the submission ✅

The submssion should follow these constraints:

  1. It should be a dictionary
  2. It should be have same keys as sample_submission
  3. It should have dictionaries for all behaviors
  4. The lengths of the arrays are same
  5. All values are intergers

You can use the helper function below to check these

In [ ]:
def validate_submission(submission, sample_submission):
    if not isinstance(submission, dict):
        print("Submission should be dict")
        return False
    
    if not submission.keys() == sample_submission.keys():
        print("Submission keys don't match")
        return False
    for behavior in submission:
      sb = submission[behavior]
      ssb = sample_submission[behavior]
      if not isinstance(sb, dict):
        print("Submission should be dict")
        return False

      if not sb.keys() == ssb.keys():
        print("Submission keys don't match")
        return False
      
      for key in sb:
        sv = sb[key]
        ssv = ssb[key]
        if not len(sv) == len(ssv):
          print(f"Submission lengths of {key} doesn't match")
          return False
      
      for key, sv in sb.items():
        if not all(isinstance(x, (np.int32, np.int64, int)) for x in list(sv)):
          print(f"Submission of {key} is not all integers")
          return False
    
    print("All tests passed")
    return True
In [ ]:
validate_submission(submission, sample_submission)

Save the prediction as npy 📨

In [ ]:
np.save("submission.npy", submission)

Submit to AIcrowd 🚀

In [ ]:
!aicrowd submission create -c mabe-task-3-learning-new-behavior -f submission.npy

Comments

You must login before you can post a comment.

Execute