Keras: Model fit_generator not pulling data samples as expected

[BoltzmannBrain]

I have a generator to pass video data frame-by-frame to a Sequential model. To train the model I'm using fit_generator described here:

  batchSize = 200
  print "Starting training..."
  model.fit_generator(
    _frameGenerator(videoPath, dataPath, batchSize),
    samples_per_epoch=5000,
    nb_epoch=2,
    callbacks=[PrintBatch()],
    verbose=args.verbosity
  )

(I've added callbacks to print log info.)

Below is my generator, that will accumulate 200 frames worth of data into X and Y -- i.e., batch size = 200.

import cv2
...
def _frameGenerator(videoPath, dataPath, batchSize):
  """
  Yield X and Y data when the batch is filled.
  """
  camera = cv2.VideoCapture(videoPath)
  width = camera.get(3)
  height = camera.get(4)
  frameCount = int(camera.get(7))  # Number of frames in the video file.

  truthData = _prepData(dataPath, frameCount)

  X = np.zeros((batchSize, 3, height, width))
  Y = np.zeros((batchSize, 1))

  batch = 0
  for frameIdx, truth in enumerate(truthData):
    ret, frame = camera.read()
    if ret is False: continue

    batchIndex = frameIdx%batchSize

    X[batchIndex] = frame
    Y[batchIndex] = truth

    if batchIndex == 0 and frameIdx != 0:
      batch += 1
      print "now yielding batch", batch
      yield X, Y

With this setup I expect 25 batches (of 200 frames each) to be passed from the generator to fit_generator, per epoch; this would be 5000 total frames per epoch -- i.e., samples_per_epoch=5000. Then for subsequent epochs, fit_generator would reinitialize the generator such that we begin training again from the start of the video. Yet this is not the case. The printed output below shows a couple oddities:

• 11 batches (2200 data samples) are passed before the model logs its first batch.
• After the first epoch is complete (after the model logs batches 0-24), the generator picks up where it left off. Shouldn't the new epoch start again from the beginning of the training dataset?

Starting training...
Epoch 1/2
now yielding batch 1
now yielding batch 2
now yielding batch 3
now yielding batch 4
now yielding batch 5
now yielding batch 6
now yielding batch 7
now yielding batch 8
now yielding batch 9
now yielding batch 10
now yielding batch 11
{'loss': 0.17430359, 'batch': 0, 'size': 200}
now yielding batch 12
{'loss': 3737.8875, 'batch': 1, 'size': 200}
now yielding batch 13
{'loss': 57103.0, 'batch': 2, 'size': 200}
now yielding batch 14
{'loss': 0.75166047, 'batch': 3, 'size': 200}
now yielding batch 15
{'loss': 0.74765885, 'batch': 4, 'size': 200}
now yielding batch 16
{'loss': 0.65361518, 'batch': 5, 'size': 200}
now yielding batch 17
{'loss': 2213.9226, 'batch': 6, 'size': 200}
now yielding batch 18
{'loss': 1.0152926, 'batch': 7, 'size': 200}
now yielding batch 19
{'loss': 1.0978817, 'batch': 8, 'size': 200}
now yielding batch 20
{'loss': 1.2816809, 'batch': 9, 'size': 200}
now yielding batch 21
{'loss': 1.0457447, 'batch': 10, 'size': 200}
now yielding batch 22
{'loss': 0.81737024, 'batch': 11, 'size': 200}
now yielding batch 23
{'loss': 0.76998961, 'batch': 12, 'size': 200}
now yielding batch 24
{'loss': 1.7846264, 'batch': 13, 'size': 200}
now yielding batch 25
{'loss': 2.209909, 'batch': 14, 'size': 200}
now yielding batch 26
{'loss': 2.4140291, 'batch': 15, 'size': 200}
now yielding batch 27
{'loss': 2.1680498, 'batch': 16, 'size': 200}
now yielding batch 28
{'loss': 2.0289741, 'batch': 17, 'size': 200}
now yielding batch 29
{'loss': 1.3026924, 'batch': 18, 'size': 200}
now yielding batch 30
{'loss': 0.94716966, 'batch': 19, 'size': 200}
now yielding batch 31
{'loss': 0.87124979, 'batch': 20, 'size': 200}
now yielding batch 32
{'loss': 0.91141301, 'batch': 21, 'size': 200}
now yielding batch 33
{'loss': 0.97045374, 'batch': 22, 'size': 200}
now yielding batch 34
{'loss': 0.81511378, 'batch': 23, 'size': 200}
now yielding batch 35
{'loss': 0.67119628, 'batch': 24, 'size': 200}
526s - loss: 2523.2104
Epoch 2/2
now yielding batch 36
{'loss': 0.54455316, 'batch': 0, 'size': 200}
now yielding batch 37
{'loss': 0.48123521, 'batch': 1, 'size': 200}
now yielding batch 38
{'loss': 0.41967782, 'batch': 2, 'size': 200}
now yielding batch 39
{'loss': 0.41290641, 'batch': 3, 'size': 200}
now yielding batch 40
{'loss': 0.40834817, 'batch': 4, 'size': 200}
now yielding batch 41
{'loss': 0.41192448, 'batch': 5, 'size': 200}
now yielding batch 42
{'loss': 0.33654153, 'batch': 6, 'size': 200}
now yielding batch 43
{'loss': 0.39498168, 'batch': 7, 'size': 200}

If there is something incorrect in my understanding of fit_generator please explain. I've gone through the documentation, this example, and these related issues. When I get my project running I'll be more than happy to include this as an example for running video data 😄

I'm using Keras v1.0.7 with the TensorFlow backend.

[BoltzmannBrain]

I've also posted this on SO.

[BoltzmannBrain]

As a temporary fix I'm manually iterating over the epochs and calling model.fit(), as shown in #107.

[almathie]

The data generator and the fitting are run in parallel. It's important to note that the generator is just that : a generator that has no idea how the data it generates is going to be used and at what epoch. It should just keep generating data forever as needed.

What happens in your log is :

1. The data generator generates 10 batch of data. One batch has 200 samples. 10 batch are generated because 10 is the default value for the max_q_size param of the fit_generator function
2. The model starts fitting the first batch (batch 0). No output there.
3. The data generator needs to replace the first batch that is being trained on. Generates the 11th batch, and outputs the log.
4. The model finishes fitting the first batch and starts fitting the second one. New batch is generated by the generator to keep the queue full.
   5 The keeps on until 25 batches are done being seen by the model. Batch 24 is the last one. At that point the model has seen 25 * 200 samples = 5000 samples which is what you defined as "one epoch". You then get the standard "epoch done" status with the time needed and the loss.
   6 Second epoch starts on the next available batch from the queue. As soon as this batch is being used a new one is generated by the generator as usual. And so on. This will keep working like that until nb_epoch epoch are done.

Everything seems to be working as intended. I suggest to close the issue.

[juliohm]

I am having a hard time trying to decipher the hidden conventions in Keras. This is basically a problem with the Python programming language that doesn't distinguish row, column vectors, and that doesn't restrict dimensions very well. Some functions in Keras expect lists, some expect Numpy arrays, this inconsistency is quite hard to grasp.

Suppose I have a feature vector of size n (a time series) and a response of size m (another time series), what the generator is expected to output?