tensorflow-fpga/doku/layer/conv2d_2.md

2D convolution with activation

TF equivalent

layers.Conv2d(
  filters,
  kernel_size=5,
  strides=(1, 1),
  padding='valid',
  data_format='channels_last',
  dilation_rate=(1, 1),
  activation='relu',
  use_bias=False,
  trainable=True
)

Tensor sizes

Input:

• from TF: [batchSize, imageY, imageX, channels]
• to FPGA: [imageY, imageX, channels]

Filter:

• from TF: [kernelY, kernelX, channels, outputChannels]
• to FPGA: [kernelY, kernelX, channels]

Output:

• from FPGA: [imageY2, imageX2]
• to TF: [batchSize, imageY2, imageX2, outputChannels]

Parallelization

• (batchSize * outputChannels) jobs will be created.
• layer can be called in parallel (functional model)

for sample in range(batchSize):
  for outputChannel in range(outputChannels):
    output[sample][outputChannel] = job(
      input[sample], 
      kernel[outputChannel]
    )

Constraints

  imageX  = imageY  = 228
  kernelX = kernelY = 5
  imageX2 = imageY2 = 224
  channels = 3

Job lengths

• to FPGA: 5 * 5 * 3 + 228 * 228 * 3 = 156027
• from FPGA: 224 * 224 = 50176

Gradient

same as layers.Conv2d: conv2d_backprop in tensorflow/tensorflow/python/ops/nn_ops.py:2290