yielding each layer information to initialize `layer`
(model, binary)
| 60 | return layers, meta |
| 61 | |
| 62 | def cfg_yielder(model, binary): |
| 63 | """ |
| 64 | yielding each layer information to initialize `layer` |
| 65 | """ |
| 66 | layers, meta = parser(model); yield meta; |
| 67 | h, w, c = meta['inp_size']; l = w * h * c |
| 68 | |
| 69 | # Start yielding |
| 70 | flat = False # flag for 1st dense layer |
| 71 | conv = '.conv.' in model |
| 72 | for i, d in enumerate(layers): |
| 73 | #----------------------------------------------------- |
| 74 | if d['type'] == '[crop]': |
| 75 | yield ['crop', i] |
| 76 | #----------------------------------------------------- |
| 77 | elif d['type'] == '[local]': |
| 78 | n = d.get('filters', 1) |
| 79 | size = d.get('size', 1) |
| 80 | stride = d.get('stride', 1) |
| 81 | pad = d.get('pad', 0) |
| 82 | activation = d.get('activation', 'logistic') |
| 83 | w_ = (w - 1 - (1 - pad) * (size - 1)) // stride + 1 |
| 84 | h_ = (h - 1 - (1 - pad) * (size - 1)) // stride + 1 |
| 85 | yield ['local', i, size, c, n, stride, |
| 86 | pad, w_, h_, activation] |
| 87 | if activation != 'linear': yield [activation, i] |
| 88 | w, h, c = w_, h_, n |
| 89 | l = w * h * c |
| 90 | #----------------------------------------------------- |
| 91 | elif d['type'] == '[convolutional]': |
| 92 | n = d.get('filters', 1) |
| 93 | size = d.get('size', 1) |
| 94 | stride = d.get('stride', 1) |
| 95 | pad = d.get('pad', 0) |
| 96 | padding = d.get('padding', 0) |
| 97 | if pad: padding = size // 2 |
| 98 | activation = d.get('activation', 'logistic') |
| 99 | batch_norm = d.get('batch_normalize', 0) or conv |
| 100 | yield ['convolutional', i, size, c, n, |
| 101 | stride, padding, batch_norm, |
| 102 | activation] |
| 103 | if activation != 'linear': yield [activation, i] |
| 104 | w_ = (w + 2 * padding - size) // stride + 1 |
| 105 | h_ = (h + 2 * padding - size) // stride + 1 |
| 106 | w, h, c = w_, h_, n |
| 107 | l = w * h * c |
| 108 | #----------------------------------------------------- |
| 109 | elif d['type'] == '[maxpool]': |
| 110 | stride = d.get('stride', 1) |
| 111 | size = d.get('size', stride) |
| 112 | padding = d.get('padding', (size-1) // 2) |
| 113 | yield ['maxpool', i, size, stride, padding] |
| 114 | w_ = (w + 2*padding) // d['stride'] |
| 115 | h_ = (h + 2*padding) // d['stride'] |
| 116 | w, h = w_, h_ |
| 117 | l = w * h * c |
| 118 | #----------------------------------------------------- |
| 119 | elif d['type'] == '[avgpool]': |