Update yad2k script to Keras 2 API.

README.md

@@ -19,8 +19,10 @@ Original paper: [YOLO9000: Better, Faster, Stronger](https://arxiv.org/abs/1612.
 - [Keras](https://github.com/fchollet/keras)
 - [Tensorflow](https://www.tensorflow.org/)
 - [Numpy](http://www.numpy.org/)
+- [h5py](http://www.h5py.org/) (For Keras model serialization.)
 - [Pillow](https://pillow.readthedocs.io/) (For rendering test results.)
 - [Python 3](https://www.python.org/)
+- [pydot-ng](https://github.com/pydot/pydot-ng) (Optional for plotting model.)
 
 ### Installation
 ```bash
@@ -31,21 +33,21 @@ cd yad2k
 conda env create -f environment.yml
 source activate yad2k
 # [Option 2] Install everything globaly.
-pip install numpy
-
+pip install numpy h5py pillow
 pip install tensorflow-gpu  # CPU-only: conda install -c conda-forge tensorflow
 pip install keras # Possibly older release: conda install keras
 ```
 
 ## Quick Start
 
-- Download Darknet model weights from the [official YOLO website](http://pjreddie.com/darknet/yolo/).
+- Download Darknet model cfg and weights from the [official YOLO website](http://pjreddie.com/darknet/yolo/).
 - Convert the Darknet YOLO_v2 model to a Keras model.
 - Test the converted model on the small test set in `images/`.
 
 ```bash
 wget http://pjreddie.com/media/files/yolo.weights
-./yad2k.py cfg/yolo.cfg yolo.weights model_data/yolo.h5
+wget https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolo.cfg
+./yad2k.py yolo.cfg yolo.weights model_data/yolo.h5
 ./test_yolo.py model_data/yolo.h5  # output in images/out/
 ```
 
@@ -65,10 +67,11 @@ YAD2K assumes the Keras backend is Tensorflow. In particular for YOLO_v2 models
 
 `yad2k/models` contains reference implementations of Darknet-19 and YOLO_v2.
 
+`train_overfit` is a sample training script that overfits a YOLO_v2 model to a single image from the Pascal VOC dataset.
+
 ## Known Issues and TODOs
 
-- Add YOLO_v2 loss function. (In-progress implementation on branch allanzelener/initial_loss_implementation).
-- Script to train YOLO_v2 reference model.
+- Expand sample training script to train YOLO_v2 reference model on full dataset.
 - Support for additional Darknet layer types.
 - Tuck away the Tensorflow dependencies with Keras wrappers where possible.
 - YOLO_v2 model does not support fully convolutional mode. Current implementation assumes 1:1 aspect ratio images.

yad2k.py

@@ -13,9 +13,10 @@
 
 import numpy as np
 from keras import backend as K
-from keras.layers import (Convolution2D, GlobalAveragePooling2D, Input, Lambda,
-                          MaxPooling2D, merge)
+from keras.layers import (Conv2D, GlobalAveragePooling2D, Input, Lambda,
+                          MaxPooling2D)
 from keras.layers.advanced_activations import LeakyReLU
+from keras.layers.merge import concatenate
 from keras.layers.normalization import BatchNormalization
 from keras.models import Model
 from keras.regularizers import l2
@@ -111,8 +112,8 @@ def _main(args):
             activation = cfg_parser[section]['activation']
             batch_normalize = 'batch_normalize' in cfg_parser[section]
 
-            # border_mode='same' is equivalent to Darknet pad=1
-            border_mode = 'same' if pad == 1 else 'valid'
+            # padding='same' is equivalent to Darknet pad=1
+            padding = 'same' if pad == 1 else 'valid'
 
             # Setting weights.
             # Darknet serializes convolutional weights as:
@@ -175,16 +176,14 @@ def _main(args):
                         activation, section))
 
             # Create Conv2D layer
-            conv_layer = (Convolution2D(
-                filters,
-                size,
-                size,
-                border_mode=border_mode,
-                subsample=(stride, stride),
-                bias=not batch_normalize,
+            conv_layer = (Conv2D(
+                filters, (size, size),
+                strides=(stride, stride),
+                kernel_regularizer=l2(weight_decay),
+                use_bias=not batch_normalize,
                 weights=conv_weights,
                 activation=act_fn,
-                W_regularizer=l2(weight_decay)))(prev_layer)
+                padding=padding))(prev_layer)
 
             if batch_normalize:
                 conv_layer = (BatchNormalization(
@@ -203,9 +202,9 @@ def _main(args):
             stride = int(cfg_parser[section]['stride'])
             all_layers.append(
                 MaxPooling2D(
+                    padding='same',
                     pool_size=(size, size),
-                    strides=(stride, stride),
-                    border_mode='same')(prev_layer))
+                    strides=(stride, stride))(prev_layer))
             prev_layer = all_layers[-1]
 
         elif section.startswith('avgpool'):
@@ -218,10 +217,10 @@ def _main(args):
             ids = [int(i) for i in cfg_parser[section]['layers'].split(',')]
             layers = [all_layers[i] for i in ids]
             if len(layers) > 1:
-                print('Merging layers:', layers)
-                merge_layer = merge(layers, mode='concat')
-                all_layers.append(merge_layer)
-                prev_layer = merge_layer
+                print('Concatenating route layers:', layers)
+                concatenate_layer = concatenate(layers)
+                all_layers.append(concatenate_layer)
+                prev_layer = concatenate_layer
             else:
                 skip_layer = layers[0]  # only one layer to route
                 all_layers.append(skip_layer)
@@ -250,7 +249,7 @@ def _main(args):
                 'Unsupported section header type: {}'.format(section))
 
     # Create and save model.
-    model = Model(input=all_layers[0], output=all_layers[-1])
+    model = Model(inputs=all_layers[0], outputs=all_layers[-1])
     print(model.summary())
     model.save('{}'.format(output_path))
     print('Saved Keras model to {}'.format(output_path))