keras.backend.resize_images

Author: Oceania2018

src/TensorFlowNET.Core/APIs/tf.image.cs

@@ -60,7 +60,11 @@ public Tensor resize_images(Tensor images, Tensor size, string method = ResizeMe
 
             public Tensor resize_images_v2(Tensor images, TensorShape size, string method = ResizeMethod.BILINEAR, bool preserve_aspect_ratio = false, bool antialias = false,
                 string name = null)
-                => image_ops_impl.resize_images(images, tf.constant(size.dims), method, preserve_aspect_ratio, antialias, name);
+                => image_ops_impl.resize_images_v2(images, size, method, preserve_aspect_ratio, antialias, name);
+
+            public Tensor resize_images_v2(Tensor images, Tensor size, string method = ResizeMethod.BILINEAR, bool preserve_aspect_ratio = false, bool antialias = false,
+                string name = null)
+                => image_ops_impl.resize_images_v2(images, size, method, preserve_aspect_ratio, antialias, name);
 
             public Tensor resize_images_with_pad(Tensor image, int target_height, int target_width, string method, bool antialias)
                 => image_ops_impl.resize_images_with_pad(image, target_height, target_width, method, antialias);
@@ -209,6 +213,9 @@ public Tensor extract_glimpse(Tensor input, Tensor size, Tensor offsets, bool ce
             public Tensor resize(Tensor image, TensorShape size, string method = ResizeMethod.BILINEAR)
                 => image_ops_impl.resize_images_v2(image, size, method: method);
 
+            public Tensor resize(Tensor image, Tensor size, string method = ResizeMethod.BILINEAR)
+                => image_ops_impl.resize_images_v2(image, size, method: method);
+
             public Tensor resize_bilinear(Tensor images, Tensor size, bool align_corners = false, bool half_pixel_centers = false, string name = null)
                 => gen_image_ops.resize_bilinear(images, size, align_corners: align_corners, half_pixel_centers: half_pixel_centers, name: name);
 

src/TensorFlowNET.Core/Operations/gen_array_ops.cs

@@ -175,18 +175,32 @@ public static Tensor pad(Tensor input, Tensor paddings, string name = null)
         {
             if (tf.Context.executing_eagerly())
             {
-                var results = tf.Runner.TFE_FastPathExecute(tf.Context, tf.Context.DeviceName,
+                /*var results = tf.Runner.TFE_FastPathExecute(tf.Context, tf.Context.DeviceName,
                     "Pad", name,
                     null,
                     input, paddings);
-                return results[0];
+                return results[0];*/
+                return pad_eager_fallback(input, paddings, name: name, ctx: tf.Context);
             }
 
             var _op = tf.OpDefLib._apply_op_helper("Pad", name: name, args: new { input, paddings });
 
             return _op.output;
         }
 
+        private static Tensor pad_eager_fallback(Tensor inputs, Tensor padding, string name = null, Context ctx = null)
+        {
+            var (_attr_T, input) = tf.Runner.ArgsToMatchingEager(ctx, args: new[] { inputs });
+            var (_attr_Tpaddings, paddings) = tf.Runner.ArgsToMatchingEager(ctx, default_dtype: tf.int32, args: new[] { padding });
+            var _inputs_flat = input.concat(paddings);
+            var _attrs = new object[] { "T", _attr_T, "Tpaddings", _attr_Tpaddings };
+
+            var results = tf.Runner.Execute(ctx, "Pad", 1, _inputs_flat, _attrs, name: name);
+            if (tf.Runner.MustRecordGradient())
+                tf.Runner.RecordGradient("Pad", _inputs_flat, _attrs, results);
+            return results[0];
+        }
+
         public static Tensor pack(Tensor[] values, int axis = 0, string name = null)
         {
             if (tf.Context.executing_eagerly())

src/TensorFlowNET.Core/Operations/image_ops_impl.cs

@@ -2210,7 +2210,7 @@ public static Tensor convert_image_dtype(Tensor image, TF_DataType dtype, bool s
         /// <param name="antialias"></param>
         /// <param name="name"></param>
         /// <returns></returns>
-        public static Tensor resize_images_v2(Tensor images, TensorShape size, string method = ResizeMethod.BILINEAR,
+        public static Tensor resize_images_v2<T>(Tensor images, T size, string method = ResizeMethod.BILINEAR,
             bool preserve_aspect_ratio = false,
             bool antialias = false,
             string name = null)

src/TensorFlowNET.Keras/BackendImpl.cs

@@ -183,5 +183,48 @@ public Tensor categorical_crossentropy(Tensor target, Tensor output, bool from_l
 
             throw new NotImplementedException("");
         }
+
+        /// <summary>
+        /// Resizes the images contained in a 4D tensor.
+        /// </summary>
+        /// <param name="x"></param>
+        /// <param name="height_factor"></param>
+        /// <param name="width_factor"></param>
+        /// <param name="data_format"></param>
+        /// <param name="interpolation"></param>
+        /// <returns></returns>
+        public Tensor resize_images(Tensor x, int height_factor, int width_factor, 
+            string data_format, string interpolation = "nearest")
+        {
+            var (rows, cols) = (0, 0);
+            if (data_format == "channels_first")
+                (rows, cols) = (2, 3);
+            else if (data_format == "channels_last")
+                (rows, cols) = (1, 2);
+            else
+                throw new ValueError($"Invalid `data_format` argument: {data_format}");
+
+            var original_shape = x.shape;
+            var new_shape = array_ops.shape(x)[new Slice(rows, cols + 1)];
+            new_shape *= constant_op.constant(np.array(height_factor, width_factor));
+
+            if (data_format == "channels_first")
+                // x = permute_dimensions(x, [0, 2, 3, 1]);
+                throw new NotImplementedException("");
+            if (interpolation == "nearest")
+                x = tf.image.resize_images_v2(x, new_shape, method: ResizeMethod.NEAREST_NEIGHBOR);
+
+            if (data_format == "channels_first")
+                // x = permute_dimensions(x, [0, 3, 1, 2]);
+                throw new NotImplementedException("");
+
+            int new_height = original_shape[rows] < 0 ? -1 : original_shape[rows] * height_factor;
+            int new_width = original_shape[cols] < 0 ? -1 : original_shape[cols] * width_factor;
+
+            TensorShape output_shape = data_format == "channels_first" ?
+                (-1, -1, new_height, new_width) : (-1, new_height, new_width, -1);
+            x.set_shape(output_shape);
+            return x;
+        }
     }
 }