rocAL - cmn fix

rocAL/docs/examples/image_processing/inference_pipeline.py

@@ -0,0 +1,135 @@
+# Copyright (c) 2018 - 2023 Advanced Micro Devices, Inc. All rights reserved.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in
+# all copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+# THE SOFTWARE.
+
+import sys
+from tkinter import W
+from amd.rocal.pipeline import pipeline_def
+import numpy as np
+import rocal_pybind as b
+from amd.rocal.plugin.pytorch import ROCALClassificationIterator
+import amd.rocal.fn as fn
+import amd.rocal.types as types
+import matplotlib.gridspec as gridspec
+import matplotlib.pyplot as plt
+
+
+seed = 1549361629
+image_dir = "../../../../data/images/AMD-tinyDataSet/"
+batch_size = 4
+gpu_id = 0
+
+def show_images(image_batch, device):
+    columns = 4
+    rows = (batch_size + 1) // (columns)
+    #fig = plt.figure(figsize = (32,(32 // columns) * rows))
+    gs = gridspec.GridSpec(rows, columns)
+    for j in range(rows*columns):
+        #print('\n Display image: ', j)
+        plt.subplot(gs[j])
+        img = image_batch[j]
+        plt.axis("off")
+        if device == "cpu":
+            plt.imshow(img)
+        else:
+            plt.imshow(img.cpu())
+    plt.show()
+
+def show_images(image_batch, image_batch1, device):
+    columns = 4
+    rows = (batch_size + 1) // (columns)
+    #fig = plt.figure(figsize = (32,(32 // columns) * rows))
+    gs = gridspec.GridSpec(rows, columns*2)
+    for j in range(rows*columns):
+        #print('\n Display image: ', j)
+        k = j*2
+        plt.subplot(gs[k])
+        img = image_batch[j]
+        plt.axis("off")
+        if device == "cpu":
+            plt.imshow(img)
+        else:
+            plt.imshow(img.cpu())
+
+        plt.subplot(gs[k+1])
+        img = image_batch1[j]
+        plt.axis("off")
+        if device == "cpu":
+            plt.imshow(img)
+        else:
+            plt.imshow(img.cpu())
+    plt.show()
+
+
+
+def show_pipeline_output(pipe, device):
+    pipe.build()
+    data_loader = ROCALClassificationIterator(pipe, device)
+    images = next(iter(data_loader))
+    show_images(images[0], device)
+
+def show_pipeline_outputs(pipe0, pipe1, device):
+    pipe0.build()
+    pipe1.build()
+    data_loader = ROCALClassificationIterator(pipe0, device)
+    data_loader1 = ROCALClassificationIterator(pipe1, device)
+    images = next(iter(data_loader))
+    images1 = next(iter(data_loader1))
+    show_images(images[0], images1[0], device)
+
+@pipeline_def(seed=seed)
+def inference_pipeline(device="cpu", path=image_dir):
+    jpegs, labels = fn.readers.file(file_root=path, shard_id=0, num_shards=1, random_shuffle=False)
+    images = fn.decoders.image(jpegs, file_root=path, max_decoded_width=1024, max_decoded_height=1024, device="cpu", output_type=types.RGB, shard_id=0, num_shards=1, random_shuffle=False)
+    images_res = fn.resize(images, scaling_mode=types.SCALING_MODE_NOT_SMALLER, interpolation_type=types.TRIANGULAR_INTERPOLATION, resize_shorter=256)
+    return fn.centre_crop(images_res, crop=(224, 224))
+
+@pipeline_def(seed=seed)
+def inference_pipeline_cmn(device="cpu", path=image_dir):
+    jpegs, labels = fn.readers.file(file_root=path, shard_id=0, num_shards=1, random_shuffle=False)
+    images = fn.decoders.image(jpegs, file_root=path, max_decoded_width=1024, max_decoded_height=1024, device="cpu", output_type=types.RGB, shard_id=0, num_shards=1, random_shuffle=False)
+    images_res = fn.resize(images, scaling_mode=types.SCALING_MODE_NOT_SMALLER, interpolation_type=types.TRIANGULAR_INTERPOLATION, resize_shorter=256)
+    return fn.crop_mirror_normalize(images_res , device="cpu",
+                                        output_dtype=types.FLOAT,
+                                        output_layout=types.NHWC,
+                                        crop=(224, 224),
+                                        mirror=0,
+                                        image_type=types.RGB,
+                                        mean=[0,0,0],
+                                        std=[255.0,255.0,255.0])
+
+def main():
+    print ('Optional arguments: <cpu/gpu image_folder>')
+    bs = batch_size
+    rocal_device = "cpu"
+    img_folder = image_dir
+    if  len(sys.argv) > 1:
+      if(sys.argv[1] == "gpu"):
+          rocal_device = "gpu"
+    if  len(sys.argv) > 2:
+      img_folder = sys.argv[2]
+
+    pipe = inference_pipeline(batch_size=bs, num_threads=1, device_id=gpu_id, rocal_cpu=True, tensor_layout=types.NHWC,
+                                  reverse_channels=True, multiplier = [0.00392,0.00392,0.00392], device=rocal_device, path=img_folder)
+    pipe1 = inference_pipeline_cmn(batch_size=bs, num_threads=1, device_id=gpu_id, rocal_cpu=True, tensor_layout=types.NHWC,
+                                  reverse_channels=True, multiplier = [0.00392,0.00392,0.00392], device=rocal_device, path=img_folder)
+    show_pipeline_outputs(pipe, pipe1, device=rocal_device)
+
+if __name__ == '__main__':
+    main()

rocAL/rocAL/include/parameters/parameter_crop.h

@@ -45,7 +45,7 @@ class CropParam
 // V Y directoin
 public:
     CropParam() = delete;
-    CropParam(unsigned int batch_size): batch_size(batch_size), _random(false), _is_center_crop(false)
+    CropParam(unsigned int batch_size): batch_size(batch_size), _random(false), _is_fixed_crop(false)
     {
         x_drift_factor     = default_x_drift_factor();
         y_drift_factor     = default_y_drift_factor();
@@ -58,7 +58,7 @@ class CropParam
         in_height = in_height_;
     }
     void set_random() {_random = true;}
-    void set_center_crop() { _is_center_crop = true; }
+    void set_fixed_crop(float anchor_x, float anchor_y) { _is_fixed_crop = true; _random = false; _crop_anchor[0] = anchor_x; _crop_anchor[1] = anchor_y;}
     void set_x_drift_factor(Parameter<float>* x_drift);
     void set_y_drift_factor(Parameter<float>* y_drift);
     std::vector<uint32_t> in_width, in_height;
@@ -83,7 +83,8 @@ class CropParam
     Parameter<float>* default_x_drift_factor();
     Parameter<float>* default_y_drift_factor();
     std::vector<uint32_t> x1_arr_val, y1_arr_val, croph_arr_val, cropw_arr_val, x2_arr_val, y2_arr_val;
-    bool _random, _is_center_crop;
+    bool _random, _is_fixed_crop;
+    float _crop_anchor [2] = {0.5, 0.5};
     virtual void fill_crop_dims(){};
     void update_crop_array();
 };

rocAL/rocAL/source/api/rocal_api_augmentation.cpp

@@ -475,6 +475,7 @@ rocalResize(
         unsigned resize_shorter,
         unsigned resize_longer,
         RocalResizeInterpolationType interpolation_type) {
+
     Image* output = nullptr;
     if ((p_context == nullptr) || (p_input == nullptr)) {
         ERR("Invalid ROCAL context or invalid input image")
@@ -1747,7 +1748,7 @@ rocalCropFixed(
         ImageInfo output_info = input->info();
         output_info.width(crop_width);
         output_info.height(crop_height);
-        output = context->master_graph->create_image(input->info(), is_output);
+        output = context->master_graph->create_image(output_info, is_output);
         output->reset_image_roi();
         std::shared_ptr<CropNode> crop_node =  context->master_graph->add_node<CropNode>({input}, {output});
         crop_node->init(crop_height, crop_width, crop_pos_x, crop_pos_y);
@@ -1786,7 +1787,7 @@ rocalCropCenterFixed(
         ImageInfo output_info = input->info();
         output_info.width(crop_width);
         output_info.height(crop_height);
-        output = context->master_graph->create_image(input->info(), is_output);
+        output = context->master_graph->create_image(output_info, is_output);
         output->reset_image_roi();
         std::shared_ptr<CropNode> crop_node =  context->master_graph->add_node<CropNode>({input}, {output});
         crop_node->init(crop_height, crop_width);

rocAL/rocAL/source/augmentations/geometry_augmentations/node_crop.cpp

@@ -80,7 +80,7 @@ void CropNode::init(unsigned int crop_h, unsigned int crop_w)
     _crop_param->crop_h = crop_h;
     _crop_param->x1 = 0; 
     _crop_param->y1 = 0;
-    _crop_param->set_center_crop();
+    _crop_param->set_fixed_crop(0.5, 0.5);    // for center_crop
 }
 
 

rocAL/rocAL/source/augmentations/geometry_augmentations/node_crop_mirror_normalize.cpp

@@ -77,12 +77,14 @@ void CropMirrorNormalizeNode::update_node()
     _mirror.update_array();
 }
 
-void CropMirrorNormalizeNode::init(int crop_h, int crop_w, float start_x, float start_y, float mean, float std_dev, IntParam *mirror)
+void CropMirrorNormalizeNode::init(int crop_h, int crop_w, float anchor_x, float anchor_y, float mean, float std_dev, IntParam *mirror)
 {
-    _crop_param->x1 = start_x;
-    _crop_param->y1 = start_y;
+    // current implementation does a fixed crop with specified dims and anchor
+    _crop_param->x1 = 0;
+    _crop_param->y1 = 0;
     _crop_param->crop_h = crop_h;
     _crop_param->crop_w = crop_w;
+    _crop_param->set_fixed_crop(anchor_x, anchor_y);
     _mean = mean;
     _std_dev = std_dev;
     _mirror.set_param(core(mirror));

rocAL/rocAL/source/parameters/parameter_rocal_crop.cpp

@@ -52,9 +52,9 @@ void RocalCropParam::fill_crop_dims() {
             // Evaluating user given crop
             cropw_arr_val[img_idx] = (crop_w <= in_width[img_idx] && crop_w > 0) ? crop_w : in_width[img_idx];
             croph_arr_val[img_idx] = (crop_h <= in_height[img_idx] && crop_h > 0) ? crop_h : in_height[img_idx];
-            if (_is_center_crop) {
-                x1_arr_val[img_idx] = static_cast<size_t>(0.5 * (in_width[img_idx] - cropw_arr_val[img_idx]));
-                y1_arr_val[img_idx] = static_cast<size_t>(0.5 * (in_height[img_idx] - croph_arr_val[img_idx]));
+            if (_is_fixed_crop) {
+                x1_arr_val[img_idx] = static_cast<size_t>(_crop_anchor[0] * (in_width[img_idx] - cropw_arr_val[img_idx]));
+                y1_arr_val[img_idx] = static_cast<size_t>(_crop_anchor[1] * (in_height[img_idx] - croph_arr_val[img_idx]));
             } else {
               x1_arr_val[img_idx] = (x1 >= in_width[img_idx]) ? 0 : x1;
               y1_arr_val[img_idx] = (y1 >= in_height[img_idx]) ? 0 : y1;

rocAL/rocAL_pybind/amd/rocal/fn.py

@@ -424,7 +424,7 @@ def crop_mirror_normalize(*inputs, bytes_per_sample_hint=0, crop=[0, 0], crop_d=
             mirror = b.CreateIntParameter(1)
 
     # pybind call arguments
-    kwargs_pybind = {"input_image0": inputs[0], "crop_depth":crop_depth, "crop_height":crop_height, "crop_width":crop_width, "start_x":0, "start_y":0, "start_z":0, "mean":mean, "std_dev":std,
+    kwargs_pybind = {"input_image0": inputs[0], "crop_depth":crop_depth, "crop_height":crop_height, "crop_width":crop_width, "start_x":crop_pos_x, "start_y":crop_pos_y, "start_z":crop_pos_z, "mean":mean, "std_dev":std,
                      "is_output": False, "mirror": mirror}
     cmn = b.CropMirrorNormalize(Pipeline._current_pipeline._handle ,*(kwargs_pybind.values()))
     Pipeline._current_pipeline._tensor_layout = output_layout