Skip to content

Commit

Permalink
rocAL - Tf pets training (#947)
Browse files Browse the repository at this point in the history 
* Zen DNN - Docker & Tests (#924)

* Zen DNN - Docker Updates

* Zen DNN - Sample Updates

* Codacy - Fix

* Zen DNN - Cleanup

* Zen DNN - single layer sample

* Rocal Updates (#921)

* rocal updates for tf training

* updates for rocal

* tf updates and pytorch bug fixes

* repo name change

* Update README.md

* dockerfile update

* [rocAL] Fix rocAL Pybind build issue.

* [rocAL] Remove unused function in pipeline.

* [rocAL] Change rocAL pybind installation from setup.py to wheel.
setup.py install is deprecated in python 3.9

* [rocAL] Make TF pets example dataset compatible with tf2.

* [rocAL] Change getImageLabels() compatible with tf.

* [rocAL] Add fix to pick wheel from dist installation folder.
Remove the old installation files in conda environment.

* [rocAL] Remove commented statement.

Co-authored-by: shobana-mcw <shobana@multicorewareinc.com>

* Docker Update (#928)

* turboJPEG version update

* turboJPEG version update

* turboJPEG version update

* TurboJPEG version update

* Update mivisionx-opencl-on-ubuntu20.dockerfile

* Update zenDNN-HIP.dockerfile

* Update level-5.dockerfile

* Update level-5.dockerfile

* Zen DNN Updates Sync

Co-authored-by: Kiriti Gowda <kiriti.nageshgowda@amd.com>

* AMD OpenVX Custom Extension - implementation (#925)

* custom node implementation files

* fix build errors

* custom extension changes for working implementation

* add README and documentation

* update readme

* fix codacy issues and CPU flow

* fix cadacy warning

* Addressed review comments

* minor change

* fix formating

* amd_migraphx - update readme for extension (#929)

* amd_custom - fixes build issue (#935)

* fixes build issue

* Update CMakeLists.txt

* tf_pets_v2

* code_cleanup

* minor code cleanup

* migraphx extension - update the readme(#936)

* vx_amd_migraphx - tests (#923)

* batch size support for migraphx

* changing to accept tensors of all batch sizes

* creates file with results

* bug fix

* changes to singular test cases - mnist and resnet50

* readme updates

* resolving PR comments

* resolving PR comments

* resolving PR comments

* Readme update to reflect tot

* formatting

* fixing typo

* readme update

* readme update

* OpenVX HIP backend - report correct number of CUs for gfx10+ in the logs (#930)

* PyTorch docker file - add argument for specifying version (#938)

* add argument for specifying pytorch version for building docker file

* add readme for pytorch

* rocAL - Fix ROCAL_USE_USER_GIVEN_SIZE_RESTRICTED (#940)

* Docker - codacy fix for pr937 (#942)

* rocAL - Adding NCHW FP16 SIMD kernel (#926)

* Adding NCHW FP16 SIMD kernel for normalization and buffer copy

* Fixed some codestyle issues with FP16 kernel

* Using FMA SSE instruction for multiply-add ops

* Removed extra spaces

* Adding fma flag to rocAL CMakeLists

* Adding FP16 intrinsics for buffer copies

* Setting rounding mode to _MM_FROUND_TO_ZERO

* rocAL  - README updates for video unit test (#939)

* Add README support for video unit test

* Update Readme for video unit test

* Minor test_suite fix

* Update video unit test Readme

* Update Readme

* Updated README with the explation of test cases and arguments

* Minor fix

* Add test case samples to video unit test README

Also add images for README

* Modify sample images

* Change sample image dimension

* Minor README changes

* Minor README changes

* Minor change

* Minor fix to handle relative input path in video unit test

* Resolve codacy warnings

* Minor change

* Add correct video reader outputs

* MIVisionX - cmake cleanup (#943)

* OpenCV EXT - Updates & Tests (#944)

* OpenCV - Readme updates

* Updates - Readme & Tests

* OpenCV - Tests Added

* Updates

* ZenDNN - model compiler (#941)

* model compiler - zendnn - mnist layers

* bug fix + lrn

* layers: batch norm, sum ; bug fixes

* fixes lgtm errors

* bug fixes

* codacy fixes

* bug fix

* codacy fixes

* Update train_withROCAL_withTFRecordReader.py

* Resolved PR comments

Co-authored-by: Kiriti Gowda <kiritigowda@gmail.com>
Co-authored-by: LakshmiKumar23 <lakshmi.kumar@amd.com>
Co-authored-by: shobana-mcw <shobana@multicorewareinc.com>
Co-authored-by: Kiriti Gowda <kiriti.nageshgowda@amd.com>
Co-authored-by: Rajy Rawther <Rajy.MeeyakhanRawther@amd.com>
Co-authored-by: Aryan Salmanpour <aryan.salmanpour@amd.com>
Co-authored-by: root <root@jenkins-worker-rocm-amd-104.local.lan>
Co-authored-by: Sundar Rajan Vaithiyanathan <99159823+SundarRajan28@users.noreply.github.com>
Co-authored-by: Fiona-MCW <70996026+fiona-gladwin@users.noreply.github.com>
  • Loading branch information
@swetha097 @kiritigowda
@LakshmiKumar23 @shobana-mcw @rrawther @AryanSalmanpour @SundarRajan28 @fiona-gladwin
10 people authored Sep 22, 2022
1 parent 1ab79b2 commit c6071b6
Show file tree
Hide file tree
Showing 2 changed files with 35 additions and 14 deletions.
20 changes: 15 additions & 5 deletions rocAL/rocAL_pybind/amd/rocal/plugin/tf.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -49,9 +49,16 @@ def __init__(self, pipeline, tensor_layout = types.NCHW, reverse_channels = Fals
self.loader._name = self.loader._reader
color_format = b.getOutputColorFormat(self.loader._handle)
self.p = (1 if (color_format == int(types.GRAY)) else 3)

self.out = np.zeros(( self.bs*self.n, self.p, int(self.h/self.bs), self.w,), dtype = "uint8")

if self.tensor_dtype == types.FLOAT:
data_type="float32"
elif self.tensor_dtype == types.FLOAT16:
data_type="float16"

if(types.NHWC == self.tensor_format):
self.out = np.zeros(( self.bs*self.n, int(self.h/self.bs), self.w, self.p), dtype = data_type)
else:
self.out = np.zeros(( self.bs*self.n, self.p, int(self.h/self.bs), self.w), dtype = data_type)

def next(self):
return self.__next__()

Expand All @@ -68,8 +75,11 @@ def __next__(self):
if self.loader.run() != 0:
self.reset()
raise StopIteration

self.loader.copyImage(self.out)

if(types.NCHW == self.tensor_format):
self.loader.copyToTensorNCHW(self.out, self.multiplier, self.offset, self.reverse_channels, int(self.tensor_dtype))
else:
self.loader.copyToTensorNHWC(self.out, self.multiplier, self.offset, self.reverse_channels, int(self.tensor_dtype))

if(self.loader._name == "TFRecordReaderDetection"):
self.bbox_list =[]
Expand Down
29 changes: 20 additions & 9 deletions ...rocAL_pybind/example/new_api/tf_petsTrainingExample/train_withROCAL_withTFRecordReader.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,9 +4,10 @@
import amd.rocal.fn as fn
import amd.rocal.types as types

import tensorflow as tf
import tensorflow.compat.v1 as tf
tf.compat.v1.disable_v2_behavior()


import numpy as np
import tensorflow_hub as hub

Expand Down Expand Up @@ -104,7 +105,7 @@ def main():
}


trainPipe = Pipeline(batch_size=TRAIN_BATCH_SIZE, num_threads=1, rocal_cpu=RUN_ON_HOST)
trainPipe = Pipeline(batch_size=TRAIN_BATCH_SIZE, num_threads=1, rocal_cpu=RUN_ON_HOST, tensor_layout = types.NHWC)
with trainPipe:
inputs = fn.readers.tfrecord(path=TRAIN_RECORDS_DIR, index_path = "", reader_type=TFRecordReaderType, user_feature_key_map=featureKeyMap,
features={
Expand All @@ -117,11 +118,17 @@ def main():
images = fn.decoders.image(jpegs, user_feature_key_map=featureKeyMap, output_type=types.RGB, path=TRAIN_RECORDS_DIR)
resized = fn.resize(images, resize_x=crop_size[0], resize_y=crop_size[1])
flip_coin = fn.random.coin_flip(probability=0.5)
cmn_images = fn.crop_mirror_normalize(resized, crop=(crop_size[1], crop_size[0]), mean=[0,0,0], std=[255,255,255], mirror=flip_coin, output_dtype=types.FLOAT, output_layout=types.NCHW, pad_output=False)
cmn_images = fn.crop_mirror_normalize(resized, crop=(crop_size[1], crop_size[0]),
mean=[0,0,0],
std=[255,255,255],
mirror=flip_coin,
output_dtype=types.FLOAT,
output_layout=types.NHWC,
pad_output=False)
trainPipe.set_outputs(cmn_images)
trainPipe.build()

valPipe = Pipeline(batch_size=TRAIN_BATCH_SIZE, num_threads=1, rocal_cpu=RUN_ON_HOST)
valPipe = Pipeline(batch_size=TRAIN_BATCH_SIZE, num_threads=1, rocal_cpu=RUN_ON_HOST, tensor_layout = types.NHWC)
with valPipe:
inputs = fn.readers.tfrecord(path=VAL_RECORDS_DIR, index_path = "", reader_type=TFRecordReaderType, user_feature_key_map=featureKeyMap,
features={
Expand All @@ -134,7 +141,13 @@ def main():
images = fn.decoders.image(jpegs, user_feature_key_map=featureKeyMap, output_type=types.RGB, path=VAL_RECORDS_DIR)
resized = fn.resize(images, resize_x=crop_size[0], resize_y=crop_size[1])
flip_coin = fn.random.coin_flip(probability=0.5)
cmn_images = fn.crop_mirror_normalize(resized, crop=(crop_size[1], crop_size[0]), mean=[0,0,0], std=[255,255,255], mirror=flip_coin, output_dtype=types.FLOAT, output_layout=types.NCHW, pad_output=False)
cmn_images = fn.crop_mirror_normalize(resized, crop=(crop_size[1], crop_size[0]),
mean=[0,0,0],
std=[255,255,255],
mirror=flip_coin,
output_dtype=types.FLOAT,
output_layout=types.NHWC,
pad_output=False)
valPipe.set_outputs(cmn_images)
valPipe.build()

Expand All @@ -148,23 +161,21 @@ def main():
while i < NUM_TRAIN_STEPS:

for t, (train_image_ndArray, train_label_ndArray) in enumerate(trainIterator, 0):
train_image_ndArray_transposed = np.transpose(train_image_ndArray, [0, 2, 3, 1])
train_label_one_hot_list = get_label_one_hot(train_label_ndArray)
train_loss, _, train_accuracy = sess.run(
[cross_entropy_mean, train_op, accuracy],
feed_dict={decoded_images: train_image_ndArray_transposed, labels: train_label_one_hot_list})
feed_dict={decoded_images: train_image_ndArray, labels: train_label_one_hot_list})
print ("Step :: %s\tTrain Loss :: %.2f\tTrain Accuracy :: %.2f%%\t" % (i, train_loss, (train_accuracy * 100)))
is_final_step = (i == (NUM_TRAIN_STEPS - 1))
if i % EVAL_EVERY == 0 or is_final_step:
mean_acc = 0
mean_loss = 0
print("\n\n-------------------------------------------------------------------------------- BEGIN VALIDATION --------------------------------------------------------------------------------")
for j, (val_image_ndArray, val_label_ndArray) in enumerate(valIterator, 0):
val_image_ndArray_transposed = np.transpose(val_image_ndArray, [0, 2, 3, 1])
val_label_one_hot_list = get_label_one_hot(val_label_ndArray)
val_loss, val_accuracy, val_prediction, val_target, correct_predicate = sess.run(
[cross_entropy_mean, accuracy, prediction, correct_label, correct_prediction],
feed_dict={decoded_images: val_image_ndArray_transposed, labels: val_label_one_hot_list})
feed_dict={decoded_images: val_image_ndArray, labels: val_label_one_hot_list})
mean_acc += val_accuracy
mean_loss += val_loss
num_correct_predicate = 0
Expand Down

0 comments on commit c6071b6

Please sign in to comment.