TensorFlow.NET/src/TensorFlowNET.Core/Gradients/nn_grad.cs at master · MSavameri/TensorFlow.NET · GitHub

Repository navigation

History

464 lines (417 loc) · 17.9 KB

Raw

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

462

463

464

/*****************************************************************************

Licensed under the Apache License, Version 2.0 (the "License");

you may not use this file except in compliance with the License.

You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software

distributed under the License is distributed on an "AS IS" BASIS,

WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

See the License for the specific language governing permissions and

limitations under the License.

******************************************************************************/

using System;

using System.Diagnostics;

using System.Linq;

using Tensorflow.Operations;

using static Tensorflow.Binding;

namespace Tensorflow.Gradients

{

/// <summary>

///

/// </summary>

[RegisterGradient("math_grad")]

public class nn_grad

{

/// <summary>

/// Return the gradients for the 2 inputs of bias_op.

/// </summary>

/// <param name="op"></param>

/// <param name="grads"></param>

/// <returns></returns>

[RegisterGradient("BiasAdd")]

public static Tensor[] _BiasAddGrad(Operation op, Tensor[] grads)

{

var grad = grads[0];

string data_format = op.get_attr("data_format")?.ToString();

var bias_add_grad = gen_nn_ops.bias_add_grad(out_backprop: grad, data_format: data_format);

return new Tensor[] { grad, bias_add_grad };

}

[RegisterGradient("Relu")]

public static Tensor[] _ReluGrad(Operation op, Tensor[] grads)

{

return new Tensor[] { gen_nn_ops.relu_grad(grads[0], op.outputs[0]) };

}

[RegisterGradient("LeakyRelu")]

public static Tensor[] _LeakyReluGrad(Operation op, Tensor[] grads)

{

var grad = grads[0];

var x = op.inputs[0];

var alpha = (float)op.get_attr("alpha");

return new Tensor[] { gen_nn_ops.leaky_relu_grad(grad, x, alpha: alpha) };

}

/// <summary>

/// The derivative of the softmax nonlinearity.

/// </summary>

/// <param name="op"></param>

/// <param name="grads"></param>

/// <returns></returns>

[RegisterGradient("Softmax")]

public static Tensor[] _SoftmaxGrad(Operation op, Tensor[] grads)

{

var grad_softmax = grads[0];

var softmax = op.outputs[0];

var mul = grad_softmax * softmax;

var sum_channels = math_ops.reduce_sum(mul, axis: constant_op.constant(-1), keepdims: true);

var sub = grad_softmax - sum_channels;

return new Tensor[] { sub * softmax };

}

/// <summary>

/// Gradient function for SoftmaxCrossEntropyWithLogits.

/// </summary>

/// <param name="op"></param>

/// <param name="grads"></param>

/// <returns></returns>

[RegisterGradient("SoftmaxCrossEntropyWithLogits")]

public static Tensor[] _SoftmaxCrossEntropyWithLogitsGrad(Operation op, Tensor[] grads)

{

var grad_loss = grads[0];

var grad_grad = grads[1];

var softmax_grad = op.outputs[1];

var grad = _BroadcastMul(grad_loss, softmax_grad);

var logits = op.inputs[0];

if (grad_grad != null && !IsZero(grad_grad))

{

throw new NotImplementedException("_SoftmaxCrossEntropyWithLogitsGrad");

}

return new Tensor[]

{

grad,

_BroadcastMul(grad_loss, -nn_ops.log_softmax(logits))

};

}

[RegisterGradient("SparseSoftmaxCrossEntropyWithLogits")]

public static Tensor[] _SparseSoftmaxCrossEntropyWithLogitsGrad(Operation op, Tensor[] grads)

{

var sparse_softmax_grad_without_gradient = array_ops.prevent_gradient(

op.outputs[1],

message: "Currently there is no way to take the second " +

"derivative of sparse_softmax_cross_entropy_with_logits due to the fused " +

"implementation's interaction with tf.gradients()");

var grad_0 = grads[0];

return new Tensor[]

{

_BroadcastMul(grad_0, sparse_softmax_grad_without_gradient),

null

};

}

[RegisterGradient("Softplus")]

public static Tensor[] _SoftplusGrad(Operation op, Tensor[] grads)

{

var grad = grads[0];

var x = op.inputs[0];

var softplus = grad * math_ops.sigmoid(x);

return new Tensor[] { softplus };

}

[RegisterGradient("SquaredDifference")]

public static Tensor[] _SquaredDifferenceGrad(Operation op, Tensor[] grads)

{

Tensor x = op.inputs[0];

Tensor y = op.inputs[1];

var grad = grads[0];

var scale = ops.convert_to_tensor(2.0f, dtype: x.dtype);

var x_grad = math_ops.scalar_mul(scale, grad) * (x - y);

if (math_grad._ShapesFullySpecifiedAndEqual(x, y, grad))

{

return new Tensor[] { x_grad, -x_grad };

}

var broadcast_info = math_grad.SmartBroadcastGradientArgs(x, y, grad);

Debug.Assert(broadcast_info.Length == 2);

var (sx, rx, must_reduce_x) = broadcast_info[0];

var (sy, ry, must_reduce_y) = broadcast_info[1];

Tensor gx, gy;

if (must_reduce_x)

{

gx = array_ops.reshape(math_ops.reduce_sum(x_grad, rx), sx);

}

else

{

gx = x_grad;

}

if (must_reduce_y)

{

gy = -array_ops.reshape(math_ops.reduce_sum(x_grad, ry), sy);

}

else

{

gy = -x_grad;

}

return new Tensor[] { gx, gy };

}

/// <summary>

/// The derivatives for deconvolution.

/// </summary>

/// <param name="op">The Deconvolution op.</param>

/// <param name="grads">The tensor representing the gradient w.r.t. the output</param>

/// <returns>The gradients w.r.t. the input and the filter</returns>

[RegisterGradient("Conv2DBackpropInput")]

public static Tensor[] _Conv2DBackpropInputGrad(Operation op, Tensor[] grads)

{

var grad = grads[0];

var dilations = op.get_attr_list<int>("dilations");

var strides = op.get_attr_list<int>("strides");

var padding = op.get_attr<string>("padding");

var explicit_paddings = op.get_attr_list<int>("explicit_paddings");

var use_cudnn_on_gpu = op.get_attr<bool>("use_cudnn_on_gpu");

var data_format = op.get_attr<string>("data_format");

return new Tensor[]

{

gen_nn_ops.conv2d_backprop_filter(grad, array_ops.shape(op.inputs[1]), op.inputs[2],

strides, padding,

use_cudnn_on_gpu: use_cudnn_on_gpu,

explicit_paddings: explicit_paddings,

dilations: dilations,

data_format: data_format),

gen_nn_ops.conv2d(grad, op.inputs[1], strides, padding,

use_cudnn_on_gpu, explicit_paddings, data_format, dilations)

};

}

/// <summary>

/// Gradient function for Conv2D.

/// </summary>

/// <param name="op"></param>

/// <param name="grads"></param>

/// <returns></returns>

[RegisterGradient("Conv2D")]

public static Tensor[] _Conv2DGrad(Operation op, Tensor[] grads)

{

var dilations = op.get_attr_list<int>("dilations");

var strides = op.get_attr_list<int>("strides");

var padding = op.get_attr<string>("padding");

var explicit_paddings = op.get_attr_list<int>("explicit_paddings");

var use_cudnn_on_gpu = op.get_attr<bool>("use_cudnn_on_gpu");

var data_format = op.get_attr<string>("data_format");

var shape = gen_array_ops.shape_n(new Tensor[] { op.inputs[0], op.inputs[1] });

return new Tensor[]

{

gen_nn_ops.conv2d_backprop_input(shape[0], op.inputs[1], grads[0],

strides, padding, use_cudnn_on_gpu, explicit_paddings,

dilations: dilations,

data_format: data_format),

gen_nn_ops.conv2d_backprop_filter(op.inputs[0], shape[1], grads[0],

strides, padding,

dilations: dilations,

explicit_paddings: explicit_paddings,

use_cudnn_on_gpu: use_cudnn_on_gpu,

data_format: data_format)

};

}

/// <summary>

/// Gradient function for Conv2D.

/// </summary>

/// <param name="op"></param>

/// <param name="grads"></param>

/// <returns></returns>

[RegisterGradient("DepthwiseConv2dNative")]

public static Tensor[] _DepthwiseConv2DGrad(Operation op, Tensor[] grads)

{

var dilations = op.get_attr_list<int>("dilations");

var strides = op.get_attr_list<int>("strides");

var padding = op.get_attr<string>("padding");

var explicit_paddings = op.get_attr_list<int>("explicit_paddings");

var data_format = op.get_attr<string>("data_format");

var shape = gen_array_ops.shape_n(new Tensor[] { op.inputs[0], op.inputs[1] });

return new Tensor[]

{

gen_nn_ops.depthwise_conv2d_native_backprop_input(

shape[0], op.inputs[1], grads[0],

strides, padding, explicit_paddings,

dilations: dilations,

data_format: data_format),

gen_nn_ops.depthwise_conv2d_native_backprop_filter(op.inputs[0], shape[1], grads[0],

strides, padding,

dilations: dilations,

explicit_paddings: explicit_paddings,

data_format: data_format)

};

}

[RegisterGradient("FusedBatchNorm")]

public static Tensor[] _FusedBatchNormGrad(Operation op, Tensor[] grads)

=> _BaseFusedBatchNormGrad(op, 0, grads);

[RegisterGradient("FusedBatchNormV2")]

public static Tensor[] _FusedBatchNormV2Grad(Operation op, Tensor[] grads)

=> _BaseFusedBatchNormGrad(op, 1, grads);

[RegisterGradient("FusedBatchNormV3")]

public static Tensor[] _FusedBatchNormV3Grad(Operation op, Tensor[] grads)

=> _BaseFusedBatchNormGrad(op, 2, grads);

/// <summary>

/// Return the gradients for the 3 inputs of BatchNorm.

/// </summary>

/// <param name="op"></param>

/// <param name="version"></param>

/// <param name="grads"></param>

/// <returns></returns>

public static Tensor[] _BaseFusedBatchNormGrad(Operation op, int version, Tensor[] grads)

{

var x = op.inputs[0];

var grad_y = grads[0];

var scale = op.inputs[1];

var epsilon = op.get_attr<float>("epsilon");

var data_format = op.get_attr<string>("data_format");

var is_training = op.get_attr<bool>("is_training");

Func<FusedBatchNormParams, Tensor[]> grad_fun = (p) =>

{

if(version == 2)

{

return gen_nn_ops.fused_batch_norm_grad_v3(p.YBackprop, p.X, p.Scale,

p.ReserveSpace1, p.ReserveSpace2, p.ReserveSpace3, p.Epsilon,

p.DataFormat, p.IsTraining, p.Name);

}

else if(version == 1)

{

return gen_nn_ops.fused_batch_norm_grad_v2(p.YBackprop, p.X, p.Scale,

p.ReserveSpace1, p.ReserveSpace2, p.Epsilon, p.DataFormat,

p.IsTraining, p.Name);

}

else

{

return gen_nn_ops.fused_batch_norm_grad(p.YBackprop, p.X, p.Scale,

p.ReserveSpace1, p.ReserveSpace2, p.Epsilon, p.DataFormat,

p.IsTraining, p.Name);

}

};

if (is_training)

{

return grad_fun(new FusedBatchNormParams

{

YBackprop = grad_y,

X = x,

Scale = scale,

ReserveSpace1 = op.outputs[3],

ReserveSpace2 = op.outputs[4],

ReserveSpace3 = version == 2 ? op.outputs[5] : null,

Epsilon = epsilon,

DataFormat = data_format,

IsTraining = is_training

});

}

else

{

var pop_mean = op.inputs[3];

var pop_var = op.inputs[4];

if (data_format == "NCHW")

throw new NotImplementedException("");

var results = grad_fun(new FusedBatchNormParams

{

YBackprop = grad_y,

X = x,

Scale = scale,

ReserveSpace1 = pop_mean,

ReserveSpace2 = pop_var,

ReserveSpace3 = version == 2 ? op.outputs[5] : null,

Epsilon = epsilon,

DataFormat = data_format,

IsTraining = is_training

});

var (dx, dscale, doffset) = (results[0], results[1], results[2]);

if (data_format == "NCHW")

throw new NotImplementedException("");

return new Tensor[]

{

dx,

dscale,

doffset,

null,

null

};

}

[RegisterGradient("BatchNormWithGlobalNormalization")]

public static Tensor _BatchNormWithGlobalNormalizationGrad(Operation op, Tensor[] grads)

{

throw new NotImplementedException("BatchNormWithGlobalNormalization");

}

private static bool IsZero(Tensor g)

{

if (new string[] { "ZerosLike", "Zeros" }.Contains(g.op.type))

return true;

throw new NotImplementedException("IsZero");

}

private static Tensor _BroadcastMul(Tensor vec, Tensor mat)

{

vec = array_ops.expand_dims(vec, -1);

return vec * mat;

}

[RegisterGradient("MaxPool")]

public static Tensor[] _MaxPoolGrad(Operation op, Tensor[] grads)

{

var grad = grads[0];

return new Tensor[]

{

gen_nn_ops.max_pool_grad(

op.inputs[0],

op.outputs[0],

grad,

op.get_attr_list<int>("ksize"),

op.get_attr_list<int>("strides"),

padding: op.get_attr("padding").ToString(),

data_format: op.get_attr("data_format").ToString())

};

}

[RegisterGradient("AvgPool")]

public static Tensor[] _AvgPoolGrad(Operation op, Tensor[] grads)

{

Tensor grad = grads[0];

return new Tensor[]

{

gen_nn_ops.avg_pool_grad(

array_ops.shape(op.inputs[0]),

grad,

op.get_attr_list<int>("ksize"),

op.get_attr_list<int>("strides"),

op.get_attr<string>("padding"),

op.get_attr<string>("data_format"))

};

}

/// <summary>

/// Return the gradients for TopK.

/// </summary>

/// <param name="op"></param>

/// <param name="grads"></param>

/// <returns></returns>

[RegisterGradient("TopK")]

public static Tensor[] _TopKGrad(Operation op, Tensor[] grads)

{

var grad = grads[0];

var _ = grads[1];

var in_shape = array_ops.shape(op.inputs[0]);

var ind_shape = array_ops.shape(op.outputs[1]);

// int32 is not supported on GPU hence up-casting

var cast = math_ops.cast(ind_shape, TF_DataType.TF_INT64);

var size = array_ops.size(ind_shape) - 1;

var ind_lastdim = array_ops.gather(cast, size);

// Flatten indices to 2D.

var stack = array_ops.stack(new object[] { -1L, ind_lastdim });

var ind_2d = array_ops.reshape(op.outputs[1], stack);

var in_lastdim = array_ops.gather(math_ops.cast(in_shape, TF_DataType.TF_INT64),

array_ops.size(in_shape) - 1);

var outerdim = array_ops.shape(ind_2d).slice(0);

// Compute linear indices(flattened to 1D).

var cast1 = math_ops.cast(outerdim, TF_DataType.TF_INT64);

var range2 = math_ops.range(tf.constant(0L), cast1 * in_lastdim, in_lastdim);

var dim2 = array_ops.expand_dims(range2, -1);

var cast2 = math_ops.cast(dim2, TF_DataType.TF_INT32);

var ind = array_ops.reshape(ind_2d + cast2, new int[] { -1 });

// Substitute grad to appropriate locations and fill the rest with zeros,

// finally reshaping it to the original input shape.

var scatter = gen_array_ops.scatter_nd(array_ops.expand_dims(ind, -1),

array_ops.reshape(grad, new int[] { -1 }),

math_ops.reduce_prod(in_shape));

return new Tensor[]

{

array_ops.reshape(scatter, in_shape),

array_ops.zeros(new int[0], dtype: TF_DataType.TF_INT32)

};

}

TensorFlow.NET/src/TensorFlowNET.Core/Gradients/nn_grad.cs at master · MSavameri/TensorFlow.NET · GitHub

Repository navigation

Expand file tree

Breadcrumbs

nn_grad.cs

Latest commit

History

Breadcrumbs

nn_grad.cs

File metadata and controls