diff --git a/tfjs-core/src/kernel_names.ts b/tfjs-core/src/kernel_names.ts
index d8139e7eed2..4e9a822453f 100644
--- a/tfjs-core/src/kernel_names.ts
+++ b/tfjs-core/src/kernel_names.ts
@@ -18,6 +18,8 @@
 // tslint:disable: variable-name
 // Unfortunately just enabling PascalCase per file (tslint:enable:
 // allow-pascal-case) doesn't work.
+import {ExplicitPadding} from '../src/ops/conv_util';
+
 import {NamedTensorInfoMap, TensorInfo} from './kernel_registry';
 import {DataType, PixelData} from './types';
 
@@ -104,7 +106,7 @@ export const Conv2D = 'Conv2D';
 export type Conv2DInputs = Pick<NamedTensorInfoMap, 'x'|'filter'>;
 export interface Conv2DAttrs {
   strides: [number, number]|number;
-  pad: 'valid'|'same'|number;
+  pad: 'valid'|'same'|number|ExplicitPadding;
   dataFormat: 'NHWC'|'NCHW';
   dilations: [number, number]|number;
   dimRoundingMode?: 'floor'|'round'|'ceil';
@@ -114,7 +116,7 @@ export const Conv2DBackpropFilter = 'Conv2DBackpropFilter';
 export type Conv2DBackpropFilterInputs = Pick<NamedTensorInfoMap, 'x'|'dy'>;
 export interface Conv2DBackpropFilterAttrs {
   strides: [number, number]|number;
-  pad: 'valid'|'same'|number;
+  pad: 'valid'|'same'|number|ExplicitPadding;
   dataFormat: 'NHWC'|'NCHW';
   dimRoundingMode?: 'floor'|'round'|'ceil';
 }
@@ -123,7 +125,7 @@ export const Conv2DBackpropInput = 'Conv2DBackpropInput';
 export type Conv2DBackpropInputInputs = Pick<NamedTensorInfoMap, 'dy'|'filter'>;
 export interface Conv2DBackpropInputAttrs {
   strides: [number, number]|number;
-  pad: 'valid'|'same'|number;
+  pad: 'valid'|'same'|number|ExplicitPadding;
   dataFormat: 'NHWC'|'NCHW';
   dimRoundingMode?: 'floor'|'round'|'ceil';
 }
diff --git a/tfjs-core/src/ops/conv1d.ts b/tfjs-core/src/ops/conv1d.ts
index 8b38efd6fdb..eb6746eaede 100644
--- a/tfjs-core/src/ops/conv1d.ts
+++ b/tfjs-core/src/ops/conv1d.ts
@@ -54,7 +54,8 @@ import {op} from './operation';
 /** @doc {heading: 'Operations', subheading: 'Convolution'} */
 function conv1d_<T extends Tensor2D|Tensor3D>(
     x: T|TensorLike, filter: Tensor3D|TensorLike, stride: number,
-    pad: 'valid'|'same'|number, dataFormat: 'NWC'|'NCW' = 'NWC', dilation = 1,
+    pad: 'valid'|'same'|number|conv_util.ExplicitPadding,
+    dataFormat: 'NWC'|'NCW' = 'NWC', dilation = 1,
     dimRoundingMode?: 'floor'|'round'|'ceil'): T {
   const $x = convertToTensor(x, 'x', 'conv1d');
   const $filter = convertToTensor(filter, 'filter', 'conv1d');
diff --git a/tfjs-core/src/ops/conv1d_test.ts b/tfjs-core/src/ops/conv1d_test.ts
index 3b57b01f9d8..1fc75ed7482 100644
--- a/tfjs-core/src/ops/conv1d_test.ts
+++ b/tfjs-core/src/ops/conv1d_test.ts
@@ -21,6 +21,26 @@ import {expectArraysClose} from '../test_util';
 import {Rank} from '../types';
 
 describeWithFlags('conv1d', ALL_ENVS, () => {
+  it('conv1d input=2x2x1,d2=1,f=1,s=1,d=1,p=explicit', async () => {
+    const inputDepth = 1;
+    const inputShape: [number, number, number] = [2, 2, inputDepth];
+    const outputDepth = 1;
+    const fSize = 1;
+    const pad =
+        [[0, 0], [0, 0], [0, 0], [0, 0]] as tf.backend_util.ExplicitPadding;
+    const stride = 1;
+    const dataFormat = 'NWC';
+    const dilation = 1;
+
+    const x = tf.tensor3d([1, 2, 3, 4], inputShape);
+    const w = tf.tensor3d([3], [fSize, inputDepth, outputDepth]);
+
+    const result = tf.conv1d(x, w, stride, pad, dataFormat, dilation);
+
+    expect(result.shape).toEqual([2, 2, 1]);
+    expectArraysClose(await result.data(), [3, 6, 9, 12]);
+  });
+
   it('conv1d input=2x2x1,d2=1,f=1,s=1,d=1,p=same', async () => {
     const inputDepth = 1;
     const inputShape: [number, number, number] = [2, 2, inputDepth];
diff --git a/tfjs-core/src/ops/conv2d.ts b/tfjs-core/src/ops/conv2d.ts
index cfa028ddebb..569bd84a6ad 100644
--- a/tfjs-core/src/ops/conv2d.ts
+++ b/tfjs-core/src/ops/conv2d.ts
@@ -61,7 +61,8 @@ import {op} from './operation';
 /** @doc {heading: 'Operations', subheading: 'Convolution'} */
 function conv2d_<T extends Tensor3D|Tensor4D>(
     x: T|TensorLike, filter: Tensor4D|TensorLike,
-    strides: [number, number]|number, pad: 'valid'|'same'|number,
+    strides: [number, number]|number,
+    pad: 'valid'|'same'|number|conv_util.ExplicitPadding,
     dataFormat: 'NHWC'|'NCHW' = 'NHWC',
     dilations: [number, number]|number = [1, 1],
     dimRoundingMode?: 'floor'|'round'|'ceil'): T {
diff --git a/tfjs-core/src/ops/conv2d_backprop_filter.ts b/tfjs-core/src/ops/conv2d_backprop_filter.ts
index cd60cb5df42..d4cf4ad931b 100644
--- a/tfjs-core/src/ops/conv2d_backprop_filter.ts
+++ b/tfjs-core/src/ops/conv2d_backprop_filter.ts
@@ -49,7 +49,8 @@ import {op} from './operation';
  */
 function conv2DBackpropFilter_<T extends Tensor3D|Tensor4D>(
     x: T, dy: T, filterShape: [number, number, number, number],
-    strides: [number, number]|number, pad: 'valid'|'same'|number,
+    strides: [number, number]|number,
+    pad: 'valid'|'same'|number|conv_util.ExplicitPadding,
     dataFormat: 'NHWC'|'NCHW' = 'NHWC',
     dimRoundingMode?: 'floor'|'round'|'ceil'): Tensor4D {
   let x4D = x as Tensor4D;
diff --git a/tfjs-core/src/ops/conv2d_backprop_input.ts b/tfjs-core/src/ops/conv2d_backprop_input.ts
index 2749d3f390e..1e50ccba95c 100644
--- a/tfjs-core/src/ops/conv2d_backprop_input.ts
+++ b/tfjs-core/src/ops/conv2d_backprop_input.ts
@@ -53,7 +53,8 @@ import {op} from './operation';
 function conv2DBackpropInput_<T extends Tensor3D|Tensor4D>(
     xShape: [number, number, number, number]|[number, number, number], dy: T,
     filter: Tensor4D, strides: [number, number]|number,
-    pad: 'valid'|'same'|number, dataFormat: 'NHWC'|'NCHW' = 'NHWC',
+    pad: 'valid'|'same'|number|conv_util.ExplicitPadding,
+    dataFormat: 'NHWC'|'NCHW' = 'NHWC',
     dimRoundingMode?: 'floor'|'round'|'ceil'): T {
   util.assert(
       xShape.length === dy.rank,
diff --git a/tfjs-core/src/ops/conv2d_test.ts b/tfjs-core/src/ops/conv2d_test.ts
index bec5ec5b077..be71e9ebbd2 100644
--- a/tfjs-core/src/ops/conv2d_test.ts
+++ b/tfjs-core/src/ops/conv2d_test.ts
@@ -166,6 +166,26 @@ describeWithFlags('conv2d', ALL_ENVS, () => {
     expectArraysClose(resultData, [133, 66, 200, 102, 108, 58, 56, 58]);
   });
 
+  it('x=[4,2,1] f=[4,2,1,1] s=1 d=1 p=explicit', async () => {
+    const inputDepth = 1;
+    const outputDepth = 1;
+    const pad =
+        [[0, 0], [1, 2], [0, 1], [0, 0]] as tf.backend_util.ExplicitPadding;
+    const stride = 1;
+    const dataFormat = 'NHWC';
+    const dilation = 1;
+
+    const x = tf.tensor3d([1, 2, 3, 4, 5, 6, 7, 8], [4, 2, inputDepth]);
+    const w =
+        tf.tensor4d([3, 1, 5, 0, 2, 7, 8, 9], [4, 2, inputDepth, outputDepth]);
+
+    const result = tf.conv2d(x, w, stride, pad, dataFormat, dilation);
+
+    const resultData = await result.data();
+    expect(result.shape).toEqual([4, 2, 1]);
+    expectArraysClose(resultData, [133, 66, 200, 102, 108, 58, 56, 58]);
+  });
+
   it('x=[2,2,1] f=[2,2,1,1] s=1 d=1 p=same', async () => {
     const inputDepth = 1;
     const inputShape: [number, number, number] = [2, 2, inputDepth];
@@ -208,6 +228,28 @@ describeWithFlags('conv2d', ALL_ENVS, () => {
     expectArraysClose(resultData, [20, 26, 13, 12]);
   });
 
+  it('x=[1,2,2] f=[2,2,1,1] s=1 d=1 p=explicit NCHW', async () => {
+    const inputDepth = 1;
+    const inputShape: [number, number, number] = [inputDepth, 2, 2];
+    const outputDepth = 1;
+    const fSize = 2;
+    const pad =
+        [[0, 0], [0, 0], [0, 1], [0, 1]] as tf.backend_util.ExplicitPadding;
+    const stride = 1;
+    const dataFormat = 'NCHW';
+    const dilation = 1;
+
+    const x = tf.tensor3d([1, 2, 3, 4], inputShape);
+    const w =
+        tf.tensor4d([3, 1, 5, 0], [fSize, fSize, inputDepth, outputDepth]);
+
+    const result = tf.conv2d(x, w, stride, pad, dataFormat, dilation);
+
+    const resultData = await result.data();
+    expect(result.shape).toEqual([1, 2, 2]);
+    expectArraysClose(resultData, [20, 26, 13, 12]);
+  });
+
   it('x=[2,2,2] f=[2,2,2,1] s=1 d=1 p=same NCHW', async () => {
     const inputDepth = 2;
     const inputShape: [number, number, number] = [inputDepth, 2, 2];
diff --git a/tfjs-core/src/ops/conv_util.ts b/tfjs-core/src/ops/conv_util.ts
index d9249b8754a..cffae54dfc4 100644
--- a/tfjs-core/src/ops/conv_util.ts
+++ b/tfjs-core/src/ops/conv_util.ts
@@ -17,7 +17,15 @@
 
 import * as util from '../util';
 
-type PadType = 'SAME'|'VALID'|'NUMBER';
+type PadType = 'SAME'|'VALID'|'NUMBER'|'EXPLICIT';
+
+// For NHWC should be in the following form:
+//  [[0, 0], [pad_top,pad_bottom], [pad_left, pad_right], [0, 0]]
+// For NCHW should be in the following form:
+//  [[0, 0], [0, 0], [pad_top,pad_bottom], [pad_left, pad_right]]
+// Reference: https://www.tensorflow.org/api_docs/python/tf/nn/conv2d
+export type ExplicitPadding =
+    [[number, number], [number, number], [number, number], [number, number]];
 
 export type PadInfo = {
   top: number,
@@ -126,8 +134,8 @@ export function computeConv2DInfo(
     inShape: [number, number, number, number],
     filterShape: [number, number, number, number],
     strides: number|[number, number], dilations: number|[number, number],
-    pad: 'same'|'valid'|number, roundingMode?: 'floor'|'round'|'ceil',
-    depthwise = false,
+    pad: 'same'|'valid'|number|ExplicitPadding,
+    roundingMode?: 'floor'|'round'|'ceil', depthwise = false,
     dataFormat: 'channelsFirst'|'channelsLast' = 'channelsLast'): Conv2DInfo {
   let [batchSize, inHeight, inWidth, inChannels] = [-1, -1, -1, -1];
   if (dataFormat === 'channelsLast') {
@@ -148,7 +156,7 @@ export function computeConv2DInfo(
       getEffectiveFilterSize(filterWidth, dilationWidth);
   const {padInfo, outHeight, outWidth} = getPadAndOutInfo(
       pad, inHeight, inWidth, strideHeight, strideWidth, effectiveFilterHeight,
-      effectiveFilterWidth, roundingMode);
+      effectiveFilterWidth, roundingMode, dataFormat);
 
   const outChannels = depthwise ? filterChannels * inChannels : filterChannels;
 
@@ -399,10 +407,12 @@ function getEffectiveFilterSize(filterSize: number, dilation: number) {
 }
 
 function getPadAndOutInfo(
-    pad: 'same'|'valid'|number, inHeight: number, inWidth: number,
-    strideHeight: number, strideWidth: number, filterHeight: number,
-    filterWidth: number, roundingMode?: 'floor'|'round'|'ceil'):
-    {padInfo: PadInfo, outHeight: number, outWidth: number} {
+    pad: 'same'|'valid'|number|ExplicitPadding, inHeight: number,
+    inWidth: number, strideHeight: number, strideWidth: number,
+    filterHeight: number, filterWidth: number,
+    roundingMode: 'floor'|'round'|'ceil',
+    dataFormat: 'channelsFirst'|
+    'channelsLast'): {padInfo: PadInfo, outHeight: number, outWidth: number} {
   let padInfo: PadInfo;
   let outHeight: number;
   let outWidth: number;
@@ -430,6 +440,20 @@ function getPadAndOutInfo(
     padInfo = {top: 0, bottom: 0, left: 0, right: 0, type: 'VALID'};
     outHeight = Math.ceil((inHeight - filterHeight + 1) / strideHeight);
     outWidth = Math.ceil((inWidth - filterWidth + 1) / strideWidth);
+  } else if (typeof pad === 'object') {
+    const top = dataFormat === 'channelsLast' ? pad[1][0] : pad[2][0];
+    const bottom = dataFormat === 'channelsLast' ? pad[1][1] : pad[2][1];
+    const left = dataFormat === 'channelsLast' ? pad[2][0] : pad[3][0];
+    const right = dataFormat === 'channelsLast' ? pad[2][1] : pad[3][1];
+    const padType = (top === 0 && bottom === 0 && left === 0 && right === 0) ?
+        'VALID' :
+        'EXPLICIT';
+    padInfo = {top, bottom, left, right, type: padType};
+    outHeight = conditionalRound(
+        (inHeight - filterHeight + top + bottom) / strideHeight + 1,
+        roundingMode);
+    outWidth = conditionalRound(
+        (inWidth - filterWidth + left + right) / strideWidth + 1, roundingMode);
   } else {
     throw Error(`Unknown padding parameter: ${pad}`);
   }
diff --git a/tfjs-core/src/ops/fused_ops.ts b/tfjs-core/src/ops/fused_ops.ts
index 7fe81be0c45..5bc1e66fa49 100644
--- a/tfjs-core/src/ops/fused_ops.ts
+++ b/tfjs-core/src/ops/fused_ops.ts
@@ -330,7 +330,7 @@ function fusedConv2d_<T extends Tensor3D|Tensor4D>({
   x: T|TensorLike,
   filter: Tensor4D|TensorLike,
   strides: [number, number]|number,
-  pad: 'valid'|'same'|number,
+  pad: 'valid'|'same'|number|conv_util.ExplicitPadding,
   dataFormat?: 'NHWC'|'NCHW',
   dilations?: [number, number]|number,
   dimRoundingMode?: 'floor'|'round'|'ceil',
diff --git a/tfjs-core/src/ops/fused_test.ts b/tfjs-core/src/ops/fused_test.ts
index 4e297d66078..643daef7a5d 100644
--- a/tfjs-core/src/ops/fused_test.ts
+++ b/tfjs-core/src/ops/fused_test.ts
@@ -920,6 +920,27 @@ describeWithFlags('fused conv2d', ALL_ENVS, () => {
     expectArraysClose(await result.data(), expected);
   });
 
+  it('basic with explicit padding', async () => {
+    const inputDepth = 1;
+    const outputDepth = 1;
+    const pad =
+        [[0, 0], [1, 2], [0, 1], [0, 0]] as tf.backend_util.ExplicitPadding;
+    const stride = 1;
+    const dataFormat = 'NHWC';
+    const dilation = 1;
+
+    const x = tf.tensor3d([1, 2, 3, 4, 5, 6, 7, 8], [4, 2, inputDepth]);
+    const w =
+        tf.tensor4d([3, 1, 5, 0, 2, 7, 8, 9], [4, 2, inputDepth, outputDepth]);
+
+    const result = tf.fused.conv2d(
+        {x, filter: w, strides: stride, pad, dataFormat, dilations: dilation});
+
+    const resultData = await result.data();
+    expect(result.shape).toEqual([4, 2, 1]);
+    expectArraysClose(resultData, [133, 66, 200, 102, 108, 58, 56, 58]);
+  });
+
   it('basic with elu', async () => {
     const inputDepth = 2;
     const inShape: [number, number, number, number] = [2, 2, 2, inputDepth];
diff --git a/tfjs-core/src/public/chained_ops/conv1d.ts b/tfjs-core/src/public/chained_ops/conv1d.ts
index 1e76d9a4f80..cf885db6320 100644
--- a/tfjs-core/src/public/chained_ops/conv1d.ts
+++ b/tfjs-core/src/public/chained_ops/conv1d.ts
@@ -15,6 +15,7 @@
  * =============================================================================
  */
 import {conv1d} from '../../ops/conv1d';
+import {ExplicitPadding} from '../../ops/conv_util';
 import {Tensor, Tensor2D, Tensor3D} from '../../tensor';
 import {Rank, TensorLike3D} from '../../types';
 
@@ -22,15 +23,15 @@ declare module '../../tensor' {
   interface Tensor<R extends Rank = Rank> {
     conv1d<T extends Tensor2D|Tensor3D>(
         filter: Tensor3D|TensorLike3D, stride: number,
-        pad: 'valid'|'same'|number, dataFormat?: 'NWC'|'NCW', dilation?: number,
-        dimRoundingMode?: 'floor'|'round'|'ceil'): T;
+        pad: 'valid'|'same'|number|ExplicitPadding, dataFormat?: 'NWC'|'NCW',
+        dilation?: number, dimRoundingMode?: 'floor'|'round'|'ceil'): T;
   }
 }
 
 Tensor.prototype.conv1d = function<T extends Tensor2D|Tensor3D>(
-    filter: Tensor3D|TensorLike3D, stride: number, pad: 'valid'|'same'|number,
-    dataFormat?: 'NWC'|'NCW', dilation?: number,
-    dimRoundingMode?: 'floor'|'round'|'ceil'): T {
+    filter: Tensor3D|TensorLike3D, stride: number,
+    pad: 'valid'|'same'|number|ExplicitPadding, dataFormat?: 'NWC'|'NCW',
+    dilation?: number, dimRoundingMode?: 'floor'|'round'|'ceil'): T {
   this.throwIfDisposed();
   return conv1d(
              this, filter, stride, pad, dataFormat, dilation,
diff --git a/tfjs-node/src/nodejs_kernel_backend.ts b/tfjs-node/src/nodejs_kernel_backend.ts
index b8dbf89d839..0e0ee048697 100644
--- a/tfjs-node/src/nodejs_kernel_backend.ts
+++ b/tfjs-node/src/nodejs_kernel_backend.ts
@@ -846,7 +846,8 @@ export class NodeJSKernelBackend extends KernelBackend {
 
   conv2d(x: Tensor4D, filter: Tensor4D, convInfo: backend_util.Conv2DInfo):
       Tensor4D {
-    if (convInfo.padInfo.type !== 'VALID' && convInfo.padInfo.type !== 'SAME') {
+    if (convInfo.padInfo.type !== 'VALID' && convInfo.padInfo.type !== 'SAME' &&
+        convInfo.padInfo.type !== 'EXPLICIT') {
       throw new Error(
           `TF Backend supports only 'valid' and 'same' padding ` +
           `while padding was ${convInfo.padInfo.type}`);
@@ -867,6 +868,18 @@ export class NodeJSKernelBackend extends KernelBackend {
       {name: 'use_cudnn_on_gpu', type: this.binding.TF_ATTR_BOOL, value: true},
       {name: 'dilations', type: this.binding.TF_ATTR_INT, value: dilations},
     ];
+    if (padding === 'EXPLICIT') {
+      const padValue = [
+        convInfo.padInfo.top, convInfo.padInfo.bottom, convInfo.padInfo.left,
+        convInfo.padInfo.right
+      ];
+      opAttrs.push({
+        name: 'explicit_paddings',
+        type: this.binding.TF_ATTR_INT,
+        value: dataFormat === 'NHWC' ? [0, 0, ...padValue, 0, 0] :
+                                       [0, 0, 0, 0, ...padValue]
+      });
+    }
     return this.executeSingleOutput('Conv2D', opAttrs, [x, filter]) as Tensor4D;
   }
 
diff --git a/tfjs-node/src/run_tests.ts b/tfjs-node/src/run_tests.ts
index dd5ff25830b..7ddc27c8dd0 100644
--- a/tfjs-node/src/run_tests.ts
+++ b/tfjs-node/src/run_tests.ts
@@ -73,6 +73,7 @@ const IGNORE_LIST: string[] = [
   'diag test-tensorflow {} bool',
   // See https://github.com/tensorflow/tfjs/issues/1891
   'conv2d test-tensorflow {} x=[2,1,2,2] f=[1,1,1,1] s=1 d=1 p=0 NCHW',
+  'conv2d test-tensorflow {} x=[1,2,2] f=[2,2,1,1] s=1 d=1 p=explicit NCHW',
   'conv2d test-tensorflow {} x=[1,2,2] f=[2,2,1,1] s=1 d=1 p=same NCHW',
   'conv2d test-tensorflow {} x=[2,2,2] f=[2,2,2,1] s=1 d=1 p=same NCHW',
   'conv2d test-tensorflow {} x=[2,1,2,2] f=[2,2,1,1] s=1 d=1 p=same NCHW',