First demo commit, v2

README.md

@@ -7,6 +7,7 @@ sdk: gradio
 sdk_version: 3.1.3
 app_file: app.py
 pinned: false
+python_version: 3.7
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,26 @@
+import os
+from models import Noise2Same
+import gradio as gr
+
+os.system("mkdir trained_models/denoising_ImageNet")
+os.system("cd trained_models/denoising_ImageNet; gdown https://drive.google.com/uc?id=1asrwULW1lDFasystBc3UfShh5EeTHpkW; gdown https://drive.google.com/uc?id=1Re1ER7KtujBunN0-74QmYrrOx77WpVXK; gdown https://drive.google.com/uc?id=1QdlyUPUKyyGtqD0zBrj5F7qQZtmUELSu; gdown https://drive.google.com/uc?id=1LQsYR26ldHebcdQtP2zt4Mh-ZH9vXQ2S; gdown https://drive.google.com/uc?id=1AxTDD4dS0DtzmBywjGyeJYgDrw-XjYbc; gdown https://drive.google.com/uc?id=1w4UdNAbOjvWSL0Jgbq8_hCniaxqsbLaQ; cd ../..")
+os.system("wget -O arch.png https://i.imgur.com/NruRABn.png")
+os.system("wget -O parrot.png https://i.imgur.com/zdji3xv.png")
+os.system("wget -O lion.png https://i.imgur.com/qNT0lJJ.png")
+
+model = Noise2Same('trained_models/', 'denoising_ImageNet', dim=2, in_channels=3)
+
+def norm(x):
+    x = (x-x.min())/(x.max()-x.min())
+    return x
+    
+def predict(img):
+  pred = model.predict(img.astype('float32'))
+  return norm(pred)
+
+img = gr.inputs.Image()
+
+title = "Noise2Same: Optimizing A Self-Supervised Bound for Image Denoising"
+description = "Interactive demo of Noise2Same, an image denoising method developed by Yaochen Xie"
+
+gr.Interface(predict, "image", "image", examples=[["lion.png"], ["arch.png"], ["parrot.png"]], title=title, description=description).launch()

basic_ops.py

@@ -0,0 +1,91 @@
+import logging, os
+logging.disable(logging.WARNING)
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+
+import tensorflow as tf
+from network_configure import conf_basic_ops
+
+
+"""This script defines basic operaters.
+"""
+
+
+def convolution_2D(inputs, filters, kernel_size, strides, use_bias, name=None):
+    """Performs 2D convolution without activation function.
+    If followed by batch normalization, set use_bias=False.
+    """
+    return tf.layers.conv2d(
+                inputs=inputs,
+                filters=filters,
+                kernel_size=kernel_size,
+                strides=strides,
+                padding='same',
+                use_bias=use_bias,
+                kernel_initializer=conf_basic_ops['kernel_initializer'],
+                name=name,
+            )
+
+def convolution_3D(inputs, filters, kernel_size, strides, use_bias, name=None):
+    """Performs 3D convolution without activation function.
+    If followed by batch normalization, set use_bias=False.
+    """
+    return tf.layers.conv3d(
+                inputs=inputs,
+                filters=filters,
+                kernel_size=kernel_size,
+                strides=strides,
+                padding='same',
+                use_bias=use_bias,
+                kernel_initializer=conf_basic_ops['kernel_initializer'],
+                name=name,
+            )
+
+def transposed_convolution_2D(inputs, filters, kernel_size, strides, use_bias, name=None):
+    """Performs 2D transposed convolution without activation function.
+    If followed by batch normalization, set use_bias=False.
+    """
+    return tf.layers.conv2d_transpose(
+                inputs=inputs,
+                filters=filters,
+                kernel_size=kernel_size,
+                strides=strides,
+                padding='same',
+                use_bias=use_bias,
+                kernel_initializer=conf_basic_ops['kernel_initializer'],
+                name=name,
+            )
+
+def transposed_convolution_3D(inputs, filters, kernel_size, strides, use_bias, name=None):
+    """Performs 3D transposed convolution without activation function.
+    If followed by batch normalization, set use_bias=False.
+    """
+    return tf.layers.conv3d_transpose(
+                inputs=inputs,
+                filters=filters,
+                kernel_size=kernel_size,
+                strides=strides,
+                padding='same',
+                use_bias=use_bias,
+                kernel_initializer=conf_basic_ops['kernel_initializer'],
+                name=name,
+            )
+
+def batch_norm(inputs, training, name=None):
+    """Performs a batch normalization.
+    We set fused=True for a significant performance boost.
+    See https://www.tensorflow.org/performance/performance_guide#common_fused_ops
+    """
+    return tf.layers.batch_normalization(
+                inputs=inputs,
+                momentum=conf_basic_ops['momentum'],
+                epsilon=conf_basic_ops['epsilon'],
+                center=True,
+                scale=True,
+                training=training, 
+                fused=True,
+                name=name,
+            )
+
+def relu(inputs, name=None):
+    return tf.nn.relu(inputs, name=name) if conf_basic_ops['relu_type'] == 'relu' \
+            else tf.nn.relu6(inputs, name=name)

models.py

@@ -0,0 +1,276 @@
+import os, cv2
+import numpy as np
+from network_configure import conf_unet
+from network import *
+from utils.predict_utils import get_coord, PercentileNormalizer, PadAndCropResizer
+from utils.train_utils import augment_patch
+from utils import train_utils
+
+# UNet implementation inherited from GVTNets: https://github.com/zhengyang-wang/GVTNets
+training_config = {'base_learning_rate': 0.0004,
+                                     'lr_decay_steps':5e3, 
+                                     'lr_decay_rate':0.5, 
+                                     'lr_staircase':True}
+
+class Noise2Same(object):
+
+    def __init__(self, base_dir, name, 
+                 dim=2, in_channels=1, lmbd=None, 
+                 masking='gaussian', mask_perc=0.5,
+                 opt_config=training_config, **kwargs):
+
+        self.base_dir = base_dir # model direction
+        self.name = name # model name
+        self.dim = dim # image dimension
+        self.in_channels = in_channels # image channels
+        self.lmbd = lmbd # lambda in loss fn
+        self.masking = masking
+        self.mask_perc = mask_perc
+        
+        self.opt_config = opt_config
+        conf_unet['dimension'] = '%dD'%dim
+        self.net = UNet(conf_unet)
+        
+    def _model_fn(self, features, labels, mode):
+        conv_op = convolution_2D if self.dim==2 else convolution_3D
+        axis = {3:[1,2,3,4], 2:[1,2,3]}[self.dim]
+        
+        def image_summary(img):
+            return tf.reduce_max(img, axis=1) if self.dim == 3 else img
+        
+        # Local average excluding the center pixel (donut)
+        def mask_kernel(features):
+            kernel = (np.array([[0.5, 1.0, 0.5], [1.0, 0.0, 1.0], [0.5, 1.0, 0.5]]) 
+                      if self.dim == 2 else 
+                      np.array([[[0, 0.5, 0], [0.5, 1.0, 0.5], [0, 0.5, 0]],
+                                [[0.5, 1.0, 0.5], [1.0, 0.0, 1.0], [0.5, 1.0, 0.5]],
+                                [[0, 0.5, 0], [0.5, 1.0, 0.5], [0, 0.5, 0]]]))
+            kernel = (kernel/kernel.sum())
+            kernels = np.empty([3, 3, self.in_channels, self.in_channels])
+            for i in range(self.in_channels):
+                kernels[:,:,i,i] = kernel
+            nn_conv_op = tf.nn.conv2d if self.dim == 2 else tf.nn.conv3d
+            return nn_conv_op(features, tf.constant(kernels.astype('float32')), 
+                              [1]*self.dim+[1,1], padding='SAME')
+        
+        if not mode == tf.estimator.ModeKeys.PREDICT:
+            noise, mask = tf.split(labels, [self.in_channels, self.in_channels], -1)
+            
+            if self.masking == 'gaussian':
+                masked_features = (1 - mask) * features + mask * noise
+            elif self.masking == 'donut':
+                masked_features = (1 - mask) * features + mask * mask_kernel(features)
+            else:
+                raise NotImplementedError
+            
+            # Prediction from masked input
+            with tf.variable_scope('main_unet', reuse=tf.compat.v1.AUTO_REUSE):
+                out = self.net(masked_features, mode == tf.estimator.ModeKeys.TRAIN)
+                out = batch_norm(out, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
+                out = relu(out)
+                preds = conv_op(out, self.in_channels, 1, 1, False, name = 'out_conv')
+                
+            # Prediction from full input
+            with tf.variable_scope('main_unet', reuse=tf.compat.v1.AUTO_REUSE):
+                rawout = self.net(features, mode == tf.estimator.ModeKeys.TRAIN)
+                rawout = batch_norm(rawout, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
+                rawout = relu(rawout)
+                rawpreds = conv_op(rawout, self.in_channels, 1, 1, False, name = 'out_conv')
+            
+            # Loss components
+            rec_mse = tf.reduce_mean(tf.square(rawpreds - features), axis=None)
+            inv_mse = tf.reduce_sum(tf.square(rawpreds - preds) * mask) / tf.reduce_sum(mask)
+            bsp_mse = tf.reduce_sum(tf.square(features - preds) * mask) / tf.reduce_sum(mask)
+
+            # Tensorboard display
+            tf.summary.image('1_inputs', image_summary(features), max_outputs=3)
+            tf.summary.image('2_raw_predictions', image_summary(rawpreds), max_outputs=3)
+            tf.summary.image('3_mask', image_summary(mask), max_outputs=3)
+            tf.summary.image('4_masked_predictions', image_summary(preds), max_outputs=3)
+            tf.summary.image('5_difference', image_summary(rawpreds-preds), max_outputs=3)
+            tf.summary.image('6_rec_error', image_summary(preds-features), max_outputs=3)
+            tf.summary.scalar('reconstruction', rec_mse, family='loss_metric') 
+            tf.summary.scalar('invariance', inv_mse, family='loss_metric') 
+            tf.summary.scalar('blind_spot', bsp_mse, family='loss_metric')
+                
+        else:
+            with tf.variable_scope('main_unet'):
+                out = self.net(features, mode == tf.estimator.ModeKeys.TRAIN)
+                out = batch_norm(out, mode == tf.estimator.ModeKeys.TRAIN, 'unet_out')
+                out = relu(out)
+                preds = conv_op(out, self.in_channels, 1, 1, False, name = 'out_conv')
+            return tf.estimator.EstimatorSpec(mode=mode, predictions=preds)
+        
+        lmbd = 2 if self.lmbd is None else self.lmbd
+        loss = rec_mse + lmbd*tf.sqrt(inv_mse)
+
+        if mode == tf.estimator.ModeKeys.TRAIN:
+            global_step = tf.train.get_or_create_global_step()
+            learning_rate = tf.train.exponential_decay(self.opt_config['base_learning_rate'], 
+                                                       global_step, 
+                                                       self.opt_config['lr_decay_steps'], 
+                                                       self.opt_config['lr_decay_rate'], 
+                                                       self.opt_config['lr_staircase'])
+            optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
+
+            update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='main_unet')
+            with tf.control_dependencies(update_ops):
+                train_op = optimizer.minimize(loss, global_step)
+        else:
+            train_op = None
+        
+        metrics = {'loss_metric/invariance':tf.metrics.mean(inv_mse),
+                              'loss_metric/blind_spot':tf.metrics.mean(bsp_mse), 
+                              'loss_metric/reconstruction':tf.metrics.mean(rec_mse)}
+        
+        return tf.estimator.EstimatorSpec(mode=mode, predictions=preds, loss=loss, train_op=train_op, 
+                                          eval_metric_ops=metrics)
+
+
+    def _input_fn(self, sources, patch_size, batch_size, is_train=True):
+        # Stratified sampling inherited from Noise2Void: https://github.com/juglab/n2v
+        get_stratified_coords = getattr(train_utils, 'get_stratified_coords%dD'%self.dim)
+        rand_float_coords = getattr(train_utils, 'rand_float_coords%dD'%self.dim)
+        
+        def generator():
+            while(True):
+                source = sources[np.random.randint(len(sources))]
+                valid_shape = source.shape[:-1] - np.array(patch_size)
+                if any([s<=0 for s in valid_shape]):
+                    source_patch = augment_patch(source)
+                else:
+                    coords = [np.random.randint(0, shape_i+1) for shape_i in valid_shape]
+                    s = tuple([slice(coord, coord+size) for coord, size in zip(coords, patch_size)])
+                    source_patch = augment_patch(source[s])
+                
+                mask = np.zeros_like(source_patch)
+                for c in range(self.in_channels):
+                    boxsize = np.round(np.sqrt(100/self.mask_perc)).astype(np.int)
+                    maskcoords = get_stratified_coords(rand_float_coords(boxsize), 
+                                                       box_size=boxsize, shape=tuple(patch_size))
+                    indexing = maskcoords + (c,)
+                    mask[indexing] = 1.0
+
+                noise_patch = np.concatenate([np.random.normal(0, 0.2, source_patch.shape), mask], axis=-1)
+                yield source_patch, noise_patch
+                
+        def generator_val():
+            for idx in range(len(sources)):
+                source_patch = sources[idx]
+                patch_size = source_patch.shape[:-1]
+                boxsize = np.round(np.sqrt(100/self.mask_perc)).astype(np.int)
+                maskcoords = get_stratified_coords(rand_float_coords(boxsize), 
+                                                   box_size=boxsize, shape=tuple(patch_size))
+                indexing = maskcoords + (0,)
+                mask = np.zeros_like(source_patch)
+                mask[indexing] = 1.0
+                noise_patch = np.concatenate([np.random.normal(0, 0.2, source_patch.shape), mask], axis=-1)
+                yield source_patch, noise_patch
+
+        output_types = (tf.float32, tf.float32)
+        output_shapes = (tf.TensorShape(list(patch_size) + [self.in_channels]), 
+                                             tf.TensorShape(list(patch_size) + [self.in_channels*2]))
+        gen = generator if is_train else generator_val
+        dataset = tf.data.Dataset.from_generator(gen, output_types=output_types, output_shapes=output_shapes)
+        dataset = dataset.batch(batch_size).prefetch(tf.data.experimental.AUTOTUNE)
+
+        return dataset
+
+
+    def train(self, source_lst, patch_size, validation=None, batch_size=64, save_steps=1000, log_steps=200, steps=50000):
+        assert len(patch_size)==self.dim
+        assert len(source_lst[0].shape)==self.dim + 1
+        assert source_lst[0].shape[-1]==self.in_channels
+
+        ses_config = tf.ConfigProto()
+        ses_config.gpu_options.allow_growth = True
+
+        run_config = tf.estimator.RunConfig(model_dir=self.base_dir+'/'+self.name, 
+                                            save_checkpoints_steps=save_steps,
+                                            session_config=ses_config, 
+                                            log_step_count_steps=log_steps,
+                                            save_summary_steps=log_steps,
+                                            keep_checkpoint_max=2)
+
+        estimator = tf.estimator.Estimator(model_fn=self._model_fn, 
+                                             model_dir=self.base_dir+'/'+self.name, 
+                                             config=run_config)
+        
+        input_fn = lambda: self._input_fn(source_lst, patch_size, batch_size=batch_size)
+        
+        if validation is not None:
+            train_spec = tf.estimator.TrainSpec(input_fn=input_fn, max_steps=steps)
+            val_input_fn = lambda: self._input_fn(validation.astype('float32'), 
+                                                  validation.shape[1:-1], 
+                                                  batch_size=4, 
+                                                  is_train=False)
+            eval_spec = tf.estimator.EvalSpec(input_fn=val_input_fn, throttle_secs=120)
+            tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
+        else:
+            estimator.train(input_fn=input_fn, steps=steps)
+            
+
+    # Used for single image prediction
+    def predict(self, image, resizer=PadAndCropResizer(), checkpoint_path=None,
+               im_mean=None, im_std=None):
+
+        tf.logging.set_verbosity(tf.logging.ERROR)
+        estimator = tf.estimator.Estimator(model_fn=self._model_fn, 
+                                            model_dir=self.base_dir+'/'+self.name)
+        
+        im_mean, im_std = ((image.mean(), image.std()) if im_mean is None or im_std is None else (im_mean, im_std)) 
+        image = (image - im_mean)/im_std
+        if self.in_channels == 1:
+            image = resizer.before(image, 2 ** (self.net.depth), exclude=None)
+            input_fn = tf.estimator.inputs.numpy_input_fn(x=image[None, ..., None], batch_size=1, num_epochs=1, shuffle=False)
+            image = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0][..., 0]
+            image = resizer.after(image, exclude=None)
+        else:
+            image = resizer.before(image, 2 ** (self.net.depth), exclude=-1)
+            input_fn = tf.estimator.inputs.numpy_input_fn(x=image[None], batch_size=1, num_epochs=1, shuffle=False)
+            image = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0]
+            image = resizer.after(image, exclude=-1)
+        image = image*im_std + im_mean
+
+        return image
+    
+    # Used for batch images prediction
+    def batch_predict(self, images, resizer=PadAndCropResizer(), checkpoint_path=None,
+               im_mean=None, im_std=None, batch_size=32):
+
+        tf.logging.set_verbosity(tf.logging.ERROR)
+        estimator = tf.estimator.Estimator(model_fn=self._model_fn, 
+                                            model_dir=self.base_dir+'/'+self.name)
+        
+        im_mean, im_std = ((images.mean(), images.std()) if im_mean is None or im_std is None else (im_mean, im_std)) 
+        
+        images = (images - im_mean)/im_std
+        images = resizer.before(images, 2 ** (self.net.depth), exclude=0)
+        input_fn = tf.estimator.inputs.numpy_input_fn(x=images[ ..., None], batch_size=batch_size, num_epochs=1, shuffle=False)
+        images = np.stack(list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path)))[..., 0]
+        images = resizer.after(images, exclude=0)
+        images = images*im_std + im_mean
+
+        return images
+
+    # Used for extremely large input images
+    def crop_predict(self, image, size, margin, resizer=PadAndCropResizer(), checkpoint_path=None,
+               im_mean=None, im_std=None):
+
+        tf.logging.set_verbosity(tf.logging.ERROR)
+        estimator = tf.estimator.Estimator(model_fn=self._model_fn, 
+                                            model_dir=self.base_dir+'/'+self.name)
+        
+        im_mean, im_std = ((image.mean(), image.std()) if im_mean is None or im_std is None else (im_mean, im_std)) 
+        image = (image - im_mean)/im_std
+        out_image = np.empty(image.shape, dtype='float32')
+        for src_s, trg_s, mrg_s in get_coord(image.shape, size, margin):
+            patch = resizer.before(image[src_s], 2 ** (self.net.depth), exclude=None)
+            input_fn = tf.estimator.inputs.numpy_input_fn(x=patch[None, ..., None], batch_size=1, num_epochs=1, shuffle=False)
+            patch = list(estimator.predict(input_fn=input_fn, checkpoint_path=checkpoint_path))[0][..., 0]
+            patch = resizer.after(patch, exclude=None)
+            out_image[trg_s] = patch[mrg_s]
+            
+        image = out_image*im_std + im_mean
+
+        return image

network.py

@@ -0,0 +1,117 @@
+import logging, os
+logging.disable(logging.WARNING)
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+
+import tensorflow as tf
+from basic_ops import *
+from resnet_module import *
+
+
+"""This script generates the U-Net architecture according to conf_unet.
+"""
+
+class UNet(object):
+    def __init__(self, conf_unet):
+        self.depth = conf_unet['depth']
+        self.dimension = conf_unet['dimension']
+        self.first_output_filters = conf_unet['first_output_filters']
+        self.encoding_block_sizes = conf_unet['encoding_block_sizes']
+        self.downsampling = conf_unet['downsampling']
+        self.decoding_block_sizes = conf_unet['decoding_block_sizes']
+        self.skip_method = conf_unet['skip_method']
+
+    def __call__(self, inputs, training):
+        """Add operations to classify a batch of input images.
+
+        Args:
+            inputs: A Tensor representing a batch of input images.
+            training: A boolean. Set to True to add operations required only when
+                training the classifier.
+
+        Returns:
+            A logits Tensor with shape [<batch_size>, self.num_classes].
+        """
+
+        return self._build_network(inputs, training)
+
+
+    ################################################################################
+    # Composite blocks building the network
+    ################################################################################
+    def _build_network(self, inputs, training):
+        # first_convolution
+        if self.dimension == '2D':
+            convolution = convolution_2D
+        elif self.dimension == '3D':
+            convolution = convolution_3D
+        inputs = convolution(inputs, self.first_output_filters, 3, 1, False, 'first_convolution')
+
+        # encoding_block_1
+        with tf.variable_scope('encoding_block_1'):
+            for block_index in range(0, self.encoding_block_sizes[0]):
+                inputs = res_block(inputs, self.first_output_filters, training, self.dimension,
+                            'res_%d' % block_index)
+
+        # encoding_block_i (down) = downsampling + zero or more res_block, i = 2, 3, ..., depth
+        skip_inputs = [] # for identity skip connections
+        for i in range(2, self.depth+1):
+            skip_inputs.append(inputs)
+            with tf.variable_scope('encoding_block_%d' % i):
+                output_filters = self.first_output_filters * (2**(i-1))
+
+                # downsampling
+                downsampling_func = self._get_downsampling_function(self.downsampling[i-2])
+                inputs = downsampling_func(inputs, output_filters, training, self.dimension,
+                            'downsampling')
+
+                for block_index in range(0, self.encoding_block_sizes[i-1]):
+                    inputs = res_block(inputs, output_filters, training, self.dimension,
+                                'res_%d' % block_index)
+
+        # bottom_block = a combination of same_gto and res_block
+        with tf.variable_scope('bottom_block'):
+            output_filters = self.first_output_filters * (2**(self.depth-1))
+            for block_index in range(0, 1):
+                current_func = res_block
+                inputs = current_func(inputs, output_filters, training, self.dimension,
+                            'block_%d' % block_index)
+
+        """
+        Note: Identity skip connections are between the output of encoding_block_i and
+        the output of upsampling in decoding_block_i, i = 1, 2, ..., depth-1.
+        skip_inputs[i] is the output of encoding_block_i now.
+        len(skip_inputs) == depth - 1
+        skip_inputs[depth-2] should be combined during decoding_block_depth-1
+        skip_inputs[0] should be combined during decoding_block_1
+        """
+
+        # decoding_block_j (up) = upsampling + zero or more res_block, j = depth-1, depth-2, ..., 1
+        for j in range(self.depth-1, 0, -1):
+            with tf.variable_scope('decoding_block_%d' % j):
+                output_filters = self.first_output_filters * (2**(j-1))
+
+                # upsampling
+                upsampling_func = up_transposed_convolution
+                inputs = upsampling_func(inputs, output_filters, training, self.dimension,
+                            'upsampling')
+
+                # combine with skip connections
+                if self.skip_method == 'add':
+                    inputs = tf.add(inputs, skip_inputs[j-1])
+                elif self.skip_method == 'concat':
+                    inputs = tf.concat([inputs, skip_inputs[j-1]], axis=-1)
+
+                for block_index in range(0, self.decoding_block_sizes[self.depth-1-j]):
+                    inputs = res_block(inputs, output_filters, training, self.dimension,
+                                'res_%d' % block_index)
+
+        return inputs
+
+
+    def _get_downsampling_function(self, name):
+        if name == 'down_res_block':
+            return down_res_block
+        elif name == 'convolution':
+            return down_convolution
+        else:
+            raise ValueError("Unsupported function: %s." % (name))

network_configure.py

@@ -0,0 +1,143 @@
+import logging, os
+logging.disable(logging.WARNING)
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+
+import tensorflow as tf
+
+
+"""This is the configuration file.
+"""
+
+
+################################################################################
+# Settings for Basic Operaters
+################################################################################
+
+conf_basic_ops = dict()
+
+# kernel_initializer for convolutions and transposed convolutions
+# If None, the default initializer is the Glorot (Xavier) normal initializer.
+conf_basic_ops['kernel_initializer'] = tf.glorot_uniform_initializer()
+
+# momentum for batch normalization
+conf_basic_ops['momentum'] = 0.997
+
+# epsilon for batch normalization
+conf_basic_ops['epsilon'] = 1e-5
+
+# String options: 'relu', 'relu6'
+conf_basic_ops['relu_type'] = 'relu'
+
+################################################################################
+# Settings for Attention Modules
+################################################################################
+
+# Set the attention in same_gto
+conf_attn_same = dict()
+
+# Define the relationship between total_key_filters and output_filters.
+# total_key_filters = output_filters // key_ratio
+conf_attn_same['key_ratio'] = 1
+
+# Define the relationship between total_value_filters and output_filters.
+# total_key_filters = output_filters // value_ratio
+conf_attn_same['value_ratio'] = 1
+
+# number of heads
+conf_attn_same['num_heads'] = 2
+
+# dropout rate, 0.0 means no dropout
+conf_attn_same['dropout_rate'] = 0.0
+
+# whether to use softmax on attention_weights
+conf_attn_same['use_softmax'] = False
+
+# whether to use bias terms in input/output transformations
+conf_attn_same['use_bias'] = True
+
+# Set the attention in up_gto
+conf_attn_up = dict()
+
+conf_attn_up['key_ratio'] = 1
+conf_attn_up['value_ratio'] = 1
+conf_attn_up['num_heads'] = 2
+conf_attn_up['dropout_rate'] = 0
+conf_attn_up['use_softmax'] = False
+conf_attn_up['use_bias'] = True
+
+# Set the attention in down_gto
+conf_attn_down = dict()
+
+conf_attn_down['key_ratio'] = 1
+conf_attn_down['value_ratio'] = 1
+conf_attn_down['num_heads'] = 2
+conf_attn_down['dropout_rate'] = 0.0
+conf_attn_down['use_softmax'] = False
+conf_attn_down['use_bias'] = True
+
+################################################################################
+# Describing the U-net
+################################################################################
+
+conf_unet = dict()
+
+"""
+Describe your U-Net under the following framework:
+
+********************************************************************************************
+layers													|	output_filters
+														|
+first_convolution + encoding_block_1 (same)				|	first_output_filters
++ encoding_block_i, i = 2, 3, ..., depth. (down)		|	first_output_filters*(2**(i-1))
++ bottom_block											|	first_output_filters*(2**(depth-1))
++ decoding_block_j, j = depth-1, depth-2, ..., 1 (up)	|	first_output_filters*(2**(j-1))
++ output_layer
+********************************************************************************************
+
+Specifically,
+encoding_block_1 (same) = one or more res_block
+encoding_block_i (down) = downsampling + zero or more res_block, i = 2, 3, ..., depth-1
+encoding_block_depth (down) = downsampling
+bottom_block = a combination of same_gto and res_block
+decoding_block_j (up) = upsampling + zero or more res_block, j = depth-1, depth-2, ..., 1
+
+Identity skip connections are between the output of encoding_block_i and
+the output of upsampling in decoding_block_i, i = 1, 2, ..., depth-1.
+The combination method could be 'add' or 'concat'.
+"""
+
+# Set U-Net depth.
+conf_unet['depth'] = 3
+
+# Set the output_filters for first_convolution and encoding_block_1 (same).
+conf_unet['first_output_filters'] = 96
+
+# Set the encoding block sizes, i.e., number of res_block in encoding_block_i, i = 1, 2, ..., depth.
+# It is an integer list whose length equals to depth.
+# The first entry should be positive since encoding_block_1 = one or more res_block.
+# The last entry should be zero since encoding_block_depth (down) = downsampling.
+conf_unet['encoding_block_sizes'] = [1, 1, 0]
+
+# Set the decoding block sizes, i.e., number of res_block in decoding_block_j, j = depth-1, depth-2, ..., 1.
+# It is an integer list whose length equals to depth-1.
+conf_unet['decoding_block_sizes'] = [1, 1]
+
+# Set the downsampling methods for each encoding_block_i, i = 2, 3, ..., depth.
+# It is an string list whose length equals to depth-1.
+# String options: 'down_gto_v1', 'down_gto_v2', 'down_res_block', 'convolution'
+conf_unet['downsampling'] = ['convolution', 'convolution']
+
+# Set the combination method for identity skip connections
+# String options: 'add', 'concat'
+conf_unet['skip_method'] = 'concat'
+
+# Set the output layer
+
+
+# Check
+assert conf_unet['depth'] == len(conf_unet['encoding_block_sizes'])
+assert conf_unet['encoding_block_sizes'][0] > 0
+assert conf_unet['encoding_block_sizes'][-1] == 0
+assert conf_unet['depth'] == len(conf_unet['decoding_block_sizes']) + 1
+assert conf_unet['depth'] == len(conf_unet['downsampling']) + 1
+assert conf_unet['skip_method'] in ['add', 'concat']

requirements.txt

@@ -0,0 +1,7 @@
+tensorflow==1.15
+scipy
+scikit-image
+tifffile
+gdown
+opencv-python
+numpy

resnet_module.py

@@ -0,0 +1,101 @@
+import logging, os
+logging.disable(logging.WARNING)
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+
+import tensorflow as tf
+from basic_ops import *
+
+
+"""This script defines non-attention same-, up-, down- modules.
+Note that pre-activation is used for residual-like blocks.
+Note that the residual block could be used for downsampling.
+"""
+
+
+def res_block(inputs, output_filters, training, dimension, name):
+    """Standard residual block with pre-activation.
+
+    Args:
+        inputs: a Tensor with shape [batch, (d,) h, w, channels]
+        output_filters: an integer
+        training: a boolean for batch normalization and dropout
+        dimension: a string, dimension of inputs/outputs -- 2D, 3D
+        name: a string
+        
+    Returns:
+        A Tensor of shape [batch, (_d,) _h, _w, output_filters]
+    """
+    if dimension == '2D':
+        convolution = convolution_2D
+        kernel_size = 3
+    elif dimension == '3D':
+        convolution = convolution_3D
+        kernel_size = 3
+    else:
+        raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
+
+    with tf.variable_scope(name):
+        if inputs.shape[-1] == output_filters:
+            shortcut = inputs
+            inputs = batch_norm(inputs, training, 'batch_norm_1')
+            inputs = relu(inputs, 'relu_1')
+        else:
+            inputs = batch_norm(inputs, training, 'batch_norm_1')
+            inputs = relu(inputs, 'relu_1')
+            shortcut = convolution(inputs, output_filters, 1, 1, False, 'projection_shortcut')
+        inputs = convolution(inputs, output_filters, kernel_size, 1, False, 'convolution_1')
+        inputs = batch_norm(inputs, training, 'batch_norm_2')
+        inputs = relu(inputs, 'relu_2')
+        inputs = convolution(inputs, output_filters, kernel_size, 1, False, 'convolution_2')
+        return tf.add(shortcut, inputs)
+
+
+def down_res_block(inputs, output_filters, training, dimension, name):
+    """Standard residual block with pre-activation for downsampling."""
+    if dimension == '2D':
+        convolution = convolution_2D
+        projection_shortcut = convolution_2D
+    elif dimension == '3D':
+        convolution = convolution_3D
+        projection_shortcut = convolution_3D
+    else:
+        raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
+
+    with tf.variable_scope(name):
+        # The projection_shortcut should come after the first batch norm and ReLU.
+        inputs = batch_norm(inputs, training, 'batch_norm_1')
+        inputs = relu(inputs, 'relu_1')
+        shortcut = projection_shortcut(inputs, output_filters, 1, 2, False, 'projection_shortcut')
+        inputs = convolution(inputs, output_filters, 2, 2, False, 'convolution_1')
+        inputs = batch_norm(inputs, training, 'batch_norm_2')
+        inputs = relu(inputs, 'relu_2')
+        inputs = convolution(inputs, output_filters, 3, 1, False, 'convolution_2')
+        return tf.add(shortcut, inputs)
+
+def down_convolution(inputs, output_filters, training, dimension, name):
+    """Use a single stride 2 convolution for downsampling."""
+    if dimension == '2D':
+        convolution = convolution_2D
+        pool = tf.layers.max_pooling2d
+    elif dimension == '3D':
+        convolution = convolution_3D
+        pool = tf.layers.max_pooling3d
+    else:
+        raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
+
+    with tf.variable_scope(name):
+        inputs = convolution(inputs, output_filters, 2, 2, True, 'convolution')
+        return inputs
+
+def up_transposed_convolution(inputs, output_filters, training, dimension, name):
+    """Use a single stride 2 transposed convolution for upsampling."""
+    if dimension == '2D':
+        transposed_convolution = transposed_convolution_2D
+    elif dimension == '3D':
+        transposed_convolution = transposed_convolution_3D
+    else:
+        raise ValueError("Dimension (%s) must be 2D or 3D." % (dimension))
+
+    with tf.variable_scope(name):
+        inputs = transposed_convolution(inputs, output_filters, 2, 2, True, 'transposed_convolution')
+        return inputs

trained_models/README.md

@@ -0,0 +1,3 @@
+Checkpoints for the trained model
+-----
+Due to the limitation of repo size, we upload the model files to Google Drive. You can manually download them [here](https://drive.google.com/drive/folders/1VYMo1OoaGxoOLNx6-qIt2Wg03lsZw_kA?usp=sharing).

utils/evaluation_utils.py

@@ -0,0 +1,70 @@
+import numpy as np
+from scipy.misc import ascent
+from skimage.measure import compare_psnr, compare_mse, compare_ssim
+from .predict_utils import normalize_mi_ma
+
+def normalize(x, pmin=2, pmax=99.8, axis=None, clip=False, eps=1e-20, dtype=np.float32):
+    """Percentile-based image normalization."""
+
+    mi = np.percentile(x,pmin,axis=axis,keepdims=True)
+    ma = np.percentile(x,pmax,axis=axis,keepdims=True)
+    return normalize_mi_ma(x, mi, ma, clip=clip, eps=eps, dtype=dtype)
+
+def norm_minmse(gt, x, normalize_gt=True):
+    """
+    normalizes and affinely scales an image pair such that the MSE is minimized  
+     
+    Parameters
+    ----------
+    gt: ndarray
+        the ground truth image      
+    x: ndarray
+        the image that will be affinely scaled 
+    normalize_gt: bool
+        set to True of gt image should be normalized (default)
+    Returns
+    -------
+    gt_scaled, x_scaled 
+    """
+    if normalize_gt:
+        gt = normalize(gt, 0.1, 99.9, clip=False).astype(np.float32, copy = False)
+    x = x.astype(np.float32, copy=False) - np.mean(x)
+    gt = gt.astype(np.float32, copy=False) - np.mean(gt)
+    scale = np.cov(x.flatten(), gt.flatten())[0, 1] / np.var(x.flatten())
+    return gt, scale * x
+
+
+def get_scores(gt, x, multichan=False):
+    
+    gt_, x_ = norm_minmse(gt, x)
+    
+    mse = compare_mse(gt_, x_)
+    psnr = compare_psnr(gt_, x_, data_range = 1.)
+    ssim = compare_ssim(gt_, x_, data_range = 1., multichannel=multichan)
+    
+    return np.sqrt(mse), psnr, ssim
+
+if __name__ == '__main__':
+
+    # ground truth image
+    y = ascent().astype(np.float32)
+    # input image to compare to 
+    x1 = y + 30*np.random.normal(0,1,y.shape)
+    # a scaled and shifted version of x1
+    x2 = 2*x1+100
+    
+    # calulate mse, psnr, and ssim of the normalized/scaled images
+    mse1  = compare_mse(*norm_minmse(y, x1))
+    mse2  = compare_mse(*norm_minmse(y, x2))
+    # should be the same
+    print("MSE1  = %.6f\nMSE2  = %.6f"%(mse1, mse2))
+
+    psnr1 = compare_psnr(*norm_minmse(y, x1), data_range = 1.)
+    psnr2 = compare_psnr(*norm_minmse(y, x2), data_range = 1.)
+    # should be the same    
+    print("PSNR1 = %.6f\nPSNR2 = %.6f"%(psnr1,psnr2))
+    
+    ssim1 = compare_ssim(*norm_minmse(y, x1), data_range = 1.)
+    ssim2 = compare_ssim(*norm_minmse(y, x2), data_range = 1.)
+    # should be the same    
+    print("SSIM1 = %.6f\nSSIM2 = %.6f"%(ssim1,ssim2))

utils/predict_utils.py

@@ -0,0 +1,152 @@
+from __future__ import print_function, unicode_literals, absolute_import, division
+from six.moves import range, zip, map, reduce, filter
+
+import collections
+import warnings
+import numpy as np
+
+
+def get_coord(shape, size, margin):
+    n_tiles_i = int(np.ceil((shape[2]-size)/float(size-2*margin)))
+    n_tiles_j = int(np.ceil((shape[1]-size)/float(size-2*margin)))
+    for i in range(n_tiles_i+1):
+        src_start_i = i*(size-2*margin) if i<n_tiles_i else (shape[2]-size)
+        src_end_i = src_start_i+size
+        left_i = margin if i>0 else 0
+        right_i = margin if i<n_tiles_i else 0
+        for j in range(n_tiles_j+1):
+            src_start_j = j*(size-2*margin) if j<n_tiles_j else (shape[1]-size)
+            src_end_j = src_start_j+size
+            left_j = margin if j>0 else 0
+            right_j = margin if j<n_tiles_j else 0
+            src_s = (slice(None, None), 
+                     slice(src_start_j, src_end_j), 
+                     slice(src_start_i, src_end_i))
+            
+            trg_s = (slice(None, None), 
+                     slice(src_start_j+left_j, src_end_j-right_j), 
+                     slice(src_start_i+left_i, src_end_i-right_i))
+            
+            mrg_s = (slice(None, None), 
+                     slice(left_j, -right_j if right_j else None), 
+                     slice(left_i, -right_i if right_i else None))
+            
+            yield src_s, trg_s, mrg_s
+
+
+# Below implementation of prediction utils inherited from CARE: https://github.com/CSBDeep/CSBDeep
+# Content-Aware Image Restoration: Pushing the Limits of Fluorescence Microscopy. Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, Siân Culley, Mauricio Rocha-Martins, Fabián Segovia-Miranda, Caren Norden, Ricardo Henriques, Marino Zerial, Michele Solimena, Jochen Rink, Pavel Tomancak, Loic Royer, Florian Jug, and Eugene W. Myers. Nature Methods 15.12 (2018): 1090–1097.
+
+def _raise(e):
+    raise e
+    
+def consume(iterator):
+    collections.deque(iterator, maxlen=0)
+
+def axes_check_and_normalize(axes,length=None,disallowed=None,return_allowed=False):
+    """
+    S(ample), T(ime), C(hannel), Z, Y, X
+    """
+    allowed = 'STCZYX'
+    axes is not None or _raise(ValueError('axis cannot be None.'))
+    axes = str(axes).upper()
+    consume(a in allowed or _raise(ValueError("invalid axis '%s', must be one of %s."%(a,list(allowed)))) for a in axes)
+    disallowed is None or consume(a not in disallowed or _raise(ValueError("disallowed axis '%s'."%a)) for a in axes)
+    consume(axes.count(a)==1 or _raise(ValueError("axis '%s' occurs more than once."%a)) for a in axes)
+    length is None or len(axes)==length or _raise(ValueError('axes (%s) must be of length %d.' % (axes,length)))
+    return (axes,allowed) if return_allowed else axes
+
+
+def axes_dict(axes):
+    """
+    from axes string to dict
+    """
+    axes, allowed = axes_check_and_normalize(axes,return_allowed=True)
+    return { a: None if axes.find(a) == -1 else axes.find(a) for a in allowed }
+
+
+def normalize_mi_ma(x, mi, ma, clip=False, eps=1e-20, dtype=np.float32):
+    if dtype is not None:
+        x   = x.astype(dtype,copy=False)
+        mi  = dtype(mi) if np.isscalar(mi) else mi.astype(dtype,copy=False)
+        ma  = dtype(ma) if np.isscalar(ma) else ma.astype(dtype,copy=False)
+        eps = dtype(eps)
+    try:
+        import numexpr
+        x = numexpr.evaluate("(x - mi) / ( ma - mi + eps )")
+    except ImportError:
+        x =                   (x - mi) / ( ma - mi + eps )
+    if clip:
+        x = np.clip(x,0,1)
+    return x
+
+class PercentileNormalizer(object):
+
+    def __init__(self, pmin=2, pmax=99.8, do_after=True, dtype=np.float32, **kwargs):
+
+        (np.isscalar(pmin) and np.isscalar(pmax) and 0 <= pmin < pmax <= 100) or _raise(ValueError())
+        self.pmin = pmin
+        self.pmax = pmax
+        self._do_after = do_after
+        self.dtype = dtype
+        self.kwargs = kwargs
+
+    def before(self, img, axes):
+
+        len(axes) == img.ndim or _raise(ValueError())
+        channel = axes_dict(axes)['C']
+        axes = None if channel is None else tuple((d for d in range(img.ndim) if d != channel))
+        self.mi = np.percentile(img,self.pmin,axis=axes,keepdims=True).astype(self.dtype,copy=False)
+        self.ma = np.percentile(img,self.pmax,axis=axes,keepdims=True).astype(self.dtype,copy=False)
+        return normalize_mi_ma(img, self.mi, self.ma, dtype=self.dtype, **self.kwargs)
+
+    def after(self, img):
+
+        self.do_after or _raise(ValueError())
+        alpha = self.ma - self.mi
+        beta  = self.mi
+        return ( alpha*img+beta ).astype(self.dtype,copy=False)
+
+    def do_after(self):
+        
+        return self._do_after
+    
+    
+class PadAndCropResizer(object):
+
+    def __init__(self, mode='reflect', **kwargs):
+
+        self.mode = mode
+        self.kwargs = kwargs
+        
+    def _normalize_exclude(self, exclude, n_dim):
+        """Return normalized list of excluded axes."""
+        if exclude is None:
+            return []
+        exclude_list = [exclude] if np.isscalar(exclude) else list(exclude)
+        exclude_list = [d%n_dim for d in exclude_list]
+        len(exclude_list) == len(np.unique(exclude_list)) or _raise(ValueError())
+        all(( isinstance(d,int) and 0<=d<n_dim for d in exclude_list )) or _raise(ValueError())
+        return exclude_list
+
+    def before(self, x, div_n, exclude):
+
+        def _split(v):
+            a = v // 2
+            return a, v-a
+        exclude = self._normalize_exclude(exclude, x.ndim)
+        self.pad = [_split((div_n-s%div_n)%div_n) if (i not in exclude) else (0,0) for i,s in enumerate(x.shape)]
+        x_pad = np.pad(x, self.pad, mode=self.mode, **self.kwargs)
+        for i in exclude:
+            del self.pad[i]
+        return x_pad
+
+    def after(self, x, exclude):
+
+        pads = self.pad[:len(x.shape)]
+        crop = [slice(p[0], -p[1] if p[1]>0 else None) for p in self.pad]
+        for i in self._normalize_exclude(exclude, x.ndim):
+            crop.insert(i,slice(None))
+        len(crop) == x.ndim or _raise(ValueError())
+        return x[tuple(crop)]
+

utils/train_utils.py

@@ -0,0 +1,60 @@
+import numpy as np
+from tqdm import tqdm
+
+
+def augment_patch(patch):
+    if len(patch.shape[:-1]) == 2:
+        patch = np.rot90(patch, k=np.random.randint(4), axes=(0, 1))
+    elif len(patch.shape[:-1]) == 3:
+        patch = np.rot90(patch, k=np.random.randint(4), axes=(1, 2))
+
+    patch = np.flip(patch, axis=-2) if np.random.randint(2) else patch
+    return patch
+
+
+# Below implementation of stratified sampling inherited from Noise2Void: https://github.com/juglab/n2v
+# Noise2void: learning denoising from single noisy images. Krull, Alexander, Tim-Oliver Buchholz, and Florian Jug. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.
+
+def get_stratified_coords2D(coord_gen, box_size, shape):
+    box_count_y = int(np.ceil(shape[0] / box_size))
+    box_count_x = int(np.ceil(shape[1] / box_size))
+    x_coords = []
+    y_coords = []
+    for i in range(box_count_y):
+        for j in range(box_count_x):
+            y, x = next(coord_gen)
+            y = int(i * box_size + y)
+            x = int(j * box_size + x)
+            if (y < shape[0] and x < shape[1]):
+                y_coords.append(y)
+                x_coords.append(x)
+    return (y_coords, x_coords)
+
+
+def get_stratified_coords3D(coord_gen, box_size, shape):
+    box_count_z = int(np.ceil(shape[0] / box_size))
+    box_count_y = int(np.ceil(shape[1] / box_size))
+    box_count_x = int(np.ceil(shape[2] / box_size))
+    x_coords = []
+    y_coords = []
+    z_coords = []
+    for i in range(box_count_z):
+        for j in range(box_count_y):
+            for k in range(box_count_x):
+                z, y, x = next(coord_gen)
+                z = int(i * box_size + z)
+                y = int(j * box_size + y)
+                x = int(k * box_size + x)
+                if (z < shape[0] and y < shape[1] and x < shape[2]):
+                    z_coords.append(z)
+                    y_coords.append(y)
+                    x_coords.append(x)
+    return (z_coords, y_coords, x_coords)
+
+def rand_float_coords2D(boxsize):
+    while True:
+        yield (np.random.rand() * boxsize, np.random.rand() * boxsize)
+        
+def rand_float_coords3D(boxsize):
+    while True:
+        yield (np.random.rand() * boxsize, np.random.rand() * boxsize, np.random.rand() * boxsize)