sequence instance fixed

cifar.ipynb

@@ -2,7 +2,7 @@
  "cells": [
   {
    "cell_type": "code",
-   "execution_count": 8,
+   "execution_count": 1,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -16,7 +16,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 9,
+   "execution_count": 2,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -25,7 +25,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 10,
+   "execution_count": 3,
    "metadata": {},
    "outputs": [
     {
@@ -52,7 +52,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": 4,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -73,7 +73,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": 5,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -109,7 +109,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 3,
+   "execution_count": 6,
    "metadata": {
     "scrolled": false
    },
@@ -174,7 +174,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 4,
+   "execution_count": 7,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -190,7 +190,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 5,
+   "execution_count": 8,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -206,7 +206,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 6,
+   "execution_count": 9,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -215,7 +215,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 11,
+   "execution_count": 10,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -224,7 +224,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 12,
+   "execution_count": 11,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -236,7 +236,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 13,
+   "execution_count": 12,
    "metadata": {
     "scrolled": false
    },
@@ -245,7 +245,7 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Validation Accuracy: 0.1054 - val_loss: 5.0138 - Time: 8.601s\n"
+      "Validation Accuracy: 0.1041 - val_loss: 4.9765 - Time: 8.761s\n"
      ]
     }
    ],
@@ -257,7 +257,7 @@
    "cell_type": "code",
    "execution_count": 13,
    "metadata": {
-    "scrolled": true
+    "scrolled": false
    },
    "outputs": [
     {

nnet_gpu/layers/BatchNormalization.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 from .base_layer import *
-from . import seqinst
 from ..stream_handler import stream_maps
 
 
@@ -15,7 +14,7 @@ def __init__(self, momentum=0.9, epsilon=1e-10, name=None):
 			self.name = self.__class__.__name__
 		else:
 			self.name = name
-		input_shape = seqinst.seq_instance.get_inp_shape()
+		input_shape = self.get_inp_shape()
 		self.shape = (None, *input_shape)
 		self.batches = 1
 		self.inp_shape = (self.batches, *input_shape)

nnet_gpu/layers/__init__.py

@@ -14,4 +14,4 @@
 from .shaping import reshape
 from .upsampling import upsampling
 
-import cupyx
+import cupyx

nnet_gpu/layers/activation.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 from .base_layer import *
-from . import seqinst
 from ..stream_handler import stream_maps
 
 
@@ -14,7 +13,7 @@ def __init__(self, activation=echo, input_shape=None, name=None):
 		else:
 			self.name = name
 		if input_shape is None:
-			input_shape = seqinst.seq_instance.get_inp_shape()
+			input_shape = self.get_inp_shape()
 		self.activation = activation
 		self.shape = (None, *input_shape)
 		self.param = 0

nnet_gpu/layers/base_layer.py

@@ -2,7 +2,6 @@
 from ..functions import *
 import numpy as np
 import cupy as cp
-from . import seqinst
 
 
 class Layer:
@@ -24,6 +23,9 @@ def __call__(self, lyr):
 		lyr.output_layers.append(self)
 		return self
 
+	def get_inp_shape(self):
+		return self.input_layer.shape[1:]
+
 
 class InputLayer(Layer):  # just placeholder
 	def __init__(self, shape=None):

nnet_gpu/layers/convolution/conv2d.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 from ..base_layer import *
-from .. import seqinst
 from ...stream_handler import stream_maps
 from .conv_utils import *
 
@@ -11,7 +10,7 @@ def __init__(self, num_kernels=0, input_shape=None, kernel_size=0, kernels=None,
 		# input_shape[row,col,channels], kernels(channels,ksz[0],ksz[1],num_kernels), biases[1,num_ker], stride[row,col]
 		super().__init__()
 		if input_shape is None:
-			input_shape = seqinst.seq_instance.get_inp_shape()
+			input_shape = self.get_inp_shape()
 		if name is None:
 			self.name = self.__class__.__name__
 		else:
@@ -35,7 +34,7 @@ def __init__(self, num_kernels=0, input_shape=None, kernel_size=0, kernels=None,
 			self.kernel_size = kernels.shape[1:3]
 		self.weights = self.kernels
 		self.bias_is_not_0 = True
-		if cp.isscalar(self.biases):  # DO BETTER FIX
+		if cp.isscalar(self.biases):  # TODO: DO BETTER FIX
 			if self.biases == 0:
 				self.bias_is_not_0 = False
 		self.dilation = dilation

nnet_gpu/layers/convolution/conv2dtranspose.py

@@ -82,7 +82,7 @@ def backprop(self, grads, layer=1):
 		with self.backp_stream:
 			self.backp_stream.wait_event(self.grad_event)
 			self.d_c_w = self.d_ker.forward(grads.transpose(3, 1, 2, 0))  # [channels,row,col,batches]
-		# self.d_c_w/=self.batches		#take mean change over batches
+		# self.d_c_w/=self.batches		# take mean change over batches
 		if layer:
 			d_inputs = cp.ascontiguousarray(self.d_inp.forward(grads))
 		# assert d_inputs.shape == (self.batches,self.row,self.col,self.channels),f"{(self.batches,self.row,self.col,self.channels)},{d_inputs.shape}"

nnet_gpu/layers/dense.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 from .base_layer import *
-from . import seqinst
 from ..stream_handler import stream_maps
 
 
@@ -15,7 +14,7 @@ def __init__(self, num_out, input_shape=None, weights=None, biases=None, activat
 		else:
 			self.name = name
 		if input_shape is None:
-			self.input_shape = seqinst.seq_instance.get_inp_shape()[0]
+			self.input_shape = self.get_inp_shape()[0]
 		else:
 			self.input_shape = input_shape
 		self.activation = activation

nnet_gpu/layers/dropout.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 from .base_layer import *
-from . import seqinst
 
 
 class dropout(Layer):
@@ -12,7 +11,7 @@ def __init__(self, rate=0.2, name=None):  # rate = amount to drop
 			self.name = self.__class__.__name__
 		else:
 			self.name = name
-		input_shape = seqinst.seq_instance.get_inp_shape()
+		input_shape = self.get_inp_shape()
 		self.shape = (None, *input_shape)
 		self.batches = 1
 		self.rate = rate

nnet_gpu/layers/pooling/globalAveragePool.py

@@ -11,7 +11,7 @@ def __init__(self, input_shape=None, name=None):
 		else:
 			self.name = name
 		if input_shape is None:
-			input_shape = seqinst.seq_instance.get_inp_shape()
+			input_shape = self.get_inp_shape()
 		self.param = 0
 		self.batches = 1
 		self.row, self.col, self.channels = input_shape

nnet_gpu/layers/pooling/max_pool.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 from ..base_layer import *
-from .. import seqinst
 from ...stream_handler import stream_maps
 
 
@@ -19,7 +18,7 @@ def __init__(self, input_shape=None, ksize=(2, 2), stride=(2, 2), name=None):
 		else:
 			self.name = name
 		if input_shape is None:
-			input_shape = seqinst.seq_instance.get_inp_shape()
+			input_shape = self.get_inp_shape()
 		self.batches = 1
 		self.row, self.col, self.channels = input_shape
 		# self.rem_col=self.row%self.ksz

nnet_gpu/layers/shaping.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 from .base_layer import *
-from . import seqinst
 
 
 class flatten(Layer):
@@ -12,7 +11,7 @@ def __init__(self, name=None):
 			self.name = self.__class__.__name__
 		else:
 			self.name = name
-		input_shape = seqinst.seq_instance.get_inp_shape()
+		input_shape = self.get_inp_shape()
 		self.r, self.c, self.channels = input_shape
 		self.fsz = self.r * self.c * self.channels
 		self.shape = (None, self.fsz)
@@ -35,7 +34,7 @@ def __init__(self, target_shape, name=None):
 			self.name = self.__class__.__name__
 		else:
 			self.name = name
-		self.input_shape = seqinst.seq_instance.get_inp_shape()
+		self.input_shape = self.get_inp_shape()
 		self.target_shape = target_shape
 		tt = 1
 		for i in self.input_shape:

nnet_gpu/layers/upsampling.py

@@ -1,6 +1,5 @@
 #!/usr/bin/env python3
 from .base_layer import *
-from . import seqinst
 
 
 class upsampling(Layer):
@@ -16,7 +15,7 @@ def __init__(self, input_shape=None, ksize=(2, 2), stride=(2, 2), name=None):
 		else:
 			self.name = name
 		if input_shape is None:
-			input_shape = seqinst.seq_instance.get_inp_shape()
+			input_shape = self.get_inp_shape()
 		self.batches = 1
 		self.row, self.col, self.channels = input_shape
 		self.out_row, self.out_col = self.row * self.ksz, self.col * self.ksz

nnet_gpu/network.py

@@ -3,6 +3,7 @@
 from .functions import *
 from .optimizers import *
 from .stream_handler import stream_maps
+from .layers import Layer
 
 import pickle
 from gc import collect
@@ -11,9 +12,9 @@
 
 # TODO- In train/fit unifunc, transpose whole data of inp at once and remove from layers.
 
-class Sequential:
+class Sequential(Layer):
 	def __init__(self):
-		layers.seqinst.seq_instance = self
+		super().__init__()
 		self.sequence = []
 		self.learning_rate = 0.001
 		self.dtype = cp.float32
@@ -23,9 +24,6 @@ def add(self, obj):
 			obj(self.sequence[-1])
 		self.sequence.append(obj)
 
-	def get_inp_shape(self):
-		return self.sequence[-1].shape[1:]
-
 	def forward(self, X_inp, training=True):
 		for obj in self.sequence:
 			X_inp = obj.forward(X_inp, training=training)
@@ -61,7 +59,7 @@ def fit(self, X_inp=None, labels=None, iterator=None, batch_size=1, epochs=1, va
 		sam_time = 0
 		for epch in range(epochs):
 			print("EPOCH:", epch + 1, "/", epochs)
-			if iterator == None:
+			if iterator is None:
 				if shuffle:
 					s = cp.random.permutation(lnxinp).astype(cp.int32, copy=False)
 					X_inp = X_inp[s]